amd/llvm: switch to 3-spaces style
authorPierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Mon, 7 Sep 2020 07:56:01 +0000 (09:56 +0200)
committerVivek Pandya <vivekvpandya@gmail.com>
Mon, 7 Sep 2020 15:55:16 +0000 (21:25 +0530)
Follow-up of !4319 using the same clang-format config.

Acked-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Acked-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5310>

src/amd/llvm/ac_llvm_build.c
src/amd/llvm/ac_llvm_build.h
src/amd/llvm/ac_llvm_cull.c
src/amd/llvm/ac_llvm_cull.h
src/amd/llvm/ac_llvm_helper.cpp
src/amd/llvm/ac_llvm_util.c
src/amd/llvm/ac_llvm_util.h
src/amd/llvm/ac_nir_to_llvm.c
src/amd/llvm/ac_nir_to_llvm.h
src/amd/llvm/ac_shader_abi.h

index 79a5f8fbb2bad3f87b5163455366dee651a78640..f1ab80e7f096c625aaee1dd0d5ecbc5c72c7047e 100644 (file)
 /* based on pieces from si_pipe.c and radeon_llvm_emit.c */
 #include "ac_llvm_build.h"
 
-#include <llvm-c/Core.h>
-#include <llvm/Config/llvm-config.h>
-
-#include "c11/threads.h"
-
-#include <assert.h>
-#include <stdio.h>
-
+#include "ac_exp_param.h"
 #include "ac_llvm_util.h"
 #include "ac_shader_util.h"
-#include "ac_exp_param.h"
+#include "c11/threads.h"
+#include "shader_enums.h"
+#include "sid.h"
 #include "util/bitscan.h"
 #include "util/macros.h"
 #include "util/u_atomic.h"
 #include "util/u_math.h"
-#include "sid.h"
+#include <llvm-c/Core.h>
+#include <llvm/Config/llvm-config.h>
 
-#include "shader_enums.h"
+#include <assert.h>
+#include <stdio.h>
 
 #define AC_LLVM_INITIAL_CF_DEPTH 4
 
 /* Data for if/else/endif and bgnloop/endloop control flow structures.
  */
 struct ac_llvm_flow {
-       /* Loop exit or next part of if/else/endif. */
-       LLVMBasicBlockRef next_block;
-       LLVMBasicBlockRef loop_entry_block;
+   /* Loop exit or next part of if/else/endif. */
+   LLVMBasicBlockRef next_block;
+   LLVMBasicBlockRef loop_entry_block;
 };
 
 /* Initialize module-independent parts of the context.
  *
  * The caller is responsible for initializing ctx::module and ctx::builder.
  */
-void
-ac_llvm_context_init(struct ac_llvm_context *ctx,
-                    struct ac_llvm_compiler *compiler,
-                    enum chip_class chip_class, enum radeon_family family,
-                    enum ac_float_mode float_mode, unsigned wave_size,
-                    unsigned ballot_mask_bits)
-{
-       ctx->context = LLVMContextCreate();
-
-       ctx->chip_class = chip_class;
-       ctx->family = family;
-       ctx->wave_size = wave_size;
-       ctx->ballot_mask_bits = ballot_mask_bits;
-       ctx->float_mode = float_mode;
-       ctx->module = ac_create_module(wave_size == 32 ? compiler->tm_wave32
-                                                      : compiler->tm,
-                                      ctx->context);
-       ctx->builder = ac_create_builder(ctx->context, float_mode);
-
-       ctx->voidt = LLVMVoidTypeInContext(ctx->context);
-       ctx->i1 = LLVMInt1TypeInContext(ctx->context);
-       ctx->i8 = LLVMInt8TypeInContext(ctx->context);
-       ctx->i16 = LLVMIntTypeInContext(ctx->context, 16);
-       ctx->i32 = LLVMIntTypeInContext(ctx->context, 32);
-       ctx->i64 = LLVMIntTypeInContext(ctx->context, 64);
-       ctx->i128 = LLVMIntTypeInContext(ctx->context, 128);
-       ctx->intptr = ctx->i32;
-       ctx->f16 = LLVMHalfTypeInContext(ctx->context);
-       ctx->f32 = LLVMFloatTypeInContext(ctx->context);
-       ctx->f64 = LLVMDoubleTypeInContext(ctx->context);
-       ctx->v2i16 = LLVMVectorType(ctx->i16, 2);
-       ctx->v4i16 = LLVMVectorType(ctx->i16, 4);
-       ctx->v2f16 = LLVMVectorType(ctx->f16, 2);
-       ctx->v4f16 = LLVMVectorType(ctx->f16, 4);
-       ctx->v2i32 = LLVMVectorType(ctx->i32, 2);
-       ctx->v3i32 = LLVMVectorType(ctx->i32, 3);
-       ctx->v4i32 = LLVMVectorType(ctx->i32, 4);
-       ctx->v2f32 = LLVMVectorType(ctx->f32, 2);
-       ctx->v3f32 = LLVMVectorType(ctx->f32, 3);
-       ctx->v4f32 = LLVMVectorType(ctx->f32, 4);
-       ctx->v8i32 = LLVMVectorType(ctx->i32, 8);
-       ctx->iN_wavemask = LLVMIntTypeInContext(ctx->context, ctx->wave_size);
-       ctx->iN_ballotmask = LLVMIntTypeInContext(ctx->context, ballot_mask_bits);
-
-       ctx->i8_0 = LLVMConstInt(ctx->i8, 0, false);
-       ctx->i8_1 = LLVMConstInt(ctx->i8, 1, false);
-       ctx->i16_0 = LLVMConstInt(ctx->i16, 0, false);
-       ctx->i16_1 = LLVMConstInt(ctx->i16, 1, false);
-       ctx->i32_0 = LLVMConstInt(ctx->i32, 0, false);
-       ctx->i32_1 = LLVMConstInt(ctx->i32, 1, false);
-       ctx->i64_0 = LLVMConstInt(ctx->i64, 0, false);
-       ctx->i64_1 = LLVMConstInt(ctx->i64, 1, false);
-       ctx->i128_0 = LLVMConstInt(ctx->i128, 0, false);
-       ctx->i128_1 = LLVMConstInt(ctx->i128, 1, false);
-       ctx->f16_0 = LLVMConstReal(ctx->f16, 0.0);
-       ctx->f16_1 = LLVMConstReal(ctx->f16, 1.0);
-       ctx->f32_0 = LLVMConstReal(ctx->f32, 0.0);
-       ctx->f32_1 = LLVMConstReal(ctx->f32, 1.0);
-       ctx->f64_0 = LLVMConstReal(ctx->f64, 0.0);
-       ctx->f64_1 = LLVMConstReal(ctx->f64, 1.0);
-
-       ctx->i1false = LLVMConstInt(ctx->i1, 0, false);
-       ctx->i1true = LLVMConstInt(ctx->i1, 1, false);
-
-       ctx->range_md_kind = LLVMGetMDKindIDInContext(ctx->context,
-                                                    "range", 5);
-
-       ctx->invariant_load_md_kind = LLVMGetMDKindIDInContext(ctx->context,
-                                                              "invariant.load", 14);
-
-       ctx->uniform_md_kind = LLVMGetMDKindIDInContext(ctx->context,
-                                                       "amdgpu.uniform", 14);
-
-       ctx->empty_md = LLVMMDNodeInContext(ctx->context, NULL, 0);
-       ctx->flow = calloc(1, sizeof(*ctx->flow));
-}
-
-void
-ac_llvm_context_dispose(struct ac_llvm_context *ctx)
-{
-       free(ctx->flow->stack);
-       free(ctx->flow);
-       ctx->flow = NULL;
-}
-
-int
-ac_get_llvm_num_components(LLVMValueRef value)
-{
-       LLVMTypeRef type = LLVMTypeOf(value);
-       unsigned num_components = LLVMGetTypeKind(type) == LLVMVectorTypeKind
-                                     ? LLVMGetVectorSize(type)
-                                     : 1;
-       return num_components;
-}
-
-LLVMValueRef
-ac_llvm_extract_elem(struct ac_llvm_context *ac,
-                    LLVMValueRef value,
-                    int index)
-{
-       if (LLVMGetTypeKind(LLVMTypeOf(value)) != LLVMVectorTypeKind) {
-               assert(index == 0);
-               return value;
-       }
-
-       return LLVMBuildExtractElement(ac->builder, value,
-                                      LLVMConstInt(ac->i32, index, false), "");
-}
-
-int
-ac_get_elem_bits(struct ac_llvm_context *ctx, LLVMTypeRef type)
-{
-       if (LLVMGetTypeKind(type) == LLVMVectorTypeKind)
-               type = LLVMGetElementType(type);
-
-       if (LLVMGetTypeKind(type) == LLVMIntegerTypeKind)
-               return LLVMGetIntTypeWidth(type);
-
-       if (LLVMGetTypeKind(type) == LLVMPointerTypeKind) {
-               if (LLVMGetPointerAddressSpace(type) == AC_ADDR_SPACE_LDS)
-                       return 32;
-       }
-
-       if (type == ctx->f16)
-               return 16;
-       if (type == ctx->f32)
-               return 32;
-       if (type == ctx->f64)
-               return 64;
-
-       unreachable("Unhandled type kind in get_elem_bits");
-}
-
-unsigned
-ac_get_type_size(LLVMTypeRef type)
-{
-       LLVMTypeKind kind = LLVMGetTypeKind(type);
-
-       switch (kind) {
-       case LLVMIntegerTypeKind:
-               return LLVMGetIntTypeWidth(type) / 8;
-       case LLVMHalfTypeKind:
-               return 2;
-       case LLVMFloatTypeKind:
-               return 4;
-       case LLVMDoubleTypeKind:
-               return 8;
-       case LLVMPointerTypeKind:
-               if (LLVMGetPointerAddressSpace(type) == AC_ADDR_SPACE_CONST_32BIT)
-                       return 4;
-               return 8;
-       case LLVMVectorTypeKind:
-               return LLVMGetVectorSize(type) *
-                      ac_get_type_size(LLVMGetElementType(type));
-       case LLVMArrayTypeKind:
-               return LLVMGetArrayLength(type) *
-                      ac_get_type_size(LLVMGetElementType(type));
-       default:
-               assert(0);
-               return 0;
-       }
+void ac_llvm_context_init(struct ac_llvm_context *ctx, struct ac_llvm_compiler *compiler,
+                          enum chip_class chip_class, enum radeon_family family,
+                          enum ac_float_mode float_mode, unsigned wave_size,
+                          unsigned ballot_mask_bits)
+{
+   ctx->context = LLVMContextCreate();
+
+   ctx->chip_class = chip_class;
+   ctx->family = family;
+   ctx->wave_size = wave_size;
+   ctx->ballot_mask_bits = ballot_mask_bits;
+   ctx->float_mode = float_mode;
+   ctx->module =
+      ac_create_module(wave_size == 32 ? compiler->tm_wave32 : compiler->tm, ctx->context);
+   ctx->builder = ac_create_builder(ctx->context, float_mode);
+
+   ctx->voidt = LLVMVoidTypeInContext(ctx->context);
+   ctx->i1 = LLVMInt1TypeInContext(ctx->context);
+   ctx->i8 = LLVMInt8TypeInContext(ctx->context);
+   ctx->i16 = LLVMIntTypeInContext(ctx->context, 16);
+   ctx->i32 = LLVMIntTypeInContext(ctx->context, 32);
+   ctx->i64 = LLVMIntTypeInContext(ctx->context, 64);
+   ctx->i128 = LLVMIntTypeInContext(ctx->context, 128);
+   ctx->intptr = ctx->i32;
+   ctx->f16 = LLVMHalfTypeInContext(ctx->context);
+   ctx->f32 = LLVMFloatTypeInContext(ctx->context);
+   ctx->f64 = LLVMDoubleTypeInContext(ctx->context);
+   ctx->v2i16 = LLVMVectorType(ctx->i16, 2);
+   ctx->v4i16 = LLVMVectorType(ctx->i16, 4);
+   ctx->v2f16 = LLVMVectorType(ctx->f16, 2);
+   ctx->v4f16 = LLVMVectorType(ctx->f16, 4);
+   ctx->v2i32 = LLVMVectorType(ctx->i32, 2);
+   ctx->v3i32 = LLVMVectorType(ctx->i32, 3);
+   ctx->v4i32 = LLVMVectorType(ctx->i32, 4);
+   ctx->v2f32 = LLVMVectorType(ctx->f32, 2);
+   ctx->v3f32 = LLVMVectorType(ctx->f32, 3);
+   ctx->v4f32 = LLVMVectorType(ctx->f32, 4);
+   ctx->v8i32 = LLVMVectorType(ctx->i32, 8);
+   ctx->iN_wavemask = LLVMIntTypeInContext(ctx->context, ctx->wave_size);
+   ctx->iN_ballotmask = LLVMIntTypeInContext(ctx->context, ballot_mask_bits);
+
+   ctx->i8_0 = LLVMConstInt(ctx->i8, 0, false);
+   ctx->i8_1 = LLVMConstInt(ctx->i8, 1, false);
+   ctx->i16_0 = LLVMConstInt(ctx->i16, 0, false);
+   ctx->i16_1 = LLVMConstInt(ctx->i16, 1, false);
+   ctx->i32_0 = LLVMConstInt(ctx->i32, 0, false);
+   ctx->i32_1 = LLVMConstInt(ctx->i32, 1, false);
+   ctx->i64_0 = LLVMConstInt(ctx->i64, 0, false);
+   ctx->i64_1 = LLVMConstInt(ctx->i64, 1, false);
+   ctx->i128_0 = LLVMConstInt(ctx->i128, 0, false);
+   ctx->i128_1 = LLVMConstInt(ctx->i128, 1, false);
+   ctx->f16_0 = LLVMConstReal(ctx->f16, 0.0);
+   ctx->f16_1 = LLVMConstReal(ctx->f16, 1.0);
+   ctx->f32_0 = LLVMConstReal(ctx->f32, 0.0);
+   ctx->f32_1 = LLVMConstReal(ctx->f32, 1.0);
+   ctx->f64_0 = LLVMConstReal(ctx->f64, 0.0);
+   ctx->f64_1 = LLVMConstReal(ctx->f64, 1.0);
+
+   ctx->i1false = LLVMConstInt(ctx->i1, 0, false);
+   ctx->i1true = LLVMConstInt(ctx->i1, 1, false);
+
+   ctx->range_md_kind = LLVMGetMDKindIDInContext(ctx->context, "range", 5);
+
+   ctx->invariant_load_md_kind = LLVMGetMDKindIDInContext(ctx->context, "invariant.load", 14);
+
+   ctx->uniform_md_kind = LLVMGetMDKindIDInContext(ctx->context, "amdgpu.uniform", 14);
+
+   ctx->empty_md = LLVMMDNodeInContext(ctx->context, NULL, 0);
+   ctx->flow = calloc(1, sizeof(*ctx->flow));
+}
+
+void ac_llvm_context_dispose(struct ac_llvm_context *ctx)
+{
+   free(ctx->flow->stack);
+   free(ctx->flow);
+   ctx->flow = NULL;
+}
+
+int ac_get_llvm_num_components(LLVMValueRef value)
+{
+   LLVMTypeRef type = LLVMTypeOf(value);
+   unsigned num_components =
+      LLVMGetTypeKind(type) == LLVMVectorTypeKind ? LLVMGetVectorSize(type) : 1;
+   return num_components;
+}
+
+LLVMValueRef ac_llvm_extract_elem(struct ac_llvm_context *ac, LLVMValueRef value, int index)
+{
+   if (LLVMGetTypeKind(LLVMTypeOf(value)) != LLVMVectorTypeKind) {
+      assert(index == 0);
+      return value;
+   }
+
+   return LLVMBuildExtractElement(ac->builder, value, LLVMConstInt(ac->i32, index, false), "");
+}
+
+int ac_get_elem_bits(struct ac_llvm_context *ctx, LLVMTypeRef type)
+{
+   if (LLVMGetTypeKind(type) == LLVMVectorTypeKind)
+      type = LLVMGetElementType(type);
+
+   if (LLVMGetTypeKind(type) == LLVMIntegerTypeKind)
+      return LLVMGetIntTypeWidth(type);
+
+   if (LLVMGetTypeKind(type) == LLVMPointerTypeKind) {
+      if (LLVMGetPointerAddressSpace(type) == AC_ADDR_SPACE_LDS)
+         return 32;
+   }
+
+   if (type == ctx->f16)
+      return 16;
+   if (type == ctx->f32)
+      return 32;
+   if (type == ctx->f64)
+      return 64;
+
+   unreachable("Unhandled type kind in get_elem_bits");
+}
+
+unsigned ac_get_type_size(LLVMTypeRef type)
+{
+   LLVMTypeKind kind = LLVMGetTypeKind(type);
+
+   switch (kind) {
+   case LLVMIntegerTypeKind:
+      return LLVMGetIntTypeWidth(type) / 8;
+   case LLVMHalfTypeKind:
+      return 2;
+   case LLVMFloatTypeKind:
+      return 4;
+   case LLVMDoubleTypeKind:
+      return 8;
+   case LLVMPointerTypeKind:
+      if (LLVMGetPointerAddressSpace(type) == AC_ADDR_SPACE_CONST_32BIT)
+         return 4;
+      return 8;
+   case LLVMVectorTypeKind:
+      return LLVMGetVectorSize(type) * ac_get_type_size(LLVMGetElementType(type));
+   case LLVMArrayTypeKind:
+      return LLVMGetArrayLength(type) * ac_get_type_size(LLVMGetElementType(type));
+   default:
+      assert(0);
+      return 0;
+   }
 }
 
 static LLVMTypeRef to_integer_type_scalar(struct ac_llvm_context *ctx, LLVMTypeRef t)
 {
-       if (t == ctx->i8)
-               return ctx->i8;
-       else if (t == ctx->f16 || t == ctx->i16)
-               return ctx->i16;
-       else if (t == ctx->f32 || t == ctx->i32)
-               return ctx->i32;
-       else if (t == ctx->f64 || t == ctx->i64)
-               return ctx->i64;
-       else
-               unreachable("Unhandled integer size");
-}
-
-LLVMTypeRef
-ac_to_integer_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
-{
-       if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
-               LLVMTypeRef elem_type = LLVMGetElementType(t);
-               return LLVMVectorType(to_integer_type_scalar(ctx, elem_type),
-                                     LLVMGetVectorSize(t));
-       }
-       if (LLVMGetTypeKind(t) == LLVMPointerTypeKind) {
-               switch (LLVMGetPointerAddressSpace(t)) {
-               case AC_ADDR_SPACE_GLOBAL:
-                       return ctx->i64;
-               case AC_ADDR_SPACE_CONST_32BIT:
-               case AC_ADDR_SPACE_LDS:
-                       return ctx->i32;
-               default:
-                       unreachable("unhandled address space");
-               }
-       }
-       return to_integer_type_scalar(ctx, t);
-}
-
-LLVMValueRef
-ac_to_integer(struct ac_llvm_context *ctx, LLVMValueRef v)
-{
-       LLVMTypeRef type = LLVMTypeOf(v);
-       if (LLVMGetTypeKind(type) == LLVMPointerTypeKind) {
-               return LLVMBuildPtrToInt(ctx->builder, v, ac_to_integer_type(ctx, type), "");
-       }
-       return LLVMBuildBitCast(ctx->builder, v, ac_to_integer_type(ctx, type), "");
-}
-
-LLVMValueRef
-ac_to_integer_or_pointer(struct ac_llvm_context *ctx, LLVMValueRef v)
-{
-       LLVMTypeRef type = LLVMTypeOf(v);
-       if (LLVMGetTypeKind(type) == LLVMPointerTypeKind)
-               return v;
-       return ac_to_integer(ctx, v);
+   if (t == ctx->i8)
+      return ctx->i8;
+   else if (t == ctx->f16 || t == ctx->i16)
+      return ctx->i16;
+   else if (t == ctx->f32 || t == ctx->i32)
+      return ctx->i32;
+   else if (t == ctx->f64 || t == ctx->i64)
+      return ctx->i64;
+   else
+      unreachable("Unhandled integer size");
 }
 
-static LLVMTypeRef to_float_type_scalar(struct ac_llvm_context *ctx, LLVMTypeRef t)
+LLVMTypeRef ac_to_integer_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
+{
+   if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
+      LLVMTypeRef elem_type = LLVMGetElementType(t);
+      return LLVMVectorType(to_integer_type_scalar(ctx, elem_type), LLVMGetVectorSize(t));
+   }
+   if (LLVMGetTypeKind(t) == LLVMPointerTypeKind) {
+      switch (LLVMGetPointerAddressSpace(t)) {
+      case AC_ADDR_SPACE_GLOBAL:
+         return ctx->i64;
+      case AC_ADDR_SPACE_CONST_32BIT:
+      case AC_ADDR_SPACE_LDS:
+         return ctx->i32;
+      default:
+         unreachable("unhandled address space");
+      }
+   }
+   return to_integer_type_scalar(ctx, t);
+}
+
+LLVMValueRef ac_to_integer(struct ac_llvm_context *ctx, LLVMValueRef v)
 {
-       if (t == ctx->i8)
-               return ctx->i8;
-       else if (t == ctx->i16 || t == ctx->f16)
-               return ctx->f16;
-       else if (t == ctx->i32 || t == ctx->f32)
-               return ctx->f32;
-       else if (t == ctx->i64 || t == ctx->f64)
-               return ctx->f64;
-       else
-               unreachable("Unhandled float size");
+   LLVMTypeRef type = LLVMTypeOf(v);
+   if (LLVMGetTypeKind(type) == LLVMPointerTypeKind) {
+      return LLVMBuildPtrToInt(ctx->builder, v, ac_to_integer_type(ctx, type), "");
+   }
+   return LLVMBuildBitCast(ctx->builder, v, ac_to_integer_type(ctx, type), "");
 }
 
-LLVMTypeRef
-ac_to_float_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
+LLVMValueRef ac_to_integer_or_pointer(struct ac_llvm_context *ctx, LLVMValueRef v)
 {
-       if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
-               LLVMTypeRef elem_type = LLVMGetElementType(t);
-               return LLVMVectorType(to_float_type_scalar(ctx, elem_type),
-                                     LLVMGetVectorSize(t));
-       }
-       return to_float_type_scalar(ctx, t);
+   LLVMTypeRef type = LLVMTypeOf(v);
+   if (LLVMGetTypeKind(type) == LLVMPointerTypeKind)
+      return v;
+   return ac_to_integer(ctx, v);
 }
 
-LLVMValueRef
-ac_to_float(struct ac_llvm_context *ctx, LLVMValueRef v)
+static LLVMTypeRef to_float_type_scalar(struct ac_llvm_context *ctx, LLVMTypeRef t)
 {
-       LLVMTypeRef type = LLVMTypeOf(v);
-       return LLVMBuildBitCast(ctx->builder, v, ac_to_float_type(ctx, type), "");
+   if (t == ctx->i8)
+      return ctx->i8;
+   else if (t == ctx->i16 || t == ctx->f16)
+      return ctx->f16;
+   else if (t == ctx->i32 || t == ctx->f32)
+      return ctx->f32;
+   else if (t == ctx->i64 || t == ctx->f64)
+      return ctx->f64;
+   else
+      unreachable("Unhandled float size");
 }
 
+LLVMTypeRef ac_to_float_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
+{
+   if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
+      LLVMTypeRef elem_type = LLVMGetElementType(t);
+      return LLVMVectorType(to_float_type_scalar(ctx, elem_type), LLVMGetVectorSize(t));
+   }
+   return to_float_type_scalar(ctx, t);
+}
 
-LLVMValueRef
-ac_build_intrinsic(struct ac_llvm_context *ctx, const char *name,
-                  LLVMTypeRef return_type, LLVMValueRef *params,
-                  unsigned param_count, unsigned attrib_mask)
+LLVMValueRef ac_to_float(struct ac_llvm_context *ctx, LLVMValueRef v)
 {
-       LLVMValueRef function, call;
-       bool set_callsite_attrs = !(attrib_mask & AC_FUNC_ATTR_LEGACY);
+   LLVMTypeRef type = LLVMTypeOf(v);
+   return LLVMBuildBitCast(ctx->builder, v, ac_to_float_type(ctx, type), "");
+}
 
-       function = LLVMGetNamedFunction(ctx->module, name);
-       if (!function) {
-               LLVMTypeRef param_types[32], function_type;
-               unsigned i;
+LLVMValueRef ac_build_intrinsic(struct ac_llvm_context *ctx, const char *name,
+                                LLVMTypeRef return_type, LLVMValueRef *params, unsigned param_count,
+                                unsigned attrib_mask)
+{
+   LLVMValueRef function, call;
+   bool set_callsite_attrs = !(attrib_mask & AC_FUNC_ATTR_LEGACY);
 
-               assert(param_count <= 32);
+   function = LLVMGetNamedFunction(ctx->module, name);
+   if (!function) {
+      LLVMTypeRef param_types[32], function_type;
+      unsigned i;
 
-               for (i = 0; i < param_count; ++i) {
-                       assert(params[i]);
-                       param_types[i] = LLVMTypeOf(params[i]);
-               }
-               function_type =
-                   LLVMFunctionType(return_type, param_types, param_count, 0);
-               function = LLVMAddFunction(ctx->module, name, function_type);
+      assert(param_count <= 32);
 
-               LLVMSetFunctionCallConv(function, LLVMCCallConv);
-               LLVMSetLinkage(function, LLVMExternalLinkage);
+      for (i = 0; i < param_count; ++i) {
+         assert(params[i]);
+         param_types[i] = LLVMTypeOf(params[i]);
+      }
+      function_type = LLVMFunctionType(return_type, param_types, param_count, 0);
+      function = LLVMAddFunction(ctx->module, name, function_type);
 
-               if (!set_callsite_attrs)
-                       ac_add_func_attributes(ctx->context, function, attrib_mask);
-       }
+      LLVMSetFunctionCallConv(function, LLVMCCallConv);
+      LLVMSetLinkage(function, LLVMExternalLinkage);
 
-       call = LLVMBuildCall(ctx->builder, function, params, param_count, "");
-       if (set_callsite_attrs)
-               ac_add_func_attributes(ctx->context, call, attrib_mask);
-       return call;
+      if (!set_callsite_attrs)
+         ac_add_func_attributes(ctx->context, function, attrib_mask);
+   }
+
+   call = LLVMBuildCall(ctx->builder, function, params, param_count, "");
+   if (set_callsite_attrs)
+      ac_add_func_attributes(ctx->context, call, attrib_mask);
+   return call;
 }
 
 /**
@@ -351,59 +321,55 @@ ac_build_intrinsic(struct ac_llvm_context *ctx, const char *name,
  */
 void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize)
 {
-       LLVMTypeRef elem_type = type;
-
-       assert(bufsize >= 8);
-
-       if (LLVMGetTypeKind(type) == LLVMVectorTypeKind) {
-               int ret = snprintf(buf, bufsize, "v%u",
-                                       LLVMGetVectorSize(type));
-               if (ret < 0) {
-                       char *type_name = LLVMPrintTypeToString(type);
-                       fprintf(stderr, "Error building type name for: %s\n",
-                               type_name);
-                       LLVMDisposeMessage(type_name);
-                       return;
-               }
-               elem_type = LLVMGetElementType(type);
-               buf += ret;
-               bufsize -= ret;
-       }
-       switch (LLVMGetTypeKind(elem_type)) {
-       default: break;
-       case LLVMIntegerTypeKind:
-               snprintf(buf, bufsize, "i%d", LLVMGetIntTypeWidth(elem_type));
-               break;
-       case LLVMHalfTypeKind:
-               snprintf(buf, bufsize, "f16");
-               break;
-       case LLVMFloatTypeKind:
-               snprintf(buf, bufsize, "f32");
-               break;
-       case LLVMDoubleTypeKind:
-               snprintf(buf, bufsize, "f64");
-               break;
-       }
+   LLVMTypeRef elem_type = type;
+
+   assert(bufsize >= 8);
+
+   if (LLVMGetTypeKind(type) == LLVMVectorTypeKind) {
+      int ret = snprintf(buf, bufsize, "v%u", LLVMGetVectorSize(type));
+      if (ret < 0) {
+         char *type_name = LLVMPrintTypeToString(type);
+         fprintf(stderr, "Error building type name for: %s\n", type_name);
+         LLVMDisposeMessage(type_name);
+         return;
+      }
+      elem_type = LLVMGetElementType(type);
+      buf += ret;
+      bufsize -= ret;
+   }
+   switch (LLVMGetTypeKind(elem_type)) {
+   default:
+      break;
+   case LLVMIntegerTypeKind:
+      snprintf(buf, bufsize, "i%d", LLVMGetIntTypeWidth(elem_type));
+      break;
+   case LLVMHalfTypeKind:
+      snprintf(buf, bufsize, "f16");
+      break;
+   case LLVMFloatTypeKind:
+      snprintf(buf, bufsize, "f32");
+      break;
+   case LLVMDoubleTypeKind:
+      snprintf(buf, bufsize, "f64");
+      break;
+   }
 }
 
 /**
  * Helper function that builds an LLVM IR PHI node and immediately adds
  * incoming edges.
  */
-LLVMValueRef
-ac_build_phi(struct ac_llvm_context *ctx, LLVMTypeRef type,
-            unsigned count_incoming, LLVMValueRef *values,
-            LLVMBasicBlockRef *blocks)
+LLVMValueRef ac_build_phi(struct ac_llvm_context *ctx, LLVMTypeRef type, unsigned count_incoming,
+                          LLVMValueRef *values, LLVMBasicBlockRef *blocks)
 {
-       LLVMValueRef phi = LLVMBuildPhi(ctx->builder, type, "");
-       LLVMAddIncoming(phi, values, blocks, count_incoming);
-       return phi;
+   LLVMValueRef phi = LLVMBuildPhi(ctx->builder, type, "");
+   LLVMAddIncoming(phi, values, blocks, count_incoming);
+   return phi;
 }
 
 void ac_build_s_barrier(struct ac_llvm_context *ctx)
 {
-       ac_build_intrinsic(ctx, "llvm.amdgcn.s.barrier", ctx->voidt, NULL,
-                          0, AC_FUNC_ATTR_CONVERGENT);
+   ac_build_intrinsic(ctx, "llvm.amdgcn.s.barrier", ctx->voidt, NULL, 0, AC_FUNC_ATTR_CONVERGENT);
 }
 
 /* Prevent optimizations (at least of memory accesses) across the current
@@ -413,375 +379,328 @@ void ac_build_s_barrier(struct ac_llvm_context *ctx)
  * Optionally, a value can be passed through the inline assembly to prevent
  * LLVM from hoisting calls to ReadNone functions.
  */
-void
-ac_build_optimization_barrier(struct ac_llvm_context *ctx,
-                             LLVMValueRef *pvgpr)
-{
-       static int counter = 0;
-
-       LLVMBuilderRef builder = ctx->builder;
-       char code[16];
-
-       snprintf(code, sizeof(code), "; %d", p_atomic_inc_return(&counter));
-
-       if (!pvgpr) {
-               LLVMTypeRef ftype = LLVMFunctionType(ctx->voidt, NULL, 0, false);
-               LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "", true, false);
-               LLVMBuildCall(builder, inlineasm, NULL, 0, "");
-       } else {
-               LLVMTypeRef ftype = LLVMFunctionType(ctx->i32, &ctx->i32, 1, false);
-               LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "=v,0", true, false);
-               LLVMTypeRef type = LLVMTypeOf(*pvgpr);
-               unsigned bitsize = ac_get_elem_bits(ctx, type);
-               LLVMValueRef vgpr = *pvgpr;
-               LLVMTypeRef vgpr_type;
-               unsigned vgpr_size;
-               LLVMValueRef vgpr0;
+void ac_build_optimization_barrier(struct ac_llvm_context *ctx, LLVMValueRef *pvgpr)
+{
+   static int counter = 0;
+
+   LLVMBuilderRef builder = ctx->builder;
+   char code[16];
+
+   snprintf(code, sizeof(code), "; %d", p_atomic_inc_return(&counter));
 
-               if (bitsize < 32)
-                       vgpr = LLVMBuildZExt(ctx->builder, vgpr, ctx->i32, "");
+   if (!pvgpr) {
+      LLVMTypeRef ftype = LLVMFunctionType(ctx->voidt, NULL, 0, false);
+      LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "", true, false);
+      LLVMBuildCall(builder, inlineasm, NULL, 0, "");
+   } else {
+      LLVMTypeRef ftype = LLVMFunctionType(ctx->i32, &ctx->i32, 1, false);
+      LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "=v,0", true, false);
+      LLVMTypeRef type = LLVMTypeOf(*pvgpr);
+      unsigned bitsize = ac_get_elem_bits(ctx, type);
+      LLVMValueRef vgpr = *pvgpr;
+      LLVMTypeRef vgpr_type;
+      unsigned vgpr_size;
+      LLVMValueRef vgpr0;
 
-               vgpr_type = LLVMTypeOf(vgpr);
-               vgpr_size = ac_get_type_size(vgpr_type);
+      if (bitsize < 32)
+         vgpr = LLVMBuildZExt(ctx->builder, vgpr, ctx->i32, "");
 
-               assert(vgpr_size % 4 == 0);
+      vgpr_type = LLVMTypeOf(vgpr);
+      vgpr_size = ac_get_type_size(vgpr_type);
 
-               vgpr = LLVMBuildBitCast(builder, vgpr, LLVMVectorType(ctx->i32, vgpr_size / 4), "");
-               vgpr0 = LLVMBuildExtractElement(builder, vgpr, ctx->i32_0, "");
-               vgpr0 = LLVMBuildCall(builder, inlineasm, &vgpr0, 1, "");
-               vgpr = LLVMBuildInsertElement(builder, vgpr, vgpr0, ctx->i32_0, "");
-               vgpr = LLVMBuildBitCast(builder, vgpr, vgpr_type, "");
+      assert(vgpr_size % 4 == 0);
 
-               if (bitsize < 32)
-                       vgpr = LLVMBuildTrunc(builder, vgpr, type, "");
+      vgpr = LLVMBuildBitCast(builder, vgpr, LLVMVectorType(ctx->i32, vgpr_size / 4), "");
+      vgpr0 = LLVMBuildExtractElement(builder, vgpr, ctx->i32_0, "");
+      vgpr0 = LLVMBuildCall(builder, inlineasm, &vgpr0, 1, "");
+      vgpr = LLVMBuildInsertElement(builder, vgpr, vgpr0, ctx->i32_0, "");
+      vgpr = LLVMBuildBitCast(builder, vgpr, vgpr_type, "");
 
-               *pvgpr = vgpr;
-       }
+      if (bitsize < 32)
+         vgpr = LLVMBuildTrunc(builder, vgpr, type, "");
+
+      *pvgpr = vgpr;
+   }
 }
 
-LLVMValueRef
-ac_build_shader_clock(struct ac_llvm_context *ctx, nir_scope scope)
+LLVMValueRef ac_build_shader_clock(struct ac_llvm_context *ctx, nir_scope scope)
 {
-       const char *name = scope == NIR_SCOPE_DEVICE ? "llvm.amdgcn.s.memrealtime" : "llvm.amdgcn.s.memtime";
-       LLVMValueRef tmp = ac_build_intrinsic(ctx, name, ctx->i64, NULL, 0, 0);
-       return LLVMBuildBitCast(ctx->builder, tmp, ctx->v2i32, "");
+   const char *name =
+      scope == NIR_SCOPE_DEVICE ? "llvm.amdgcn.s.memrealtime" : "llvm.amdgcn.s.memtime";
+   LLVMValueRef tmp = ac_build_intrinsic(ctx, name, ctx->i64, NULL, 0, 0);
+   return LLVMBuildBitCast(ctx->builder, tmp, ctx->v2i32, "");
 }
 
-LLVMValueRef
-ac_build_ballot(struct ac_llvm_context *ctx,
-               LLVMValueRef value)
+LLVMValueRef ac_build_ballot(struct ac_llvm_context *ctx, LLVMValueRef value)
 {
-       const char *name;
+   const char *name;
 
-       if (LLVM_VERSION_MAJOR >= 9) {
-               if (ctx->wave_size == 64)
-                       name = "llvm.amdgcn.icmp.i64.i32";
-               else
-                       name = "llvm.amdgcn.icmp.i32.i32";
-       } else {
-               name = "llvm.amdgcn.icmp.i32";
-       }
-       LLVMValueRef args[3] = {
-               value,
-               ctx->i32_0,
-               LLVMConstInt(ctx->i32, LLVMIntNE, 0)
-       };
+   if (LLVM_VERSION_MAJOR >= 9) {
+      if (ctx->wave_size == 64)
+         name = "llvm.amdgcn.icmp.i64.i32";
+      else
+         name = "llvm.amdgcn.icmp.i32.i32";
+   } else {
+      name = "llvm.amdgcn.icmp.i32";
+   }
+   LLVMValueRef args[3] = {value, ctx->i32_0, LLVMConstInt(ctx->i32, LLVMIntNE, 0)};
 
-       /* We currently have no other way to prevent LLVM from lifting the icmp
-        * calls to a dominating basic block.
-        */
-       ac_build_optimization_barrier(ctx, &args[0]);
+   /* We currently have no other way to prevent LLVM from lifting the icmp
+    * calls to a dominating basic block.
+    */
+   ac_build_optimization_barrier(ctx, &args[0]);
 
-       args[0] = ac_to_integer(ctx, args[0]);
+   args[0] = ac_to_integer(ctx, args[0]);
 
-       return ac_build_intrinsic(ctx, name, ctx->iN_wavemask, args, 3,
-                                 AC_FUNC_ATTR_NOUNWIND |
-                                 AC_FUNC_ATTR_READNONE |
-                                 AC_FUNC_ATTR_CONVERGENT);
+   return ac_build_intrinsic(
+      ctx, name, ctx->iN_wavemask, args, 3,
+      AC_FUNC_ATTR_NOUNWIND | AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
 }
 
-LLVMValueRef ac_get_i1_sgpr_mask(struct ac_llvm_context *ctx,
-                                LLVMValueRef value)
+LLVMValueRef ac_get_i1_sgpr_mask(struct ac_llvm_context *ctx, LLVMValueRef value)
 {
-       const char *name;
-
-       if (LLVM_VERSION_MAJOR >= 9) {
-               if (ctx->wave_size == 64)
-                       name = "llvm.amdgcn.icmp.i64.i1";
-               else
-                       name = "llvm.amdgcn.icmp.i32.i1";
-       } else {
-               name = "llvm.amdgcn.icmp.i1";
-       }
-       LLVMValueRef args[3] = {
-               value,
-               ctx->i1false,
-               LLVMConstInt(ctx->i32, LLVMIntNE, 0),
-       };
+   const char *name;
+
+   if (LLVM_VERSION_MAJOR >= 9) {
+      if (ctx->wave_size == 64)
+         name = "llvm.amdgcn.icmp.i64.i1";
+      else
+         name = "llvm.amdgcn.icmp.i32.i1";
+   } else {
+      name = "llvm.amdgcn.icmp.i1";
+   }
+   LLVMValueRef args[3] = {
+      value,
+      ctx->i1false,
+      LLVMConstInt(ctx->i32, LLVMIntNE, 0),
+   };
 
-       return ac_build_intrinsic(ctx, name, ctx->iN_wavemask, args, 3,
-                                 AC_FUNC_ATTR_NOUNWIND |
-                                 AC_FUNC_ATTR_READNONE |
-                                 AC_FUNC_ATTR_CONVERGENT);
+   return ac_build_intrinsic(
+      ctx, name, ctx->iN_wavemask, args, 3,
+      AC_FUNC_ATTR_NOUNWIND | AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
 }
 
-LLVMValueRef
-ac_build_vote_all(struct ac_llvm_context *ctx, LLVMValueRef value)
+LLVMValueRef ac_build_vote_all(struct ac_llvm_context *ctx, LLVMValueRef value)
 {
-       LLVMValueRef active_set = ac_build_ballot(ctx, ctx->i32_1);
-       LLVMValueRef vote_set = ac_build_ballot(ctx, value);
-       return LLVMBuildICmp(ctx->builder, LLVMIntEQ, vote_set, active_set, "");
+   LLVMValueRef active_set = ac_build_ballot(ctx, ctx->i32_1);
+   LLVMValueRef vote_set = ac_build_ballot(ctx, value);
+   return LLVMBuildICmp(ctx->builder, LLVMIntEQ, vote_set, active_set, "");
 }
 
-LLVMValueRef
-ac_build_vote_any(struct ac_llvm_context *ctx, LLVMValueRef value)
+LLVMValueRef ac_build_vote_any(struct ac_llvm_context *ctx, LLVMValueRef value)
 {
-       LLVMValueRef vote_set = ac_build_ballot(ctx, value);
-       return LLVMBuildICmp(ctx->builder, LLVMIntNE, vote_set,
-                            LLVMConstInt(ctx->iN_wavemask, 0, 0), "");
+   LLVMValueRef vote_set = ac_build_ballot(ctx, value);
+   return LLVMBuildICmp(ctx->builder, LLVMIntNE, vote_set, LLVMConstInt(ctx->iN_wavemask, 0, 0),
+                        "");
 }
 
-LLVMValueRef
-ac_build_vote_eq(struct ac_llvm_context *ctx, LLVMValueRef value)
+LLVMValueRef ac_build_vote_eq(struct ac_llvm_context *ctx, LLVMValueRef value)
 {
-       LLVMValueRef active_set = ac_build_ballot(ctx, ctx->i32_1);
-       LLVMValueRef vote_set = ac_build_ballot(ctx, value);
+   LLVMValueRef active_set = ac_build_ballot(ctx, ctx->i32_1);
+   LLVMValueRef vote_set = ac_build_ballot(ctx, value);
 
-       LLVMValueRef all = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
-                                        vote_set, active_set, "");
-       LLVMValueRef none = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
-                                         vote_set,
-                                         LLVMConstInt(ctx->iN_wavemask, 0, 0), "");
-       return LLVMBuildOr(ctx->builder, all, none, "");
+   LLVMValueRef all = LLVMBuildICmp(ctx->builder, LLVMIntEQ, vote_set, active_set, "");
+   LLVMValueRef none =
+      LLVMBuildICmp(ctx->builder, LLVMIntEQ, vote_set, LLVMConstInt(ctx->iN_wavemask, 0, 0), "");
+   return LLVMBuildOr(ctx->builder, all, none, "");
 }
 
-LLVMValueRef
-ac_build_varying_gather_values(struct ac_llvm_context *ctx, LLVMValueRef *values,
-                              unsigned value_count, unsigned component)
+LLVMValueRef ac_build_varying_gather_values(struct ac_llvm_context *ctx, LLVMValueRef *values,
+                                            unsigned value_count, unsigned component)
 {
-       LLVMValueRef vec = NULL;
+   LLVMValueRef vec = NULL;
 
-       if (value_count == 1) {
-               return values[component];
-       } else if (!value_count)
-               unreachable("value_count is 0");
+   if (value_count == 1) {
+      return values[component];
+   } else if (!value_count)
+      unreachable("value_count is 0");
 
-       for (unsigned i = component; i < value_count + component; i++) {
-               LLVMValueRef value = values[i];
+   for (unsigned i = component; i < value_count + component; i++) {
+      LLVMValueRef value = values[i];
 
-               if (i == component)
-                       vec = LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value), value_count));
-               LLVMValueRef index = LLVMConstInt(ctx->i32, i - component, false);
-               vec = LLVMBuildInsertElement(ctx->builder, vec, value, index, "");
-       }
-       return vec;
+      if (i == component)
+         vec = LLVMGetUndef(LLVMVectorType(LLVMTypeOf(value), value_count));
+      LLVMValueRef index = LLVMConstInt(ctx->i32, i - component, false);
+      vec = LLVMBuildInsertElement(ctx->builder, vec, value, index, "");
+   }
+   return vec;
 }
 
-LLVMValueRef
-ac_build_gather_values_extended(struct ac_llvm_context *ctx,
-                               LLVMValueRef *values,
-                               unsigned value_count,
-                               unsigned value_stride,
-                               bool load,
-                               bool always_vector)
+LLVMValueRef ac_build_gather_values_extended(struct ac_llvm_context *ctx, LLVMValueRef *values,
+                                             unsigned value_count, unsigned value_stride, bool load,
+                                             bool always_vector)
 {
-       LLVMBuilderRef builder = ctx->builder;
-       LLVMValueRef vec = NULL;
-       unsigned i;
+   LLVMBuilderRef builder = ctx->builder;
+   LLVMValueRef vec = NULL;
+   unsigned i;
 
-       if (value_count == 1 && !always_vector) {
-               if (load)
-                       return LLVMBuildLoad(builder, values[0], "");
-               return values[0];
-       } else if (!value_count)
-               unreachable("value_count is 0");
+   if (value_count == 1 && !always_vector) {
+      if (load)
+         return LLVMBuildLoad(builder, values[0], "");
+      return values[0];
+   } else if (!value_count)
+      unreachable("value_count is 0");
 
-       for (i = 0; i < value_count; i++) {
-               LLVMValueRef value = values[i * value_stride];
-               if (load)
-                       value = LLVMBuildLoad(builder, value, "");
+   for (i = 0; i < value_count; i++) {
+      LLVMValueRef value = values[i * value_stride];
+      if (load)
+         value = LLVMBuildLoad(builder, value, "");
 
-               if (!i)
-                       vec = LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value), value_count));
-               LLVMValueRef index = LLVMConstInt(ctx->i32, i, false);
-               vec = LLVMBuildInsertElement(builder, vec, value, index, "");
-       }
-       return vec;
+      if (!i)
+         vec = LLVMGetUndef(LLVMVectorType(LLVMTypeOf(value), value_count));
+      LLVMValueRef index = LLVMConstInt(ctx->i32, i, false);
+      vec = LLVMBuildInsertElement(builder, vec, value, index, "");
+   }
+   return vec;
 }
 
-LLVMValueRef
-ac_build_gather_values(struct ac_llvm_context *ctx,
-                      LLVMValueRef *values,
-                      unsigned value_count)
+LLVMValueRef ac_build_gather_values(struct ac_llvm_context *ctx, LLVMValueRef *values,
+                                    unsigned value_count)
 {
-       return ac_build_gather_values_extended(ctx, values, value_count, 1, false, false);
+   return ac_build_gather_values_extended(ctx, values, value_count, 1, false, false);
 }
 
 /* Expand a scalar or vector to <dst_channels x type> by filling the remaining
  * channels with undef. Extract at most src_channels components from the input.
  */
-static LLVMValueRef
-ac_build_expand(struct ac_llvm_context *ctx,
-               LLVMValueRef value,
-               unsigned src_channels,
-               unsigned dst_channels)
+static LLVMValueRef ac_build_expand(struct ac_llvm_context *ctx, LLVMValueRef value,
+                                    unsigned src_channels, unsigned dst_channels)
 {
-       LLVMTypeRef elemtype;
-       LLVMValueRef chan[dst_channels];
+   LLVMTypeRef elemtype;
+   LLVMValueRef chan[dst_channels];
 
-       if (LLVMGetTypeKind(LLVMTypeOf(value)) == LLVMVectorTypeKind) {
-               unsigned vec_size = LLVMGetVectorSize(LLVMTypeOf(value));
+   if (LLVMGetTypeKind(LLVMTypeOf(value)) == LLVMVectorTypeKind) {
+      unsigned vec_size = LLVMGetVectorSize(LLVMTypeOf(value));
 
-               if (src_channels == dst_channels && vec_size == dst_channels)
-                       return value;
+      if (src_channels == dst_channels && vec_size == dst_channels)
+         return value;
 
-               src_channels = MIN2(src_channels, vec_size);
+      src_channels = MIN2(src_channels, vec_size);
 
-               for (unsigned i = 0; i < src_channels; i++)
-                       chan[i] = ac_llvm_extract_elem(ctx, value, i);
+      for (unsigned i = 0; i < src_channels; i++)
+         chan[i] = ac_llvm_extract_elem(ctx, value, i);
 
-               elemtype = LLVMGetElementType(LLVMTypeOf(value));
-       } else {
-               if (src_channels) {
-                       assert(src_channels == 1);
-                       chan[0] = value;
-               }
-               elemtype = LLVMTypeOf(value);
-       }
+      elemtype = LLVMGetElementType(LLVMTypeOf(value));
+   } else {
+      if (src_channels) {
+         assert(src_channels == 1);
+         chan[0] = value;
+      }
+      elemtype = LLVMTypeOf(value);
+   }
 
-       for (unsigned i = src_channels; i < dst_channels; i++)
-               chan[i] = LLVMGetUndef(elemtype);
+   for (unsigned i = src_channels; i < dst_channels; i++)
+      chan[i] = LLVMGetUndef(elemtype);
 
-       return ac_build_gather_values(ctx, chan, dst_channels);
+   return ac_build_gather_values(ctx, chan, dst_channels);
 }
 
 /* Extract components [start, start + channels) from a vector.
  */
-LLVMValueRef
-ac_extract_components(struct ac_llvm_context *ctx,
-                     LLVMValueRef value,
-                     unsigned start,
-                     unsigned channels)
+LLVMValueRef ac_extract_components(struct ac_llvm_context *ctx, LLVMValueRef value, unsigned start,
+                                   unsigned channels)
 {
-       LLVMValueRef chan[channels];
+   LLVMValueRef chan[channels];
 
-       for (unsigned i = 0; i < channels; i++)
-               chan[i] = ac_llvm_extract_elem(ctx, value, i + start);
+   for (unsigned i = 0; i < channels; i++)
+      chan[i] = ac_llvm_extract_elem(ctx, value, i + start);
 
-       return ac_build_gather_values(ctx, chan, channels);
+   return ac_build_gather_values(ctx, chan, channels);
 }
 
 /* Expand a scalar or vector to <4 x type> by filling the remaining channels
  * with undef. Extract at most num_channels components from the input.
  */
-LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
-                                    LLVMValueRef value,
-                                    unsigned num_channels)
+LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx, LLVMValueRef value,
+                                     unsigned num_channels)
 {
-       return ac_build_expand(ctx, value, num_channels, 4);
+   return ac_build_expand(ctx, value, num_channels, 4);
 }
 
 LLVMValueRef ac_build_round(struct ac_llvm_context *ctx, LLVMValueRef value)
 {
-       unsigned type_size = ac_get_type_size(LLVMTypeOf(value));
-       const char *name;
+   unsigned type_size = ac_get_type_size(LLVMTypeOf(value));
+   const char *name;
 
-       if (type_size == 2)
-               name = "llvm.rint.f16";
-       else if (type_size == 4)
-               name = "llvm.rint.f32";
-       else
-               name = "llvm.rint.f64";
+   if (type_size == 2)
+      name = "llvm.rint.f16";
+   else if (type_size == 4)
+      name = "llvm.rint.f32";
+   else
+      name = "llvm.rint.f64";
 
-       return ac_build_intrinsic(ctx, name, LLVMTypeOf(value), &value, 1,
-                                 AC_FUNC_ATTR_READNONE);
+   return ac_build_intrinsic(ctx, name, LLVMTypeOf(value), &value, 1, AC_FUNC_ATTR_READNONE);
 }
 
-LLVMValueRef
-ac_build_fdiv(struct ac_llvm_context *ctx,
-             LLVMValueRef num,
-             LLVMValueRef den)
+LLVMValueRef ac_build_fdiv(struct ac_llvm_context *ctx, LLVMValueRef num, LLVMValueRef den)
 {
-       unsigned type_size = ac_get_type_size(LLVMTypeOf(den));
-       const char *name;
+   unsigned type_size = ac_get_type_size(LLVMTypeOf(den));
+   const char *name;
 
-       /* For doubles, we need precise division to pass GLCTS. */
-       if (ctx->float_mode == AC_FLOAT_MODE_DEFAULT_OPENGL &&
-           type_size == 8)
-               return LLVMBuildFDiv(ctx->builder, num, den, "");
+   /* For doubles, we need precise division to pass GLCTS. */
+   if (ctx->float_mode == AC_FLOAT_MODE_DEFAULT_OPENGL && type_size == 8)
+      return LLVMBuildFDiv(ctx->builder, num, den, "");
 
-       if (type_size == 2)
-               name = "llvm.amdgcn.rcp.f16";
-       else if (type_size == 4)
-               name = "llvm.amdgcn.rcp.f32";
-       else
-               name = "llvm.amdgcn.rcp.f64";
+   if (type_size == 2)
+      name = "llvm.amdgcn.rcp.f16";
+   else if (type_size == 4)
+      name = "llvm.amdgcn.rcp.f32";
+   else
+      name = "llvm.amdgcn.rcp.f64";
 
-        LLVMValueRef rcp = ac_build_intrinsic(ctx, name, LLVMTypeOf(den),
-                                              &den, 1, AC_FUNC_ATTR_READNONE);
+   LLVMValueRef rcp =
+      ac_build_intrinsic(ctx, name, LLVMTypeOf(den), &den, 1, AC_FUNC_ATTR_READNONE);
 
-       return LLVMBuildFMul(ctx->builder, num, rcp, "");
+   return LLVMBuildFMul(ctx->builder, num, rcp, "");
 }
 
 /* See fast_idiv_by_const.h. */
 /* Set: increment = util_fast_udiv_info::increment ? multiplier : 0; */
-LLVMValueRef ac_build_fast_udiv(struct ac_llvm_context *ctx,
-                               LLVMValueRef num,
-                               LLVMValueRef multiplier,
-                               LLVMValueRef pre_shift,
-                               LLVMValueRef post_shift,
-                               LLVMValueRef increment)
-{
-       LLVMBuilderRef builder = ctx->builder;
-
-       num = LLVMBuildLShr(builder, num, pre_shift, "");
-       num = LLVMBuildMul(builder,
-                          LLVMBuildZExt(builder, num, ctx->i64, ""),
-                          LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
-       num = LLVMBuildAdd(builder, num,
-                          LLVMBuildZExt(builder, increment, ctx->i64, ""), "");
-       num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
-       num = LLVMBuildTrunc(builder, num, ctx->i32, "");
-       return LLVMBuildLShr(builder, num, post_shift, "");
+LLVMValueRef ac_build_fast_udiv(struct ac_llvm_context *ctx, LLVMValueRef num,
+                                LLVMValueRef multiplier, LLVMValueRef pre_shift,
+                                LLVMValueRef post_shift, LLVMValueRef increment)
+{
+   LLVMBuilderRef builder = ctx->builder;
+
+   num = LLVMBuildLShr(builder, num, pre_shift, "");
+   num = LLVMBuildMul(builder, LLVMBuildZExt(builder, num, ctx->i64, ""),
+                      LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
+   num = LLVMBuildAdd(builder, num, LLVMBuildZExt(builder, increment, ctx->i64, ""), "");
+   num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
+   num = LLVMBuildTrunc(builder, num, ctx->i32, "");
+   return LLVMBuildLShr(builder, num, post_shift, "");
 }
 
 /* See fast_idiv_by_const.h. */
 /* If num != UINT_MAX, this more efficient version can be used. */
 /* Set: increment = util_fast_udiv_info::increment; */
-LLVMValueRef ac_build_fast_udiv_nuw(struct ac_llvm_context *ctx,
-                                   LLVMValueRef num,
-                                   LLVMValueRef multiplier,
-                                   LLVMValueRef pre_shift,
-                                   LLVMValueRef post_shift,
-                                   LLVMValueRef increment)
+LLVMValueRef ac_build_fast_udiv_nuw(struct ac_llvm_context *ctx, LLVMValueRef num,
+                                    LLVMValueRef multiplier, LLVMValueRef pre_shift,
+                                    LLVMValueRef post_shift, LLVMValueRef increment)
 {
-       LLVMBuilderRef builder = ctx->builder;
+   LLVMBuilderRef builder = ctx->builder;
 
-       num = LLVMBuildLShr(builder, num, pre_shift, "");
-       num = LLVMBuildNUWAdd(builder, num, increment, "");
-       num = LLVMBuildMul(builder,
-                          LLVMBuildZExt(builder, num, ctx->i64, ""),
-                          LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
-       num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
-       num = LLVMBuildTrunc(builder, num, ctx->i32, "");
-       return LLVMBuildLShr(builder, num, post_shift, "");
+   num = LLVMBuildLShr(builder, num, pre_shift, "");
+   num = LLVMBuildNUWAdd(builder, num, increment, "");
+   num = LLVMBuildMul(builder, LLVMBuildZExt(builder, num, ctx->i64, ""),
+                      LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
+   num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
+   num = LLVMBuildTrunc(builder, num, ctx->i32, "");
+   return LLVMBuildLShr(builder, num, post_shift, "");
 }
 
 /* See fast_idiv_by_const.h. */
 /* Both operands must fit in 31 bits and the divisor must not be 1. */
-LLVMValueRef ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context *ctx,
-                                             LLVMValueRef num,
-                                             LLVMValueRef multiplier,
-                                             LLVMValueRef post_shift)
+LLVMValueRef ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context *ctx, LLVMValueRef num,
+                                              LLVMValueRef multiplier, LLVMValueRef post_shift)
 {
-       LLVMBuilderRef builder = ctx->builder;
+   LLVMBuilderRef builder = ctx->builder;
 
-       num = LLVMBuildMul(builder,
-                          LLVMBuildZExt(builder, num, ctx->i64, ""),
-                          LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
-       num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
-       num = LLVMBuildTrunc(builder, num, ctx->i32, "");
-       return LLVMBuildLShr(builder, num, post_shift, "");
+   num = LLVMBuildMul(builder, LLVMBuildZExt(builder, num, ctx->i64, ""),
+                      LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
+   num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
+   num = LLVMBuildTrunc(builder, num, ctx->i32, "");
+   return LLVMBuildLShr(builder, num, post_shift, "");
 }
 
 /* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
@@ -789,26 +708,20 @@ LLVMValueRef ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context *ctx,
  * already multiplied by two. id is the cube face number.
  */
 struct cube_selection_coords {
-       LLVMValueRef stc[2];
-       LLVMValueRef ma;
-       LLVMValueRef id;
+   LLVMValueRef stc[2];
+   LLVMValueRef ma;
+   LLVMValueRef id;
 };
 
-static void
-build_cube_intrinsic(struct ac_llvm_context *ctx,
-                    LLVMValueRef in[3],
-                    struct cube_selection_coords *out)
+static void build_cube_intrinsic(struct ac_llvm_context *ctx, LLVMValueRef in[3],
+                                 struct cube_selection_coords *out)
 {
-       LLVMTypeRef f32 = ctx->f32;
+   LLVMTypeRef f32 = ctx->f32;
 
-       out->stc[1] = ac_build_intrinsic(ctx, "llvm.amdgcn.cubetc",
-                                        f32, in, 3, AC_FUNC_ATTR_READNONE);
-       out->stc[0] = ac_build_intrinsic(ctx, "llvm.amdgcn.cubesc",
-                                        f32, in, 3, AC_FUNC_ATTR_READNONE);
-       out->ma = ac_build_intrinsic(ctx, "llvm.amdgcn.cubema",
-                                    f32, in, 3, AC_FUNC_ATTR_READNONE);
-       out->id = ac_build_intrinsic(ctx, "llvm.amdgcn.cubeid",
-                                    f32, in, 3, AC_FUNC_ATTR_READNONE);
+   out->stc[1] = ac_build_intrinsic(ctx, "llvm.amdgcn.cubetc", f32, in, 3, AC_FUNC_ATTR_READNONE);
+   out->stc[0] = ac_build_intrinsic(ctx, "llvm.amdgcn.cubesc", f32, in, 3, AC_FUNC_ATTR_READNONE);
+   out->ma = ac_build_intrinsic(ctx, "llvm.amdgcn.cubema", f32, in, 3, AC_FUNC_ATTR_READNONE);
+   out->id = ac_build_intrinsic(ctx, "llvm.amdgcn.cubeid", f32, in, 3, AC_FUNC_ATTR_READNONE);
 }
 
 /**
@@ -821,281 +734,250 @@ build_cube_intrinsic(struct ac_llvm_context *ctx,
  * the selcoords major axis.
  */
 static void build_cube_select(struct ac_llvm_context *ctx,
-                             const struct cube_selection_coords *selcoords,
-                             const LLVMValueRef *coords,
-                             LLVMValueRef *out_st,
-                             LLVMValueRef *out_ma)
-{
-       LLVMBuilderRef builder = ctx->builder;
-       LLVMTypeRef f32 = LLVMTypeOf(coords[0]);
-       LLVMValueRef is_ma_positive;
-       LLVMValueRef sgn_ma;
-       LLVMValueRef is_ma_z, is_not_ma_z;
-       LLVMValueRef is_ma_y;
-       LLVMValueRef is_ma_x;
-       LLVMValueRef sgn;
-       LLVMValueRef tmp;
-
-       is_ma_positive = LLVMBuildFCmp(builder, LLVMRealUGE,
-               selcoords->ma, LLVMConstReal(f32, 0.0), "");
-       sgn_ma = LLVMBuildSelect(builder, is_ma_positive,
-               LLVMConstReal(f32, 1.0), LLVMConstReal(f32, -1.0), "");
-
-       is_ma_z = LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 4.0), "");
-       is_not_ma_z = LLVMBuildNot(builder, is_ma_z, "");
-       is_ma_y = LLVMBuildAnd(builder, is_not_ma_z,
-               LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 2.0), ""), "");
-       is_ma_x = LLVMBuildAnd(builder, is_not_ma_z, LLVMBuildNot(builder, is_ma_y, ""), "");
-
-       /* Select sc */
-       tmp = LLVMBuildSelect(builder, is_ma_x, coords[2], coords[0], "");
-       sgn = LLVMBuildSelect(builder, is_ma_y, LLVMConstReal(f32, 1.0),
-               LLVMBuildSelect(builder, is_ma_z, sgn_ma,
-                       LLVMBuildFNeg(builder, sgn_ma, ""), ""), "");
-       out_st[0] = LLVMBuildFMul(builder, tmp, sgn, "");
-
-       /* Select tc */
-       tmp = LLVMBuildSelect(builder, is_ma_y, coords[2], coords[1], "");
-       sgn = LLVMBuildSelect(builder, is_ma_y, sgn_ma,
-               LLVMConstReal(f32, -1.0), "");
-       out_st[1] = LLVMBuildFMul(builder, tmp, sgn, "");
-
-       /* Select ma */
-       tmp = LLVMBuildSelect(builder, is_ma_z, coords[2],
-               LLVMBuildSelect(builder, is_ma_y, coords[1], coords[0], ""), "");
-       tmp = ac_build_intrinsic(ctx, "llvm.fabs.f32",
-                                ctx->f32, &tmp, 1, AC_FUNC_ATTR_READNONE);
-       *out_ma = LLVMBuildFMul(builder, tmp, LLVMConstReal(f32, 2.0), "");
-}
-
-void
-ac_prepare_cube_coords(struct ac_llvm_context *ctx,
-                      bool is_deriv, bool is_array, bool is_lod,
-                      LLVMValueRef *coords_arg,
-                      LLVMValueRef *derivs_arg)
-{
-
-       LLVMBuilderRef builder = ctx->builder;
-       struct cube_selection_coords selcoords;
-       LLVMValueRef coords[3];
-       LLVMValueRef invma;
-
-       if (is_array && !is_lod) {
-               LLVMValueRef tmp = ac_build_round(ctx, coords_arg[3]);
-
-               /* Section 8.9 (Texture Functions) of the GLSL 4.50 spec says:
-                *
-                *    "For Array forms, the array layer used will be
-                *
-                *       max(0, min(d−1, floor(layer+0.5)))
-                *
-                *     where d is the depth of the texture array and layer
-                *     comes from the component indicated in the tables below.
-                *     Workaroudn for an issue where the layer is taken from a
-                *     helper invocation which happens to fall on a different
-                *     layer due to extrapolation."
-                *
-                * GFX8 and earlier attempt to implement this in hardware by
-                * clamping the value of coords[2] = (8 * layer) + face.
-                * Unfortunately, this means that the we end up with the wrong
-                * face when clamping occurs.
-                *
-                * Clamp the layer earlier to work around the issue.
-                */
-               if (ctx->chip_class <= GFX8) {
-                       LLVMValueRef ge0;
-                       ge0 = LLVMBuildFCmp(builder, LLVMRealOGE, tmp, ctx->f32_0, "");
-                       tmp = LLVMBuildSelect(builder, ge0, tmp, ctx->f32_0, "");
-               }
-
-               coords_arg[3] = tmp;
-       }
-
-       build_cube_intrinsic(ctx, coords_arg, &selcoords);
-
-       invma = ac_build_intrinsic(ctx, "llvm.fabs.f32",
-                       ctx->f32, &selcoords.ma, 1, AC_FUNC_ATTR_READNONE);
-       invma = ac_build_fdiv(ctx, LLVMConstReal(ctx->f32, 1.0), invma);
-
-       for (int i = 0; i < 2; ++i)
-               coords[i] = LLVMBuildFMul(builder, selcoords.stc[i], invma, "");
-
-       coords[2] = selcoords.id;
-
-       if (is_deriv && derivs_arg) {
-               LLVMValueRef derivs[4];
-               int axis;
-
-               /* Convert cube derivatives to 2D derivatives. */
-               for (axis = 0; axis < 2; axis++) {
-                       LLVMValueRef deriv_st[2];
-                       LLVMValueRef deriv_ma;
-
-                       /* Transform the derivative alongside the texture
-                        * coordinate. Mathematically, the correct formula is
-                        * as follows. Assume we're projecting onto the +Z face
-                        * and denote by dx/dh the derivative of the (original)
-                        * X texture coordinate with respect to horizontal
-                        * window coordinates. The projection onto the +Z face
-                        * plane is:
-                        *
-                        *   f(x,z) = x/z
-                        *
-                        * Then df/dh = df/dx * dx/dh + df/dz * dz/dh
-                        *            = 1/z * dx/dh - x/z * 1/z * dz/dh.
-                        *
-                        * This motivatives the implementation below.
-                        *
-                        * Whether this actually gives the expected results for
-                        * apps that might feed in derivatives obtained via
-                        * finite differences is anyone's guess. The OpenGL spec
-                        * seems awfully quiet about how textureGrad for cube
-                        * maps should be handled.
-                        */
-                       build_cube_select(ctx, &selcoords, &derivs_arg[axis * 3],
-                                         deriv_st, &deriv_ma);
-
-                       deriv_ma = LLVMBuildFMul(builder, deriv_ma, invma, "");
-
-                       for (int i = 0; i < 2; ++i)
-                               derivs[axis * 2 + i] =
-                                       LLVMBuildFSub(builder,
-                                               LLVMBuildFMul(builder, deriv_st[i], invma, ""),
-                                               LLVMBuildFMul(builder, deriv_ma, coords[i], ""), "");
-               }
-
-               memcpy(derivs_arg, derivs, sizeof(derivs));
-       }
-
-       /* Shift the texture coordinate. This must be applied after the
-        * derivative calculation.
-        */
-       for (int i = 0; i < 2; ++i)
-               coords[i] = LLVMBuildFAdd(builder, coords[i], LLVMConstReal(ctx->f32, 1.5), "");
-
-       if (is_array) {
-               /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
-               /* coords_arg.w component - array_index for cube arrays */
-               coords[2] = ac_build_fmad(ctx, coords_arg[3], LLVMConstReal(ctx->f32, 8.0), coords[2]);
-       }
-
-       memcpy(coords_arg, coords, sizeof(coords));
-}
-
-
-LLVMValueRef
-ac_build_fs_interp(struct ac_llvm_context *ctx,
-                  LLVMValueRef llvm_chan,
-                  LLVMValueRef attr_number,
-                  LLVMValueRef params,
-                  LLVMValueRef i,
-                  LLVMValueRef j)
-{
-       LLVMValueRef args[5];
-       LLVMValueRef p1;
-
-       args[0] = i;
-       args[1] = llvm_chan;
-       args[2] = attr_number;
-       args[3] = params;
-
-       p1 = ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p1",
-                               ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
-
-       args[0] = p1;
-       args[1] = j;
-       args[2] = llvm_chan;
-       args[3] = attr_number;
-       args[4] = params;
+                              const struct cube_selection_coords *selcoords,
+                              const LLVMValueRef *coords, LLVMValueRef *out_st,
+                              LLVMValueRef *out_ma)
+{
+   LLVMBuilderRef builder = ctx->builder;
+   LLVMTypeRef f32 = LLVMTypeOf(coords[0]);
+   LLVMValueRef is_ma_positive;
+   LLVMValueRef sgn_ma;
+   LLVMValueRef is_ma_z, is_not_ma_z;
+   LLVMValueRef is_ma_y;
+   LLVMValueRef is_ma_x;
+   LLVMValueRef sgn;
+   LLVMValueRef tmp;
+
+   is_ma_positive = LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->ma, LLVMConstReal(f32, 0.0), "");
+   sgn_ma = LLVMBuildSelect(builder, is_ma_positive, LLVMConstReal(f32, 1.0),
+                            LLVMConstReal(f32, -1.0), "");
+
+   is_ma_z = LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 4.0), "");
+   is_not_ma_z = LLVMBuildNot(builder, is_ma_z, "");
+   is_ma_y = LLVMBuildAnd(
+      builder, is_not_ma_z,
+      LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 2.0), ""), "");
+   is_ma_x = LLVMBuildAnd(builder, is_not_ma_z, LLVMBuildNot(builder, is_ma_y, ""), "");
+
+   /* Select sc */
+   tmp = LLVMBuildSelect(builder, is_ma_x, coords[2], coords[0], "");
+   sgn = LLVMBuildSelect(
+      builder, is_ma_y, LLVMConstReal(f32, 1.0),
+      LLVMBuildSelect(builder, is_ma_z, sgn_ma, LLVMBuildFNeg(builder, sgn_ma, ""), ""), "");
+   out_st[0] = LLVMBuildFMul(builder, tmp, sgn, "");
+
+   /* Select tc */
+   tmp = LLVMBuildSelect(builder, is_ma_y, coords[2], coords[1], "");
+   sgn = LLVMBuildSelect(builder, is_ma_y, sgn_ma, LLVMConstReal(f32, -1.0), "");
+   out_st[1] = LLVMBuildFMul(builder, tmp, sgn, "");
+
+   /* Select ma */
+   tmp = LLVMBuildSelect(builder, is_ma_z, coords[2],
+                         LLVMBuildSelect(builder, is_ma_y, coords[1], coords[0], ""), "");
+   tmp = ac_build_intrinsic(ctx, "llvm.fabs.f32", ctx->f32, &tmp, 1, AC_FUNC_ATTR_READNONE);
+   *out_ma = LLVMBuildFMul(builder, tmp, LLVMConstReal(f32, 2.0), "");
+}
+
+void ac_prepare_cube_coords(struct ac_llvm_context *ctx, bool is_deriv, bool is_array, bool is_lod,
+                            LLVMValueRef *coords_arg, LLVMValueRef *derivs_arg)
+{
+
+   LLVMBuilderRef builder = ctx->builder;
+   struct cube_selection_coords selcoords;
+   LLVMValueRef coords[3];
+   LLVMValueRef invma;
+
+   if (is_array && !is_lod) {
+      LLVMValueRef tmp = ac_build_round(ctx, coords_arg[3]);
+
+      /* Section 8.9 (Texture Functions) of the GLSL 4.50 spec says:
+       *
+       *    "For Array forms, the array layer used will be
+       *
+       *       max(0, min(d−1, floor(layer+0.5)))
+       *
+       *     where d is the depth of the texture array and layer
+       *     comes from the component indicated in the tables below.
+       *     Workaroudn for an issue where the layer is taken from a
+       *     helper invocation which happens to fall on a different
+       *     layer due to extrapolation."
+       *
+       * GFX8 and earlier attempt to implement this in hardware by
+       * clamping the value of coords[2] = (8 * layer) + face.
+       * Unfortunately, this means that the we end up with the wrong
+       * face when clamping occurs.
+       *
+       * Clamp the layer earlier to work around the issue.
+       */
+      if (ctx->chip_class <= GFX8) {
+         LLVMValueRef ge0;
+         ge0 = LLVMBuildFCmp(builder, LLVMRealOGE, tmp, ctx->f32_0, "");
+         tmp = LLVMBuildSelect(builder, ge0, tmp, ctx->f32_0, "");
+      }
+
+      coords_arg[3] = tmp;
+   }
+
+   build_cube_intrinsic(ctx, coords_arg, &selcoords);
+
+   invma =
+      ac_build_intrinsic(ctx, "llvm.fabs.f32", ctx->f32, &selcoords.ma, 1, AC_FUNC_ATTR_READNONE);
+   invma = ac_build_fdiv(ctx, LLVMConstReal(ctx->f32, 1.0), invma);
+
+   for (int i = 0; i < 2; ++i)
+      coords[i] = LLVMBuildFMul(builder, selcoords.stc[i], invma, "");
+
+   coords[2] = selcoords.id;
+
+   if (is_deriv && derivs_arg) {
+      LLVMValueRef derivs[4];
+      int axis;
+
+      /* Convert cube derivatives to 2D derivatives. */
+      for (axis = 0; axis < 2; axis++) {
+         LLVMValueRef deriv_st[2];
+         LLVMValueRef deriv_ma;
+
+         /* Transform the derivative alongside the texture
+          * coordinate. Mathematically, the correct formula is
+          * as follows. Assume we're projecting onto the +Z face
+          * and denote by dx/dh the derivative of the (original)
+          * X texture coordinate with respect to horizontal
+          * window coordinates. The projection onto the +Z face
+          * plane is:
+          *
+          *   f(x,z) = x/z
+          *
+          * Then df/dh = df/dx * dx/dh + df/dz * dz/dh
+          *            = 1/z * dx/dh - x/z * 1/z * dz/dh.
+          *
+          * This motivatives the implementation below.
+          *
+          * Whether this actually gives the expected results for
+          * apps that might feed in derivatives obtained via
+          * finite differences is anyone's guess. The OpenGL spec
+          * seems awfully quiet about how textureGrad for cube
+          * maps should be handled.
+          */
+         build_cube_select(ctx, &selcoords, &derivs_arg[axis * 3], deriv_st, &deriv_ma);
+
+         deriv_ma = LLVMBuildFMul(builder, deriv_ma, invma, "");
+
+         for (int i = 0; i < 2; ++i)
+            derivs[axis * 2 + i] =
+               LLVMBuildFSub(builder, LLVMBuildFMul(builder, deriv_st[i], invma, ""),
+                             LLVMBuildFMul(builder, deriv_ma, coords[i], ""), "");
+      }
+
+      memcpy(derivs_arg, derivs, sizeof(derivs));
+   }
+
+   /* Shift the texture coordinate. This must be applied after the
+    * derivative calculation.
+    */
+   for (int i = 0; i < 2; ++i)
+      coords[i] = LLVMBuildFAdd(builder, coords[i], LLVMConstReal(ctx->f32, 1.5), "");
+
+   if (is_array) {
+      /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
+      /* coords_arg.w component - array_index for cube arrays */
+      coords[2] = ac_build_fmad(ctx, coords_arg[3], LLVMConstReal(ctx->f32, 8.0), coords[2]);
+   }
 
-       return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p2",
-                                 ctx->f32, args, 5, AC_FUNC_ATTR_READNONE);
+   memcpy(coords_arg, coords, sizeof(coords));
 }
 
-LLVMValueRef
-ac_build_fs_interp_f16(struct ac_llvm_context *ctx,
-                      LLVMValueRef llvm_chan,
-                      LLVMValueRef attr_number,
-                      LLVMValueRef params,
-                      LLVMValueRef i,
-                      LLVMValueRef j)
+LLVMValueRef ac_build_fs_interp(struct ac_llvm_context *ctx, LLVMValueRef llvm_chan,
+                                LLVMValueRef attr_number, LLVMValueRef params, LLVMValueRef i,
+                                LLVMValueRef j)
 {
-       LLVMValueRef args[6];
-       LLVMValueRef p1;
-
-       args[0] = i;
-       args[1] = llvm_chan;
-       args[2] = attr_number;
-       args[3] = ctx->i1false;
-       args[4] = params;
-
-       p1 = ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p1.f16",
-                               ctx->f32, args, 5, AC_FUNC_ATTR_READNONE);
-
-       args[0] = p1;
-       args[1] = j;
-       args[2] = llvm_chan;
-       args[3] = attr_number;
-       args[4] = ctx->i1false;
-       args[5] = params;
+   LLVMValueRef args[5];
+   LLVMValueRef p1;
+
+   args[0] = i;
+   args[1] = llvm_chan;
+   args[2] = attr_number;
+   args[3] = params;
+
+   p1 = ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p1", ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
 
-       return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p2.f16",
-                                 ctx->f16, args, 6, AC_FUNC_ATTR_READNONE);
+   args[0] = p1;
+   args[1] = j;
+   args[2] = llvm_chan;
+   args[3] = attr_number;
+   args[4] = params;
+
+   return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p2", ctx->f32, args, 5,
+                             AC_FUNC_ATTR_READNONE);
 }
 
-LLVMValueRef
-ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
-                      LLVMValueRef parameter,
-                      LLVMValueRef llvm_chan,
-                      LLVMValueRef attr_number,
-                      LLVMValueRef params)
+LLVMValueRef ac_build_fs_interp_f16(struct ac_llvm_context *ctx, LLVMValueRef llvm_chan,
+                                    LLVMValueRef attr_number, LLVMValueRef params, LLVMValueRef i,
+                                    LLVMValueRef j)
 {
-       LLVMValueRef args[4];
+   LLVMValueRef args[6];
+   LLVMValueRef p1;
+
+   args[0] = i;
+   args[1] = llvm_chan;
+   args[2] = attr_number;
+   args[3] = ctx->i1false;
+   args[4] = params;
+
+   p1 = ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p1.f16", ctx->f32, args, 5,
+                           AC_FUNC_ATTR_READNONE);
 
-       args[0] = parameter;
-       args[1] = llvm_chan;
-       args[2] = attr_number;
-       args[3] = params;
+   args[0] = p1;
+   args[1] = j;
+   args[2] = llvm_chan;
+   args[3] = attr_number;
+   args[4] = ctx->i1false;
+   args[5] = params;
 
-       return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.mov",
-                                 ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
+   return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p2.f16", ctx->f16, args, 6,
+                             AC_FUNC_ATTR_READNONE);
 }
 
-LLVMValueRef
-ac_build_gep_ptr(struct ac_llvm_context *ctx,
-                LLVMValueRef base_ptr,
-                LLVMValueRef index)
+LLVMValueRef ac_build_fs_interp_mov(struct ac_llvm_context *ctx, LLVMValueRef parameter,
+                                    LLVMValueRef llvm_chan, LLVMValueRef attr_number,
+                                    LLVMValueRef params)
 {
-       return LLVMBuildGEP(ctx->builder, base_ptr, &index, 1, "");
+   LLVMValueRef args[4];
+
+   args[0] = parameter;
+   args[1] = llvm_chan;
+   args[2] = attr_number;
+   args[3] = params;
+
+   return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.mov", ctx->f32, args, 4,
+                             AC_FUNC_ATTR_READNONE);
 }
 
-LLVMValueRef
-ac_build_gep0(struct ac_llvm_context *ctx,
-             LLVMValueRef base_ptr,
-             LLVMValueRef index)
+LLVMValueRef ac_build_gep_ptr(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
+                              LLVMValueRef index)
 {
-       LLVMValueRef indices[2] = {
-               ctx->i32_0,
-               index,
-       };
-       return LLVMBuildGEP(ctx->builder, base_ptr, indices, 2, "");
+   return LLVMBuildGEP(ctx->builder, base_ptr, &index, 1, "");
 }
 
-LLVMValueRef ac_build_pointer_add(struct ac_llvm_context *ctx, LLVMValueRef ptr,
-                                 LLVMValueRef index)
+LLVMValueRef ac_build_gep0(struct ac_llvm_context *ctx, LLVMValueRef base_ptr, LLVMValueRef index)
 {
-       return LLVMBuildPointerCast(ctx->builder,
-                                   LLVMBuildGEP(ctx->builder, ptr, &index, 1, ""),
-                                   LLVMTypeOf(ptr), "");
+   LLVMValueRef indices[2] = {
+      ctx->i32_0,
+      index,
+   };
+   return LLVMBuildGEP(ctx->builder, base_ptr, indices, 2, "");
 }
 
-void
-ac_build_indexed_store(struct ac_llvm_context *ctx,
-                      LLVMValueRef base_ptr, LLVMValueRef index,
-                      LLVMValueRef value)
+LLVMValueRef ac_build_pointer_add(struct ac_llvm_context *ctx, LLVMValueRef ptr, LLVMValueRef index)
 {
-       LLVMBuildStore(ctx->builder, value,
-                      ac_build_gep0(ctx, base_ptr, index));
+   return LLVMBuildPointerCast(ctx->builder, LLVMBuildGEP(ctx->builder, ptr, &index, 1, ""),
+                               LLVMTypeOf(ptr), "");
+}
+
+void ac_build_indexed_store(struct ac_llvm_context *ctx, LLVMValueRef base_ptr, LLVMValueRef index,
+                            LLVMValueRef value)
+{
+   LLVMBuildStore(ctx->builder, value, ac_build_gep0(ctx, base_ptr, index));
 }
 
 /**
@@ -1126,425 +1008,328 @@ ac_build_indexed_store(struct ac_llvm_context *ctx,
  *      ptr2 = LLVMBuildInBoundsGEP(ptr1, 32 / elemsize);
  *      sampler = load(ptr2); // becomes "s_load ptr1, 32" thanks to InBounds
  */
-static LLVMValueRef
-ac_build_load_custom(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
-                    LLVMValueRef index, bool uniform, bool invariant,
-                    bool no_unsigned_wraparound)
+static LLVMValueRef ac_build_load_custom(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
+                                         LLVMValueRef index, bool uniform, bool invariant,
+                                         bool no_unsigned_wraparound)
 {
-       LLVMValueRef pointer, result;
+   LLVMValueRef pointer, result;
 
-       if (no_unsigned_wraparound &&
-           LLVMGetPointerAddressSpace(LLVMTypeOf(base_ptr)) == AC_ADDR_SPACE_CONST_32BIT)
-               pointer = LLVMBuildInBoundsGEP(ctx->builder, base_ptr, &index, 1, "");
-       else
-               pointer = LLVMBuildGEP(ctx->builder, base_ptr, &index, 1, "");
+   if (no_unsigned_wraparound &&
+       LLVMGetPointerAddressSpace(LLVMTypeOf(base_ptr)) == AC_ADDR_SPACE_CONST_32BIT)
+      pointer = LLVMBuildInBoundsGEP(ctx->builder, base_ptr, &index, 1, "");
+   else
+      pointer = LLVMBuildGEP(ctx->builder, base_ptr, &index, 1, "");
 
-       if (uniform)
-               LLVMSetMetadata(pointer, ctx->uniform_md_kind, ctx->empty_md);
-       result = LLVMBuildLoad(ctx->builder, pointer, "");
-       if (invariant)
-               LLVMSetMetadata(result, ctx->invariant_load_md_kind, ctx->empty_md);
-       return result;
+   if (uniform)
+      LLVMSetMetadata(pointer, ctx->uniform_md_kind, ctx->empty_md);
+   result = LLVMBuildLoad(ctx->builder, pointer, "");
+   if (invariant)
+      LLVMSetMetadata(result, ctx->invariant_load_md_kind, ctx->empty_md);
+   return result;
 }
 
-LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
-                          LLVMValueRef index)
+LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMValueRef base_ptr, LLVMValueRef index)
 {
-       return ac_build_load_custom(ctx, base_ptr, index, false, false, false);
+   return ac_build_load_custom(ctx, base_ptr, index, false, false, false);
 }
 
-LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx,
-                                    LLVMValueRef base_ptr, LLVMValueRef index)
+LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
+                                     LLVMValueRef index)
 {
-       return ac_build_load_custom(ctx, base_ptr, index, false, true, false);
+   return ac_build_load_custom(ctx, base_ptr, index, false, true, false);
 }
 
 /* This assumes that there is no unsigned integer wraparound during the address
  * computation, excluding all GEPs within base_ptr. */
-LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx,
-                                  LLVMValueRef base_ptr, LLVMValueRef index)
+LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
+                                   LLVMValueRef index)
 {
-       return ac_build_load_custom(ctx, base_ptr, index, true, true, true);
+   return ac_build_load_custom(ctx, base_ptr, index, true, true, true);
 }
 
 /* See ac_build_load_custom() documentation. */
 LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx,
-                                  LLVMValueRef base_ptr, LLVMValueRef index)
+                                                   LLVMValueRef base_ptr, LLVMValueRef index)
 {
-       return ac_build_load_custom(ctx, base_ptr, index, true, true, false);
+   return ac_build_load_custom(ctx, base_ptr, index, true, true, false);
 }
 
-static unsigned get_load_cache_policy(struct ac_llvm_context *ctx,
-                                     unsigned cache_policy)
+static unsigned get_load_cache_policy(struct ac_llvm_context *ctx, unsigned cache_policy)
 {
-       return cache_policy |
-              (ctx->chip_class >= GFX10 && cache_policy & ac_glc ? ac_dlc : 0);
+   return cache_policy | (ctx->chip_class >= GFX10 && cache_policy & ac_glc ? ac_dlc : 0);
 }
 
-static void
-ac_build_buffer_store_common(struct ac_llvm_context *ctx,
-                            LLVMValueRef rsrc,
-                            LLVMValueRef data,
-                            LLVMValueRef vindex,
-                            LLVMValueRef voffset,
-                            LLVMValueRef soffset,
-                            unsigned cache_policy,
-                            bool use_format,
-                            bool structurized)
+static void ac_build_buffer_store_common(struct ac_llvm_context *ctx, LLVMValueRef rsrc,
+                                         LLVMValueRef data, LLVMValueRef vindex,
+                                         LLVMValueRef voffset, LLVMValueRef soffset,
+                                         unsigned cache_policy, bool use_format, bool structurized)
 {
-       LLVMValueRef args[6];
-       int idx = 0;
-       args[idx++] = data;
-       args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
-       if (structurized)
-               args[idx++] = vindex ? vindex : ctx->i32_0;
-       args[idx++] = voffset ? voffset : ctx->i32_0;
-       args[idx++] = soffset ? soffset : ctx->i32_0;
-       args[idx++] = LLVMConstInt(ctx->i32, cache_policy, 0);
-       const char *indexing_kind = structurized ? "struct" : "raw";
-       char name[256], type_name[8];
+   LLVMValueRef args[6];
+   int idx = 0;
+   args[idx++] = data;
+   args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
+   if (structurized)
+      args[idx++] = vindex ? vindex : ctx->i32_0;
+   args[idx++] = voffset ? voffset : ctx->i32_0;
+   args[idx++] = soffset ? soffset : ctx->i32_0;
+   args[idx++] = LLVMConstInt(ctx->i32, cache_policy, 0);
+   const char *indexing_kind = structurized ? "struct" : "raw";
+   char name[256], type_name[8];
 
-       ac_build_type_name_for_intr(LLVMTypeOf(data), type_name, sizeof(type_name));
+   ac_build_type_name_for_intr(LLVMTypeOf(data), type_name, sizeof(type_name));
 
-       if (use_format) {
-               snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.store.format.%s",
-                        indexing_kind, type_name);
-       } else {
-               snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.store.%s",
-                        indexing_kind, type_name);
-       }
+   if (use_format) {
+      snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.store.format.%s", indexing_kind,
+               type_name);
+   } else {
+      snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.store.%s", indexing_kind, type_name);
+   }
 
-       ac_build_intrinsic(ctx, name, ctx->voidt, args, idx,
-                          AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY);
+   ac_build_intrinsic(ctx, name, ctx->voidt, args, idx, AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY);
 }
 
-void
-ac_build_buffer_store_format(struct ac_llvm_context *ctx,
-                            LLVMValueRef rsrc,
-                            LLVMValueRef data,
-                            LLVMValueRef vindex,
-                            LLVMValueRef voffset,
-                            unsigned cache_policy)
+void ac_build_buffer_store_format(struct ac_llvm_context *ctx, LLVMValueRef rsrc, LLVMValueRef data,
+                                  LLVMValueRef vindex, LLVMValueRef voffset, unsigned cache_policy)
 {
-       ac_build_buffer_store_common(ctx, rsrc, data, vindex, voffset, NULL,
-                                    cache_policy, true, true);
+   ac_build_buffer_store_common(ctx, rsrc, data, vindex, voffset, NULL, cache_policy, true, true);
 }
 
 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
  * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
  * or v4i32 (num_channels=3,4).
  */
-void
-ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
-                           LLVMValueRef rsrc,
-                           LLVMValueRef vdata,
-                           unsigned num_channels,
-                           LLVMValueRef voffset,
-                           LLVMValueRef soffset,
-                           unsigned inst_offset,
-                           unsigned cache_policy)
-{
-       /* Split 3 channel stores, because only LLVM 9+ support 3-channel
-        * intrinsics. */
-       if (num_channels == 3 && !ac_has_vec3_support(ctx->chip_class, false)) {
-               LLVMValueRef v[3], v01;
-
-               for (int i = 0; i < 3; i++) {
-                       v[i] = LLVMBuildExtractElement(ctx->builder, vdata,
-                                       LLVMConstInt(ctx->i32, i, 0), "");
-               }
-               v01 = ac_build_gather_values(ctx, v, 2);
-
-               ac_build_buffer_store_dword(ctx, rsrc, v01, 2, voffset,
-                                           soffset, inst_offset, cache_policy);
-               ac_build_buffer_store_dword(ctx, rsrc, v[2], 1, voffset,
-                                           soffset, inst_offset + 8,
-                                           cache_policy);
-               return;
-       }
-
-       /* SWIZZLE_ENABLE requires that soffset isn't folded into voffset
-        * (voffset is swizzled, but soffset isn't swizzled).
-        * llvm.amdgcn.buffer.store doesn't have a separate soffset parameter.
-        */
-       if (!(cache_policy & ac_swizzled)) {
-               LLVMValueRef offset = soffset;
-
-               if (inst_offset)
-                       offset = LLVMBuildAdd(ctx->builder, offset,
-                                             LLVMConstInt(ctx->i32, inst_offset, 0), "");
-
-               ac_build_buffer_store_common(ctx, rsrc, ac_to_float(ctx, vdata),
-                                            ctx->i32_0, voffset, offset,
-                                            cache_policy, false, false);
-               return;
-       }
-
-       static const unsigned dfmts[] = {
-               V_008F0C_BUF_DATA_FORMAT_32,
-               V_008F0C_BUF_DATA_FORMAT_32_32,
-               V_008F0C_BUF_DATA_FORMAT_32_32_32,
-               V_008F0C_BUF_DATA_FORMAT_32_32_32_32
-       };
-       unsigned dfmt = dfmts[num_channels - 1];
-       unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
-       LLVMValueRef immoffset = LLVMConstInt(ctx->i32, inst_offset, 0);
-
-       ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset,
-                                  immoffset, num_channels, dfmt, nfmt, cache_policy);
-}
-
-static LLVMValueRef
-ac_build_buffer_load_common(struct ac_llvm_context *ctx,
-                           LLVMValueRef rsrc,
-                           LLVMValueRef vindex,
-                           LLVMValueRef voffset,
-                           LLVMValueRef soffset,
-                           unsigned num_channels,
-                           LLVMTypeRef channel_type,
-                           unsigned cache_policy,
-                           bool can_speculate,
-                           bool use_format,
-                           bool structurized)
-{
-       LLVMValueRef args[5];
-       int idx = 0;
-       args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
-       if (structurized)
-               args[idx++] = vindex ? vindex : ctx->i32_0;
-       args[idx++] = voffset ? voffset : ctx->i32_0;
-       args[idx++] = soffset ? soffset : ctx->i32_0;
-       args[idx++] = LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0);
-       unsigned func = !ac_has_vec3_support(ctx->chip_class, use_format) && num_channels == 3 ? 4 : num_channels;
-       const char *indexing_kind = structurized ? "struct" : "raw";
-       char name[256], type_name[8];
-
-       /* D16 is only supported on gfx8+ */
-       assert(!use_format ||
-              (channel_type != ctx->f16 && channel_type != ctx->i16) ||
-              ctx->chip_class >= GFX8);
-
-       LLVMTypeRef type = func > 1 ? LLVMVectorType(channel_type, func) : channel_type;
-       ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
-
-       if (use_format) {
-               snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.load.format.%s",
-                        indexing_kind, type_name);
-       } else {
-               snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.load.%s",
-                        indexing_kind, type_name);
-       }
-
-       return ac_build_intrinsic(ctx, name, type, args, idx,
-                                 ac_get_load_intr_attribs(can_speculate));
-}
-
-LLVMValueRef
-ac_build_buffer_load(struct ac_llvm_context *ctx,
-                    LLVMValueRef rsrc,
-                    int num_channels,
-                    LLVMValueRef vindex,
-                    LLVMValueRef voffset,
-                    LLVMValueRef soffset,
-                    unsigned inst_offset,
-                    unsigned cache_policy,
-                    bool can_speculate,
-                    bool allow_smem)
-{
-       LLVMValueRef offset = LLVMConstInt(ctx->i32, inst_offset, 0);
-       if (voffset)
-               offset = LLVMBuildAdd(ctx->builder, offset, voffset, "");
-       if (soffset)
-               offset = LLVMBuildAdd(ctx->builder, offset, soffset, "");
-
-       if (allow_smem && !(cache_policy & ac_slc) &&
-           (!(cache_policy & ac_glc) || ctx->chip_class >= GFX8)) {
-               assert(vindex == NULL);
-
-               LLVMValueRef result[8];
-
-               for (int i = 0; i < num_channels; i++) {
-                       if (i) {
-                               offset = LLVMBuildAdd(ctx->builder, offset,
-                                                     LLVMConstInt(ctx->i32, 4, 0), "");
-                       }
-                       LLVMValueRef args[3] = {
-                               rsrc,
-                               offset,
-                               LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0),
-                       };
-                       result[i] = ac_build_intrinsic(ctx,
-                                                      "llvm.amdgcn.s.buffer.load.f32",
-                                                      ctx->f32, args, 3,
-                                                      AC_FUNC_ATTR_READNONE);
-               }
-               if (num_channels == 1)
-                       return result[0];
-
-               if (num_channels == 3 && !ac_has_vec3_support(ctx->chip_class, false))
-                       result[num_channels++] = LLVMGetUndef(ctx->f32);
-               return ac_build_gather_values(ctx, result, num_channels);
-       }
-
-       return ac_build_buffer_load_common(ctx, rsrc, vindex,
-                                          offset, ctx->i32_0,
-                                          num_channels, ctx->f32,
-                                          cache_policy,
-                                          can_speculate, false, false);
-}
-
-LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
-                                        LLVMValueRef rsrc,
-                                        LLVMValueRef vindex,
-                                        LLVMValueRef voffset,
-                                        unsigned num_channels,
-                                        unsigned cache_policy,
-                                        bool can_speculate,
-                                        bool d16)
-{
-       return ac_build_buffer_load_common(ctx, rsrc, vindex, voffset,
-                                          ctx->i32_0, num_channels,
-                                          d16 ? ctx->f16 : ctx->f32,
-                                          cache_policy, can_speculate,
-                                          true, true);
-}
-
-static LLVMValueRef
-ac_build_tbuffer_load(struct ac_llvm_context *ctx,
-                           LLVMValueRef rsrc,
-                           LLVMValueRef vindex,
-                           LLVMValueRef voffset,
-                           LLVMValueRef soffset,
-                           LLVMValueRef immoffset,
-                           unsigned num_channels,
-                           unsigned dfmt,
-                           unsigned nfmt,
-                           unsigned cache_policy,
-                           bool can_speculate,
-                           bool structurized)
-{
-       voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
-
-       LLVMValueRef args[6];
-       int idx = 0;
-       args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
-       if (structurized)
-               args[idx++] = vindex ? vindex : ctx->i32_0;
-       args[idx++] = voffset ? voffset : ctx->i32_0;
-       args[idx++] = soffset ? soffset : ctx->i32_0;
-       args[idx++] = LLVMConstInt(ctx->i32, ac_get_tbuffer_format(ctx->chip_class, dfmt, nfmt), 0);
-       args[idx++] = LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0);
-       unsigned func = !ac_has_vec3_support(ctx->chip_class, true) && num_channels == 3 ? 4 : num_channels;
-       const char *indexing_kind = structurized ? "struct" : "raw";
-       char name[256], type_name[8];
-
-       LLVMTypeRef type = func > 1 ? LLVMVectorType(ctx->i32, func) : ctx->i32;
-       ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
-
-       snprintf(name, sizeof(name), "llvm.amdgcn.%s.tbuffer.load.%s",
-                indexing_kind, type_name);
-
-       return ac_build_intrinsic(ctx, name, type, args, idx,
-                                 ac_get_load_intr_attribs(can_speculate));
-}
-
-LLVMValueRef
-ac_build_struct_tbuffer_load(struct ac_llvm_context *ctx,
-                            LLVMValueRef rsrc,
-                            LLVMValueRef vindex,
-                            LLVMValueRef voffset,
-                            LLVMValueRef soffset,
-                            LLVMValueRef immoffset,
-                            unsigned num_channels,
-                            unsigned dfmt,
-                            unsigned nfmt,
-                            unsigned cache_policy,
-                            bool can_speculate)
-{
-       return ac_build_tbuffer_load(ctx, rsrc, vindex, voffset, soffset,
-                                    immoffset, num_channels, dfmt, nfmt,
-                                    cache_policy, can_speculate, true);
-}
-
-LLVMValueRef
-ac_build_raw_tbuffer_load(struct ac_llvm_context *ctx,
-                         LLVMValueRef rsrc,
-                         LLVMValueRef voffset,
-                         LLVMValueRef soffset,
-                         LLVMValueRef immoffset,
-                         unsigned num_channels,
-                         unsigned dfmt,
-                         unsigned nfmt,
-                         unsigned cache_policy,
-                         bool can_speculate)
-{
-       return ac_build_tbuffer_load(ctx, rsrc, NULL, voffset, soffset,
-                                    immoffset, num_channels, dfmt, nfmt,
-                                    cache_policy, can_speculate, false);
-}
-
-LLVMValueRef
-ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
-                           LLVMValueRef rsrc,
-                           LLVMValueRef voffset,
-                           LLVMValueRef soffset,
-                           LLVMValueRef immoffset,
-                           unsigned cache_policy)
-{
-       LLVMValueRef res;
-
-       if (LLVM_VERSION_MAJOR >= 9) {
-               voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
-
-               /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
-               res = ac_build_buffer_load_common(ctx, rsrc, NULL,
-                                                 voffset, soffset,
-                                                 1, ctx->i16, cache_policy,
-                                                 false, false, false);
-       } else {
-               unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_16;
-               unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
-
-               res = ac_build_raw_tbuffer_load(ctx, rsrc, voffset, soffset,
-                                               immoffset, 1, dfmt, nfmt, cache_policy,
-                                               false);
-
-               res = LLVMBuildTrunc(ctx->builder, res, ctx->i16, "");
-       }
-
-       return res;
-}
-
-LLVMValueRef
-ac_build_tbuffer_load_byte(struct ac_llvm_context *ctx,
-                          LLVMValueRef rsrc,
-                          LLVMValueRef voffset,
-                          LLVMValueRef soffset,
-                          LLVMValueRef immoffset,
-                          unsigned cache_policy)
-{
-       LLVMValueRef res;
-
-       if (LLVM_VERSION_MAJOR >= 9) {
-               voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
-
-               /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
-               res = ac_build_buffer_load_common(ctx, rsrc, NULL,
-                                                 voffset, soffset,
-                                                 1, ctx->i8, cache_policy,
-                                                 false, false, false);
-       } else {
-               unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_8;
-               unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
+void ac_build_buffer_store_dword(struct ac_llvm_context *ctx, LLVMValueRef rsrc, LLVMValueRef vdata,
+                                 unsigned num_channels, LLVMValueRef voffset, LLVMValueRef soffset,
+                                 unsigned inst_offset, unsigned cache_policy)
+{
+   /* Split 3 channel stores, because only LLVM 9+ support 3-channel
+    * intrinsics. */
+   if (num_channels == 3 && !ac_has_vec3_support(ctx->chip_class, false)) {
+      LLVMValueRef v[3], v01;
+
+      for (int i = 0; i < 3; i++) {
+         v[i] = LLVMBuildExtractElement(ctx->builder, vdata, LLVMConstInt(ctx->i32, i, 0), "");
+      }
+      v01 = ac_build_gather_values(ctx, v, 2);
+
+      ac_build_buffer_store_dword(ctx, rsrc, v01, 2, voffset, soffset, inst_offset, cache_policy);
+      ac_build_buffer_store_dword(ctx, rsrc, v[2], 1, voffset, soffset, inst_offset + 8,
+                                  cache_policy);
+      return;
+   }
+
+   /* SWIZZLE_ENABLE requires that soffset isn't folded into voffset
+    * (voffset is swizzled, but soffset isn't swizzled).
+    * llvm.amdgcn.buffer.store doesn't have a separate soffset parameter.
+    */
+   if (!(cache_policy & ac_swizzled)) {
+      LLVMValueRef offset = soffset;
+
+      if (inst_offset)
+         offset = LLVMBuildAdd(ctx->builder, offset, LLVMConstInt(ctx->i32, inst_offset, 0), "");
+
+      ac_build_buffer_store_common(ctx, rsrc, ac_to_float(ctx, vdata), ctx->i32_0, voffset, offset,
+                                   cache_policy, false, false);
+      return;
+   }
+
+   static const unsigned dfmts[] = {V_008F0C_BUF_DATA_FORMAT_32, V_008F0C_BUF_DATA_FORMAT_32_32,
+                                    V_008F0C_BUF_DATA_FORMAT_32_32_32,
+                                    V_008F0C_BUF_DATA_FORMAT_32_32_32_32};
+   unsigned dfmt = dfmts[num_channels - 1];
+   unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
+   LLVMValueRef immoffset = LLVMConstInt(ctx->i32, inst_offset, 0);
+
+   ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset, immoffset, num_channels, dfmt,
+                              nfmt, cache_policy);
+}
+
+static LLVMValueRef ac_build_buffer_load_common(struct ac_llvm_context *ctx, LLVMValueRef rsrc,
+                                                LLVMValueRef vindex, LLVMValueRef voffset,
+                                                LLVMValueRef soffset, unsigned num_channels,
+                                                LLVMTypeRef channel_type, unsigned cache_policy,
+                                                bool can_speculate, bool use_format,
+                                                bool structurized)
+{
+   LLVMValueRef args[5];
+   int idx = 0;
+   args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
+   if (structurized)
+      args[idx++] = vindex ? vindex : ctx->i32_0;
+   args[idx++] = voffset ? voffset : ctx->i32_0;
+   args[idx++] = soffset ? soffset : ctx->i32_0;
+   args[idx++] = LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0);
+   unsigned func =
+      !ac_has_vec3_support(ctx->chip_class, use_format) && num_channels == 3 ? 4 : num_channels;
+   const char *indexing_kind = structurized ? "struct" : "raw";
+   char name[256], type_name[8];
+
+   /* D16 is only supported on gfx8+ */
+   assert(!use_format || (channel_type != ctx->f16 && channel_type != ctx->i16) ||
+          ctx->chip_class >= GFX8);
+
+   LLVMTypeRef type = func > 1 ? LLVMVectorType(channel_type, func) : channel_type;
+   ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
+
+   if (use_format) {
+      snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.load.format.%s", indexing_kind,
+               type_name);
+   } else {
+      snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.load.%s", indexing_kind, type_name);
+   }
+
+   return ac_build_intrinsic(ctx, name, type, args, idx, ac_get_load_intr_attribs(can_speculate));
+}
+
+LLVMValueRef ac_build_buffer_load(struct ac_llvm_context *ctx, LLVMValueRef rsrc, int num_channels,
+                                  LLVMValueRef vindex, LLVMValueRef voffset, LLVMValueRef soffset,
+                                  unsigned inst_offset, unsigned cache_policy, bool can_speculate,
+                                  bool allow_smem)
+{
+   LLVMValueRef offset = LLVMConstInt(ctx->i32, inst_offset, 0);
+   if (voffset)
+      offset = LLVMBuildAdd(ctx->builder, offset, voffset, "");
+   if (soffset)
+      offset = LLVMBuildAdd(ctx->builder, offset, soffset, "");
+
+   if (allow_smem && !(cache_policy & ac_slc) &&
+       (!(cache_policy & ac_glc) || ctx->chip_class >= GFX8)) {
+      assert(vindex == NULL);
+
+      LLVMValueRef result[8];
+
+      for (int i = 0; i < num_channels; i++) {
+         if (i) {
+            offset = LLVMBuildAdd(ctx->builder, offset, LLVMConstInt(ctx->i32, 4, 0), "");
+         }
+         LLVMValueRef args[3] = {
+            rsrc,
+            offset,
+            LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0),
+         };
+         result[i] = ac_build_intrinsic(ctx, "llvm.amdgcn.s.buffer.load.f32", ctx->f32, args, 3,
+                                        AC_FUNC_ATTR_READNONE);
+      }
+      if (num_channels == 1)
+         return result[0];
+
+      if (num_channels == 3 && !ac_has_vec3_support(ctx->chip_class, false))
+         result[num_channels++] = LLVMGetUndef(ctx->f32);
+      return ac_build_gather_values(ctx, result, num_channels);
+   }
+
+   return ac_build_buffer_load_common(ctx, rsrc, vindex, offset, ctx->i32_0, num_channels, ctx->f32,
+                                      cache_policy, can_speculate, false, false);
+}
+
+LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx, LLVMValueRef rsrc,
+                                         LLVMValueRef vindex, LLVMValueRef voffset,
+                                         unsigned num_channels, unsigned cache_policy,
+                                         bool can_speculate, bool d16)
+{
+   return ac_build_buffer_load_common(ctx, rsrc, vindex, voffset, ctx->i32_0, num_channels,
+                                      d16 ? ctx->f16 : ctx->f32, cache_policy, can_speculate, true,
+                                      true);
+}
+
+static LLVMValueRef ac_build_tbuffer_load(struct ac_llvm_context *ctx, LLVMValueRef rsrc,
+                                          LLVMValueRef vindex, LLVMValueRef voffset,
+                                          LLVMValueRef soffset, LLVMValueRef immoffset,
+                                          unsigned num_channels, unsigned dfmt, unsigned nfmt,
+                                          unsigned cache_policy, bool can_speculate,
+                                          bool structurized)
+{
+   voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
+
+   LLVMValueRef args[6];
+   int idx = 0;
+   args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
+   if (structurized)
+      args[idx++] = vindex ? vindex : ctx->i32_0;
+   args[idx++] = voffset ? voffset : ctx->i32_0;
+   args[idx++] = soffset ? soffset : ctx->i32_0;
+   args[idx++] = LLVMConstInt(ctx->i32, ac_get_tbuffer_format(ctx->chip_class, dfmt, nfmt), 0);
+   args[idx++] = LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0);
+   unsigned func =
+      !ac_has_vec3_support(ctx->chip_class, true) && num_channels == 3 ? 4 : num_channels;
+   const char *indexing_kind = structurized ? "struct" : "raw";
+   char name[256], type_name[8];
+
+   LLVMTypeRef type = func > 1 ? LLVMVectorType(ctx->i32, func) : ctx->i32;
+   ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
+
+   snprintf(name, sizeof(name), "llvm.amdgcn.%s.tbuffer.load.%s", indexing_kind, type_name);
+
+   return ac_build_intrinsic(ctx, name, type, args, idx, ac_get_load_intr_attribs(can_speculate));
+}
 
-               res = ac_build_raw_tbuffer_load(ctx, rsrc, voffset, soffset,
-                                               immoffset, 1, dfmt, nfmt, cache_policy,
-                                               false);
+LLVMValueRef ac_build_struct_tbuffer_load(struct ac_llvm_context *ctx, LLVMValueRef rsrc,
+                                          LLVMValueRef vindex, LLVMValueRef voffset,
+                                          LLVMValueRef soffset, LLVMValueRef immoffset,
+                                          unsigned num_channels, unsigned dfmt, unsigned nfmt,
+                                          unsigned cache_policy, bool can_speculate)
+{
+   return ac_build_tbuffer_load(ctx, rsrc, vindex, voffset, soffset, immoffset, num_channels, dfmt,
+                                nfmt, cache_policy, can_speculate, true);
+}
+
+LLVMValueRef ac_build_raw_tbuffer_load(struct ac_llvm_context *ctx, LLVMValueRef rsrc,
+                                       LLVMValueRef voffset, LLVMValueRef soffset,
+                                       LLVMValueRef immoffset, unsigned num_channels, unsigned dfmt,
+                                       unsigned nfmt, unsigned cache_policy, bool can_speculate)
+{
+   return ac_build_tbuffer_load(ctx, rsrc, NULL, voffset, soffset, immoffset, num_channels, dfmt,
+                                nfmt, cache_policy, can_speculate, false);
+}
+
+LLVMValueRef ac_build_tbuffer_load_short(struct ac_llvm_context *ctx, LLVMValueRef rsrc,
+                                         LLVMValueRef voffset, LLVMValueRef soffset,
+                                         LLVMValueRef immoffset, unsigned cache_policy)
+{
+   LLVMValueRef res;
+
+   if (LLVM_VERSION_MAJOR >= 9) {
+      voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
+
+      /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
+      res = ac_build_buffer_load_common(ctx, rsrc, NULL, voffset, soffset, 1, ctx->i16,
+                                        cache_policy, false, false, false);
+   } else {
+      unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_16;
+      unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
+
+      res = ac_build_raw_tbuffer_load(ctx, rsrc, voffset, soffset, immoffset, 1, dfmt, nfmt,
+                                      cache_policy, false);
+
+      res = LLVMBuildTrunc(ctx->builder, res, ctx->i16, "");
+   }
+
+   return res;
+}
+
+LLVMValueRef ac_build_tbuffer_load_byte(struct ac_llvm_context *ctx, LLVMValueRef rsrc,
+                                        LLVMValueRef voffset, LLVMValueRef soffset,
+                                        LLVMValueRef immoffset, unsigned cache_policy)
+{
+   LLVMValueRef res;
+
+   if (LLVM_VERSION_MAJOR >= 9) {
+      voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
+
+      /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
+      res = ac_build_buffer_load_common(ctx, rsrc, NULL, voffset, soffset, 1, ctx->i8, cache_policy,
+                                        false, false, false);
+   } else {
+      unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_8;
+      unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
+
+      res = ac_build_raw_tbuffer_load(ctx, rsrc, voffset, soffset, immoffset, 1, dfmt, nfmt,
+                                      cache_policy, false);
 
-               res = LLVMBuildTrunc(ctx->builder, res, ctx->i8, "");
-       }
-
-       return res;
+      res = LLVMBuildTrunc(ctx->builder, res, ctx->i8, "");
+   }
+
+   return res;
 }
 
 /**
@@ -1553,62 +1338,63 @@ ac_build_tbuffer_load_byte(struct ac_llvm_context *ctx,
  * The input exponent is expected to be biased analogous to IEEE-754, i.e. by
  * 2^(exp_bits-1) - 1 (as defined in OpenGL and other graphics APIs).
  */
-static LLVMValueRef
-ac_ufN_to_float(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned exp_bits, unsigned mant_bits)
+static LLVMValueRef ac_ufN_to_float(struct ac_llvm_context *ctx, LLVMValueRef src,
+                                    unsigned exp_bits, unsigned mant_bits)
 {
-       assert(LLVMTypeOf(src) == ctx->i32);
+   assert(LLVMTypeOf(src) == ctx->i32);
 
-       LLVMValueRef tmp;
-       LLVMValueRef mantissa;
-       mantissa = LLVMBuildAnd(ctx->builder, src, LLVMConstInt(ctx->i32, (1 << mant_bits) - 1, false), "");
+   LLVMValueRef tmp;
+   LLVMValueRef mantissa;
+   mantissa =
+      LLVMBuildAnd(ctx->builder, src, LLVMConstInt(ctx->i32, (1 << mant_bits) - 1, false), "");
 
-       /* Converting normal numbers is just a shift + correcting the exponent bias */
-       unsigned normal_shift = 23 - mant_bits;
-       unsigned bias_shift = 127 - ((1 << (exp_bits - 1)) - 1);
-       LLVMValueRef shifted, normal;
+   /* Converting normal numbers is just a shift + correcting the exponent bias */
+   unsigned normal_shift = 23 - mant_bits;
+   unsigned bias_shift = 127 - ((1 << (exp_bits - 1)) - 1);
+   LLVMValueRef shifted, normal;
 
-       shifted = LLVMBuildShl(ctx->builder, src, LLVMConstInt(ctx->i32, normal_shift, false), "");
-       normal = LLVMBuildAdd(ctx->builder, shifted, LLVMConstInt(ctx->i32, bias_shift << 23, false), "");
+   shifted = LLVMBuildShl(ctx->builder, src, LLVMConstInt(ctx->i32, normal_shift, false), "");
+   normal =
+      LLVMBuildAdd(ctx->builder, shifted, LLVMConstInt(ctx->i32, bias_shift << 23, false), "");
 
-       /* Converting nan/inf numbers is the same, but with a different exponent update */
-       LLVMValueRef naninf;
-       naninf = LLVMBuildOr(ctx->builder, normal, LLVMConstInt(ctx->i32, 0xff << 23, false), "");
+   /* Converting nan/inf numbers is the same, but with a different exponent update */
+   LLVMValueRef naninf;
+   naninf = LLVMBuildOr(ctx->builder, normal, LLVMConstInt(ctx->i32, 0xff << 23, false), "");
 
-       /* Converting denormals is the complex case: determine the leading zeros of the
-        * mantissa to obtain the correct shift for the mantissa and exponent correction.
-        */
-       LLVMValueRef denormal;
-       LLVMValueRef params[2] = {
-               mantissa,
-               ctx->i1true, /* result can be undef when arg is 0 */
-       };
-       LLVMValueRef ctlz = ac_build_intrinsic(ctx, "llvm.ctlz.i32", ctx->i32,
-                                             params, 2, AC_FUNC_ATTR_READNONE);
+   /* Converting denormals is the complex case: determine the leading zeros of the
+    * mantissa to obtain the correct shift for the mantissa and exponent correction.
+    */
+   LLVMValueRef denormal;
+   LLVMValueRef params[2] = {
+      mantissa, ctx->i1true, /* result can be undef when arg is 0 */
+   };
+   LLVMValueRef ctlz =
+      ac_build_intrinsic(ctx, "llvm.ctlz.i32", ctx->i32, params, 2, AC_FUNC_ATTR_READNONE);
 
-       /* Shift such that the leading 1 ends up as the LSB of the exponent field. */
-       tmp = LLVMBuildSub(ctx->builder, ctlz, LLVMConstInt(ctx->i32, 8, false), "");
-       denormal = LLVMBuildShl(ctx->builder, mantissa, tmp, "");
+   /* Shift such that the leading 1 ends up as the LSB of the exponent field. */
+   tmp = LLVMBuildSub(ctx->builder, ctlz, LLVMConstInt(ctx->i32, 8, false), "");
+   denormal = LLVMBuildShl(ctx->builder, mantissa, tmp, "");
 
-       unsigned denormal_exp = bias_shift + (32 - mant_bits) - 1;
-       tmp = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, denormal_exp, false), ctlz, "");
-       tmp = LLVMBuildShl(ctx->builder, tmp, LLVMConstInt(ctx->i32, 23, false), "");
-       denormal = LLVMBuildAdd(ctx->builder, denormal, tmp, "");
+   unsigned denormal_exp = bias_shift + (32 - mant_bits) - 1;
+   tmp = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, denormal_exp, false), ctlz, "");
+   tmp = LLVMBuildShl(ctx->builder, tmp, LLVMConstInt(ctx->i32, 23, false), "");
+   denormal = LLVMBuildAdd(ctx->builder, denormal, tmp, "");
 
-       /* Select the final result. */
-       LLVMValueRef result;
+   /* Select the final result. */
+   LLVMValueRef result;
 
-       tmp = LLVMBuildICmp(ctx->builder, LLVMIntUGE, src,
-                           LLVMConstInt(ctx->i32, ((1 << exp_bits) - 1) << mant_bits, false), "");
-       result = LLVMBuildSelect(ctx->builder, tmp, naninf, normal, "");
+   tmp = LLVMBuildICmp(ctx->builder, LLVMIntUGE, src,
+                       LLVMConstInt(ctx->i32, ((1 << exp_bits) - 1) << mant_bits, false), "");
+   result = LLVMBuildSelect(ctx->builder, tmp, naninf, normal, "");
 
-       tmp = LLVMBuildICmp(ctx->builder, LLVMIntUGE, src,
-                           LLVMConstInt(ctx->i32, 1 << mant_bits, false), "");
-       result = LLVMBuildSelect(ctx->builder, tmp, result, denormal, "");
+   tmp = LLVMBuildICmp(ctx->builder, LLVMIntUGE, src, LLVMConstInt(ctx->i32, 1 << mant_bits, false),
+                       "");
+   result = LLVMBuildSelect(ctx->builder, tmp, result, denormal, "");
 
-       tmp = LLVMBuildICmp(ctx->builder, LLVMIntNE, src, ctx->i32_0, "");
-       result = LLVMBuildSelect(ctx->builder, tmp, result, ctx->i32_0, "");
+   tmp = LLVMBuildICmp(ctx->builder, LLVMIntNE, src, ctx->i32_0, "");
+   result = LLVMBuildSelect(ctx->builder, tmp, result, ctx->i32_0, "");
 
-       return ac_to_float(ctx, result);
+   return ac_to_float(ctx, result);
 }
 
 /**
@@ -1629,354 +1415,305 @@ ac_ufN_to_float(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned exp_bits
  * \param rsrc buffer resource descriptor
  * \return the resulting vector of floats or integers bitcast to <4 x i32>
  */
-LLVMValueRef
-ac_build_opencoded_load_format(struct ac_llvm_context *ctx,
-                              unsigned log_size,
-                              unsigned num_channels,
-                              unsigned format,
-                              bool reverse,
-                              bool known_aligned,
-                              LLVMValueRef rsrc,
-                              LLVMValueRef vindex,
-                              LLVMValueRef voffset,
-                              LLVMValueRef soffset,
-                              unsigned cache_policy,
-                              bool can_speculate)
-{
-       LLVMValueRef tmp;
-       unsigned load_log_size = log_size;
-       unsigned load_num_channels = num_channels;
-       if (log_size == 3) {
-               load_log_size = 2;
-               if (format == AC_FETCH_FORMAT_FLOAT) {
-                       load_num_channels = 2 * num_channels;
-               } else {
-                       load_num_channels = 1; /* 10_11_11 or 2_10_10_10 */
-               }
-       }
-
-       int log_recombine = 0;
-       if ((ctx->chip_class == GFX6 || ctx->chip_class >= GFX10) && !known_aligned) {
-               /* Avoid alignment restrictions by loading one byte at a time. */
-               load_num_channels <<= load_log_size;
-               log_recombine = load_log_size;
-               load_log_size = 0;
-       } else if (load_num_channels == 2 || load_num_channels == 4) {
-               log_recombine = -util_logbase2(load_num_channels);
-               load_num_channels = 1;
-               load_log_size += -log_recombine;
-       }
-
-       assert(load_log_size >= 2 || LLVM_VERSION_MAJOR >= 9);
-
-       LLVMValueRef loads[32]; /* up to 32 bytes */
-       for (unsigned i = 0; i < load_num_channels; ++i) {
-               tmp = LLVMBuildAdd(ctx->builder, soffset,
-                                  LLVMConstInt(ctx->i32, i << load_log_size, false), "");
-               LLVMTypeRef channel_type = load_log_size == 0 ? ctx->i8 :
-                                          load_log_size == 1 ? ctx->i16 : ctx->i32;
-               unsigned num_channels = 1 << (MAX2(load_log_size, 2) - 2);
-               loads[i] = ac_build_buffer_load_common(
-                               ctx, rsrc, vindex, voffset, tmp,
-                               num_channels, channel_type, cache_policy,
-                               can_speculate, false, true);
-               if (load_log_size >= 2)
-                       loads[i] = ac_to_integer(ctx, loads[i]);
-       }
-
-       if (log_recombine > 0) {
-               /* Recombine bytes if necessary (GFX6 only) */
-               LLVMTypeRef dst_type = log_recombine == 2 ? ctx->i32 : ctx->i16;
-
-               for (unsigned src = 0, dst = 0; src < load_num_channels; ++dst) {
-                       LLVMValueRef accum = NULL;
-                       for (unsigned i = 0; i < (1 << log_recombine); ++i, ++src) {
-                               tmp = LLVMBuildZExt(ctx->builder, loads[src], dst_type, "");
-                               if (i == 0) {
-                                       accum = tmp;
-                               } else {
-                                       tmp = LLVMBuildShl(ctx->builder, tmp,
-                                                          LLVMConstInt(dst_type, 8 * i, false), "");
-                                       accum = LLVMBuildOr(ctx->builder, accum, tmp, "");
-                               }
-                       }
-                       loads[dst] = accum;
-               }
-       } else if (log_recombine < 0) {
-               /* Split vectors of dwords */
-               if (load_log_size > 2) {
-                       assert(load_num_channels == 1);
-                       LLVMValueRef loaded = loads[0];
-                       unsigned log_split = load_log_size - 2;
-                       log_recombine += log_split;
-                       load_num_channels = 1 << log_split;
-                       load_log_size = 2;
-                       for (unsigned i = 0; i < load_num_channels; ++i) {
-                               tmp = LLVMConstInt(ctx->i32, i, false);
-                               loads[i] = LLVMBuildExtractElement(ctx->builder, loaded, tmp, "");
-                       }
-               }
-
-               /* Further split dwords and shorts if required */
-               if (log_recombine < 0) {
-                       for (unsigned src = load_num_channels,
-                                     dst = load_num_channels << -log_recombine;
-                            src > 0; --src) {
-                               unsigned dst_bits = 1 << (3 + load_log_size + log_recombine);
-                               LLVMTypeRef dst_type = LLVMIntTypeInContext(ctx->context, dst_bits);
-                               LLVMValueRef loaded = loads[src - 1];
-                               LLVMTypeRef loaded_type = LLVMTypeOf(loaded);
-                               for (unsigned i = 1 << -log_recombine; i > 0; --i, --dst) {
-                                       tmp = LLVMConstInt(loaded_type, dst_bits * (i - 1), false);
-                                       tmp = LLVMBuildLShr(ctx->builder, loaded, tmp, "");
-                                       loads[dst - 1] = LLVMBuildTrunc(ctx->builder, tmp, dst_type, "");
-                               }
-                       }
-               }
-       }
-
-       if (log_size == 3) {
-               if (format == AC_FETCH_FORMAT_FLOAT) {
-                       for (unsigned i = 0; i < num_channels; ++i) {
-                               tmp = ac_build_gather_values(ctx, &loads[2 * i], 2);
-                               loads[i] = LLVMBuildBitCast(ctx->builder, tmp, ctx->f64, "");
-                       }
-               } else if (format == AC_FETCH_FORMAT_FIXED) {
-                       /* 10_11_11_FLOAT */
-                       LLVMValueRef data = loads[0];
-                       LLVMValueRef i32_2047 = LLVMConstInt(ctx->i32, 2047, false);
-                       LLVMValueRef r = LLVMBuildAnd(ctx->builder, data, i32_2047, "");
-                       tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 11, false), "");
-                       LLVMValueRef g = LLVMBuildAnd(ctx->builder, tmp, i32_2047, "");
-                       LLVMValueRef b = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 22, false), "");
-
-                       loads[0] = ac_to_integer(ctx, ac_ufN_to_float(ctx, r, 5, 6));
-                       loads[1] = ac_to_integer(ctx, ac_ufN_to_float(ctx, g, 5, 6));
-                       loads[2] = ac_to_integer(ctx, ac_ufN_to_float(ctx, b, 5, 5));
-
-                       num_channels = 3;
-                       log_size = 2;
-                       format = AC_FETCH_FORMAT_FLOAT;
-               } else {
-                       /* 2_10_10_10 data formats */
-                       LLVMValueRef data = loads[0];
-                       LLVMTypeRef i10 = LLVMIntTypeInContext(ctx->context, 10);
-                       LLVMTypeRef i2 = LLVMIntTypeInContext(ctx->context, 2);
-                       loads[0] = LLVMBuildTrunc(ctx->builder, data, i10, "");
-                       tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 10, false), "");
-                       loads[1] = LLVMBuildTrunc(ctx->builder, tmp, i10, "");
-                       tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 20, false), "");
-                       loads[2] = LLVMBuildTrunc(ctx->builder, tmp, i10, "");
-                       tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 30, false), "");
-                       loads[3] = LLVMBuildTrunc(ctx->builder, tmp, i2, "");
-
-                       num_channels = 4;
-               }
-       }
-
-       if (format == AC_FETCH_FORMAT_FLOAT) {
-               if (log_size != 2) {
-                       for (unsigned chan = 0; chan < num_channels; ++chan) {
-                               tmp = ac_to_float(ctx, loads[chan]);
-                               if (log_size == 3)
-                                       tmp = LLVMBuildFPTrunc(ctx->builder, tmp, ctx->f32, "");
-                               else if (log_size == 1)
-                                       tmp = LLVMBuildFPExt(ctx->builder, tmp, ctx->f32, "");
-                               loads[chan] = ac_to_integer(ctx, tmp);
-                       }
-               }
-       } else if (format == AC_FETCH_FORMAT_UINT) {
-               if (log_size != 2) {
-                       for (unsigned chan = 0; chan < num_channels; ++chan)
-                               loads[chan] = LLVMBuildZExt(ctx->builder, loads[chan], ctx->i32, "");
-               }
-       } else if (format == AC_FETCH_FORMAT_SINT) {
-               if (log_size != 2) {
-                       for (unsigned chan = 0; chan < num_channels; ++chan)
-                               loads[chan] = LLVMBuildSExt(ctx->builder, loads[chan], ctx->i32, "");
-               }
-       } else {
-               bool unsign = format == AC_FETCH_FORMAT_UNORM ||
-                             format == AC_FETCH_FORMAT_USCALED ||
-                             format == AC_FETCH_FORMAT_UINT;
-
-               for (unsigned chan = 0; chan < num_channels; ++chan) {
-                       if (unsign) {
-                               tmp = LLVMBuildUIToFP(ctx->builder, loads[chan], ctx->f32, "");
-                       } else {
-                               tmp = LLVMBuildSIToFP(ctx->builder, loads[chan], ctx->f32, "");
-                       }
-
-                       LLVMValueRef scale = NULL;
-                       if (format == AC_FETCH_FORMAT_FIXED) {
-                               assert(log_size == 2);
-                               scale = LLVMConstReal(ctx->f32, 1.0 / 0x10000);
-                       } else if (format == AC_FETCH_FORMAT_UNORM) {
-                               unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(loads[chan]));
-                               scale = LLVMConstReal(ctx->f32, 1.0 / (((uint64_t)1 << bits) - 1));
-                       } else if (format == AC_FETCH_FORMAT_SNORM) {
-                               unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(loads[chan]));
-                               scale = LLVMConstReal(ctx->f32, 1.0 / (((uint64_t)1 << (bits - 1)) - 1));
-                       }
-                       if (scale)
-                               tmp = LLVMBuildFMul(ctx->builder, tmp, scale, "");
-
-                       if (format == AC_FETCH_FORMAT_SNORM) {
-                               /* Clamp to [-1, 1] */
-                               LLVMValueRef neg_one = LLVMConstReal(ctx->f32, -1.0);
-                               LLVMValueRef clamp =
-                                       LLVMBuildFCmp(ctx->builder, LLVMRealULT, tmp, neg_one, "");
-                               tmp = LLVMBuildSelect(ctx->builder, clamp, neg_one, tmp, "");
-                       }
-
-                       loads[chan] = ac_to_integer(ctx, tmp);
-               }
-       }
-
-       while (num_channels < 4) {
-               if (format == AC_FETCH_FORMAT_UINT || format == AC_FETCH_FORMAT_SINT) {
-                       loads[num_channels] = num_channels == 3 ? ctx->i32_1 : ctx->i32_0;
-               } else {
-                       loads[num_channels] = ac_to_integer(ctx, num_channels == 3 ? ctx->f32_1 : ctx->f32_0);
-               }
-               num_channels++;
-       }
-
-       if (reverse) {
-               tmp = loads[0];
-               loads[0] = loads[2];
-               loads[2] = tmp;
-       }
-
-       return ac_build_gather_values(ctx, loads, 4);
-}
-
-static void
-ac_build_tbuffer_store(struct ac_llvm_context *ctx,
-                      LLVMValueRef rsrc,
-                      LLVMValueRef vdata,
-                      LLVMValueRef vindex,
-                      LLVMValueRef voffset,
-                      LLVMValueRef soffset,
-                      LLVMValueRef immoffset,
-                      unsigned num_channels,
-                      unsigned dfmt,
-                      unsigned nfmt,
-                      unsigned cache_policy,
-                      bool structurized)
-{
-       voffset = LLVMBuildAdd(ctx->builder, voffset ? voffset : ctx->i32_0,
-                              immoffset, "");
-
-       LLVMValueRef args[7];
-       int idx = 0;
-       args[idx++] = vdata;
-       args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
-       if (structurized)
-               args[idx++] = vindex ? vindex : ctx->i32_0;
-       args[idx++] = voffset ? voffset : ctx->i32_0;
-       args[idx++] = soffset ? soffset : ctx->i32_0;
-       args[idx++] = LLVMConstInt(ctx->i32, ac_get_tbuffer_format(ctx->chip_class, dfmt, nfmt), 0);
-       args[idx++] = LLVMConstInt(ctx->i32, cache_policy, 0);
-       unsigned func = !ac_has_vec3_support(ctx->chip_class, true) && num_channels == 3 ? 4 : num_channels;
-       const char *indexing_kind = structurized ? "struct" : "raw";
-       char name[256], type_name[8];
-
-       LLVMTypeRef type = func > 1 ? LLVMVectorType(ctx->i32, func) : ctx->i32;
-       ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
-
-       snprintf(name, sizeof(name), "llvm.amdgcn.%s.tbuffer.store.%s",
-                indexing_kind, type_name);
-
-       ac_build_intrinsic(ctx, name, ctx->voidt, args, idx,
-                          AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY);
-}
-
-void
-ac_build_struct_tbuffer_store(struct ac_llvm_context *ctx,
-                             LLVMValueRef rsrc,
-                             LLVMValueRef vdata,
-                             LLVMValueRef vindex,
-                             LLVMValueRef voffset,
-                             LLVMValueRef soffset,
-                             LLVMValueRef immoffset,
-                             unsigned num_channels,
-                             unsigned dfmt,
-                             unsigned nfmt,
-                             unsigned cache_policy)
-{
-       ac_build_tbuffer_store(ctx, rsrc, vdata, vindex, voffset, soffset,
-                              immoffset, num_channels, dfmt, nfmt, cache_policy,
-                              true);
-}
-
-void
-ac_build_raw_tbuffer_store(struct ac_llvm_context *ctx,
-                          LLVMValueRef rsrc,
-                          LLVMValueRef vdata,
-                          LLVMValueRef voffset,
-                          LLVMValueRef soffset,
-                          LLVMValueRef immoffset,
-                          unsigned num_channels,
-                          unsigned dfmt,
-                          unsigned nfmt,
-                          unsigned cache_policy)
-{
-       ac_build_tbuffer_store(ctx, rsrc, vdata, NULL, voffset, soffset,
-                              immoffset, num_channels, dfmt, nfmt, cache_policy,
-                              false);
-}
-
-void
-ac_build_tbuffer_store_short(struct ac_llvm_context *ctx,
-                            LLVMValueRef rsrc,
-                            LLVMValueRef vdata,
-                            LLVMValueRef voffset,
-                            LLVMValueRef soffset,
-                            unsigned cache_policy)
-{
-       vdata = LLVMBuildBitCast(ctx->builder, vdata, ctx->i16, "");
-
-       if (LLVM_VERSION_MAJOR >= 9) {
-               /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
-               ac_build_buffer_store_common(ctx, rsrc, vdata, NULL,
-                                            voffset, soffset, cache_policy,
-                                            false, false);
-       } else {
-               unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_16;
-               unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
-
-               vdata = LLVMBuildZExt(ctx->builder, vdata, ctx->i32, "");
-
-               ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset,
-                                          ctx->i32_0, 1, dfmt, nfmt, cache_policy);
-       }
-}
-
-void
-ac_build_tbuffer_store_byte(struct ac_llvm_context *ctx,
-                           LLVMValueRef rsrc,
-                           LLVMValueRef vdata,
-                           LLVMValueRef voffset,
-                           LLVMValueRef soffset,
-                           unsigned cache_policy)
-{
-       vdata = LLVMBuildBitCast(ctx->builder, vdata, ctx->i8, "");
-
-       if (LLVM_VERSION_MAJOR >= 9) {
-               /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
-               ac_build_buffer_store_common(ctx, rsrc, vdata, NULL,
-                                            voffset, soffset, cache_policy,
-                                            false, false);
-       } else {
-               unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_8;
-               unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
-
-               vdata = LLVMBuildZExt(ctx->builder, vdata, ctx->i32, "");
-
-               ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset,
-                                          ctx->i32_0, 1, dfmt, nfmt, cache_policy);
-       }
+LLVMValueRef ac_build_opencoded_load_format(struct ac_llvm_context *ctx, unsigned log_size,
+                                            unsigned num_channels, unsigned format, bool reverse,
+                                            bool known_aligned, LLVMValueRef rsrc,
+                                            LLVMValueRef vindex, LLVMValueRef voffset,
+                                            LLVMValueRef soffset, unsigned cache_policy,
+                                            bool can_speculate)
+{
+   LLVMValueRef tmp;
+   unsigned load_log_size = log_size;
+   unsigned load_num_channels = num_channels;
+   if (log_size == 3) {
+      load_log_size = 2;
+      if (format == AC_FETCH_FORMAT_FLOAT) {
+         load_num_channels = 2 * num_channels;
+      } else {
+         load_num_channels = 1; /* 10_11_11 or 2_10_10_10 */
+      }
+   }
+
+   int log_recombine = 0;
+   if ((ctx->chip_class == GFX6 || ctx->chip_class >= GFX10) && !known_aligned) {
+      /* Avoid alignment restrictions by loading one byte at a time. */
+      load_num_channels <<= load_log_size;
+      log_recombine = load_log_size;
+      load_log_size = 0;
+   } else if (load_num_channels == 2 || load_num_channels == 4) {
+      log_recombine = -util_logbase2(load_num_channels);
+      load_num_channels = 1;
+      load_log_size += -log_recombine;
+   }
+
+   assert(load_log_size >= 2 || LLVM_VERSION_MAJOR >= 9);
+
+   LLVMValueRef loads[32]; /* up to 32 bytes */
+   for (unsigned i = 0; i < load_num_channels; ++i) {
+      tmp =
+         LLVMBuildAdd(ctx->builder, soffset, LLVMConstInt(ctx->i32, i << load_log_size, false), "");
+      LLVMTypeRef channel_type =
+         load_log_size == 0 ? ctx->i8 : load_log_size == 1 ? ctx->i16 : ctx->i32;
+      unsigned num_channels = 1 << (MAX2(load_log_size, 2) - 2);
+      loads[i] =
+         ac_build_buffer_load_common(ctx, rsrc, vindex, voffset, tmp, num_channels, channel_type,
+                                     cache_policy, can_speculate, false, true);
+      if (load_log_size >= 2)
+         loads[i] = ac_to_integer(ctx, loads[i]);
+   }
+
+   if (log_recombine > 0) {
+      /* Recombine bytes if necessary (GFX6 only) */
+      LLVMTypeRef dst_type = log_recombine == 2 ? ctx->i32 : ctx->i16;
+
+      for (unsigned src = 0, dst = 0; src < load_num_channels; ++dst) {
+         LLVMValueRef accum = NULL;
+         for (unsigned i = 0; i < (1 << log_recombine); ++i, ++src) {
+            tmp = LLVMBuildZExt(ctx->builder, loads[src], dst_type, "");
+            if (i == 0) {
+               accum = tmp;
+            } else {
+               tmp = LLVMBuildShl(ctx->builder, tmp, LLVMConstInt(dst_type, 8 * i, false), "");
+               accum = LLVMBuildOr(ctx->builder, accum, tmp, "");
+            }
+         }
+         loads[dst] = accum;
+      }
+   } else if (log_recombine < 0) {
+      /* Split vectors of dwords */
+      if (load_log_size > 2) {
+         assert(load_num_channels == 1);
+         LLVMValueRef loaded = loads[0];
+         unsigned log_split = load_log_size - 2;
+         log_recombine += log_split;
+         load_num_channels = 1 << log_split;
+         load_log_size = 2;
+         for (unsigned i = 0; i < load_num_channels; ++i) {
+            tmp = LLVMConstInt(ctx->i32, i, false);
+            loads[i] = LLVMBuildExtractElement(ctx->builder, loaded, tmp, "");
+         }
+      }
+
+      /* Further split dwords and shorts if required */
+      if (log_recombine < 0) {
+         for (unsigned src = load_num_channels, dst = load_num_channels << -log_recombine; src > 0;
+              --src) {
+            unsigned dst_bits = 1 << (3 + load_log_size + log_recombine);
+            LLVMTypeRef dst_type = LLVMIntTypeInContext(ctx->context, dst_bits);
+            LLVMValueRef loaded = loads[src - 1];
+            LLVMTypeRef loaded_type = LLVMTypeOf(loaded);
+            for (unsigned i = 1 << -log_recombine; i > 0; --i, --dst) {
+               tmp = LLVMConstInt(loaded_type, dst_bits * (i - 1), false);
+               tmp = LLVMBuildLShr(ctx->builder, loaded, tmp, "");
+               loads[dst - 1] = LLVMBuildTrunc(ctx->builder, tmp, dst_type, "");
+            }
+         }
+      }
+   }
+
+   if (log_size == 3) {
+      if (format == AC_FETCH_FORMAT_FLOAT) {
+         for (unsigned i = 0; i < num_channels; ++i) {
+            tmp = ac_build_gather_values(ctx, &loads[2 * i], 2);
+            loads[i] = LLVMBuildBitCast(ctx->builder, tmp, ctx->f64, "");
+         }
+      } else if (format == AC_FETCH_FORMAT_FIXED) {
+         /* 10_11_11_FLOAT */
+         LLVMValueRef data = loads[0];
+         LLVMValueRef i32_2047 = LLVMConstInt(ctx->i32, 2047, false);
+         LLVMValueRef r = LLVMBuildAnd(ctx->builder, data, i32_2047, "");
+         tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 11, false), "");
+         LLVMValueRef g = LLVMBuildAnd(ctx->builder, tmp, i32_2047, "");
+         LLVMValueRef b = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 22, false), "");
+
+         loads[0] = ac_to_integer(ctx, ac_ufN_to_float(ctx, r, 5, 6));
+         loads[1] = ac_to_integer(ctx, ac_ufN_to_float(ctx, g, 5, 6));
+         loads[2] = ac_to_integer(ctx, ac_ufN_to_float(ctx, b, 5, 5));
+
+         num_channels = 3;
+         log_size = 2;
+         format = AC_FETCH_FORMAT_FLOAT;
+      } else {
+         /* 2_10_10_10 data formats */
+         LLVMValueRef data = loads[0];
+         LLVMTypeRef i10 = LLVMIntTypeInContext(ctx->context, 10);
+         LLVMTypeRef i2 = LLVMIntTypeInContext(ctx->context, 2);
+         loads[0] = LLVMBuildTrunc(ctx->builder, data, i10, "");
+         tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 10, false), "");
+         loads[1] = LLVMBuildTrunc(ctx->builder, tmp, i10, "");
+         tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 20, false), "");
+         loads[2] = LLVMBuildTrunc(ctx->builder, tmp, i10, "");
+         tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 30, false), "");
+         loads[3] = LLVMBuildTrunc(ctx->builder, tmp, i2, "");
+
+         num_channels = 4;
+      }
+   }
+
+   if (format == AC_FETCH_FORMAT_FLOAT) {
+      if (log_size != 2) {
+         for (unsigned chan = 0; chan < num_channels; ++chan) {
+            tmp = ac_to_float(ctx, loads[chan]);
+            if (log_size == 3)
+               tmp = LLVMBuildFPTrunc(ctx->builder, tmp, ctx->f32, "");
+            else if (log_size == 1)
+               tmp = LLVMBuildFPExt(ctx->builder, tmp, ctx->f32, "");
+            loads[chan] = ac_to_integer(ctx, tmp);
+         }
+      }
+   } else if (format == AC_FETCH_FORMAT_UINT) {
+      if (log_size != 2) {
+         for (unsigned chan = 0; chan < num_channels; ++chan)
+            loads[chan] = LLVMBuildZExt(ctx->builder, loads[chan], ctx->i32, "");
+      }
+   } else if (format == AC_FETCH_FORMAT_SINT) {
+      if (log_size != 2) {
+         for (unsigned chan = 0; chan < num_channels; ++chan)
+            loads[chan] = LLVMBuildSExt(ctx->builder, loads[chan], ctx->i32, "");
+      }
+   } else {
+      bool unsign = format == AC_FETCH_FORMAT_UNORM || format == AC_FETCH_FORMAT_USCALED ||
+                    format == AC_FETCH_FORMAT_UINT;
+
+      for (unsigned chan = 0; chan < num_channels; ++chan) {
+         if (unsign) {
+            tmp = LLVMBuildUIToFP(ctx->builder, loads[chan], ctx->f32, "");
+         } else {
+            tmp = LLVMBuildSIToFP(ctx->builder, loads[chan], ctx->f32, "");
+         }
+
+         LLVMValueRef scale = NULL;
+         if (format == AC_FETCH_FORMAT_FIXED) {
+            assert(log_size == 2);
+            scale = LLVMConstReal(ctx->f32, 1.0 / 0x10000);
+         } else if (format == AC_FETCH_FORMAT_UNORM) {
+            unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(loads[chan]));
+            scale = LLVMConstReal(ctx->f32, 1.0 / (((uint64_t)1 << bits) - 1));
+         } else if (format == AC_FETCH_FORMAT_SNORM) {
+            unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(loads[chan]));
+            scale = LLVMConstReal(ctx->f32, 1.0 / (((uint64_t)1 << (bits - 1)) - 1));
+         }
+         if (scale)
+            tmp = LLVMBuildFMul(ctx->builder, tmp, scale, "");
+
+         if (format == AC_FETCH_FORMAT_SNORM) {
+            /* Clamp to [-1, 1] */
+            LLVMValueRef neg_one = LLVMConstReal(ctx->f32, -1.0);
+            LLVMValueRef clamp = LLVMBuildFCmp(ctx->builder, LLVMRealULT, tmp, neg_one, "");
+            tmp = LLVMBuildSelect(ctx->builder, clamp, neg_one, tmp, "");
+         }
+
+         loads[chan] = ac_to_integer(ctx, tmp);
+      }
+   }
+
+   while (num_channels < 4) {
+      if (format == AC_FETCH_FORMAT_UINT || format == AC_FETCH_FORMAT_SINT) {
+         loads[num_channels] = num_channels == 3 ? ctx->i32_1 : ctx->i32_0;
+      } else {
+         loads[num_channels] = ac_to_integer(ctx, num_channels == 3 ? ctx->f32_1 : ctx->f32_0);
+      }
+      num_channels++;
+   }
+
+   if (reverse) {
+      tmp = loads[0];
+      loads[0] = loads[2];
+      loads[2] = tmp;
+   }
+
+   return ac_build_gather_values(ctx, loads, 4);
+}
+
+static void ac_build_tbuffer_store(struct ac_llvm_context *ctx, LLVMValueRef rsrc,
+                                   LLVMValueRef vdata, LLVMValueRef vindex, LLVMValueRef voffset,
+                                   LLVMValueRef soffset, LLVMValueRef immoffset,
+                                   unsigned num_channels, unsigned dfmt, unsigned nfmt,
+                                   unsigned cache_policy, bool structurized)
+{
+   voffset = LLVMBuildAdd(ctx->builder, voffset ? voffset : ctx->i32_0, immoffset, "");
+
+   LLVMValueRef args[7];
+   int idx = 0;
+   args[idx++] = vdata;
+   args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
+   if (structurized)
+      args[idx++] = vindex ? vindex : ctx->i32_0;
+   args[idx++] = voffset ? voffset : ctx->i32_0;
+   args[idx++] = soffset ? soffset : ctx->i32_0;
+   args[idx++] = LLVMConstInt(ctx->i32, ac_get_tbuffer_format(ctx->chip_class, dfmt, nfmt), 0);
+   args[idx++] = LLVMConstInt(ctx->i32, cache_policy, 0);
+   unsigned func =
+      !ac_has_vec3_support(ctx->chip_class, true) && num_channels == 3 ? 4 : num_channels;
+   const char *indexing_kind = structurized ? "struct" : "raw";
+   char name[256], type_name[8];
+
+   LLVMTypeRef type = func > 1 ? LLVMVectorType(ctx->i32, func) : ctx->i32;
+   ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
+
+   snprintf(name, sizeof(name), "llvm.amdgcn.%s.tbuffer.store.%s", indexing_kind, type_name);
+
+   ac_build_intrinsic(ctx, name, ctx->voidt, args, idx, AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY);
+}
+
+void ac_build_struct_tbuffer_store(struct ac_llvm_context *ctx, LLVMValueRef rsrc,
+                                   LLVMValueRef vdata, LLVMValueRef vindex, LLVMValueRef voffset,
+                                   LLVMValueRef soffset, LLVMValueRef immoffset,
+                                   unsigned num_channels, unsigned dfmt, unsigned nfmt,
+                                   unsigned cache_policy)
+{
+   ac_build_tbuffer_store(ctx, rsrc, vdata, vindex, voffset, soffset, immoffset, num_channels, dfmt,
+                          nfmt, cache_policy, true);
+}
+
+void ac_build_raw_tbuffer_store(struct ac_llvm_context *ctx, LLVMValueRef rsrc, LLVMValueRef vdata,
+                                LLVMValueRef voffset, LLVMValueRef soffset, LLVMValueRef immoffset,
+                                unsigned num_channels, unsigned dfmt, unsigned nfmt,
+                                unsigned cache_policy)
+{
+   ac_build_tbuffer_store(ctx, rsrc, vdata, NULL, voffset, soffset, immoffset, num_channels, dfmt,
+                          nfmt, cache_policy, false);
+}
+
+void ac_build_tbuffer_store_short(struct ac_llvm_context *ctx, LLVMValueRef rsrc,
+                                  LLVMValueRef vdata, LLVMValueRef voffset, LLVMValueRef soffset,
+                                  unsigned cache_policy)
+{
+   vdata = LLVMBuildBitCast(ctx->builder, vdata, ctx->i16, "");
+
+   if (LLVM_VERSION_MAJOR >= 9) {
+      /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
+      ac_build_buffer_store_common(ctx, rsrc, vdata, NULL, voffset, soffset, cache_policy, false,
+                                   false);
+   } else {
+      unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_16;
+      unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
+
+      vdata = LLVMBuildZExt(ctx->builder, vdata, ctx->i32, "");
+
+      ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset, ctx->i32_0, 1, dfmt, nfmt,
+                                 cache_policy);
+   }
+}
+
+void ac_build_tbuffer_store_byte(struct ac_llvm_context *ctx, LLVMValueRef rsrc, LLVMValueRef vdata,
+                                 LLVMValueRef voffset, LLVMValueRef soffset, unsigned cache_policy)
+{
+   vdata = LLVMBuildBitCast(ctx->builder, vdata, ctx->i8, "");
+
+   if (LLVM_VERSION_MAJOR >= 9) {
+      /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
+      ac_build_buffer_store_common(ctx, rsrc, vdata, NULL, voffset, soffset, cache_policy, false,
+                                   false);
+   } else {
+      unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_8;
+      unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
+
+      vdata = LLVMBuildZExt(ctx->builder, vdata, ctx->i32, "");
+
+      ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset, ctx->i32_0, 1, dfmt, nfmt,
+                                 cache_policy);
+   }
 }
 /**
  * Set range metadata on an instruction.  This can only be used on load and
@@ -1985,40 +1722,37 @@ ac_build_tbuffer_store_byte(struct ac_llvm_context *ctx,
  * \p lo is the minimum value inclusive.
  * \p hi is the maximum value exclusive.
  */
-static void set_range_metadata(struct ac_llvm_context *ctx,
-                              LLVMValueRef value, unsigned lo, unsigned hi)
+static void set_range_metadata(struct ac_llvm_context *ctx, LLVMValueRef value, unsigned lo,
+                               unsigned hi)
 {
-       LLVMValueRef range_md, md_args[2];
-       LLVMTypeRef type = LLVMTypeOf(value);
-       LLVMContextRef context = LLVMGetTypeContext(type);
+   LLVMValueRef range_md, md_args[2];
+   LLVMTypeRef type = LLVMTypeOf(value);
+   LLVMContextRef context = LLVMGetTypeContext(type);
 
-       md_args[0] = LLVMConstInt(type, lo, false);
-       md_args[1] = LLVMConstInt(type, hi, false);
-       range_md = LLVMMDNodeInContext(context, md_args, 2);
-       LLVMSetMetadata(value, ctx->range_md_kind, range_md);
+   md_args[0] = LLVMConstInt(type, lo, false);
+   md_args[1] = LLVMConstInt(type, hi, false);
+   range_md = LLVMMDNodeInContext(context, md_args, 2);
+   LLVMSetMetadata(value, ctx->range_md_kind, range_md);
 }
 
-LLVMValueRef
-ac_get_thread_id(struct ac_llvm_context *ctx)
+LLVMValueRef ac_get_thread_id(struct ac_llvm_context *ctx)
 {
-       LLVMValueRef tid;
+   LLVMValueRef tid;
 
-       LLVMValueRef tid_args[2];
-       tid_args[0] = LLVMConstInt(ctx->i32, 0xffffffff, false);
-       tid_args[1] = ctx->i32_0;
-       tid_args[1] = ac_build_intrinsic(ctx,
-                                        "llvm.amdgcn.mbcnt.lo", ctx->i32,
-                                        tid_args, 2, AC_FUNC_ATTR_READNONE);
+   LLVMValueRef tid_args[2];
+   tid_args[0] = LLVMConstInt(ctx->i32, 0xffffffff, false);
+   tid_args[1] = ctx->i32_0;
+   tid_args[1] =
+      ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.lo", ctx->i32, tid_args, 2, AC_FUNC_ATTR_READNONE);
 
-       if (ctx->wave_size == 32) {
-               tid = tid_args[1];
-       } else {
-               tid = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi",
-                                        ctx->i32, tid_args,
-                                        2, AC_FUNC_ATTR_READNONE);
-       }
-       set_range_metadata(ctx, tid, 0, ctx->wave_size);
-       return tid;
+   if (ctx->wave_size == 32) {
+      tid = tid_args[1];
+   } else {
+      tid = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi", ctx->i32, tid_args, 2,
+                               AC_FUNC_ATTR_READNONE);
+   }
+   set_range_metadata(ctx, tid, 0, ctx->wave_size);
+   return tid;
 }
 
 /*
@@ -2045,1506 +1779,1395 @@ ac_get_thread_id(struct ac_llvm_context *ctx)
  * Adding 1 yields the TID of the pixel to the right of the left pixel, and
  * adding 2 yields the TID of the pixel below the top pixel.
  */
-LLVMValueRef
-ac_build_ddxy(struct ac_llvm_context *ctx,
-             uint32_t mask,
-             int idx,
-             LLVMValueRef val)
-{
-       unsigned tl_lanes[4], trbl_lanes[4];
-       char name[32], type[8];
-       LLVMValueRef tl, trbl;
-       LLVMTypeRef result_type;
-       LLVMValueRef result;
-
-       result_type = ac_to_float_type(ctx, LLVMTypeOf(val));
-
-       if (result_type == ctx->f16)
-               val = LLVMBuildZExt(ctx->builder, val, ctx->i32, "");
-       else if (result_type == ctx->v2f16)
-               val = LLVMBuildBitCast(ctx->builder, val, ctx->i32, "");
-
-       for (unsigned i = 0; i < 4; ++i) {
-               tl_lanes[i] = i & mask;
-               trbl_lanes[i] = (i & mask) + idx;
-       }
-
-       tl = ac_build_quad_swizzle(ctx, val,
-                                  tl_lanes[0], tl_lanes[1],
-                                  tl_lanes[2], tl_lanes[3]);
-       trbl = ac_build_quad_swizzle(ctx, val,
-                                    trbl_lanes[0], trbl_lanes[1],
-                                    trbl_lanes[2], trbl_lanes[3]);
-
-       if (result_type == ctx->f16) {
-               tl = LLVMBuildTrunc(ctx->builder, tl, ctx->i16, "");
-               trbl = LLVMBuildTrunc(ctx->builder, trbl, ctx->i16, "");
-       }
-
-       tl = LLVMBuildBitCast(ctx->builder, tl, result_type, "");
-       trbl = LLVMBuildBitCast(ctx->builder, trbl, result_type, "");
-       result = LLVMBuildFSub(ctx->builder, trbl, tl, "");
-
-       ac_build_type_name_for_intr(result_type, type, sizeof(type));
-       snprintf(name, sizeof(name), "llvm.amdgcn.wqm.%s", type);
-
-       return ac_build_intrinsic(ctx, name, result_type, &result, 1, 0);
-}
-
-void
-ac_build_sendmsg(struct ac_llvm_context *ctx,
-                uint32_t msg,
-                LLVMValueRef wave_id)
-{
-       LLVMValueRef args[2];
-       args[0] = LLVMConstInt(ctx->i32, msg, false);
-       args[1] = wave_id;
-       ac_build_intrinsic(ctx, "llvm.amdgcn.s.sendmsg", ctx->voidt, args, 2, 0);
-}
-
-LLVMValueRef
-ac_build_imsb(struct ac_llvm_context *ctx,
-             LLVMValueRef arg,
-             LLVMTypeRef dst_type)
-{
-       LLVMValueRef msb = ac_build_intrinsic(ctx, "llvm.amdgcn.sffbh.i32",
-                                             dst_type, &arg, 1,
-                                             AC_FUNC_ATTR_READNONE);
-
-       /* The HW returns the last bit index from MSB, but NIR/TGSI wants
-        * the index from LSB. Invert it by doing "31 - msb". */
-       msb = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, 31, false),
-                          msb, "");
-
-       LLVMValueRef all_ones = LLVMConstInt(ctx->i32, -1, true);
-       LLVMValueRef cond = LLVMBuildOr(ctx->builder,
-                                       LLVMBuildICmp(ctx->builder, LLVMIntEQ,
-                                                     arg, ctx->i32_0, ""),
-                                       LLVMBuildICmp(ctx->builder, LLVMIntEQ,
-                                                     arg, all_ones, ""), "");
-
-       return LLVMBuildSelect(ctx->builder, cond, all_ones, msb, "");
-}
-
-LLVMValueRef
-ac_build_umsb(struct ac_llvm_context *ctx,
-             LLVMValueRef arg,
-             LLVMTypeRef dst_type)
-{
-       const char *intrin_name;
-       LLVMTypeRef type;
-       LLVMValueRef highest_bit;
-       LLVMValueRef zero;
-       unsigned bitsize;
-
-       bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(arg));
-       switch (bitsize) {
-       case 64:
-               intrin_name = "llvm.ctlz.i64";
-               type = ctx->i64;
-               highest_bit = LLVMConstInt(ctx->i64, 63, false);
-               zero = ctx->i64_0;
-               break;
-       case 32:
-               intrin_name = "llvm.ctlz.i32";
-               type = ctx->i32;
-               highest_bit = LLVMConstInt(ctx->i32, 31, false);
-               zero = ctx->i32_0;
-               break;
-       case 16:
-               intrin_name = "llvm.ctlz.i16";
-               type = ctx->i16;
-               highest_bit = LLVMConstInt(ctx->i16, 15, false);
-               zero = ctx->i16_0;
-               break;
-       case 8:
-               intrin_name = "llvm.ctlz.i8";
-               type = ctx->i8;
-               highest_bit = LLVMConstInt(ctx->i8, 7, false);
-               zero = ctx->i8_0;
-               break;
-       default:
-               unreachable(!"invalid bitsize");
-               break;
-       }
+LLVMValueRef ac_build_ddxy(struct ac_llvm_context *ctx, uint32_t mask, int idx, LLVMValueRef val)
+{
+   unsigned tl_lanes[4], trbl_lanes[4];
+   char name[32], type[8];
+   LLVMValueRef tl, trbl;
+   LLVMTypeRef result_type;
+   LLVMValueRef result;
+
+   result_type = ac_to_float_type(ctx, LLVMTypeOf(val));
+
+   if (result_type == ctx->f16)
+      val = LLVMBuildZExt(ctx->builder, val, ctx->i32, "");
+   else if (result_type == ctx->v2f16)
+      val = LLVMBuildBitCast(ctx->builder, val, ctx->i32, "");
+
+   for (unsigned i = 0; i < 4; ++i) {
+      tl_lanes[i] = i & mask;
+      trbl_lanes[i] = (i & mask) + idx;
+   }
+
+   tl = ac_build_quad_swizzle(ctx, val, tl_lanes[0], tl_lanes[1], tl_lanes[2], tl_lanes[3]);
+   trbl =
+      ac_build_quad_swizzle(ctx, val, trbl_lanes[0], trbl_lanes[1], trbl_lanes[2], trbl_lanes[3]);
+
+   if (result_type == ctx->f16) {
+      tl = LLVMBuildTrunc(ctx->builder, tl, ctx->i16, "");
+      trbl = LLVMBuildTrunc(ctx->builder, trbl, ctx->i16, "");
+   }
+
+   tl = LLVMBuildBitCast(ctx->builder, tl, result_type, "");
+   trbl = LLVMBuildBitCast(ctx->builder, trbl, result_type, "");
+   result = LLVMBuildFSub(ctx->builder, trbl, tl, "");
+
+   ac_build_type_name_for_intr(result_type, type, sizeof(type));
+   snprintf(name, sizeof(name), "llvm.amdgcn.wqm.%s", type);
+
+   return ac_build_intrinsic(ctx, name, result_type, &result, 1, 0);
+}
 
-       LLVMValueRef params[2] = {
-               arg,
-               ctx->i1true,
-       };
+void ac_build_sendmsg(struct ac_llvm_context *ctx, uint32_t msg, LLVMValueRef wave_id)
+{
+   LLVMValueRef args[2];
+   args[0] = LLVMConstInt(ctx->i32, msg, false);
+   args[1] = wave_id;
+   ac_build_intrinsic(ctx, "llvm.amdgcn.s.sendmsg", ctx->voidt, args, 2, 0);
+}
+
+LLVMValueRef ac_build_imsb(struct ac_llvm_context *ctx, LLVMValueRef arg, LLVMTypeRef dst_type)
+{
+   LLVMValueRef msb =
+      ac_build_intrinsic(ctx, "llvm.amdgcn.sffbh.i32", dst_type, &arg, 1, AC_FUNC_ATTR_READNONE);
+
+   /* The HW returns the last bit index from MSB, but NIR/TGSI wants
+    * the index from LSB. Invert it by doing "31 - msb". */
+   msb = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, 31, false), msb, "");
+
+   LLVMValueRef all_ones = LLVMConstInt(ctx->i32, -1, true);
+   LLVMValueRef cond =
+      LLVMBuildOr(ctx->builder, LLVMBuildICmp(ctx->builder, LLVMIntEQ, arg, ctx->i32_0, ""),
+                  LLVMBuildICmp(ctx->builder, LLVMIntEQ, arg, all_ones, ""), "");
+
+   return LLVMBuildSelect(ctx->builder, cond, all_ones, msb, "");
+}
+
+LLVMValueRef ac_build_umsb(struct ac_llvm_context *ctx, LLVMValueRef arg, LLVMTypeRef dst_type)
+{
+   const char *intrin_name;
+   LLVMTypeRef type;
+   LLVMValueRef highest_bit;
+   LLVMValueRef zero;
+   unsigned bitsize;
+
+   bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(arg));
+   switch (bitsize) {
+   case 64:
+      intrin_name = "llvm.ctlz.i64";
+      type = ctx->i64;
+      highest_bit = LLVMConstInt(ctx->i64, 63, false);
+      zero = ctx->i64_0;
+      break;
+   case 32:
+      intrin_name = "llvm.ctlz.i32";
+      type = ctx->i32;
+      highest_bit = LLVMConstInt(ctx->i32, 31, false);
+      zero = ctx->i32_0;
+      break;
+   case 16:
+      intrin_name = "llvm.ctlz.i16";
+      type = ctx->i16;
+      highest_bit = LLVMConstInt(ctx->i16, 15, false);
+      zero = ctx->i16_0;
+      break;
+   case 8:
+      intrin_name = "llvm.ctlz.i8";
+      type = ctx->i8;
+      highest_bit = LLVMConstInt(ctx->i8, 7, false);
+      zero = ctx->i8_0;
+      break;
+   default:
+      unreachable(!"invalid bitsize");
+      break;
+   }
 
-       LLVMValueRef msb = ac_build_intrinsic(ctx, intrin_name, type,
-                                             params, 2,
-                                             AC_FUNC_ATTR_READNONE);
+   LLVMValueRef params[2] = {
+      arg,
+      ctx->i1true,
+   };
 
-       /* The HW returns the last bit index from MSB, but TGSI/NIR wants
-        * the index from LSB. Invert it by doing "31 - msb". */
-       msb = LLVMBuildSub(ctx->builder, highest_bit, msb, "");
+   LLVMValueRef msb = ac_build_intrinsic(ctx, intrin_name, type, params, 2, AC_FUNC_ATTR_READNONE);
 
-       if (bitsize == 64) {
-               msb = LLVMBuildTrunc(ctx->builder, msb, ctx->i32, "");
-       } else if (bitsize < 32) {
-               msb = LLVMBuildSExt(ctx->builder, msb, ctx->i32, "");
-       }
+   /* The HW returns the last bit index from MSB, but TGSI/NIR wants
+    * the index from LSB. Invert it by doing "31 - msb". */
+   msb = LLVMBuildSub(ctx->builder, highest_bit, msb, "");
 
-       /* check for zero */
-       return LLVMBuildSelect(ctx->builder,
-                              LLVMBuildICmp(ctx->builder, LLVMIntEQ, arg, zero, ""),
-                              LLVMConstInt(ctx->i32, -1, true), msb, "");
+   if (bitsize == 64) {
+      msb = LLVMBuildTrunc(ctx->builder, msb, ctx->i32, "");
+   } else if (bitsize < 32) {
+      msb = LLVMBuildSExt(ctx->builder, msb, ctx->i32, "");
+   }
+
+   /* check for zero */
+   return LLVMBuildSelect(ctx->builder, LLVMBuildICmp(ctx->builder, LLVMIntEQ, arg, zero, ""),
+                          LLVMConstInt(ctx->i32, -1, true), msb, "");
 }
 
-LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
-                          LLVMValueRef b)
+LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b)
 {
-       char name[64], type[64];
+   char name[64], type[64];
 
-       ac_build_type_name_for_intr(LLVMTypeOf(a), type, sizeof(type));
-       snprintf(name, sizeof(name), "llvm.minnum.%s", type);
-       LLVMValueRef args[2] = {a, b};
-       return ac_build_intrinsic(ctx, name, LLVMTypeOf(a), args, 2,
-                                 AC_FUNC_ATTR_READNONE);
+   ac_build_type_name_for_intr(LLVMTypeOf(a), type, sizeof(type));
+   snprintf(name, sizeof(name), "llvm.minnum.%s", type);
+   LLVMValueRef args[2] = {a, b};
+   return ac_build_intrinsic(ctx, name, LLVMTypeOf(a), args, 2, AC_FUNC_ATTR_READNONE);
 }
 
-LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
-                          LLVMValueRef b)
+LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b)
 {
-       char name[64], type[64];
+   char name[64], type[64];
 
-       ac_build_type_name_for_intr(LLVMTypeOf(a), type, sizeof(type));
-       snprintf(name, sizeof(name), "llvm.maxnum.%s", type);
-       LLVMValueRef args[2] = {a, b};
-       return ac_build_intrinsic(ctx, name, LLVMTypeOf(a), args, 2,
-                                 AC_FUNC_ATTR_READNONE);
+   ac_build_type_name_for_intr(LLVMTypeOf(a), type, sizeof(type));
+   snprintf(name, sizeof(name), "llvm.maxnum.%s", type);
+   LLVMValueRef args[2] = {a, b};
+   return ac_build_intrinsic(ctx, name, LLVMTypeOf(a), args, 2, AC_FUNC_ATTR_READNONE);
 }
 
-LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
-                          LLVMValueRef b)
+LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b)
 {
-       LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSLE, a, b, "");
-       return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
+   LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSLE, a, b, "");
+   return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
 }
 
-LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
-                          LLVMValueRef b)
+LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b)
 {
-       LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, a, b, "");
-       return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
+   LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, a, b, "");
+   return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
 }
 
-LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a,
-                          LLVMValueRef b)
+LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b)
 {
-       LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntULE, a, b, "");
-       return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
+   LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntULE, a, b, "");
+   return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
 }
 
-LLVMValueRef ac_build_umax(struct ac_llvm_context *ctx, LLVMValueRef a,
-                          LLVMValueRef b)
+LLVMValueRef ac_build_umax(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b)
 {
-       LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntUGE, a, b, "");
-       return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
+   LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntUGE, a, b, "");
+   return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
 }
 
 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value)
 {
-       LLVMTypeRef t = LLVMTypeOf(value);
-       return ac_build_fmin(ctx, ac_build_fmax(ctx, value, LLVMConstReal(t, 0.0)),
-                            LLVMConstReal(t, 1.0));
+   LLVMTypeRef t = LLVMTypeOf(value);
+   return ac_build_fmin(ctx, ac_build_fmax(ctx, value, LLVMConstReal(t, 0.0)),
+                        LLVMConstReal(t, 1.0));
 }
 
 void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a)
 {
-       LLVMValueRef args[9];
+   LLVMValueRef args[9];
 
-       args[0] = LLVMConstInt(ctx->i32, a->target, 0);
-       args[1] = LLVMConstInt(ctx->i32, a->enabled_channels, 0);
+   args[0] = LLVMConstInt(ctx->i32, a->target, 0);
+   args[1] = LLVMConstInt(ctx->i32, a->enabled_channels, 0);
 
-       if (a->compr) {
-               args[2] = LLVMBuildBitCast(ctx->builder, a->out[0],
-                               ctx->v2i16, "");
-               args[3] = LLVMBuildBitCast(ctx->builder, a->out[1],
-                               ctx->v2i16, "");
-               args[4] = LLVMConstInt(ctx->i1, a->done, 0);
-               args[5] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
+   if (a->compr) {
+      args[2] = LLVMBuildBitCast(ctx->builder, a->out[0], ctx->v2i16, "");
+      args[3] = LLVMBuildBitCast(ctx->builder, a->out[1], ctx->v2i16, "");
+      args[4] = LLVMConstInt(ctx->i1, a->done, 0);
+      args[5] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
 
-               ac_build_intrinsic(ctx, "llvm.amdgcn.exp.compr.v2i16",
-                                  ctx->voidt, args, 6, 0);
-       } else {
-               args[2] = a->out[0];
-               args[3] = a->out[1];
-               args[4] = a->out[2];
-               args[5] = a->out[3];
-               args[6] = LLVMConstInt(ctx->i1, a->done, 0);
-               args[7] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
+      ac_build_intrinsic(ctx, "llvm.amdgcn.exp.compr.v2i16", ctx->voidt, args, 6, 0);
+   } else {
+      args[2] = a->out[0];
+      args[3] = a->out[1];
+      args[4] = a->out[2];
+      args[5] = a->out[3];
+      args[6] = LLVMConstInt(ctx->i1, a->done, 0);
+      args[7] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
 
-               ac_build_intrinsic(ctx, "llvm.amdgcn.exp.f32",
-                                  ctx->voidt, args, 8, 0);
-       }
+      ac_build_intrinsic(ctx, "llvm.amdgcn.exp.f32", ctx->voidt, args, 8, 0);
+   }
 }
 
 void ac_build_export_null(struct ac_llvm_context *ctx)
 {
-       struct ac_export_args args;
+   struct ac_export_args args;
 
-       args.enabled_channels = 0x0; /* enabled channels */
-       args.valid_mask = 1; /* whether the EXEC mask is valid */
-       args.done = 1; /* DONE bit */
-       args.target = V_008DFC_SQ_EXP_NULL;
-       args.compr = 0; /* COMPR flag (0 = 32-bit export) */
-       args.out[0] = LLVMGetUndef(ctx->f32); /* R */
-       args.out[1] = LLVMGetUndef(ctx->f32); /* G */
-       args.out[2] = LLVMGetUndef(ctx->f32); /* B */
-       args.out[3] = LLVMGetUndef(ctx->f32); /* A */
+   args.enabled_channels = 0x0; /* enabled channels */
+   args.valid_mask = 1;         /* whether the EXEC mask is valid */
+   args.done = 1;               /* DONE bit */
+   args.target = V_008DFC_SQ_EXP_NULL;
+   args.compr = 0;                       /* COMPR flag (0 = 32-bit export) */
+   args.out[0] = LLVMGetUndef(ctx->f32); /* R */
+   args.out[1] = LLVMGetUndef(ctx->f32); /* G */
+   args.out[2] = LLVMGetUndef(ctx->f32); /* B */
+   args.out[3] = LLVMGetUndef(ctx->f32); /* A */
 
-       ac_build_export(ctx, &args);
+   ac_build_export(ctx, &args);
 }
 
 static unsigned ac_num_coords(enum ac_image_dim dim)
 {
-       switch (dim) {
-       case ac_image_1d:
-               return 1;
-       case ac_image_2d:
-       case ac_image_1darray:
-                return 2;
-       case ac_image_3d:
-       case ac_image_cube:
-       case ac_image_2darray:
-       case ac_image_2dmsaa:
-               return 3;
-       case ac_image_2darraymsaa:
-               return 4;
-       default:
-               unreachable("ac_num_coords: bad dim");
-       }
+   switch (dim) {
+   case ac_image_1d:
+      return 1;
+   case ac_image_2d:
+   case ac_image_1darray:
+      return 2;
+   case ac_image_3d:
+   case ac_image_cube:
+   case ac_image_2darray:
+   case ac_image_2dmsaa:
+      return 3;
+   case ac_image_2darraymsaa:
+      return 4;
+   default:
+      unreachable("ac_num_coords: bad dim");
+   }
 }
 
 static unsigned ac_num_derivs(enum ac_image_dim dim)
 {
-       switch (dim) {
-       case ac_image_1d:
-       case ac_image_1darray:
-               return 2;
-       case ac_image_2d:
-       case ac_image_2darray:
-       case ac_image_cube:
-               return 4;
-       case ac_image_3d:
-               return 6;
-       case ac_image_2dmsaa:
-       case ac_image_2darraymsaa:
-       default:
-               unreachable("derivatives not supported");
-       }
+   switch (dim) {
+   case ac_image_1d:
+   case ac_image_1darray:
+      return 2;
+   case ac_image_2d:
+   case ac_image_2darray:
+   case ac_image_cube:
+      return 4;
+   case ac_image_3d:
+      return 6;
+   case ac_image_2dmsaa:
+   case ac_image_2darraymsaa:
+   default:
+      unreachable("derivatives not supported");
+   }
 }
 
 static const char *get_atomic_name(enum ac_atomic_op op)
 {
-       switch (op) {
-       case ac_atomic_swap: return "swap";
-       case ac_atomic_add: return "add";
-       case ac_atomic_sub: return "sub";
-       case ac_atomic_smin: return "smin";
-       case ac_atomic_umin: return "umin";
-       case ac_atomic_smax: return "smax";
-       case ac_atomic_umax: return "umax";
-       case ac_atomic_and: return "and";
-       case ac_atomic_or: return "or";
-       case ac_atomic_xor: return "xor";
-       case ac_atomic_inc_wrap: return "inc";
-       case ac_atomic_dec_wrap: return "dec";
-       }
-       unreachable("bad atomic op");
-}
-
-LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
-                                  struct ac_image_args *a)
-{
-       const char *overload[3] = { "", "", "" };
-       unsigned num_overloads = 0;
-       LLVMValueRef args[18];
-       unsigned num_args = 0;
-       enum ac_image_dim dim = a->dim;
-
-       assert(!a->lod || a->lod == ctx->i32_0 || a->lod == ctx->f32_0 ||
-              !a->level_zero);
-       assert((a->opcode != ac_image_get_resinfo && a->opcode != ac_image_load_mip &&
-               a->opcode != ac_image_store_mip) ||
-              a->lod);
-       assert(a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
-              (!a->compare && !a->offset));
-       assert((a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
-               a->opcode == ac_image_get_lod) ||
-              !a->bias);
-       assert((a->bias ? 1 : 0) +
-              (a->lod ? 1 : 0) +
-              (a->level_zero ? 1 : 0) +
-              (a->derivs[0] ? 1 : 0) <= 1);
-       assert((a->min_lod ? 1 : 0) +
-              (a->lod ? 1 : 0) +
-              (a->level_zero ? 1 : 0) <= 1);
-       assert(!a->d16 || (ctx->chip_class >= GFX8 &&
-                          a->opcode != ac_image_atomic &&
-                          a->opcode != ac_image_atomic_cmpswap &&
-                          a->opcode != ac_image_get_lod &&
-                          a->opcode != ac_image_get_resinfo));
-
-       if (a->opcode == ac_image_get_lod) {
-               switch (dim) {
-               case ac_image_1darray:
-                       dim = ac_image_1d;
-                       break;
-               case ac_image_2darray:
-               case ac_image_cube:
-                       dim = ac_image_2d;
-                       break;
-               default:
-                       break;
-               }
-       }
-
-       bool sample = a->opcode == ac_image_sample ||
-                     a->opcode == ac_image_gather4 ||
-                     a->opcode == ac_image_get_lod;
-       bool atomic = a->opcode == ac_image_atomic ||
-                     a->opcode == ac_image_atomic_cmpswap;
-       bool load = a->opcode == ac_image_sample ||
-                   a->opcode == ac_image_gather4 ||
-                   a->opcode == ac_image_load ||
-                   a->opcode == ac_image_load_mip;
-       LLVMTypeRef coord_type = sample ? ctx->f32 : ctx->i32;
-
-       if (atomic || a->opcode == ac_image_store || a->opcode == ac_image_store_mip) {
-               args[num_args++] = a->data[0];
-               if (a->opcode == ac_image_atomic_cmpswap)
-                       args[num_args++] = a->data[1];
-       }
-
-       if (!atomic)
-               args[num_args++] = LLVMConstInt(ctx->i32, a->dmask, false);
-
-       if (a->offset)
-               args[num_args++] = ac_to_integer(ctx, a->offset);
-       if (a->bias) {
-               args[num_args++] = ac_to_float(ctx, a->bias);
-               overload[num_overloads++] = ".f32";
-       }
-       if (a->compare)
-               args[num_args++] = ac_to_float(ctx, a->compare);
-       if (a->derivs[0]) {
-               unsigned count = ac_num_derivs(dim);
-               for (unsigned i = 0; i < count; ++i)
-                       args[num_args++] = ac_to_float(ctx, a->derivs[i]);
-               overload[num_overloads++] = ".f32";
-       }
-       unsigned num_coords =
-               a->opcode != ac_image_get_resinfo ? ac_num_coords(dim) : 0;
-       for (unsigned i = 0; i < num_coords; ++i)
-               args[num_args++] = LLVMBuildBitCast(ctx->builder, a->coords[i], coord_type, "");
-       if (a->lod)
-               args[num_args++] = LLVMBuildBitCast(ctx->builder, a->lod, coord_type, "");
-       if (a->min_lod)
-               args[num_args++] = LLVMBuildBitCast(ctx->builder, a->min_lod, coord_type, "");
-
-       overload[num_overloads++] = sample ? ".f32" : ".i32";
-
-       args[num_args++] = a->resource;
-       if (sample) {
-               args[num_args++] = a->sampler;
-               args[num_args++] = LLVMConstInt(ctx->i1, a->unorm, false);
-       }
-
-       args[num_args++] = ctx->i32_0; /* texfailctrl */
-       args[num_args++] = LLVMConstInt(ctx->i32,
-                                       load ? get_load_cache_policy(ctx, a->cache_policy) :
-                                              a->cache_policy, false);
-
-       const char *name;
-       const char *atomic_subop = "";
-       switch (a->opcode) {
-       case ac_image_sample: name = "sample"; break;
-       case ac_image_gather4: name = "gather4"; break;
-       case ac_image_load: name = "load"; break;
-       case ac_image_load_mip: name = "load.mip"; break;
-       case ac_image_store: name = "store"; break;
-       case ac_image_store_mip: name = "store.mip"; break;
-       case ac_image_atomic:
-               name = "atomic.";
-               atomic_subop = get_atomic_name(a->atomic);
-               break;
-       case ac_image_atomic_cmpswap:
-               name = "atomic.";
-               atomic_subop = "cmpswap";
-               break;
-       case ac_image_get_lod: name = "getlod"; break;
-       case ac_image_get_resinfo: name = "getresinfo"; break;
-       default: unreachable("invalid image opcode");
-       }
-
-       const char *dimname;
-       switch (dim) {
-       case ac_image_1d: dimname = "1d"; break;
-       case ac_image_2d: dimname = "2d"; break;
-       case ac_image_3d: dimname = "3d"; break;
-       case ac_image_cube: dimname = "cube"; break;
-       case ac_image_1darray: dimname = "1darray"; break;
-       case ac_image_2darray: dimname = "2darray"; break;
-       case ac_image_2dmsaa: dimname = "2dmsaa"; break;
-       case ac_image_2darraymsaa: dimname = "2darraymsaa"; break;
-       default: unreachable("invalid dim");
-       }
-
-       bool lod_suffix =
-               a->lod && (a->opcode == ac_image_sample || a->opcode == ac_image_gather4);
-       char intr_name[96];
-       snprintf(intr_name, sizeof(intr_name),
-                "llvm.amdgcn.image.%s%s" /* base name */
-                "%s%s%s%s" /* sample/gather modifiers */
-                ".%s.%s%s%s%s", /* dimension and type overloads */
-                name, atomic_subop,
-                a->compare ? ".c" : "",
-                a->bias ? ".b" :
-                lod_suffix ? ".l" :
-                a->derivs[0] ? ".d" :
-                a->level_zero ? ".lz" : "",
-                a->min_lod ? ".cl" : "",
-                a->offset ? ".o" : "",
-                dimname,
-                atomic ? "i32" : (a->d16 ? "v4f16" : "v4f32"),
-                overload[0], overload[1], overload[2]);
-
-       LLVMTypeRef retty;
-       if (atomic)
-               retty = ctx->i32;
-       else if (a->opcode == ac_image_store || a->opcode == ac_image_store_mip)
-               retty = ctx->voidt;
-       else
-               retty = a->d16 ? ctx->v4f16 : ctx->v4f32;
-
-       LLVMValueRef result =
-               ac_build_intrinsic(ctx, intr_name, retty, args, num_args,
-                                  a->attributes);
-       if (!sample && !atomic && retty != ctx->voidt)
-               result = ac_to_integer(ctx, result);
-
-       return result;
-}
-
-LLVMValueRef ac_build_image_get_sample_count(struct ac_llvm_context *ctx,
-                                            LLVMValueRef rsrc)
-{
-       LLVMValueRef samples;
-
-       /* Read the samples from the descriptor directly.
-        * Hardware doesn't have any instruction for this.
-        */
-       samples = LLVMBuildExtractElement(ctx->builder, rsrc,
-                                         LLVMConstInt(ctx->i32, 3, 0), "");
-       samples = LLVMBuildLShr(ctx->builder, samples,
-                               LLVMConstInt(ctx->i32, 16, 0), "");
-       samples = LLVMBuildAnd(ctx->builder, samples,
-                              LLVMConstInt(ctx->i32, 0xf, 0), "");
-       samples = LLVMBuildShl(ctx->builder, ctx->i32_1,
-                              samples, "");
-       return samples;
-}
-
-LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
-                                   LLVMValueRef args[2])
-{
-       return ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pkrtz", ctx->v2f16,
-                                 args, 2, AC_FUNC_ATTR_READNONE);
-}
-
-LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
-                                    LLVMValueRef args[2])
-{
-       LLVMValueRef res =
-               ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.i16",
-                                  ctx->v2i16, args, 2,
-                                  AC_FUNC_ATTR_READNONE);
-       return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
-}
-
-LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
-                                    LLVMValueRef args[2])
-{
-       LLVMValueRef res =
-               ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.u16",
-                                  ctx->v2i16, args, 2,
-                                  AC_FUNC_ATTR_READNONE);
-       return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
+   switch (op) {
+   case ac_atomic_swap:
+      return "swap";
+   case ac_atomic_add:
+      return "add";
+   case ac_atomic_sub:
+      return "sub";
+   case ac_atomic_smin:
+      return "smin";
+   case ac_atomic_umin:
+      return "umin";
+   case ac_atomic_smax:
+      return "smax";
+   case ac_atomic_umax:
+      return "umax";
+   case ac_atomic_and:
+      return "and";
+   case ac_atomic_or:
+      return "or";
+   case ac_atomic_xor:
+      return "xor";
+   case ac_atomic_inc_wrap:
+      return "inc";
+   case ac_atomic_dec_wrap:
+      return "dec";
+   }
+   unreachable("bad atomic op");
+}
+
+LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx, struct ac_image_args *a)
+{
+   const char *overload[3] = {"", "", ""};
+   unsigned num_overloads = 0;
+   LLVMValueRef args[18];
+   unsigned num_args = 0;
+   enum ac_image_dim dim = a->dim;
+
+   assert(!a->lod || a->lod == ctx->i32_0 || a->lod == ctx->f32_0 || !a->level_zero);
+   assert((a->opcode != ac_image_get_resinfo && a->opcode != ac_image_load_mip &&
+           a->opcode != ac_image_store_mip) ||
+          a->lod);
+   assert(a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
+          (!a->compare && !a->offset));
+   assert((a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
+           a->opcode == ac_image_get_lod) ||
+          !a->bias);
+   assert((a->bias ? 1 : 0) + (a->lod ? 1 : 0) + (a->level_zero ? 1 : 0) + (a->derivs[0] ? 1 : 0) <=
+          1);
+   assert((a->min_lod ? 1 : 0) + (a->lod ? 1 : 0) + (a->level_zero ? 1 : 0) <= 1);
+   assert(!a->d16 || (ctx->chip_class >= GFX8 && a->opcode != ac_image_atomic &&
+                      a->opcode != ac_image_atomic_cmpswap && a->opcode != ac_image_get_lod &&
+                      a->opcode != ac_image_get_resinfo));
+
+   if (a->opcode == ac_image_get_lod) {
+      switch (dim) {
+      case ac_image_1darray:
+         dim = ac_image_1d;
+         break;
+      case ac_image_2darray:
+      case ac_image_cube:
+         dim = ac_image_2d;
+         break;
+      default:
+         break;
+      }
+   }
+
+   bool sample = a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
+                 a->opcode == ac_image_get_lod;
+   bool atomic = a->opcode == ac_image_atomic || a->opcode == ac_image_atomic_cmpswap;
+   bool load = a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
+               a->opcode == ac_image_load || a->opcode == ac_image_load_mip;
+   LLVMTypeRef coord_type = sample ? ctx->f32 : ctx->i32;
+
+   if (atomic || a->opcode == ac_image_store || a->opcode == ac_image_store_mip) {
+      args[num_args++] = a->data[0];
+      if (a->opcode == ac_image_atomic_cmpswap)
+         args[num_args++] = a->data[1];
+   }
+
+   if (!atomic)
+      args[num_args++] = LLVMConstInt(ctx->i32, a->dmask, false);
+
+   if (a->offset)
+      args[num_args++] = ac_to_integer(ctx, a->offset);
+   if (a->bias) {
+      args[num_args++] = ac_to_float(ctx, a->bias);
+      overload[num_overloads++] = ".f32";
+   }
+   if (a->compare)
+      args[num_args++] = ac_to_float(ctx, a->compare);
+   if (a->derivs[0]) {
+      unsigned count = ac_num_derivs(dim);
+      for (unsigned i = 0; i < count; ++i)
+         args[num_args++] = ac_to_float(ctx, a->derivs[i]);
+      overload[num_overloads++] = ".f32";
+   }
+   unsigned num_coords = a->opcode != ac_image_get_resinfo ? ac_num_coords(dim) : 0;
+   for (unsigned i = 0; i < num_coords; ++i)
+      args[num_args++] = LLVMBuildBitCast(ctx->builder, a->coords[i], coord_type, "");
+   if (a->lod)
+      args[num_args++] = LLVMBuildBitCast(ctx->builder, a->lod, coord_type, "");
+   if (a->min_lod)
+      args[num_args++] = LLVMBuildBitCast(ctx->builder, a->min_lod, coord_type, "");
+
+   overload[num_overloads++] = sample ? ".f32" : ".i32";
+
+   args[num_args++] = a->resource;
+   if (sample) {
+      args[num_args++] = a->sampler;
+      args[num_args++] = LLVMConstInt(ctx->i1, a->unorm, false);
+   }
+
+   args[num_args++] = ctx->i32_0; /* texfailctrl */
+   args[num_args++] = LLVMConstInt(
+      ctx->i32, load ? get_load_cache_policy(ctx, a->cache_policy) : a->cache_policy, false);
+
+   const char *name;
+   const char *atomic_subop = "";
+   switch (a->opcode) {
+   case ac_image_sample:
+      name = "sample";
+      break;
+   case ac_image_gather4:
+      name = "gather4";
+      break;
+   case ac_image_load:
+      name = "load";
+      break;
+   case ac_image_load_mip:
+      name = "load.mip";
+      break;
+   case ac_image_store:
+      name = "store";
+      break;
+   case ac_image_store_mip:
+      name = "store.mip";
+      break;
+   case ac_image_atomic:
+      name = "atomic.";
+      atomic_subop = get_atomic_name(a->atomic);
+      break;
+   case ac_image_atomic_cmpswap:
+      name = "atomic.";
+      atomic_subop = "cmpswap";
+      break;
+   case ac_image_get_lod:
+      name = "getlod";
+      break;
+   case ac_image_get_resinfo:
+      name = "getresinfo";
+      break;
+   default:
+      unreachable("invalid image opcode");
+   }
+
+   const char *dimname;
+   switch (dim) {
+   case ac_image_1d:
+      dimname = "1d";
+      break;
+   case ac_image_2d:
+      dimname = "2d";
+      break;
+   case ac_image_3d:
+      dimname = "3d";
+      break;
+   case ac_image_cube:
+      dimname = "cube";
+      break;
+   case ac_image_1darray:
+      dimname = "1darray";
+      break;
+   case ac_image_2darray:
+      dimname = "2darray";
+      break;
+   case ac_image_2dmsaa:
+      dimname = "2dmsaa";
+      break;
+   case ac_image_2darraymsaa:
+      dimname = "2darraymsaa";
+      break;
+   default:
+      unreachable("invalid dim");
+   }
+
+   bool lod_suffix = a->lod && (a->opcode == ac_image_sample || a->opcode == ac_image_gather4);
+   char intr_name[96];
+   snprintf(intr_name, sizeof(intr_name),
+            "llvm.amdgcn.image.%s%s" /* base name */
+            "%s%s%s%s"               /* sample/gather modifiers */
+            ".%s.%s%s%s%s",          /* dimension and type overloads */
+            name, atomic_subop, a->compare ? ".c" : "",
+            a->bias ? ".b" : lod_suffix ? ".l" : a->derivs[0] ? ".d" : a->level_zero ? ".lz" : "",
+            a->min_lod ? ".cl" : "", a->offset ? ".o" : "", dimname,
+            atomic ? "i32" : (a->d16 ? "v4f16" : "v4f32"), overload[0], overload[1], overload[2]);
+
+   LLVMTypeRef retty;
+   if (atomic)
+      retty = ctx->i32;
+   else if (a->opcode == ac_image_store || a->opcode == ac_image_store_mip)
+      retty = ctx->voidt;
+   else
+      retty = a->d16 ? ctx->v4f16 : ctx->v4f32;
+
+   LLVMValueRef result = ac_build_intrinsic(ctx, intr_name, retty, args, num_args, a->attributes);
+   if (!sample && !atomic && retty != ctx->voidt)
+      result = ac_to_integer(ctx, result);
+
+   return result;
+}
+
+LLVMValueRef ac_build_image_get_sample_count(struct ac_llvm_context *ctx, LLVMValueRef rsrc)
+{
+   LLVMValueRef samples;
+
+   /* Read the samples from the descriptor directly.
+    * Hardware doesn't have any instruction for this.
+    */
+   samples = LLVMBuildExtractElement(ctx->builder, rsrc, LLVMConstInt(ctx->i32, 3, 0), "");
+   samples = LLVMBuildLShr(ctx->builder, samples, LLVMConstInt(ctx->i32, 16, 0), "");
+   samples = LLVMBuildAnd(ctx->builder, samples, LLVMConstInt(ctx->i32, 0xf, 0), "");
+   samples = LLVMBuildShl(ctx->builder, ctx->i32_1, samples, "");
+   return samples;
+}
+
+LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx, LLVMValueRef args[2])
+{
+   return ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pkrtz", ctx->v2f16, args, 2,
+                             AC_FUNC_ATTR_READNONE);
+}
+
+LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx, LLVMValueRef args[2])
+{
+   LLVMValueRef res = ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.i16", ctx->v2i16, args, 2,
+                                         AC_FUNC_ATTR_READNONE);
+   return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
+}
+
+LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx, LLVMValueRef args[2])
+{
+   LLVMValueRef res = ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.u16", ctx->v2i16, args, 2,
+                                         AC_FUNC_ATTR_READNONE);
+   return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
 }
 
 /* The 8-bit and 10-bit clamping is for HW workarounds. */
-LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
-                                LLVMValueRef args[2], unsigned bits, bool hi)
-{
-       assert(bits == 8 || bits == 10 || bits == 16);
-
-       LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
-               bits == 8 ? 127 : bits == 10 ? 511 : 32767, 0);
-       LLVMValueRef min_rgb = LLVMConstInt(ctx->i32,
-               bits == 8 ? -128 : bits == 10 ? -512 : -32768, 0);
-       LLVMValueRef max_alpha =
-               bits != 10 ? max_rgb : ctx->i32_1;
-       LLVMValueRef min_alpha =
-               bits != 10 ? min_rgb : LLVMConstInt(ctx->i32, -2, 0);
-
-       /* Clamp. */
-       if (bits != 16) {
-               for (int i = 0; i < 2; i++) {
-                       bool alpha = hi && i == 1;
-                       args[i] = ac_build_imin(ctx, args[i],
-                                               alpha ? max_alpha : max_rgb);
-                       args[i] = ac_build_imax(ctx, args[i],
-                                               alpha ? min_alpha : min_rgb);
-               }
-       }
-
-       LLVMValueRef res =
-               ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.i16",
-                                  ctx->v2i16, args, 2,
-                                  AC_FUNC_ATTR_READNONE);
-       return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
+LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx, LLVMValueRef args[2], unsigned bits,
+                                 bool hi)
+{
+   assert(bits == 8 || bits == 10 || bits == 16);
+
+   LLVMValueRef max_rgb = LLVMConstInt(ctx->i32, bits == 8 ? 127 : bits == 10 ? 511 : 32767, 0);
+   LLVMValueRef min_rgb = LLVMConstInt(ctx->i32, bits == 8 ? -128 : bits == 10 ? -512 : -32768, 0);
+   LLVMValueRef max_alpha = bits != 10 ? max_rgb : ctx->i32_1;
+   LLVMValueRef min_alpha = bits != 10 ? min_rgb : LLVMConstInt(ctx->i32, -2, 0);
+
+   /* Clamp. */
+   if (bits != 16) {
+      for (int i = 0; i < 2; i++) {
+         bool alpha = hi && i == 1;
+         args[i] = ac_build_imin(ctx, args[i], alpha ? max_alpha : max_rgb);
+         args[i] = ac_build_imax(ctx, args[i], alpha ? min_alpha : min_rgb);
+      }
+   }
+
+   LLVMValueRef res =
+      ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.i16", ctx->v2i16, args, 2, AC_FUNC_ATTR_READNONE);
+   return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
 }
 
 /* The 8-bit and 10-bit clamping is for HW workarounds. */
-LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
-                                LLVMValueRef args[2], unsigned bits, bool hi)
+LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx, LLVMValueRef args[2], unsigned bits,
+                                 bool hi)
 {
-       assert(bits == 8 || bits == 10 || bits == 16);
+   assert(bits == 8 || bits == 10 || bits == 16);
 
-       LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
-               bits == 8 ? 255 : bits == 10 ? 1023 : 65535, 0);
-       LLVMValueRef max_alpha =
-               bits != 10 ? max_rgb : LLVMConstInt(ctx->i32, 3, 0);
+   LLVMValueRef max_rgb = LLVMConstInt(ctx->i32, bits == 8 ? 255 : bits == 10 ? 1023 : 65535, 0);
+   LLVMValueRef max_alpha = bits != 10 ? max_rgb : LLVMConstInt(ctx->i32, 3, 0);
 
-       /* Clamp. */
-       if (bits != 16) {
-               for (int i = 0; i < 2; i++) {
-                       bool alpha = hi && i == 1;
-                       args[i] = ac_build_umin(ctx, args[i],
-                                               alpha ? max_alpha : max_rgb);
-               }
-       }
+   /* Clamp. */
+   if (bits != 16) {
+      for (int i = 0; i < 2; i++) {
+         bool alpha = hi && i == 1;
+         args[i] = ac_build_umin(ctx, args[i], alpha ? max_alpha : max_rgb);
+      }
+   }
 
-       LLVMValueRef res =
-               ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.u16",
-                                  ctx->v2i16, args, 2,
-                                  AC_FUNC_ATTR_READNONE);
-       return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
+   LLVMValueRef res =
+      ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.u16", ctx->v2i16, args, 2, AC_FUNC_ATTR_READNONE);
+   return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
 }
 
 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1)
 {
-       return ac_build_intrinsic(ctx, "llvm.amdgcn.wqm.vote", ctx->i1,
-                                 &i1, 1, AC_FUNC_ATTR_READNONE);
+   return ac_build_intrinsic(ctx, "llvm.amdgcn.wqm.vote", ctx->i1, &i1, 1, AC_FUNC_ATTR_READNONE);
 }
 
 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1)
 {
-       ac_build_intrinsic(ctx, "llvm.amdgcn.kill", ctx->voidt,
-                          &i1, 1, 0);
+   ac_build_intrinsic(ctx, "llvm.amdgcn.kill", ctx->voidt, &i1, 1, 0);
 }
 
-LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
-                         LLVMValueRef offset, LLVMValueRef width,
-                         bool is_signed)
+LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input, LLVMValueRef offset,
+                          LLVMValueRef width, bool is_signed)
 {
-       LLVMValueRef args[] = {
-               input,
-               offset,
-               width,
-       };
-
-       return ac_build_intrinsic(ctx, is_signed ? "llvm.amdgcn.sbfe.i32" :
-                                                  "llvm.amdgcn.ubfe.i32",
-                                 ctx->i32, args, 3, AC_FUNC_ATTR_READNONE);
+   LLVMValueRef args[] = {
+      input,
+      offset,
+      width,
+   };
 
+   return ac_build_intrinsic(ctx, is_signed ? "llvm.amdgcn.sbfe.i32" : "llvm.amdgcn.ubfe.i32",
+                             ctx->i32, args, 3, AC_FUNC_ATTR_READNONE);
 }
 
-LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
-                          LLVMValueRef s1, LLVMValueRef s2)
+LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0, LLVMValueRef s1,
+                           LLVMValueRef s2)
 {
-       return LLVMBuildAdd(ctx->builder,
-                           LLVMBuildMul(ctx->builder, s0, s1, ""), s2, "");
+   return LLVMBuildAdd(ctx->builder, LLVMBuildMul(ctx->builder, s0, s1, ""), s2, "");
 }
 
-LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0,
-                          LLVMValueRef s1, LLVMValueRef s2)
+LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0, LLVMValueRef s1,
+                           LLVMValueRef s2)
 {
-       /* FMA is better on GFX10, because it has FMA units instead of MUL-ADD units. */
-       if (ctx->chip_class >= GFX10) {
-               return ac_build_intrinsic(ctx, "llvm.fma.f32", ctx->f32,
-                                         (LLVMValueRef []) {s0, s1, s2}, 3,
-                                         AC_FUNC_ATTR_READNONE);
-       }
+   /* FMA is better on GFX10, because it has FMA units instead of MUL-ADD units. */
+   if (ctx->chip_class >= GFX10) {
+      return ac_build_intrinsic(ctx, "llvm.fma.f32", ctx->f32, (LLVMValueRef[]){s0, s1, s2}, 3,
+                                AC_FUNC_ATTR_READNONE);
+   }
 
-       return LLVMBuildFAdd(ctx->builder,
-                            LLVMBuildFMul(ctx->builder, s0, s1, ""), s2, "");
+   return LLVMBuildFAdd(ctx->builder, LLVMBuildFMul(ctx->builder, s0, s1, ""), s2, "");
 }
 
 void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned wait_flags)
 {
-       if (!wait_flags)
-               return;
-
-       unsigned lgkmcnt = 63;
-       unsigned vmcnt = ctx->chip_class >= GFX9 ? 63 : 15;
-       unsigned vscnt = 63;
-
-       if (wait_flags & AC_WAIT_LGKM)
-               lgkmcnt = 0;
-       if (wait_flags & AC_WAIT_VLOAD)
-               vmcnt = 0;
-
-       if (wait_flags & AC_WAIT_VSTORE) {
-               if (ctx->chip_class >= GFX10)
-                       vscnt = 0;
-               else
-                       vmcnt = 0;
-       }
-
-       /* There is no intrinsic for vscnt(0), so use a fence. */
-       if ((wait_flags & AC_WAIT_LGKM &&
-            wait_flags & AC_WAIT_VLOAD &&
-            wait_flags & AC_WAIT_VSTORE) ||
-           vscnt == 0) {
-               LLVMBuildFence(ctx->builder, LLVMAtomicOrderingRelease, false, "");
-               return;
-       }
-
-       unsigned simm16 = (lgkmcnt << 8) |
-                         (7 << 4) | /* expcnt */
-                         (vmcnt & 0xf) |
-                         ((vmcnt >> 4) << 14);
-
-       LLVMValueRef args[1] = {
-               LLVMConstInt(ctx->i32, simm16, false),
-       };
-       ac_build_intrinsic(ctx, "llvm.amdgcn.s.waitcnt",
-                          ctx->voidt, args, 1, 0);
-}
-
-LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
-                           unsigned bitsize)
-{
-       LLVMTypeRef type;
-       char *intr;
-
-       if (bitsize == 16) {
-               intr = "llvm.amdgcn.fract.f16";
-               type = ctx->f16;
-       } else if (bitsize == 32) {
-               intr = "llvm.amdgcn.fract.f32";
-               type = ctx->f32;
-       } else {
-               intr = "llvm.amdgcn.fract.f64";
-               type = ctx->f64;
-       }
-
-       LLVMValueRef params[] = {
-               src0,
-       };
-       return ac_build_intrinsic(ctx, intr, type, params, 1,
-                                 AC_FUNC_ATTR_READNONE);
+   if (!wait_flags)
+      return;
+
+   unsigned lgkmcnt = 63;
+   unsigned vmcnt = ctx->chip_class >= GFX9 ? 63 : 15;
+   unsigned vscnt = 63;
+
+   if (wait_flags & AC_WAIT_LGKM)
+      lgkmcnt = 0;
+   if (wait_flags & AC_WAIT_VLOAD)
+      vmcnt = 0;
+
+   if (wait_flags & AC_WAIT_VSTORE) {
+      if (ctx->chip_class >= GFX10)
+         vscnt = 0;
+      else
+         vmcnt = 0;
+   }
+
+   /* There is no intrinsic for vscnt(0), so use a fence. */
+   if ((wait_flags & AC_WAIT_LGKM && wait_flags & AC_WAIT_VLOAD && wait_flags & AC_WAIT_VSTORE) ||
+       vscnt == 0) {
+      LLVMBuildFence(ctx->builder, LLVMAtomicOrderingRelease, false, "");
+      return;
+   }
+
+   unsigned simm16 = (lgkmcnt << 8) | (7 << 4) | /* expcnt */
+                     (vmcnt & 0xf) | ((vmcnt >> 4) << 14);
+
+   LLVMValueRef args[1] = {
+      LLVMConstInt(ctx->i32, simm16, false),
+   };
+   ac_build_intrinsic(ctx, "llvm.amdgcn.s.waitcnt", ctx->voidt, args, 1, 0);
+}
+
+LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0, unsigned bitsize)
+{
+   LLVMTypeRef type;
+   char *intr;
+
+   if (bitsize == 16) {
+      intr = "llvm.amdgcn.fract.f16";
+      type = ctx->f16;
+   } else if (bitsize == 32) {
+      intr = "llvm.amdgcn.fract.f32";
+      type = ctx->f32;
+   } else {
+      intr = "llvm.amdgcn.fract.f64";
+      type = ctx->f64;
+   }
+
+   LLVMValueRef params[] = {
+      src0,
+   };
+   return ac_build_intrinsic(ctx, intr, type, params, 1, AC_FUNC_ATTR_READNONE);
 }
 
 LLVMValueRef ac_const_uint_vec(struct ac_llvm_context *ctx, LLVMTypeRef type, uint64_t value)
 {
 
-       if (LLVMGetTypeKind(type) == LLVMVectorTypeKind) {
-               LLVMValueRef scalar = LLVMConstInt(LLVMGetElementType(type), value, 0);
-               unsigned vec_size = LLVMGetVectorSize(type);
-               LLVMValueRef *scalars = alloca(vec_size * sizeof(LLVMValueRef*));
+   if (LLVMGetTypeKind(type) == LLVMVectorTypeKind) {
+      LLVMValueRef scalar = LLVMConstInt(LLVMGetElementType(type), value, 0);
+      unsigned vec_size = LLVMGetVectorSize(type);
+      LLVMValueRef *scalars = alloca(vec_size * sizeof(LLVMValueRef *));
 
-               for (unsigned i = 0; i < vec_size; i++)
-                       scalars[i] = scalar;
-               return LLVMConstVector(scalars, vec_size);
-       }
-       return LLVMConstInt(type, value, 0);
+      for (unsigned i = 0; i < vec_size; i++)
+         scalars[i] = scalar;
+      return LLVMConstVector(scalars, vec_size);
+   }
+   return LLVMConstInt(type, value, 0);
 }
 
 LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0)
 {
-       LLVMTypeRef type = LLVMTypeOf(src0);
-       LLVMValueRef val;
+   LLVMTypeRef type = LLVMTypeOf(src0);
+   LLVMValueRef val;
 
-       /* v_med3 is selected only when max is first. (LLVM bug?) */
-       val = ac_build_imax(ctx, src0, ac_const_uint_vec(ctx, type, -1));
-       return ac_build_imin(ctx, val, ac_const_uint_vec(ctx, type, 1));
+   /* v_med3 is selected only when max is first. (LLVM bug?) */
+   val = ac_build_imax(ctx, src0, ac_const_uint_vec(ctx, type, -1));
+   return ac_build_imin(ctx, val, ac_const_uint_vec(ctx, type, 1));
 }
 
 static LLVMValueRef ac_eliminate_negative_zero(struct ac_llvm_context *ctx, LLVMValueRef val)
 {
-       ac_enable_signed_zeros(ctx);
-       /* (val + 0) converts negative zero to positive zero. */
-       val = LLVMBuildFAdd(ctx->builder, val, LLVMConstNull(LLVMTypeOf(val)), "");
-       ac_disable_signed_zeros(ctx);
-       return val;
+   ac_enable_signed_zeros(ctx);
+   /* (val + 0) converts negative zero to positive zero. */
+   val = LLVMBuildFAdd(ctx->builder, val, LLVMConstNull(LLVMTypeOf(val)), "");
+   ac_disable_signed_zeros(ctx);
+   return val;
 }
 
 LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src)
 {
-       LLVMTypeRef type = LLVMTypeOf(src);
-       LLVMValueRef pos, neg, dw[2], val;
-       unsigned bitsize = ac_get_elem_bits(ctx, type);
-
-       /* The standard version leads to this:
-        *   v_cmp_ngt_f32_e64 s[0:1], s4, 0                       ; D40B0000 00010004
-        *   v_cndmask_b32_e64 v4, 1.0, s4, s[0:1]                 ; D5010004 000008F2
-        *   v_cmp_le_f32_e32 vcc, 0, v4                           ; 7C060880
-        *   v_cndmask_b32_e32 v4, -1.0, v4, vcc                   ; 020808F3
-        *
-        * The isign version:
-        *   v_add_f32_e64 v4, s4, 0                               ; D5030004 00010004
-        *   v_med3_i32 v4, v4, -1, 1                              ; D5580004 02058304
-        *   v_cvt_f32_i32_e32 v4, v4                              ; 7E080B04
-        *
-        * (src0 + 0) converts negative zero to positive zero.
-        * After that, int(fsign(x)) == isign(floatBitsToInt(x)).
-        *
-        * For FP64, use the standard version, which doesn't suffer from the huge DP rate
-        * reduction. (FP64 comparisons are as fast as int64 comparisons)
-        */
-       if (bitsize == 16 || bitsize == 32) {
-               val = ac_to_integer(ctx, ac_eliminate_negative_zero(ctx, src));
-               val = ac_build_isign(ctx, val);
-               return LLVMBuildSIToFP(ctx->builder, val, type, "");
-       }
-
-       assert(bitsize == 64);
-       pos = LLVMBuildFCmp(ctx->builder, LLVMRealOGT, src, ctx->f64_0, "");
-       neg = LLVMBuildFCmp(ctx->builder, LLVMRealOLT, src, ctx->f64_0, "");
-       dw[0] = ctx->i32_0;
-       dw[1] = LLVMBuildSelect(ctx->builder, pos, LLVMConstInt(ctx->i32, 0x3FF00000, 0),
-                               LLVMBuildSelect(ctx->builder, neg,
-                                               LLVMConstInt(ctx->i32, 0xBFF00000, 0),
-                                               ctx->i32_0, ""), "");
-       return LLVMBuildBitCast(ctx->builder, ac_build_gather_values(ctx, dw, 2), ctx->f64, "");
+   LLVMTypeRef type = LLVMTypeOf(src);
+   LLVMValueRef pos, neg, dw[2], val;
+   unsigned bitsize = ac_get_elem_bits(ctx, type);
+
+   /* The standard version leads to this:
+    *   v_cmp_ngt_f32_e64 s[0:1], s4, 0                       ; D40B0000 00010004
+    *   v_cndmask_b32_e64 v4, 1.0, s4, s[0:1]                 ; D5010004 000008F2
+    *   v_cmp_le_f32_e32 vcc, 0, v4                           ; 7C060880
+    *   v_cndmask_b32_e32 v4, -1.0, v4, vcc                   ; 020808F3
+    *
+    * The isign version:
+    *   v_add_f32_e64 v4, s4, 0                               ; D5030004 00010004
+    *   v_med3_i32 v4, v4, -1, 1                              ; D5580004 02058304
+    *   v_cvt_f32_i32_e32 v4, v4                              ; 7E080B04
+    *
+    * (src0 + 0) converts negative zero to positive zero.
+    * After that, int(fsign(x)) == isign(floatBitsToInt(x)).
+    *
+    * For FP64, use the standard version, which doesn't suffer from the huge DP rate
+    * reduction. (FP64 comparisons are as fast as int64 comparisons)
+    */
+   if (bitsize == 16 || bitsize == 32) {
+      val = ac_to_integer(ctx, ac_eliminate_negative_zero(ctx, src));
+      val = ac_build_isign(ctx, val);
+      return LLVMBuildSIToFP(ctx->builder, val, type, "");
+   }
+
+   assert(bitsize == 64);
+   pos = LLVMBuildFCmp(ctx->builder, LLVMRealOGT, src, ctx->f64_0, "");
+   neg = LLVMBuildFCmp(ctx->builder, LLVMRealOLT, src, ctx->f64_0, "");
+   dw[0] = ctx->i32_0;
+   dw[1] = LLVMBuildSelect(
+      ctx->builder, pos, LLVMConstInt(ctx->i32, 0x3FF00000, 0),
+      LLVMBuildSelect(ctx->builder, neg, LLVMConstInt(ctx->i32, 0xBFF00000, 0), ctx->i32_0, ""),
+      "");
+   return LLVMBuildBitCast(ctx->builder, ac_build_gather_values(ctx, dw, 2), ctx->f64, "");
 }
 
 LLVMValueRef ac_build_bit_count(struct ac_llvm_context *ctx, LLVMValueRef src0)
 {
-       LLVMValueRef result;
-       unsigned bitsize;
-
-       bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
-
-       switch (bitsize) {
-       case 128:
-               result = ac_build_intrinsic(ctx, "llvm.ctpop.i128", ctx->i128,
-                                           (LLVMValueRef []) { src0 }, 1,
-                                           AC_FUNC_ATTR_READNONE);
-               result = LLVMBuildTrunc(ctx->builder, result, ctx->i32, "");
-               break;
-       case 64:
-               result = ac_build_intrinsic(ctx, "llvm.ctpop.i64", ctx->i64,
-                                           (LLVMValueRef []) { src0 }, 1,
-                                           AC_FUNC_ATTR_READNONE);
-
-               result = LLVMBuildTrunc(ctx->builder, result, ctx->i32, "");
-               break;
-       case 32:
-               result = ac_build_intrinsic(ctx, "llvm.ctpop.i32", ctx->i32,
-                                           (LLVMValueRef []) { src0 }, 1,
-                                           AC_FUNC_ATTR_READNONE);
-               break;
-       case 16:
-               result = ac_build_intrinsic(ctx, "llvm.ctpop.i16", ctx->i16,
-                                           (LLVMValueRef []) { src0 }, 1,
-                                           AC_FUNC_ATTR_READNONE);
-
-               result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
-               break;
-       case 8:
-               result = ac_build_intrinsic(ctx, "llvm.ctpop.i8", ctx->i8,
-                                           (LLVMValueRef []) { src0 }, 1,
-                                           AC_FUNC_ATTR_READNONE);
-
-               result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
-               break;
-       default:
-               unreachable(!"invalid bitsize");
-               break;
-       }
-
-       return result;
-}
-
-LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx,
-                                      LLVMValueRef src0)
-{
-       LLVMValueRef result;
-       unsigned bitsize;
-
-       bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
-
-       switch (bitsize) {
-       case 64:
-               result = ac_build_intrinsic(ctx, "llvm.bitreverse.i64", ctx->i64,
-                                           (LLVMValueRef []) { src0 }, 1,
-                                           AC_FUNC_ATTR_READNONE);
-
-               result = LLVMBuildTrunc(ctx->builder, result, ctx->i32, "");
-               break;
-       case 32:
-               result = ac_build_intrinsic(ctx, "llvm.bitreverse.i32", ctx->i32,
-                                           (LLVMValueRef []) { src0 }, 1,
-                                           AC_FUNC_ATTR_READNONE);
-               break;
-       case 16:
-               result = ac_build_intrinsic(ctx, "llvm.bitreverse.i16", ctx->i16,
-                                           (LLVMValueRef []) { src0 }, 1,
-                                           AC_FUNC_ATTR_READNONE);
-
-               result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
-               break;
-       case 8:
-               result = ac_build_intrinsic(ctx, "llvm.bitreverse.i8", ctx->i8,
-                                           (LLVMValueRef []) { src0 }, 1,
-                                           AC_FUNC_ATTR_READNONE);
-
-               result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
-               break;
-       default:
-               unreachable(!"invalid bitsize");
-               break;
-       }
-
-       return result;
-}
-
-#define AC_EXP_TARGET          0
+   LLVMValueRef result;
+   unsigned bitsize;
+
+   bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
+
+   switch (bitsize) {
+   case 128:
+      result = ac_build_intrinsic(ctx, "llvm.ctpop.i128", ctx->i128, (LLVMValueRef[]){src0}, 1,
+                                  AC_FUNC_ATTR_READNONE);
+      result = LLVMBuildTrunc(ctx->builder, result, ctx->i32, "");
+      break;
+   case 64:
+      result = ac_build_intrinsic(ctx, "llvm.ctpop.i64", ctx->i64, (LLVMValueRef[]){src0}, 1,
+                                  AC_FUNC_ATTR_READNONE);
+
+      result = LLVMBuildTrunc(ctx->builder, result, ctx->i32, "");
+      break;
+   case 32:
+      result = ac_build_intrinsic(ctx, "llvm.ctpop.i32", ctx->i32, (LLVMValueRef[]){src0}, 1,
+                                  AC_FUNC_ATTR_READNONE);
+      break;
+   case 16:
+      result = ac_build_intrinsic(ctx, "llvm.ctpop.i16", ctx->i16, (LLVMValueRef[]){src0}, 1,
+                                  AC_FUNC_ATTR_READNONE);
+
+      result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
+      break;
+   case 8:
+      result = ac_build_intrinsic(ctx, "llvm.ctpop.i8", ctx->i8, (LLVMValueRef[]){src0}, 1,
+                                  AC_FUNC_ATTR_READNONE);
+
+      result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
+      break;
+   default:
+      unreachable(!"invalid bitsize");
+      break;
+   }
+
+   return result;
+}
+
+LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx, LLVMValueRef src0)
+{
+   LLVMValueRef result;
+   unsigned bitsize;
+
+   bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
+
+   switch (bitsize) {
+   case 64:
+      result = ac_build_intrinsic(ctx, "llvm.bitreverse.i64", ctx->i64, (LLVMValueRef[]){src0}, 1,
+                                  AC_FUNC_ATTR_READNONE);
+
+      result = LLVMBuildTrunc(ctx->builder, result, ctx->i32, "");
+      break;
+   case 32:
+      result = ac_build_intrinsic(ctx, "llvm.bitreverse.i32", ctx->i32, (LLVMValueRef[]){src0}, 1,
+                                  AC_FUNC_ATTR_READNONE);
+      break;
+   case 16:
+      result = ac_build_intrinsic(ctx, "llvm.bitreverse.i16", ctx->i16, (LLVMValueRef[]){src0}, 1,
+                                  AC_FUNC_ATTR_READNONE);
+
+      result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
+      break;
+   case 8:
+      result = ac_build_intrinsic(ctx, "llvm.bitreverse.i8", ctx->i8, (LLVMValueRef[]){src0}, 1,
+                                  AC_FUNC_ATTR_READNONE);
+
+      result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
+      break;
+   default:
+      unreachable(!"invalid bitsize");
+      break;
+   }
+
+   return result;
+}
+
+#define AC_EXP_TARGET           0
 #define AC_EXP_ENABLED_CHANNELS 1
-#define AC_EXP_OUT0            2
+#define AC_EXP_OUT0             2
 
-enum ac_ir_type {
-       AC_IR_UNDEF,
-       AC_IR_CONST,
-       AC_IR_VALUE,
+enum ac_ir_type
+{
+   AC_IR_UNDEF,
+   AC_IR_CONST,
+   AC_IR_VALUE,
 };
 
-struct ac_vs_exp_chan
-{
-       LLVMValueRef value;
-       float const_float;
-       enum ac_ir_type type;
+struct ac_vs_exp_chan {
+   LLVMValueRef value;
+   float const_float;
+   enum ac_ir_type type;
 };
 
 struct ac_vs_exp_inst {
-       unsigned offset;
-       LLVMValueRef inst;
-       struct ac_vs_exp_chan chan[4];
+   unsigned offset;
+   LLVMValueRef inst;
+   struct ac_vs_exp_chan chan[4];
 };
 
 struct ac_vs_exports {
-       unsigned num;
-       struct ac_vs_exp_inst exp[VARYING_SLOT_MAX];
+   unsigned num;
+   struct ac_vs_exp_inst exp[VARYING_SLOT_MAX];
 };
 
 /* Return true if the PARAM export has been eliminated. */
-static bool ac_eliminate_const_output(uint8_t *vs_output_param_offset,
-                                     uint32_t num_outputs,
-                                     struct ac_vs_exp_inst *exp)
-{
-       unsigned i, default_val; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
-       bool is_zero[4] = {}, is_one[4] = {};
-
-       for (i = 0; i < 4; i++) {
-               /* It's a constant expression. Undef outputs are eliminated too. */
-               if (exp->chan[i].type == AC_IR_UNDEF) {
-                       is_zero[i] = true;
-                       is_one[i] = true;
-               } else if (exp->chan[i].type == AC_IR_CONST) {
-                       if (exp->chan[i].const_float == 0)
-                               is_zero[i] = true;
-                       else if (exp->chan[i].const_float == 1)
-                               is_one[i] = true;
-                       else
-                               return false; /* other constant */
-               } else
-                       return false;
-       }
-
-       /* Only certain combinations of 0 and 1 can be eliminated. */
-       if (is_zero[0] && is_zero[1] && is_zero[2])
-               default_val = is_zero[3] ? 0 : 1;
-       else if (is_one[0] && is_one[1] && is_one[2])
-               default_val = is_zero[3] ? 2 : 3;
-       else
-               return false;
-
-       /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
-       LLVMInstructionEraseFromParent(exp->inst);
-
-       /* Change OFFSET to DEFAULT_VAL. */
-       for (i = 0; i < num_outputs; i++) {
-               if (vs_output_param_offset[i] == exp->offset) {
-                       vs_output_param_offset[i] =
-                               AC_EXP_PARAM_DEFAULT_VAL_0000 + default_val;
-                       break;
-               }
-       }
-       return true;
+static bool ac_eliminate_const_output(uint8_t *vs_output_param_offset, uint32_t num_outputs,
+                                      struct ac_vs_exp_inst *exp)
+{
+   unsigned i, default_val; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
+   bool is_zero[4] = {}, is_one[4] = {};
+
+   for (i = 0; i < 4; i++) {
+      /* It's a constant expression. Undef outputs are eliminated too. */
+      if (exp->chan[i].type == AC_IR_UNDEF) {
+         is_zero[i] = true;
+         is_one[i] = true;
+      } else if (exp->chan[i].type == AC_IR_CONST) {
+         if (exp->chan[i].const_float == 0)
+            is_zero[i] = true;
+         else if (exp->chan[i].const_float == 1)
+            is_one[i] = true;
+         else
+            return false; /* other constant */
+      } else
+         return false;
+   }
+
+   /* Only certain combinations of 0 and 1 can be eliminated. */
+   if (is_zero[0] && is_zero[1] && is_zero[2])
+      default_val = is_zero[3] ? 0 : 1;
+   else if (is_one[0] && is_one[1] && is_one[2])
+      default_val = is_zero[3] ? 2 : 3;
+   else
+      return false;
+
+   /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
+   LLVMInstructionEraseFromParent(exp->inst);
+
+   /* Change OFFSET to DEFAULT_VAL. */
+   for (i = 0; i < num_outputs; i++) {
+      if (vs_output_param_offset[i] == exp->offset) {
+         vs_output_param_offset[i] = AC_EXP_PARAM_DEFAULT_VAL_0000 + default_val;
+         break;
+      }
+   }
+   return true;
 }
 
 static bool ac_eliminate_duplicated_output(struct ac_llvm_context *ctx,
-                                          uint8_t *vs_output_param_offset,
-                                          uint32_t num_outputs,
-                                          struct ac_vs_exports *processed,
-                                          struct ac_vs_exp_inst *exp)
-{
-       unsigned p, copy_back_channels = 0;
-
-       /* See if the output is already in the list of processed outputs.
-        * The LLVMValueRef comparison relies on SSA.
-        */
-       for (p = 0; p < processed->num; p++) {
-               bool different = false;
-
-               for (unsigned j = 0; j < 4; j++) {
-                       struct ac_vs_exp_chan *c1 = &processed->exp[p].chan[j];
-                       struct ac_vs_exp_chan *c2 = &exp->chan[j];
-
-                       /* Treat undef as a match. */
-                       if (c2->type == AC_IR_UNDEF)
-                               continue;
-
-                       /* If c1 is undef but c2 isn't, we can copy c2 to c1
-                        * and consider the instruction duplicated.
-                        */
-                       if (c1->type == AC_IR_UNDEF) {
-                               copy_back_channels |= 1 << j;
-                               continue;
-                       }
-
-                       /* Test whether the channels are not equal. */
-                       if (c1->type != c2->type ||
-                           (c1->type == AC_IR_CONST &&
-                            c1->const_float != c2->const_float) ||
-                           (c1->type == AC_IR_VALUE &&
-                            c1->value != c2->value)) {
-                               different = true;
-                               break;
-                       }
-               }
-               if (!different)
-                       break;
-
-               copy_back_channels = 0;
-       }
-       if (p == processed->num)
-               return false;
-
-       /* If a match was found, but the matching export has undef where the new
-        * one has a normal value, copy the normal value to the undef channel.
-        */
-       struct ac_vs_exp_inst *match = &processed->exp[p];
-
-       /* Get current enabled channels mask. */
-       LLVMValueRef arg = LLVMGetOperand(match->inst, AC_EXP_ENABLED_CHANNELS);
-       unsigned enabled_channels = LLVMConstIntGetZExtValue(arg);
-
-       while (copy_back_channels) {
-               unsigned chan = u_bit_scan(&copy_back_channels);
-
-               assert(match->chan[chan].type == AC_IR_UNDEF);
-               LLVMSetOperand(match->inst, AC_EXP_OUT0 + chan,
-                              exp->chan[chan].value);
-               match->chan[chan] = exp->chan[chan];
-
-               /* Update number of enabled channels because the original mask
-                * is not always 0xf.
-                */
-               enabled_channels |= (1 << chan);
-               LLVMSetOperand(match->inst, AC_EXP_ENABLED_CHANNELS,
-                              LLVMConstInt(ctx->i32, enabled_channels, 0));
-       }
-
-       /* The PARAM export is duplicated. Kill it. */
-       LLVMInstructionEraseFromParent(exp->inst);
-
-       /* Change OFFSET to the matching export. */
-       for (unsigned i = 0; i < num_outputs; i++) {
-               if (vs_output_param_offset[i] == exp->offset) {
-                       vs_output_param_offset[i] = match->offset;
-                       break;
-               }
-       }
-       return true;
-}
-
-void ac_optimize_vs_outputs(struct ac_llvm_context *ctx,
-                           LLVMValueRef main_fn,
-                           uint8_t *vs_output_param_offset,
-                           uint32_t num_outputs,
-                           uint32_t skip_output_mask,
-                           uint8_t *num_param_exports)
-{
-       LLVMBasicBlockRef bb;
-       bool removed_any = false;
-       struct ac_vs_exports exports;
-
-       exports.num = 0;
-
-       /* Process all LLVM instructions. */
-       bb = LLVMGetFirstBasicBlock(main_fn);
-       while (bb) {
-               LLVMValueRef inst = LLVMGetFirstInstruction(bb);
-
-               while (inst) {
-                       LLVMValueRef cur = inst;
-                       inst = LLVMGetNextInstruction(inst);
-                       struct ac_vs_exp_inst exp;
-
-                       if (LLVMGetInstructionOpcode(cur) != LLVMCall)
-                               continue;
-
-                       LLVMValueRef callee = ac_llvm_get_called_value(cur);
-
-                       if (!ac_llvm_is_function(callee))
-                               continue;
-
-                       const char *name = LLVMGetValueName(callee);
-                       unsigned num_args = LLVMCountParams(callee);
-
-                       /* Check if this is an export instruction. */
-                       if ((num_args != 9 && num_args != 8) ||
-                           (strcmp(name, "llvm.SI.export") &&
-                            strcmp(name, "llvm.amdgcn.exp.f32")))
-                               continue;
-
-                       LLVMValueRef arg = LLVMGetOperand(cur, AC_EXP_TARGET);
-                       unsigned target = LLVMConstIntGetZExtValue(arg);
-
-                       if (target < V_008DFC_SQ_EXP_PARAM)
-                               continue;
-
-                       target -= V_008DFC_SQ_EXP_PARAM;
-
-                       /* Parse the instruction. */
-                       memset(&exp, 0, sizeof(exp));
-                       exp.offset = target;
-                       exp.inst = cur;
-
-                       for (unsigned i = 0; i < 4; i++) {
-                               LLVMValueRef v = LLVMGetOperand(cur, AC_EXP_OUT0 + i);
-
-                               exp.chan[i].value = v;
-
-                               if (LLVMIsUndef(v)) {
-                                       exp.chan[i].type = AC_IR_UNDEF;
-                               } else if (LLVMIsAConstantFP(v)) {
-                                       LLVMBool loses_info;
-                                       exp.chan[i].type = AC_IR_CONST;
-                                       exp.chan[i].const_float =
-                                               LLVMConstRealGetDouble(v, &loses_info);
-                               } else {
-                                       exp.chan[i].type = AC_IR_VALUE;
-                               }
-                       }
-
-                       /* Eliminate constant and duplicated PARAM exports. */
-                       if (!((1u << target) & skip_output_mask) &&
-                            (ac_eliminate_const_output(vs_output_param_offset,
-                                                      num_outputs, &exp) ||
-                            ac_eliminate_duplicated_output(ctx,
-                                                           vs_output_param_offset,
-                                                           num_outputs, &exports,
-                                                           &exp))) {
-                               removed_any = true;
-                       } else {
-                               exports.exp[exports.num++] = exp;
-                       }
-               }
-               bb = LLVMGetNextBasicBlock(bb);
-       }
-
-       /* Remove holes in export memory due to removed PARAM exports.
-        * This is done by renumbering all PARAM exports.
-        */
-       if (removed_any) {
-               uint8_t old_offset[VARYING_SLOT_MAX];
-               unsigned out, i;
-
-               /* Make a copy of the offsets. We need the old version while
-                * we are modifying some of them. */
-               memcpy(old_offset, vs_output_param_offset,
-                      sizeof(old_offset));
-
-               for (i = 0; i < exports.num; i++) {
-                       unsigned offset = exports.exp[i].offset;
-
-                       /* Update vs_output_param_offset. Multiple outputs can
-                        * have the same offset.
-                        */
-                       for (out = 0; out < num_outputs; out++) {
-                               if (old_offset[out] == offset)
-                                       vs_output_param_offset[out] = i;
-                       }
-
-                       /* Change the PARAM offset in the instruction. */
-                       LLVMSetOperand(exports.exp[i].inst, AC_EXP_TARGET,
-                                      LLVMConstInt(ctx->i32,
-                                                   V_008DFC_SQ_EXP_PARAM + i, 0));
-               }
-               *num_param_exports = exports.num;
-       }
+                                           uint8_t *vs_output_param_offset, uint32_t num_outputs,
+                                           struct ac_vs_exports *processed,
+                                           struct ac_vs_exp_inst *exp)
+{
+   unsigned p, copy_back_channels = 0;
+
+   /* See if the output is already in the list of processed outputs.
+    * The LLVMValueRef comparison relies on SSA.
+    */
+   for (p = 0; p < processed->num; p++) {
+      bool different = false;
+
+      for (unsigned j = 0; j < 4; j++) {
+         struct ac_vs_exp_chan *c1 = &processed->exp[p].chan[j];
+         struct ac_vs_exp_chan *c2 = &exp->chan[j];
+
+         /* Treat undef as a match. */
+         if (c2->type == AC_IR_UNDEF)
+            continue;
+
+         /* If c1 is undef but c2 isn't, we can copy c2 to c1
+          * and consider the instruction duplicated.
+          */
+         if (c1->type == AC_IR_UNDEF) {
+            copy_back_channels |= 1 << j;
+            continue;
+         }
+
+         /* Test whether the channels are not equal. */
+         if (c1->type != c2->type ||
+             (c1->type == AC_IR_CONST && c1->const_float != c2->const_float) ||
+             (c1->type == AC_IR_VALUE && c1->value != c2->value)) {
+            different = true;
+            break;
+         }
+      }
+      if (!different)
+         break;
+
+      copy_back_channels = 0;
+   }
+   if (p == processed->num)
+      return false;
+
+   /* If a match was found, but the matching export has undef where the new
+    * one has a normal value, copy the normal value to the undef channel.
+    */
+   struct ac_vs_exp_inst *match = &processed->exp[p];
+
+   /* Get current enabled channels mask. */
+   LLVMValueRef arg = LLVMGetOperand(match->inst, AC_EXP_ENABLED_CHANNELS);
+   unsigned enabled_channels = LLVMConstIntGetZExtValue(arg);
+
+   while (copy_back_channels) {
+      unsigned chan = u_bit_scan(&copy_back_channels);
+
+      assert(match->chan[chan].type == AC_IR_UNDEF);
+      LLVMSetOperand(match->inst, AC_EXP_OUT0 + chan, exp->chan[chan].value);
+      match->chan[chan] = exp->chan[chan];
+
+      /* Update number of enabled channels because the original mask
+       * is not always 0xf.
+       */
+      enabled_channels |= (1 << chan);
+      LLVMSetOperand(match->inst, AC_EXP_ENABLED_CHANNELS,
+                     LLVMConstInt(ctx->i32, enabled_channels, 0));
+   }
+
+   /* The PARAM export is duplicated. Kill it. */
+   LLVMInstructionEraseFromParent(exp->inst);
+
+   /* Change OFFSET to the matching export. */
+   for (unsigned i = 0; i < num_outputs; i++) {
+      if (vs_output_param_offset[i] == exp->offset) {
+         vs_output_param_offset[i] = match->offset;
+         break;
+      }
+   }
+   return true;
+}
+
+void ac_optimize_vs_outputs(struct ac_llvm_context *ctx, LLVMValueRef main_fn,
+                            uint8_t *vs_output_param_offset, uint32_t num_outputs,
+                            uint32_t skip_output_mask, uint8_t *num_param_exports)
+{
+   LLVMBasicBlockRef bb;
+   bool removed_any = false;
+   struct ac_vs_exports exports;
+
+   exports.num = 0;
+
+   /* Process all LLVM instructions. */
+   bb = LLVMGetFirstBasicBlock(main_fn);
+   while (bb) {
+      LLVMValueRef inst = LLVMGetFirstInstruction(bb);
+
+      while (inst) {
+         LLVMValueRef cur = inst;
+         inst = LLVMGetNextInstruction(inst);
+         struct ac_vs_exp_inst exp;
+
+         if (LLVMGetInstructionOpcode(cur) != LLVMCall)
+            continue;
+
+         LLVMValueRef callee = ac_llvm_get_called_value(cur);
+
+         if (!ac_llvm_is_function(callee))
+            continue;
+
+         const char *name = LLVMGetValueName(callee);
+         unsigned num_args = LLVMCountParams(callee);
+
+         /* Check if this is an export instruction. */
+         if ((num_args != 9 && num_args != 8) ||
+             (strcmp(name, "llvm.SI.export") && strcmp(name, "llvm.amdgcn.exp.f32")))
+            continue;
+
+         LLVMValueRef arg = LLVMGetOperand(cur, AC_EXP_TARGET);
+         unsigned target = LLVMConstIntGetZExtValue(arg);
+
+         if (target < V_008DFC_SQ_EXP_PARAM)
+            continue;
+
+         target -= V_008DFC_SQ_EXP_PARAM;
+
+         /* Parse the instruction. */
+         memset(&exp, 0, sizeof(exp));
+         exp.offset = target;
+         exp.inst = cur;
+
+         for (unsigned i = 0; i < 4; i++) {
+            LLVMValueRef v = LLVMGetOperand(cur, AC_EXP_OUT0 + i);
+
+            exp.chan[i].value = v;
+
+            if (LLVMIsUndef(v)) {
+               exp.chan[i].type = AC_IR_UNDEF;
+            } else if (LLVMIsAConstantFP(v)) {
+               LLVMBool loses_info;
+               exp.chan[i].type = AC_IR_CONST;
+               exp.chan[i].const_float = LLVMConstRealGetDouble(v, &loses_info);
+            } else {
+               exp.chan[i].type = AC_IR_VALUE;
+            }
+         }
+
+         /* Eliminate constant and duplicated PARAM exports. */
+         if (!((1u << target) & skip_output_mask) &&
+             (ac_eliminate_const_output(vs_output_param_offset, num_outputs, &exp) ||
+              ac_eliminate_duplicated_output(ctx, vs_output_param_offset, num_outputs, &exports,
+                                             &exp))) {
+            removed_any = true;
+         } else {
+            exports.exp[exports.num++] = exp;
+         }
+      }
+      bb = LLVMGetNextBasicBlock(bb);
+   }
+
+   /* Remove holes in export memory due to removed PARAM exports.
+    * This is done by renumbering all PARAM exports.
+    */
+   if (removed_any) {
+      uint8_t old_offset[VARYING_SLOT_MAX];
+      unsigned out, i;
+
+      /* Make a copy of the offsets. We need the old version while
+       * we are modifying some of them. */
+      memcpy(old_offset, vs_output_param_offset, sizeof(old_offset));
+
+      for (i = 0; i < exports.num; i++) {
+         unsigned offset = exports.exp[i].offset;
+
+         /* Update vs_output_param_offset. Multiple outputs can
+          * have the same offset.
+          */
+         for (out = 0; out < num_outputs; out++) {
+            if (old_offset[out] == offset)
+               vs_output_param_offset[out] = i;
+         }
+
+         /* Change the PARAM offset in the instruction. */
+         LLVMSetOperand(exports.exp[i].inst, AC_EXP_TARGET,
+                        LLVMConstInt(ctx->i32, V_008DFC_SQ_EXP_PARAM + i, 0));
+      }
+      *num_param_exports = exports.num;
+   }
 }
 
 void ac_init_exec_full_mask(struct ac_llvm_context *ctx)
 {
-       LLVMValueRef full_mask = LLVMConstInt(ctx->i64, ~0ull, 0);
-       ac_build_intrinsic(ctx,
-                          "llvm.amdgcn.init.exec", ctx->voidt,
-                          &full_mask, 1, AC_FUNC_ATTR_CONVERGENT);
+   LLVMValueRef full_mask = LLVMConstInt(ctx->i64, ~0ull, 0);
+   ac_build_intrinsic(ctx, "llvm.amdgcn.init.exec", ctx->voidt, &full_mask, 1,
+                      AC_FUNC_ATTR_CONVERGENT);
 }
 
 void ac_declare_lds_as_pointer(struct ac_llvm_context *ctx)
 {
-       unsigned lds_size = ctx->chip_class >= GFX7 ? 65536 : 32768;
-       ctx->lds = LLVMBuildIntToPtr(ctx->builder, ctx->i32_0,
-                                    LLVMPointerType(LLVMArrayType(ctx->i32, lds_size / 4), AC_ADDR_SPACE_LDS),
-                                    "lds");
-}
-
-LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
-                        LLVMValueRef dw_addr)
-{
-       return LLVMBuildLoad(ctx->builder, ac_build_gep0(ctx, ctx->lds, dw_addr), "");
-}
-
-void ac_lds_store(struct ac_llvm_context *ctx,
-                 LLVMValueRef dw_addr,
-                 LLVMValueRef value)
-{
-       value = ac_to_integer(ctx, value);
-       ac_build_indexed_store(ctx, ctx->lds,
-                              dw_addr, value);
-}
-
-LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx,
-                        LLVMTypeRef dst_type,
-                        LLVMValueRef src0)
-{
-       unsigned src0_bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
-       const char *intrin_name;
-       LLVMTypeRef type;
-       LLVMValueRef zero;
-
-       switch (src0_bitsize) {
-       case 64:
-               intrin_name = "llvm.cttz.i64";
-               type = ctx->i64;
-               zero = ctx->i64_0;
-               break;
-       case 32:
-               intrin_name = "llvm.cttz.i32";
-               type = ctx->i32;
-               zero = ctx->i32_0;
-               break;
-       case 16:
-               intrin_name = "llvm.cttz.i16";
-               type = ctx->i16;
-               zero = ctx->i16_0;
-               break;
-       case 8:
-               intrin_name = "llvm.cttz.i8";
-               type = ctx->i8;
-               zero = ctx->i8_0;
-               break;
-       default:
-               unreachable(!"invalid bitsize");
-       }
-
-       LLVMValueRef params[2] = {
-               src0,
-
-               /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
-                * add special code to check for x=0. The reason is that
-                * the LLVM behavior for x=0 is different from what we
-                * need here. However, LLVM also assumes that ffs(x) is
-                * in [0, 31], but GLSL expects that ffs(0) = -1, so
-                * a conditional assignment to handle 0 is still required.
-                *
-                * The hardware already implements the correct behavior.
-                */
-               ctx->i1true,
-       };
-
-       LLVMValueRef lsb = ac_build_intrinsic(ctx, intrin_name, type,
-                                             params, 2,
-                                             AC_FUNC_ATTR_READNONE);
-
-       if (src0_bitsize == 64) {
-               lsb = LLVMBuildTrunc(ctx->builder, lsb, ctx->i32, "");
-       } else if (src0_bitsize < 32) {
-               lsb = LLVMBuildSExt(ctx->builder, lsb, ctx->i32, "");
-       }
-
-       /* TODO: We need an intrinsic to skip this conditional. */
-       /* Check for zero: */
-       return LLVMBuildSelect(ctx->builder, LLVMBuildICmp(ctx->builder,
-                                                          LLVMIntEQ, src0,
-                                                          zero, ""),
-                              LLVMConstInt(ctx->i32, -1, 0), lsb, "");
+   unsigned lds_size = ctx->chip_class >= GFX7 ? 65536 : 32768;
+   ctx->lds = LLVMBuildIntToPtr(
+      ctx->builder, ctx->i32_0,
+      LLVMPointerType(LLVMArrayType(ctx->i32, lds_size / 4), AC_ADDR_SPACE_LDS), "lds");
+}
+
+LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx, LLVMValueRef dw_addr)
+{
+   return LLVMBuildLoad(ctx->builder, ac_build_gep0(ctx, ctx->lds, dw_addr), "");
+}
+
+void ac_lds_store(struct ac_llvm_context *ctx, LLVMValueRef dw_addr, LLVMValueRef value)
+{
+   value = ac_to_integer(ctx, value);
+   ac_build_indexed_store(ctx, ctx->lds, dw_addr, value);
+}
+
+LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx, LLVMTypeRef dst_type, LLVMValueRef src0)
+{
+   unsigned src0_bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
+   const char *intrin_name;
+   LLVMTypeRef type;
+   LLVMValueRef zero;
+
+   switch (src0_bitsize) {
+   case 64:
+      intrin_name = "llvm.cttz.i64";
+      type = ctx->i64;
+      zero = ctx->i64_0;
+      break;
+   case 32:
+      intrin_name = "llvm.cttz.i32";
+      type = ctx->i32;
+      zero = ctx->i32_0;
+      break;
+   case 16:
+      intrin_name = "llvm.cttz.i16";
+      type = ctx->i16;
+      zero = ctx->i16_0;
+      break;
+   case 8:
+      intrin_name = "llvm.cttz.i8";
+      type = ctx->i8;
+      zero = ctx->i8_0;
+      break;
+   default:
+      unreachable(!"invalid bitsize");
+   }
+
+   LLVMValueRef params[2] = {
+      src0,
+
+      /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
+       * add special code to check for x=0. The reason is that
+       * the LLVM behavior for x=0 is different from what we
+       * need here. However, LLVM also assumes that ffs(x) is
+       * in [0, 31], but GLSL expects that ffs(0) = -1, so
+       * a conditional assignment to handle 0 is still required.
+       *
+       * The hardware already implements the correct behavior.
+       */
+      ctx->i1true,
+   };
+
+   LLVMValueRef lsb = ac_build_intrinsic(ctx, intrin_name, type, params, 2, AC_FUNC_ATTR_READNONE);
+
+   if (src0_bitsize == 64) {
+      lsb = LLVMBuildTrunc(ctx->builder, lsb, ctx->i32, "");
+   } else if (src0_bitsize < 32) {
+      lsb = LLVMBuildSExt(ctx->builder, lsb, ctx->i32, "");
+   }
+
+   /* TODO: We need an intrinsic to skip this conditional. */
+   /* Check for zero: */
+   return LLVMBuildSelect(ctx->builder, LLVMBuildICmp(ctx->builder, LLVMIntEQ, src0, zero, ""),
+                          LLVMConstInt(ctx->i32, -1, 0), lsb, "");
 }
 
 LLVMTypeRef ac_array_in_const_addr_space(LLVMTypeRef elem_type)
 {
-       return LLVMPointerType(elem_type, AC_ADDR_SPACE_CONST);
+   return LLVMPointerType(elem_type, AC_ADDR_SPACE_CONST);
 }
 
 LLVMTypeRef ac_array_in_const32_addr_space(LLVMTypeRef elem_type)
 {
-       return LLVMPointerType(elem_type, AC_ADDR_SPACE_CONST_32BIT);
+   return LLVMPointerType(elem_type, AC_ADDR_SPACE_CONST_32BIT);
 }
 
-static struct ac_llvm_flow *
-get_current_flow(struct ac_llvm_context *ctx)
+static struct ac_llvm_flow *get_current_flow(struct ac_llvm_context *ctx)
 {
-       if (ctx->flow->depth > 0)
-               return &ctx->flow->stack[ctx->flow->depth - 1];
-       return NULL;
+   if (ctx->flow->depth > 0)
+      return &ctx->flow->stack[ctx->flow->depth - 1];
+   return NULL;
 }
 
-static struct ac_llvm_flow *
-get_innermost_loop(struct ac_llvm_context *ctx)
+static struct ac_llvm_flow *get_innermost_loop(struct ac_llvm_context *ctx)
 {
-       for (unsigned i = ctx->flow->depth; i > 0; --i) {
-               if (ctx->flow->stack[i - 1].loop_entry_block)
-                       return &ctx->flow->stack[i - 1];
-       }
-       return NULL;
+   for (unsigned i = ctx->flow->depth; i > 0; --i) {
+      if (ctx->flow->stack[i - 1].loop_entry_block)
+         return &ctx->flow->stack[i - 1];
+   }
+   return NULL;
 }
 
-static struct ac_llvm_flow *
-push_flow(struct ac_llvm_context *ctx)
+static struct ac_llvm_flow *push_flow(struct ac_llvm_context *ctx)
 {
-       struct ac_llvm_flow *flow;
+   struct ac_llvm_flow *flow;
 
-       if (ctx->flow->depth >= ctx->flow->depth_max) {
-               unsigned new_max = MAX2(ctx->flow->depth << 1,
-                                       AC_LLVM_INITIAL_CF_DEPTH);
+   if (ctx->flow->depth >= ctx->flow->depth_max) {
+      unsigned new_max = MAX2(ctx->flow->depth << 1, AC_LLVM_INITIAL_CF_DEPTH);
 
-               ctx->flow->stack = realloc(ctx->flow->stack, new_max * sizeof(*ctx->flow->stack));
-               ctx->flow->depth_max = new_max;
-       }
+      ctx->flow->stack = realloc(ctx->flow->stack, new_max * sizeof(*ctx->flow->stack));
+      ctx->flow->depth_max = new_max;
+   }
 
-       flow = &ctx->flow->stack[ctx->flow->depth];
-       ctx->flow->depth++;
+   flow = &ctx->flow->stack[ctx->flow->depth];
+   ctx->flow->depth++;
 
-       flow->next_block = NULL;
-       flow->loop_entry_block = NULL;
-       return flow;
+   flow->next_block = NULL;
+   flow->loop_entry_block = NULL;
+   return flow;
 }
 
-static void set_basicblock_name(LLVMBasicBlockRef bb, const char *base,
-                               int label_id)
+static void set_basicblock_name(LLVMBasicBlockRef bb, const char *base, int label_id)
 {
-       char buf[32];
-       snprintf(buf, sizeof(buf), "%s%d", base, label_id);
-       LLVMSetValueName(LLVMBasicBlockAsValue(bb), buf);
+   char buf[32];
+   snprintf(buf, sizeof(buf), "%s%d", base, label_id);
+   LLVMSetValueName(LLVMBasicBlockAsValue(bb), buf);
 }
 
 /* Append a basic block at the level of the parent flow.
  */
-static LLVMBasicBlockRef append_basic_block(struct ac_llvm_context *ctx,
-                                           const char *name)
+static LLVMBasicBlockRef append_basic_block(struct ac_llvm_context *ctx, const char *name)
 {
-       assert(ctx->flow->depth >= 1);
+   assert(ctx->flow->depth >= 1);
 
-       if (ctx->flow->depth >= 2) {
-               struct ac_llvm_flow *flow = &ctx->flow->stack[ctx->flow->depth - 2];
+   if (ctx->flow->depth >= 2) {
+      struct ac_llvm_flow *flow = &ctx->flow->stack[ctx->flow->depth - 2];
 
-               return LLVMInsertBasicBlockInContext(ctx->context,
-                                                    flow->next_block, name);
-       }
+      return LLVMInsertBasicBlockInContext(ctx->context, flow->next_block, name);
+   }
 
-       LLVMValueRef main_fn =
-               LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx->builder));
-       return LLVMAppendBasicBlockInContext(ctx->context, main_fn, name);
+   LLVMValueRef main_fn = LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx->builder));
+   return LLVMAppendBasicBlockInContext(ctx->context, main_fn, name);
 }
 
 /* Emit a branch to the given default target for the current block if
  * applicable -- that is, if the current block does not already contain a
  * branch from a break or continue.
  */
-static void emit_default_branch(LLVMBuilderRef builder,
-                               LLVMBasicBlockRef target)
+static void emit_default_branch(LLVMBuilderRef builder, LLVMBasicBlockRef target)
 {
-       if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(builder)))
-                LLVMBuildBr(builder, target);
+   if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(builder)))
+      LLVMBuildBr(builder, target);
 }
 
 void ac_build_bgnloop(struct ac_llvm_context *ctx, int label_id)
 {
-       struct ac_llvm_flow *flow = push_flow(ctx);
-       flow->loop_entry_block = append_basic_block(ctx, "LOOP");
-       flow->next_block = append_basic_block(ctx, "ENDLOOP");
-       set_basicblock_name(flow->loop_entry_block, "loop", label_id);
-       LLVMBuildBr(ctx->builder, flow->loop_entry_block);
-       LLVMPositionBuilderAtEnd(ctx->builder, flow->loop_entry_block);
+   struct ac_llvm_flow *flow = push_flow(ctx);
+   flow->loop_entry_block = append_basic_block(ctx, "LOOP");
+   flow->next_block = append_basic_block(ctx, "ENDLOOP");
+   set_basicblock_name(flow->loop_entry_block, "loop", label_id);
+   LLVMBuildBr(ctx->builder, flow->loop_entry_block);
+   LLVMPositionBuilderAtEnd(ctx->builder, flow->loop_entry_block);
 }
 
 void ac_build_break(struct ac_llvm_context *ctx)
 {
-       struct ac_llvm_flow *flow = get_innermost_loop(ctx);
-       LLVMBuildBr(ctx->builder, flow->next_block);
+   struct ac_llvm_flow *flow = get_innermost_loop(ctx);
+   LLVMBuildBr(ctx->builder, flow->next_block);
 }
 
 void ac_build_continue(struct ac_llvm_context *ctx)
 {
-       struct ac_llvm_flow *flow = get_innermost_loop(ctx);
-       LLVMBuildBr(ctx->builder, flow->loop_entry_block);
+   struct ac_llvm_flow *flow = get_innermost_loop(ctx);
+   LLVMBuildBr(ctx->builder, flow->loop_entry_block);
 }
 
 void ac_build_else(struct ac_llvm_context *ctx, int label_id)
 {
-       struct ac_llvm_flow *current_branch = get_current_flow(ctx);
-       LLVMBasicBlockRef endif_block;
+   struct ac_llvm_flow *current_branch = get_current_flow(ctx);
+   LLVMBasicBlockRef endif_block;
 
-       assert(!current_branch->loop_entry_block);
+   assert(!current_branch->loop_entry_block);
 
-       endif_block = append_basic_block(ctx, "ENDIF");
-       emit_default_branch(ctx->builder, endif_block);
+   endif_block = append_basic_block(ctx, "ENDIF");
+   emit_default_branch(ctx->builder, endif_block);
 
-       LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
-       set_basicblock_name(current_branch->next_block, "else", label_id);
+   LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
+   set_basicblock_name(current_branch->next_block, "else", label_id);
 
-       current_branch->next_block = endif_block;
+   current_branch->next_block = endif_block;
 }
 
 void ac_build_endif(struct ac_llvm_context *ctx, int label_id)
 {
-       struct ac_llvm_flow *current_branch = get_current_flow(ctx);
+   struct ac_llvm_flow *current_branch = get_current_flow(ctx);
 
-       assert(!current_branch->loop_entry_block);
+   assert(!current_branch->loop_entry_block);
 
-       emit_default_branch(ctx->builder, current_branch->next_block);
-       LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
-       set_basicblock_name(current_branch->next_block, "endif", label_id);
+   emit_default_branch(ctx->builder, current_branch->next_block);
+   LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
+   set_basicblock_name(current_branch->next_block, "endif", label_id);
 
-       ctx->flow->depth--;
+   ctx->flow->depth--;
 }
 
 void ac_build_endloop(struct ac_llvm_context *ctx, int label_id)
 {
-       struct ac_llvm_flow *current_loop = get_current_flow(ctx);
+   struct ac_llvm_flow *current_loop = get_current_flow(ctx);
 
-       assert(current_loop->loop_entry_block);
+   assert(current_loop->loop_entry_block);
 
-       emit_default_branch(ctx->builder, current_loop->loop_entry_block);
+   emit_default_branch(ctx->builder, current_loop->loop_entry_block);
 
-       LLVMPositionBuilderAtEnd(ctx->builder, current_loop->next_block);
-       set_basicblock_name(current_loop->next_block, "endloop", label_id);
-       ctx->flow->depth--;
+   LLVMPositionBuilderAtEnd(ctx->builder, current_loop->next_block);
+   set_basicblock_name(current_loop->next_block, "endloop", label_id);
+   ctx->flow->depth--;
 }
 
 void ac_build_ifcc(struct ac_llvm_context *ctx, LLVMValueRef cond, int label_id)
 {
-       struct ac_llvm_flow *flow = push_flow(ctx);
-       LLVMBasicBlockRef if_block;
+   struct ac_llvm_flow *flow = push_flow(ctx);
+   LLVMBasicBlockRef if_block;
 
-       if_block = append_basic_block(ctx, "IF");
-       flow->next_block = append_basic_block(ctx, "ELSE");
-       set_basicblock_name(if_block, "if", label_id);
-       LLVMBuildCondBr(ctx->builder, cond, if_block, flow->next_block);
-       LLVMPositionBuilderAtEnd(ctx->builder, if_block);
+   if_block = append_basic_block(ctx, "IF");
+   flow->next_block = append_basic_block(ctx, "ELSE");
+   set_basicblock_name(if_block, "if", label_id);
+   LLVMBuildCondBr(ctx->builder, cond, if_block, flow->next_block);
+   LLVMPositionBuilderAtEnd(ctx->builder, if_block);
 }
 
-void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value,
-                int label_id)
+void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value, int label_id)
 {
-       LLVMValueRef cond = LLVMBuildFCmp(ctx->builder, LLVMRealUNE,
-                                         value, ctx->f32_0, "");
-       ac_build_ifcc(ctx, cond, label_id);
+   LLVMValueRef cond = LLVMBuildFCmp(ctx->builder, LLVMRealUNE, value, ctx->f32_0, "");
+   ac_build_ifcc(ctx, cond, label_id);
 }
 
-void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value,
-                 int label_id)
+void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value, int label_id)
 {
-       LLVMValueRef cond = LLVMBuildICmp(ctx->builder, LLVMIntNE,
-                                         ac_to_integer(ctx, value),
-                                         ctx->i32_0, "");
-       ac_build_ifcc(ctx, cond, label_id);
+   LLVMValueRef cond =
+      LLVMBuildICmp(ctx->builder, LLVMIntNE, ac_to_integer(ctx, value), ctx->i32_0, "");
+   ac_build_ifcc(ctx, cond, label_id);
 }
 
-LLVMValueRef ac_build_alloca_undef(struct ac_llvm_context *ac, LLVMTypeRef type,
-                            const char *name)
+LLVMValueRef ac_build_alloca_undef(struct ac_llvm_context *ac, LLVMTypeRef type, const char *name)
 {
-       LLVMBuilderRef builder = ac->builder;
-       LLVMBasicBlockRef current_block = LLVMGetInsertBlock(builder);
-       LLVMValueRef function = LLVMGetBasicBlockParent(current_block);
-       LLVMBasicBlockRef first_block = LLVMGetEntryBasicBlock(function);
-       LLVMValueRef first_instr = LLVMGetFirstInstruction(first_block);
-       LLVMBuilderRef first_builder = LLVMCreateBuilderInContext(ac->context);
-       LLVMValueRef res;
+   LLVMBuilderRef builder = ac->builder;
+   LLVMBasicBlockRef current_block = LLVMGetInsertBlock(builder);
+   LLVMValueRef function = LLVMGetBasicBlockParent(current_block);
+   LLVMBasicBlockRef first_block = LLVMGetEntryBasicBlock(function);
+   LLVMValueRef first_instr = LLVMGetFirstInstruction(first_block);
+   LLVMBuilderRef first_builder = LLVMCreateBuilderInContext(ac->context);
+   LLVMValueRef res;
 
-       if (first_instr) {
-               LLVMPositionBuilderBefore(first_builder, first_instr);
-       } else {
-               LLVMPositionBuilderAtEnd(first_builder, first_block);
-       }
+   if (first_instr) {
+      LLVMPositionBuilderBefore(first_builder, first_instr);
+   } else {
+      LLVMPositionBuilderAtEnd(first_builder, first_block);
+   }
 
-       res = LLVMBuildAlloca(first_builder, type, name);
-       LLVMDisposeBuilder(first_builder);
-       return res;
+   res = LLVMBuildAlloca(first_builder, type, name);
+   LLVMDisposeBuilder(first_builder);
+   return res;
 }
 
-LLVMValueRef ac_build_alloca(struct ac_llvm_context *ac,
-                                  LLVMTypeRef type, const char *name)
+LLVMValueRef ac_build_alloca(struct ac_llvm_context *ac, LLVMTypeRef type, const char *name)
 {
-       LLVMValueRef ptr = ac_build_alloca_undef(ac, type, name);
-       LLVMBuildStore(ac->builder, LLVMConstNull(type), ptr);
-       return ptr;
+   LLVMValueRef ptr = ac_build_alloca_undef(ac, type, name);
+   LLVMBuildStore(ac->builder, LLVMConstNull(type), ptr);
+   return ptr;
 }
 
-LLVMValueRef ac_cast_ptr(struct ac_llvm_context *ctx, LLVMValueRef ptr,
-                         LLVMTypeRef type)
+LLVMValueRef ac_cast_ptr(struct ac_llvm_context *ctx, LLVMValueRef ptr, LLVMTypeRef type)
 {
-       int addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
-       return LLVMBuildBitCast(ctx->builder, ptr,
-                               LLVMPointerType(type, addr_space), "");
+   int addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
+   return LLVMBuildBitCast(ctx->builder, ptr, LLVMPointerType(type, addr_space), "");
 }
 
-LLVMValueRef ac_trim_vector(struct ac_llvm_context *ctx, LLVMValueRef value,
-                           unsigned count)
+LLVMValueRef ac_trim_vector(struct ac_llvm_context *ctx, LLVMValueRef value, unsigned count)
 {
-       unsigned num_components = ac_get_llvm_num_components(value);
-       if (count == num_components)
-               return value;
+   unsigned num_components = ac_get_llvm_num_components(value);
+   if (count == num_components)
+      return value;
 
-       LLVMValueRef masks[MAX2(count, 2)];
-       masks[0] = ctx->i32_0;
-       masks[1] = ctx->i32_1;
-       for (unsigned i = 2; i < count; i++)
-               masks[i] = LLVMConstInt(ctx->i32, i, false);
+   LLVMValueRef masks[MAX2(count, 2)];
+   masks[0] = ctx->i32_0;
+   masks[1] = ctx->i32_1;
+   for (unsigned i = 2; i < count; i++)
+      masks[i] = LLVMConstInt(ctx->i32, i, false);
 
-       if (count == 1)
-               return LLVMBuildExtractElement(ctx->builder, value, masks[0],
-                                              "");
+   if (count == 1)
+      return LLVMBuildExtractElement(ctx->builder, value, masks[0], "");
 
-       LLVMValueRef swizzle = LLVMConstVector(masks, count);
-       return LLVMBuildShuffleVector(ctx->builder, value, value, swizzle, "");
+   LLVMValueRef swizzle = LLVMConstVector(masks, count);
+   return LLVMBuildShuffleVector(ctx->builder, value, value, swizzle, "");
 }
 
-LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param,
-                            unsigned rshift, unsigned bitwidth)
+LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param, unsigned rshift,
+                             unsigned bitwidth)
 {
-       LLVMValueRef value = param;
-       if (rshift)
-               value = LLVMBuildLShr(ctx->builder, value,
-                                     LLVMConstInt(ctx->i32, rshift, false), "");
+   LLVMValueRef value = param;
+   if (rshift)
+      value = LLVMBuildLShr(ctx->builder, value, LLVMConstInt(ctx->i32, rshift, false), "");
 
-       if (rshift + bitwidth < 32) {
-               unsigned mask = (1 << bitwidth) - 1;
-               value = LLVMBuildAnd(ctx->builder, value,
-                                    LLVMConstInt(ctx->i32, mask, false), "");
-       }
-       return value;
+   if (rshift + bitwidth < 32) {
+      unsigned mask = (1 << bitwidth) - 1;
+      value = LLVMBuildAnd(ctx->builder, value, LLVMConstInt(ctx->i32, mask, false), "");
+   }
+   return value;
 }
 
 /* Adjust the sample index according to FMASK.
@@ -3561,108 +3184,96 @@ LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param,
  * The sample index should be adjusted as follows:
  *   addr[sample_index] = (fmask >> (addr[sample_index] * 4)) & 0xF;
  */
-void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask,
-                             LLVMValueRef *addr, bool is_array_tex)
-{
-       struct ac_image_args fmask_load = {};
-       fmask_load.opcode = ac_image_load;
-       fmask_load.resource = fmask;
-       fmask_load.dmask = 0xf;
-       fmask_load.dim = is_array_tex ? ac_image_2darray : ac_image_2d;
-       fmask_load.attributes = AC_FUNC_ATTR_READNONE;
-
-       fmask_load.coords[0] = addr[0];
-       fmask_load.coords[1] = addr[1];
-       if (is_array_tex)
-               fmask_load.coords[2] = addr[2];
-
-       LLVMValueRef fmask_value = ac_build_image_opcode(ac, &fmask_load);
-       fmask_value = LLVMBuildExtractElement(ac->builder, fmask_value,
-                                             ac->i32_0, "");
-
-       /* Apply the formula. */
-       unsigned sample_chan = is_array_tex ? 3 : 2;
-       LLVMValueRef final_sample;
-       final_sample = LLVMBuildMul(ac->builder, addr[sample_chan],
-                                   LLVMConstInt(ac->i32, 4, 0), "");
-       final_sample = LLVMBuildLShr(ac->builder, fmask_value, final_sample, "");
-       /* Mask the sample index by 0x7, because 0x8 means an unknown value
-        * with EQAA, so those will map to 0. */
-       final_sample = LLVMBuildAnd(ac->builder, final_sample,
-                                   LLVMConstInt(ac->i32, 0x7, 0), "");
-
-       /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
-        * resource descriptor is 0 (invalid).
-        */
-       LLVMValueRef tmp;
-       tmp = LLVMBuildBitCast(ac->builder, fmask, ac->v8i32, "");
-       tmp = LLVMBuildExtractElement(ac->builder, tmp, ac->i32_1, "");
-       tmp = LLVMBuildICmp(ac->builder, LLVMIntNE, tmp, ac->i32_0, "");
-
-       /* Replace the MSAA sample index. */
-       addr[sample_chan] = LLVMBuildSelect(ac->builder, tmp, final_sample,
-                                           addr[sample_chan], "");
-}
-
-static LLVMValueRef
-_ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src,
-                 LLVMValueRef lane, bool with_opt_barrier)
-{
-       LLVMTypeRef type = LLVMTypeOf(src);
-       LLVMValueRef result;
-
-       if (with_opt_barrier)
-               ac_build_optimization_barrier(ctx, &src);
-
-       src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
-       if (lane)
-               lane = LLVMBuildZExt(ctx->builder, lane, ctx->i32, "");
-
-       result = ac_build_intrinsic(ctx,
-                       lane == NULL ? "llvm.amdgcn.readfirstlane" : "llvm.amdgcn.readlane",
-                       ctx->i32, (LLVMValueRef []) { src, lane },
-                       lane == NULL ? 1 : 2,
-                       AC_FUNC_ATTR_READNONE |
-                       AC_FUNC_ATTR_CONVERGENT);
-
-       return LLVMBuildTrunc(ctx->builder, result, type, "");
-}
-
-static LLVMValueRef
-ac_build_readlane_common(struct ac_llvm_context *ctx,
-                        LLVMValueRef src, LLVMValueRef lane,
-                        bool with_opt_barrier)
-{
-       LLVMTypeRef src_type = LLVMTypeOf(src);
-       src = ac_to_integer(ctx, src);
-       unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
-       LLVMValueRef ret;
-
-       if (bits > 32) {
-               assert(bits % 32 == 0);
-               LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
-               LLVMValueRef src_vector =
-                       LLVMBuildBitCast(ctx->builder, src, vec_type, "");
-               ret = LLVMGetUndef(vec_type);
-               for (unsigned i = 0; i < bits / 32; i++) {
-                       LLVMValueRef ret_comp;
-
-                       src = LLVMBuildExtractElement(ctx->builder, src_vector,
-                                               LLVMConstInt(ctx->i32, i, 0), "");
-
-                       ret_comp = _ac_build_readlane(ctx, src, lane,
-                                                     with_opt_barrier);
-
-                       ret = LLVMBuildInsertElement(ctx->builder, ret, ret_comp,
-                                               LLVMConstInt(ctx->i32, i, 0), "");
-               }
-       } else {
-               ret = _ac_build_readlane(ctx, src, lane, with_opt_barrier);
-       }
-
-       if (LLVMGetTypeKind(src_type) == LLVMPointerTypeKind)
-               return LLVMBuildIntToPtr(ctx->builder, ret, src_type, "");
-       return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
+void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask, LLVMValueRef *addr,
+                              bool is_array_tex)
+{
+   struct ac_image_args fmask_load = {};
+   fmask_load.opcode = ac_image_load;
+   fmask_load.resource = fmask;
+   fmask_load.dmask = 0xf;
+   fmask_load.dim = is_array_tex ? ac_image_2darray : ac_image_2d;
+   fmask_load.attributes = AC_FUNC_ATTR_READNONE;
+
+   fmask_load.coords[0] = addr[0];
+   fmask_load.coords[1] = addr[1];
+   if (is_array_tex)
+      fmask_load.coords[2] = addr[2];
+
+   LLVMValueRef fmask_value = ac_build_image_opcode(ac, &fmask_load);
+   fmask_value = LLVMBuildExtractElement(ac->builder, fmask_value, ac->i32_0, "");
+
+   /* Apply the formula. */
+   unsigned sample_chan = is_array_tex ? 3 : 2;
+   LLVMValueRef final_sample;
+   final_sample = LLVMBuildMul(ac->builder, addr[sample_chan], LLVMConstInt(ac->i32, 4, 0), "");
+   final_sample = LLVMBuildLShr(ac->builder, fmask_value, final_sample, "");
+   /* Mask the sample index by 0x7, because 0x8 means an unknown value
+    * with EQAA, so those will map to 0. */
+   final_sample = LLVMBuildAnd(ac->builder, final_sample, LLVMConstInt(ac->i32, 0x7, 0), "");
+
+   /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
+    * resource descriptor is 0 (invalid).
+    */
+   LLVMValueRef tmp;
+   tmp = LLVMBuildBitCast(ac->builder, fmask, ac->v8i32, "");
+   tmp = LLVMBuildExtractElement(ac->builder, tmp, ac->i32_1, "");
+   tmp = LLVMBuildICmp(ac->builder, LLVMIntNE, tmp, ac->i32_0, "");
+
+   /* Replace the MSAA sample index. */
+   addr[sample_chan] = LLVMBuildSelect(ac->builder, tmp, final_sample, addr[sample_chan], "");
+}
+
+static LLVMValueRef _ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src,
+                                       LLVMValueRef lane, bool with_opt_barrier)
+{
+   LLVMTypeRef type = LLVMTypeOf(src);
+   LLVMValueRef result;
+
+   if (with_opt_barrier)
+      ac_build_optimization_barrier(ctx, &src);
+
+   src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
+   if (lane)
+      lane = LLVMBuildZExt(ctx->builder, lane, ctx->i32, "");
+
+   result =
+      ac_build_intrinsic(ctx, lane == NULL ? "llvm.amdgcn.readfirstlane" : "llvm.amdgcn.readlane",
+                         ctx->i32, (LLVMValueRef[]){src, lane}, lane == NULL ? 1 : 2,
+                         AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
+
+   return LLVMBuildTrunc(ctx->builder, result, type, "");
+}
+
+static LLVMValueRef ac_build_readlane_common(struct ac_llvm_context *ctx, LLVMValueRef src,
+                                             LLVMValueRef lane, bool with_opt_barrier)
+{
+   LLVMTypeRef src_type = LLVMTypeOf(src);
+   src = ac_to_integer(ctx, src);
+   unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
+   LLVMValueRef ret;
+
+   if (bits > 32) {
+      assert(bits % 32 == 0);
+      LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
+      LLVMValueRef src_vector = LLVMBuildBitCast(ctx->builder, src, vec_type, "");
+      ret = LLVMGetUndef(vec_type);
+      for (unsigned i = 0; i < bits / 32; i++) {
+         LLVMValueRef ret_comp;
+
+         src = LLVMBuildExtractElement(ctx->builder, src_vector, LLVMConstInt(ctx->i32, i, 0), "");
+
+         ret_comp = _ac_build_readlane(ctx, src, lane, with_opt_barrier);
+
+         ret =
+            LLVMBuildInsertElement(ctx->builder, ret, ret_comp, LLVMConstInt(ctx->i32, i, 0), "");
+      }
+   } else {
+      ret = _ac_build_readlane(ctx, src, lane, with_opt_barrier);
+   }
+
+   if (LLVMGetTypeKind(src_type) == LLVMPointerTypeKind)
+      return LLVMBuildIntToPtr(ctx->builder, ret, src_type, "");
+   return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
 }
 
 /**
@@ -3676,429 +3287,433 @@ ac_build_readlane_common(struct ac_llvm_context *ctx,
  * @param lane - id of the lane or NULL for the first active lane
  * @return value of the lane
  */
-LLVMValueRef ac_build_readlane_no_opt_barrier(struct ac_llvm_context *ctx,
-                                             LLVMValueRef src, LLVMValueRef lane)
-{
-       return ac_build_readlane_common(ctx, src, lane, false);
-}
-
-
-LLVMValueRef
-ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane)
-{
-       return ac_build_readlane_common(ctx, src, lane, true);
-}
-
-LLVMValueRef
-ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value, LLVMValueRef lane)
-{
-       return ac_build_intrinsic(ctx, "llvm.amdgcn.writelane", ctx->i32,
-                                 (LLVMValueRef []) {value, lane, src}, 3,
-                                 AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
-}
-
-LLVMValueRef
-ac_build_mbcnt(struct ac_llvm_context *ctx, LLVMValueRef mask)
-{
-       if (ctx->wave_size == 32) {
-               return ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.lo", ctx->i32,
-                                         (LLVMValueRef []) { mask, ctx->i32_0 },
-                                         2, AC_FUNC_ATTR_READNONE);
-       }
-       LLVMValueRef mask_vec = LLVMBuildBitCast(ctx->builder, mask, ctx->v2i32, "");
-       LLVMValueRef mask_lo = LLVMBuildExtractElement(ctx->builder, mask_vec,
-                                                      ctx->i32_0, "");
-       LLVMValueRef mask_hi = LLVMBuildExtractElement(ctx->builder, mask_vec,
-                                                      ctx->i32_1, "");
-       LLVMValueRef val =
-               ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.lo", ctx->i32,
-                                  (LLVMValueRef []) { mask_lo, ctx->i32_0 },
-                                  2, AC_FUNC_ATTR_READNONE);
-       val = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi", ctx->i32,
-                                (LLVMValueRef []) { mask_hi, val },
-                                2, AC_FUNC_ATTR_READNONE);
-       return val;
-}
-
-enum dpp_ctrl {
-       _dpp_quad_perm = 0x000,
-       _dpp_row_sl = 0x100,
-       _dpp_row_sr = 0x110,
-       _dpp_row_rr = 0x120,
-       dpp_wf_sl1 = 0x130,
-       dpp_wf_rl1 = 0x134,
-       dpp_wf_sr1 = 0x138,
-       dpp_wf_rr1 = 0x13C,
-       dpp_row_mirror = 0x140,
-       dpp_row_half_mirror = 0x141,
-       dpp_row_bcast15 = 0x142,
-       dpp_row_bcast31 = 0x143
+LLVMValueRef ac_build_readlane_no_opt_barrier(struct ac_llvm_context *ctx, LLVMValueRef src,
+                                              LLVMValueRef lane)
+{
+   return ac_build_readlane_common(ctx, src, lane, false);
+}
+
+LLVMValueRef ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane)
+{
+   return ac_build_readlane_common(ctx, src, lane, true);
+}
+
+LLVMValueRef ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value,
+                                LLVMValueRef lane)
+{
+   return ac_build_intrinsic(ctx, "llvm.amdgcn.writelane", ctx->i32,
+                             (LLVMValueRef[]){value, lane, src}, 3,
+                             AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
+}
+
+LLVMValueRef ac_build_mbcnt(struct ac_llvm_context *ctx, LLVMValueRef mask)
+{
+   if (ctx->wave_size == 32) {
+      return ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.lo", ctx->i32,
+                                (LLVMValueRef[]){mask, ctx->i32_0}, 2, AC_FUNC_ATTR_READNONE);
+   }
+   LLVMValueRef mask_vec = LLVMBuildBitCast(ctx->builder, mask, ctx->v2i32, "");
+   LLVMValueRef mask_lo = LLVMBuildExtractElement(ctx->builder, mask_vec, ctx->i32_0, "");
+   LLVMValueRef mask_hi = LLVMBuildExtractElement(ctx->builder, mask_vec, ctx->i32_1, "");
+   LLVMValueRef val =
+      ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.lo", ctx->i32,
+                         (LLVMValueRef[]){mask_lo, ctx->i32_0}, 2, AC_FUNC_ATTR_READNONE);
+   val = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi", ctx->i32, (LLVMValueRef[]){mask_hi, val},
+                            2, AC_FUNC_ATTR_READNONE);
+   return val;
+}
+
+enum dpp_ctrl
+{
+   _dpp_quad_perm = 0x000,
+   _dpp_row_sl = 0x100,
+   _dpp_row_sr = 0x110,
+   _dpp_row_rr = 0x120,
+   dpp_wf_sl1 = 0x130,
+   dpp_wf_rl1 = 0x134,
+   dpp_wf_sr1 = 0x138,
+   dpp_wf_rr1 = 0x13C,
+   dpp_row_mirror = 0x140,
+   dpp_row_half_mirror = 0x141,
+   dpp_row_bcast15 = 0x142,
+   dpp_row_bcast31 = 0x143
 };
 
-static inline enum dpp_ctrl
-dpp_quad_perm(unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3)
-{
-       assert(lane0 < 4 && lane1 < 4 && lane2 < 4 && lane3 < 4);
-       return _dpp_quad_perm | lane0 | (lane1 << 2) | (lane2 << 4) | (lane3 << 6);
-}
-
-static inline enum dpp_ctrl
-dpp_row_sl(unsigned amount)
-{
-       assert(amount > 0 && amount < 16);
-       return _dpp_row_sl | amount;
-}
-
-static inline enum dpp_ctrl
-dpp_row_sr(unsigned amount)
-{
-       assert(amount > 0 && amount < 16);
-       return _dpp_row_sr | amount;
-}
-
-static LLVMValueRef
-_ac_build_dpp(struct ac_llvm_context *ctx, LLVMValueRef old, LLVMValueRef src,
-             enum dpp_ctrl dpp_ctrl, unsigned row_mask, unsigned bank_mask,
-             bool bound_ctrl)
-{
-       LLVMTypeRef type = LLVMTypeOf(src);
-       LLVMValueRef res;
-
-       old = LLVMBuildZExt(ctx->builder, old, ctx->i32, "");
-       src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
-
-       res = ac_build_intrinsic(ctx, "llvm.amdgcn.update.dpp.i32", ctx->i32,
-                                (LLVMValueRef[]) {
-                                       old, src,
-                                       LLVMConstInt(ctx->i32, dpp_ctrl, 0),
-                                       LLVMConstInt(ctx->i32, row_mask, 0),
-                                       LLVMConstInt(ctx->i32, bank_mask, 0),
-                                       LLVMConstInt(ctx->i1, bound_ctrl, 0) },
-                                6, AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
-
-       return LLVMBuildTrunc(ctx->builder, res, type, "");
-}
-
-static LLVMValueRef
-ac_build_dpp(struct ac_llvm_context *ctx, LLVMValueRef old, LLVMValueRef src,
-            enum dpp_ctrl dpp_ctrl, unsigned row_mask, unsigned bank_mask,
-            bool bound_ctrl)
-{
-       LLVMTypeRef src_type = LLVMTypeOf(src);
-       src = ac_to_integer(ctx, src);
-       old = ac_to_integer(ctx, old);
-       unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
-       LLVMValueRef ret;
-       if (bits > 32) {
-               assert(bits % 32 == 0);
-               LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
-               LLVMValueRef src_vector =
-                       LLVMBuildBitCast(ctx->builder, src, vec_type, "");
-               LLVMValueRef old_vector =
-                       LLVMBuildBitCast(ctx->builder, old, vec_type, "");
-               ret = LLVMGetUndef(vec_type);
-               for (unsigned i = 0; i < bits / 32; i++) {
-                       src = LLVMBuildExtractElement(ctx->builder, src_vector,
-                                                     LLVMConstInt(ctx->i32, i,
-                                                                  0), "");
-                       old = LLVMBuildExtractElement(ctx->builder, old_vector,
-                                                     LLVMConstInt(ctx->i32, i,
-                                                                  0), "");
-                       LLVMValueRef ret_comp = _ac_build_dpp(ctx, old, src,
-                                                             dpp_ctrl,
-                                                             row_mask,
-                                                             bank_mask,
-                                                             bound_ctrl);
-                       ret = LLVMBuildInsertElement(ctx->builder, ret,
-                                                    ret_comp,
-                                                    LLVMConstInt(ctx->i32, i,
-                                                                 0), "");
-               }
-       } else {
-               ret = _ac_build_dpp(ctx, old, src, dpp_ctrl, row_mask,
-                                   bank_mask, bound_ctrl);
-       }
-       return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
-}
-
-static LLVMValueRef
-_ac_build_permlane16(struct ac_llvm_context *ctx, LLVMValueRef src, uint64_t sel,
-                    bool exchange_rows, bool bound_ctrl)
-{
-       LLVMTypeRef type = LLVMTypeOf(src);
-       LLVMValueRef result;
-
-       src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
-
-       LLVMValueRef args[6] = {
-               src,
-               src,
-               LLVMConstInt(ctx->i32, sel, false),
-               LLVMConstInt(ctx->i32, sel >> 32, false),
-               ctx->i1true, /* fi */
-               bound_ctrl ? ctx->i1true : ctx->i1false,
-       };
-
-       result = ac_build_intrinsic(ctx, exchange_rows ? "llvm.amdgcn.permlanex16"
-                                                      : "llvm.amdgcn.permlane16",
-                                   ctx->i32, args, 6,
-                                   AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
-
-       return LLVMBuildTrunc(ctx->builder, result, type, "");
-}
-
-static LLVMValueRef
-ac_build_permlane16(struct ac_llvm_context *ctx, LLVMValueRef src, uint64_t sel,
-                   bool exchange_rows, bool bound_ctrl)
-{
-       LLVMTypeRef src_type = LLVMTypeOf(src);
-       src = ac_to_integer(ctx, src);
-       unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
-       LLVMValueRef ret;
-       if (bits > 32) {
-               assert(bits % 32 == 0);
-               LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
-               LLVMValueRef src_vector =
-                       LLVMBuildBitCast(ctx->builder, src, vec_type, "");
-               ret = LLVMGetUndef(vec_type);
-               for (unsigned i = 0; i < bits / 32; i++) {
-                       src = LLVMBuildExtractElement(ctx->builder, src_vector,
-                                                     LLVMConstInt(ctx->i32, i,
-                                                                  0), "");
-                       LLVMValueRef ret_comp =
-                               _ac_build_permlane16(ctx, src, sel,
-                                                    exchange_rows,
-                                                    bound_ctrl);
-                       ret = LLVMBuildInsertElement(ctx->builder, ret,
-                                                    ret_comp,
-                                                    LLVMConstInt(ctx->i32, i,
-                                                                 0), "");
-               }
-       } else {
-               ret = _ac_build_permlane16(ctx, src, sel, exchange_rows,
-                                          bound_ctrl);
-       }
-       return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
-}
-
-static inline unsigned
-ds_pattern_bitmode(unsigned and_mask, unsigned or_mask, unsigned xor_mask)
-{
-       assert(and_mask < 32 && or_mask < 32 && xor_mask < 32);
-       return and_mask | (or_mask << 5) | (xor_mask << 10);
-}
-
-static LLVMValueRef
-_ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask)
-{
-       LLVMTypeRef src_type = LLVMTypeOf(src);
-       LLVMValueRef ret;
-
-       src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
-
-       ret = ac_build_intrinsic(ctx, "llvm.amdgcn.ds.swizzle", ctx->i32,
-                                (LLVMValueRef []) {
-                                       src, LLVMConstInt(ctx->i32, mask, 0) },
-                                2, AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
-
-       return LLVMBuildTrunc(ctx->builder, ret, src_type, "");
-}
-
-LLVMValueRef
-ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask)
-{
-       LLVMTypeRef src_type = LLVMTypeOf(src);
-       src = ac_to_integer(ctx, src);
-       unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
-       LLVMValueRef ret;
-       if (bits > 32) {
-               assert(bits % 32 == 0);
-               LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
-               LLVMValueRef src_vector =
-                       LLVMBuildBitCast(ctx->builder, src, vec_type, "");
-               ret = LLVMGetUndef(vec_type);
-               for (unsigned i = 0; i < bits / 32; i++) {
-                       src = LLVMBuildExtractElement(ctx->builder, src_vector,
-                                                     LLVMConstInt(ctx->i32, i,
-                                                                  0), "");
-                       LLVMValueRef ret_comp = _ac_build_ds_swizzle(ctx, src,
-                                                                    mask);
-                       ret = LLVMBuildInsertElement(ctx->builder, ret,
-                                                    ret_comp,
-                                                    LLVMConstInt(ctx->i32, i,
-                                                                 0), "");
-               }
-       } else {
-               ret = _ac_build_ds_swizzle(ctx, src, mask);
-       }
-       return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
-}
-
-static LLVMValueRef
-ac_build_wwm(struct ac_llvm_context *ctx, LLVMValueRef src)
-{
-       LLVMTypeRef src_type = LLVMTypeOf(src);
-       unsigned bitsize = ac_get_elem_bits(ctx, src_type);
-       char name[32], type[8];
-       LLVMValueRef ret;
-
-       src = ac_to_integer(ctx, src);
-
-       if (bitsize < 32)
-               src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
-
-       ac_build_type_name_for_intr(LLVMTypeOf(src), type, sizeof(type));
-       snprintf(name, sizeof(name), "llvm.amdgcn.wwm.%s", type);
-       ret = ac_build_intrinsic(ctx, name, LLVMTypeOf(src),
-                                (LLVMValueRef []) { src }, 1,
-                                AC_FUNC_ATTR_READNONE);
-
-       if (bitsize < 32)
-               ret = LLVMBuildTrunc(ctx->builder, ret,
-                                    ac_to_integer_type(ctx, src_type), "");
-
-       return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
-}
-
-static LLVMValueRef
-ac_build_set_inactive(struct ac_llvm_context *ctx, LLVMValueRef src,
-                     LLVMValueRef inactive)
-{
-       char name[33], type[8];
-       LLVMTypeRef src_type = LLVMTypeOf(src);
-       unsigned bitsize = ac_get_elem_bits(ctx, src_type);
-       src = ac_to_integer(ctx, src);
-       inactive = ac_to_integer(ctx, inactive);
-
-       if (bitsize < 32) {
-               src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
-               inactive = LLVMBuildZExt(ctx->builder, inactive, ctx->i32, "");
-       }
-
-       ac_build_type_name_for_intr(LLVMTypeOf(src), type, sizeof(type));
-       snprintf(name, sizeof(name), "llvm.amdgcn.set.inactive.%s", type);
-       LLVMValueRef ret =
-               ac_build_intrinsic(ctx, name,
-                                       LLVMTypeOf(src), (LLVMValueRef []) {
-                                       src, inactive }, 2,
-                                       AC_FUNC_ATTR_READNONE |
-                                       AC_FUNC_ATTR_CONVERGENT);
-       if (bitsize < 32)
-               ret = LLVMBuildTrunc(ctx->builder, ret, src_type, "");
-
-       return ret;
-}
-
-static LLVMValueRef
-get_reduction_identity(struct ac_llvm_context *ctx, nir_op op, unsigned type_size)
-{
-       if (type_size == 1) {
-               switch (op) {
-               case nir_op_iadd: return ctx->i8_0;
-               case nir_op_imul: return ctx->i8_1;
-               case nir_op_imin: return LLVMConstInt(ctx->i8, INT8_MAX, 0);
-               case nir_op_umin: return LLVMConstInt(ctx->i8, UINT8_MAX, 0);
-               case nir_op_imax: return LLVMConstInt(ctx->i8, INT8_MIN, 0);
-               case nir_op_umax: return ctx->i8_0;
-               case nir_op_iand: return LLVMConstInt(ctx->i8, -1, 0);
-               case nir_op_ior: return ctx->i8_0;
-               case nir_op_ixor: return ctx->i8_0;
-               default:
-                       unreachable("bad reduction intrinsic");
-               }
-       } else if (type_size == 2) {
-               switch (op) {
-               case nir_op_iadd: return ctx->i16_0;
-               case nir_op_fadd: return ctx->f16_0;
-               case nir_op_imul: return ctx->i16_1;
-               case nir_op_fmul: return ctx->f16_1;
-               case nir_op_imin: return LLVMConstInt(ctx->i16, INT16_MAX, 0);
-               case nir_op_umin: return LLVMConstInt(ctx->i16, UINT16_MAX, 0);
-               case nir_op_fmin: return LLVMConstReal(ctx->f16, INFINITY);
-               case nir_op_imax: return LLVMConstInt(ctx->i16, INT16_MIN, 0);
-               case nir_op_umax: return ctx->i16_0;
-               case nir_op_fmax: return LLVMConstReal(ctx->f16, -INFINITY);
-               case nir_op_iand: return LLVMConstInt(ctx->i16, -1, 0);
-               case nir_op_ior: return ctx->i16_0;
-               case nir_op_ixor: return ctx->i16_0;
-               default:
-                       unreachable("bad reduction intrinsic");
-               }
-       } else if (type_size == 4) {
-               switch (op) {
-               case nir_op_iadd: return ctx->i32_0;
-               case nir_op_fadd: return ctx->f32_0;
-               case nir_op_imul: return ctx->i32_1;
-               case nir_op_fmul: return ctx->f32_1;
-               case nir_op_imin: return LLVMConstInt(ctx->i32, INT32_MAX, 0);
-               case nir_op_umin: return LLVMConstInt(ctx->i32, UINT32_MAX, 0);
-               case nir_op_fmin: return LLVMConstReal(ctx->f32, INFINITY);
-               case nir_op_imax: return LLVMConstInt(ctx->i32, INT32_MIN, 0);
-               case nir_op_umax: return ctx->i32_0;
-               case nir_op_fmax: return LLVMConstReal(ctx->f32, -INFINITY);
-               case nir_op_iand: return LLVMConstInt(ctx->i32, -1, 0);
-               case nir_op_ior: return ctx->i32_0;
-               case nir_op_ixor: return ctx->i32_0;
-               default:
-                       unreachable("bad reduction intrinsic");
-               }
-       } else { /* type_size == 64bit */
-               switch (op) {
-               case nir_op_iadd: return ctx->i64_0;
-               case nir_op_fadd: return ctx->f64_0;
-               case nir_op_imul: return ctx->i64_1;
-               case nir_op_fmul: return ctx->f64_1;
-               case nir_op_imin: return LLVMConstInt(ctx->i64, INT64_MAX, 0);
-               case nir_op_umin: return LLVMConstInt(ctx->i64, UINT64_MAX, 0);
-               case nir_op_fmin: return LLVMConstReal(ctx->f64, INFINITY);
-               case nir_op_imax: return LLVMConstInt(ctx->i64, INT64_MIN, 0);
-               case nir_op_umax: return ctx->i64_0;
-               case nir_op_fmax: return LLVMConstReal(ctx->f64, -INFINITY);
-               case nir_op_iand: return LLVMConstInt(ctx->i64, -1, 0);
-               case nir_op_ior: return ctx->i64_0;
-               case nir_op_ixor: return ctx->i64_0;
-               default:
-                       unreachable("bad reduction intrinsic");
-               }
-       }
-}
-
-static LLVMValueRef
-ac_build_alu_op(struct ac_llvm_context *ctx, LLVMValueRef lhs, LLVMValueRef rhs, nir_op op)
-{
-       bool _64bit = ac_get_type_size(LLVMTypeOf(lhs)) == 8;
-       bool _32bit = ac_get_type_size(LLVMTypeOf(lhs)) == 4;
-       switch (op) {
-       case nir_op_iadd: return LLVMBuildAdd(ctx->builder, lhs, rhs, "");
-       case nir_op_fadd: return LLVMBuildFAdd(ctx->builder, lhs, rhs, "");
-       case nir_op_imul: return LLVMBuildMul(ctx->builder, lhs, rhs, "");
-       case nir_op_fmul: return LLVMBuildFMul(ctx->builder, lhs, rhs, "");
-       case nir_op_imin: return LLVMBuildSelect(ctx->builder,
-                                       LLVMBuildICmp(ctx->builder, LLVMIntSLT, lhs, rhs, ""),
-                                       lhs, rhs, "");
-       case nir_op_umin: return LLVMBuildSelect(ctx->builder,
-                                       LLVMBuildICmp(ctx->builder, LLVMIntULT, lhs, rhs, ""),
-                                       lhs, rhs, "");
-       case nir_op_fmin: return ac_build_intrinsic(ctx,
-                                       _64bit ? "llvm.minnum.f64" : _32bit ? "llvm.minnum.f32" : "llvm.minnum.f16",
-                                       _64bit ? ctx->f64 : _32bit ? ctx->f32 : ctx->f16,
-                                       (LLVMValueRef[]){lhs, rhs}, 2, AC_FUNC_ATTR_READNONE);
-       case nir_op_imax: return LLVMBuildSelect(ctx->builder,
-                                       LLVMBuildICmp(ctx->builder, LLVMIntSGT, lhs, rhs, ""),
-                                       lhs, rhs, "");
-       case nir_op_umax: return LLVMBuildSelect(ctx->builder,
-                                       LLVMBuildICmp(ctx->builder, LLVMIntUGT, lhs, rhs, ""),
-                                       lhs, rhs, "");
-       case nir_op_fmax: return ac_build_intrinsic(ctx,
-                                       _64bit ? "llvm.maxnum.f64" : _32bit ? "llvm.maxnum.f32" : "llvm.maxnum.f16",
-                                       _64bit ? ctx->f64 : _32bit ? ctx->f32 : ctx->f16,
-                                       (LLVMValueRef[]){lhs, rhs}, 2, AC_FUNC_ATTR_READNONE);
-       case nir_op_iand: return LLVMBuildAnd(ctx->builder, lhs, rhs, "");
-       case nir_op_ior: return LLVMBuildOr(ctx->builder, lhs, rhs, "");
-       case nir_op_ixor: return LLVMBuildXor(ctx->builder, lhs, rhs, "");
-       default:
-               unreachable("bad reduction intrinsic");
-       }
+static inline enum dpp_ctrl dpp_quad_perm(unsigned lane0, unsigned lane1, unsigned lane2,
+                                          unsigned lane3)
+{
+   assert(lane0 < 4 && lane1 < 4 && lane2 < 4 && lane3 < 4);
+   return _dpp_quad_perm | lane0 | (lane1 << 2) | (lane2 << 4) | (lane3 << 6);
+}
+
+static inline enum dpp_ctrl dpp_row_sl(unsigned amount)
+{
+   assert(amount > 0 && amount < 16);
+   return _dpp_row_sl | amount;
+}
+
+static inline enum dpp_ctrl dpp_row_sr(unsigned amount)
+{
+   assert(amount > 0 && amount < 16);
+   return _dpp_row_sr | amount;
+}
+
+static LLVMValueRef _ac_build_dpp(struct ac_llvm_context *ctx, LLVMValueRef old, LLVMValueRef src,
+                                  enum dpp_ctrl dpp_ctrl, unsigned row_mask, unsigned bank_mask,
+                                  bool bound_ctrl)
+{
+   LLVMTypeRef type = LLVMTypeOf(src);
+   LLVMValueRef res;
+
+   old = LLVMBuildZExt(ctx->builder, old, ctx->i32, "");
+   src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
+
+   res = ac_build_intrinsic(
+      ctx, "llvm.amdgcn.update.dpp.i32", ctx->i32,
+      (LLVMValueRef[]){old, src, LLVMConstInt(ctx->i32, dpp_ctrl, 0),
+                       LLVMConstInt(ctx->i32, row_mask, 0), LLVMConstInt(ctx->i32, bank_mask, 0),
+                       LLVMConstInt(ctx->i1, bound_ctrl, 0)},
+      6, AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
+
+   return LLVMBuildTrunc(ctx->builder, res, type, "");
+}
+
+static LLVMValueRef ac_build_dpp(struct ac_llvm_context *ctx, LLVMValueRef old, LLVMValueRef src,
+                                 enum dpp_ctrl dpp_ctrl, unsigned row_mask, unsigned bank_mask,
+                                 bool bound_ctrl)
+{
+   LLVMTypeRef src_type = LLVMTypeOf(src);
+   src = ac_to_integer(ctx, src);
+   old = ac_to_integer(ctx, old);
+   unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
+   LLVMValueRef ret;
+   if (bits > 32) {
+      assert(bits % 32 == 0);
+      LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
+      LLVMValueRef src_vector = LLVMBuildBitCast(ctx->builder, src, vec_type, "");
+      LLVMValueRef old_vector = LLVMBuildBitCast(ctx->builder, old, vec_type, "");
+      ret = LLVMGetUndef(vec_type);
+      for (unsigned i = 0; i < bits / 32; i++) {
+         src = LLVMBuildExtractElement(ctx->builder, src_vector, LLVMConstInt(ctx->i32, i, 0), "");
+         old = LLVMBuildExtractElement(ctx->builder, old_vector, LLVMConstInt(ctx->i32, i, 0), "");
+         LLVMValueRef ret_comp =
+            _ac_build_dpp(ctx, old, src, dpp_ctrl, row_mask, bank_mask, bound_ctrl);
+         ret =
+            LLVMBuildInsertElement(ctx->builder, ret, ret_comp, LLVMConstInt(ctx->i32, i, 0), "");
+      }
+   } else {
+      ret = _ac_build_dpp(ctx, old, src, dpp_ctrl, row_mask, bank_mask, bound_ctrl);
+   }
+   return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
+}
+
+static LLVMValueRef _ac_build_permlane16(struct ac_llvm_context *ctx, LLVMValueRef src,
+                                         uint64_t sel, bool exchange_rows, bool bound_ctrl)
+{
+   LLVMTypeRef type = LLVMTypeOf(src);
+   LLVMValueRef result;
+
+   src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
+
+   LLVMValueRef args[6] = {
+      src,
+      src,
+      LLVMConstInt(ctx->i32, sel, false),
+      LLVMConstInt(ctx->i32, sel >> 32, false),
+      ctx->i1true, /* fi */
+      bound_ctrl ? ctx->i1true : ctx->i1false,
+   };
+
+   result =
+      ac_build_intrinsic(ctx, exchange_rows ? "llvm.amdgcn.permlanex16" : "llvm.amdgcn.permlane16",
+                         ctx->i32, args, 6, AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
+
+   return LLVMBuildTrunc(ctx->builder, result, type, "");
+}
+
+static LLVMValueRef ac_build_permlane16(struct ac_llvm_context *ctx, LLVMValueRef src, uint64_t sel,
+                                        bool exchange_rows, bool bound_ctrl)
+{
+   LLVMTypeRef src_type = LLVMTypeOf(src);
+   src = ac_to_integer(ctx, src);
+   unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
+   LLVMValueRef ret;
+   if (bits > 32) {
+      assert(bits % 32 == 0);
+      LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
+      LLVMValueRef src_vector = LLVMBuildBitCast(ctx->builder, src, vec_type, "");
+      ret = LLVMGetUndef(vec_type);
+      for (unsigned i = 0; i < bits / 32; i++) {
+         src = LLVMBuildExtractElement(ctx->builder, src_vector, LLVMConstInt(ctx->i32, i, 0), "");
+         LLVMValueRef ret_comp = _ac_build_permlane16(ctx, src, sel, exchange_rows, bound_ctrl);
+         ret =
+            LLVMBuildInsertElement(ctx->builder, ret, ret_comp, LLVMConstInt(ctx->i32, i, 0), "");
+      }
+   } else {
+      ret = _ac_build_permlane16(ctx, src, sel, exchange_rows, bound_ctrl);
+   }
+   return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
+}
+
+static inline unsigned ds_pattern_bitmode(unsigned and_mask, unsigned or_mask, unsigned xor_mask)
+{
+   assert(and_mask < 32 && or_mask < 32 && xor_mask < 32);
+   return and_mask | (or_mask << 5) | (xor_mask << 10);
+}
+
+static LLVMValueRef _ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src,
+                                         unsigned mask)
+{
+   LLVMTypeRef src_type = LLVMTypeOf(src);
+   LLVMValueRef ret;
+
+   src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
+
+   ret = ac_build_intrinsic(ctx, "llvm.amdgcn.ds.swizzle", ctx->i32,
+                            (LLVMValueRef[]){src, LLVMConstInt(ctx->i32, mask, 0)}, 2,
+                            AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
+
+   return LLVMBuildTrunc(ctx->builder, ret, src_type, "");
+}
+
+LLVMValueRef ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask)
+{
+   LLVMTypeRef src_type = LLVMTypeOf(src);
+   src = ac_to_integer(ctx, src);
+   unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
+   LLVMValueRef ret;
+   if (bits > 32) {
+      assert(bits % 32 == 0);
+      LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
+      LLVMValueRef src_vector = LLVMBuildBitCast(ctx->builder, src, vec_type, "");
+      ret = LLVMGetUndef(vec_type);
+      for (unsigned i = 0; i < bits / 32; i++) {
+         src = LLVMBuildExtractElement(ctx->builder, src_vector, LLVMConstInt(ctx->i32, i, 0), "");
+         LLVMValueRef ret_comp = _ac_build_ds_swizzle(ctx, src, mask);
+         ret =
+            LLVMBuildInsertElement(ctx->builder, ret, ret_comp, LLVMConstInt(ctx->i32, i, 0), "");
+      }
+   } else {
+      ret = _ac_build_ds_swizzle(ctx, src, mask);
+   }
+   return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
+}
+
+static LLVMValueRef ac_build_wwm(struct ac_llvm_context *ctx, LLVMValueRef src)
+{
+   LLVMTypeRef src_type = LLVMTypeOf(src);
+   unsigned bitsize = ac_get_elem_bits(ctx, src_type);
+   char name[32], type[8];
+   LLVMValueRef ret;
+
+   src = ac_to_integer(ctx, src);
+
+   if (bitsize < 32)
+      src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
+
+   ac_build_type_name_for_intr(LLVMTypeOf(src), type, sizeof(type));
+   snprintf(name, sizeof(name), "llvm.amdgcn.wwm.%s", type);
+   ret = ac_build_intrinsic(ctx, name, LLVMTypeOf(src), (LLVMValueRef[]){src}, 1,
+                            AC_FUNC_ATTR_READNONE);
+
+   if (bitsize < 32)
+      ret = LLVMBuildTrunc(ctx->builder, ret, ac_to_integer_type(ctx, src_type), "");
+
+   return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
+}
+
+static LLVMValueRef ac_build_set_inactive(struct ac_llvm_context *ctx, LLVMValueRef src,
+                                          LLVMValueRef inactive)
+{
+   char name[33], type[8];
+   LLVMTypeRef src_type = LLVMTypeOf(src);
+   unsigned bitsize = ac_get_elem_bits(ctx, src_type);
+   src = ac_to_integer(ctx, src);
+   inactive = ac_to_integer(ctx, inactive);
+
+   if (bitsize < 32) {
+      src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
+      inactive = LLVMBuildZExt(ctx->builder, inactive, ctx->i32, "");
+   }
+
+   ac_build_type_name_for_intr(LLVMTypeOf(src), type, sizeof(type));
+   snprintf(name, sizeof(name), "llvm.amdgcn.set.inactive.%s", type);
+   LLVMValueRef ret =
+      ac_build_intrinsic(ctx, name, LLVMTypeOf(src), (LLVMValueRef[]){src, inactive}, 2,
+                         AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
+   if (bitsize < 32)
+      ret = LLVMBuildTrunc(ctx->builder, ret, src_type, "");
+
+   return ret;
+}
+
+static LLVMValueRef get_reduction_identity(struct ac_llvm_context *ctx, nir_op op,
+                                           unsigned type_size)
+{
+   if (type_size == 1) {
+      switch (op) {
+      case nir_op_iadd:
+         return ctx->i8_0;
+      case nir_op_imul:
+         return ctx->i8_1;
+      case nir_op_imin:
+         return LLVMConstInt(ctx->i8, INT8_MAX, 0);
+      case nir_op_umin:
+         return LLVMConstInt(ctx->i8, UINT8_MAX, 0);
+      case nir_op_imax:
+         return LLVMConstInt(ctx->i8, INT8_MIN, 0);
+      case nir_op_umax:
+         return ctx->i8_0;
+      case nir_op_iand:
+         return LLVMConstInt(ctx->i8, -1, 0);
+      case nir_op_ior:
+         return ctx->i8_0;
+      case nir_op_ixor:
+         return ctx->i8_0;
+      default:
+         unreachable("bad reduction intrinsic");
+      }
+   } else if (type_size == 2) {
+      switch (op) {
+      case nir_op_iadd:
+         return ctx->i16_0;
+      case nir_op_fadd:
+         return ctx->f16_0;
+      case nir_op_imul:
+         return ctx->i16_1;
+      case nir_op_fmul:
+         return ctx->f16_1;
+      case nir_op_imin:
+         return LLVMConstInt(ctx->i16, INT16_MAX, 0);
+      case nir_op_umin:
+         return LLVMConstInt(ctx->i16, UINT16_MAX, 0);
+      case nir_op_fmin:
+         return LLVMConstReal(ctx->f16, INFINITY);
+      case nir_op_imax:
+         return LLVMConstInt(ctx->i16, INT16_MIN, 0);
+      case nir_op_umax:
+         return ctx->i16_0;
+      case nir_op_fmax:
+         return LLVMConstReal(ctx->f16, -INFINITY);
+      case nir_op_iand:
+         return LLVMConstInt(ctx->i16, -1, 0);
+      case nir_op_ior:
+         return ctx->i16_0;
+      case nir_op_ixor:
+         return ctx->i16_0;
+      default:
+         unreachable("bad reduction intrinsic");
+      }
+   } else if (type_size == 4) {
+      switch (op) {
+      case nir_op_iadd:
+         return ctx->i32_0;
+      case nir_op_fadd:
+         return ctx->f32_0;
+      case nir_op_imul:
+         return ctx->i32_1;
+      case nir_op_fmul:
+         return ctx->f32_1;
+      case nir_op_imin:
+         return LLVMConstInt(ctx->i32, INT32_MAX, 0);
+      case nir_op_umin:
+         return LLVMConstInt(ctx->i32, UINT32_MAX, 0);
+      case nir_op_fmin:
+         return LLVMConstReal(ctx->f32, INFINITY);
+      case nir_op_imax:
+         return LLVMConstInt(ctx->i32, INT32_MIN, 0);
+      case nir_op_umax:
+         return ctx->i32_0;
+      case nir_op_fmax:
+         return LLVMConstReal(ctx->f32, -INFINITY);
+      case nir_op_iand:
+         return LLVMConstInt(ctx->i32, -1, 0);
+      case nir_op_ior:
+         return ctx->i32_0;
+      case nir_op_ixor:
+         return ctx->i32_0;
+      default:
+         unreachable("bad reduction intrinsic");
+      }
+   } else { /* type_size == 64bit */
+      switch (op) {
+      case nir_op_iadd:
+         return ctx->i64_0;
+      case nir_op_fadd:
+         return ctx->f64_0;
+      case nir_op_imul:
+         return ctx->i64_1;
+      case nir_op_fmul:
+         return ctx->f64_1;
+      case nir_op_imin:
+         return LLVMConstInt(ctx->i64, INT64_MAX, 0);
+      case nir_op_umin:
+         return LLVMConstInt(ctx->i64, UINT64_MAX, 0);
+      case nir_op_fmin:
+         return LLVMConstReal(ctx->f64, INFINITY);
+      case nir_op_imax:
+         return LLVMConstInt(ctx->i64, INT64_MIN, 0);
+      case nir_op_umax:
+         return ctx->i64_0;
+      case nir_op_fmax:
+         return LLVMConstReal(ctx->f64, -INFINITY);
+      case nir_op_iand:
+         return LLVMConstInt(ctx->i64, -1, 0);
+      case nir_op_ior:
+         return ctx->i64_0;
+      case nir_op_ixor:
+         return ctx->i64_0;
+      default:
+         unreachable("bad reduction intrinsic");
+      }
+   }
+}
+
+static LLVMValueRef ac_build_alu_op(struct ac_llvm_context *ctx, LLVMValueRef lhs, LLVMValueRef rhs,
+                                    nir_op op)
+{
+   bool _64bit = ac_get_type_size(LLVMTypeOf(lhs)) == 8;
+   bool _32bit = ac_get_type_size(LLVMTypeOf(lhs)) == 4;
+   switch (op) {
+   case nir_op_iadd:
+      return LLVMBuildAdd(ctx->builder, lhs, rhs, "");
+   case nir_op_fadd:
+      return LLVMBuildFAdd(ctx->builder, lhs, rhs, "");
+   case nir_op_imul:
+      return LLVMBuildMul(ctx->builder, lhs, rhs, "");
+   case nir_op_fmul:
+      return LLVMBuildFMul(ctx->builder, lhs, rhs, "");
+   case nir_op_imin:
+      return LLVMBuildSelect(ctx->builder, LLVMBuildICmp(ctx->builder, LLVMIntSLT, lhs, rhs, ""),
+                             lhs, rhs, "");
+   case nir_op_umin:
+      return LLVMBuildSelect(ctx->builder, LLVMBuildICmp(ctx->builder, LLVMIntULT, lhs, rhs, ""),
+                             lhs, rhs, "");
+   case nir_op_fmin:
+      return ac_build_intrinsic(
+         ctx, _64bit ? "llvm.minnum.f64" : _32bit ? "llvm.minnum.f32" : "llvm.minnum.f16",
+         _64bit ? ctx->f64 : _32bit ? ctx->f32 : ctx->f16, (LLVMValueRef[]){lhs, rhs}, 2,
+         AC_FUNC_ATTR_READNONE);
+   case nir_op_imax:
+      return LLVMBuildSelect(ctx->builder, LLVMBuildICmp(ctx->builder, LLVMIntSGT, lhs, rhs, ""),
+                             lhs, rhs, "");
+   case nir_op_umax:
+      return LLVMBuildSelect(ctx->builder, LLVMBuildICmp(ctx->builder, LLVMIntUGT, lhs, rhs, ""),
+                             lhs, rhs, "");
+   case nir_op_fmax:
+      return ac_build_intrinsic(
+         ctx, _64bit ? "llvm.maxnum.f64" : _32bit ? "llvm.maxnum.f32" : "llvm.maxnum.f16",
+         _64bit ? ctx->f64 : _32bit ? ctx->f32 : ctx->f16, (LLVMValueRef[]){lhs, rhs}, 2,
+         AC_FUNC_ATTR_READNONE);
+   case nir_op_iand:
+      return LLVMBuildAnd(ctx->builder, lhs, rhs, "");
+   case nir_op_ior:
+      return LLVMBuildOr(ctx->builder, lhs, rhs, "");
+   case nir_op_ixor:
+      return LLVMBuildXor(ctx->builder, lhs, rhs, "");
+   default:
+      unreachable("bad reduction intrinsic");
+   }
 }
 
 /**
@@ -4108,297 +3723,292 @@ ac_build_alu_op(struct ac_llvm_context *ctx, LLVMValueRef lhs, LLVMValueRef rhs,
  *     prefix of this many threads
  * \return src, shifted 1 lane up, and identity shifted into lane 0.
  */
-static LLVMValueRef
-ac_wavefront_shift_right_1(struct ac_llvm_context *ctx, LLVMValueRef src,
-                           LLVMValueRef identity, unsigned maxprefix)
-{
-       if (ctx->chip_class >= GFX10) {
-               /* wavefront shift_right by 1 on GFX10 (emulate dpp_wf_sr1) */
-               LLVMValueRef active, tmp1, tmp2;
-               LLVMValueRef tid = ac_get_thread_id(ctx);
-
-               tmp1 = ac_build_dpp(ctx, identity, src, dpp_row_sr(1), 0xf, 0xf, false);
-
-               tmp2 = ac_build_permlane16(ctx, src, (uint64_t)~0, true, false);
-
-               if (maxprefix > 32) {
-                       active = LLVMBuildICmp(ctx->builder, LLVMIntEQ, tid,
-                                              LLVMConstInt(ctx->i32, 32, false), "");
-
-                       tmp2 = LLVMBuildSelect(ctx->builder, active,
-                                              ac_build_readlane(ctx, src,
-                                                                LLVMConstInt(ctx->i32, 31, false)),
-                                              tmp2, "");
-
-                       active = LLVMBuildOr(ctx->builder, active,
-                                            LLVMBuildICmp(ctx->builder, LLVMIntEQ,
-                                                          LLVMBuildAnd(ctx->builder, tid,
-                                                                       LLVMConstInt(ctx->i32, 0x1f, false), ""),
-                                                          LLVMConstInt(ctx->i32, 0x10, false), ""), "");
-                       return LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
-               } else if (maxprefix > 16) {
-                       active = LLVMBuildICmp(ctx->builder, LLVMIntEQ, tid,
-                                              LLVMConstInt(ctx->i32, 16, false), "");
-
-                       return LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
-               }
-       } else if (ctx->chip_class >= GFX8) {
-               return ac_build_dpp(ctx, identity, src, dpp_wf_sr1, 0xf, 0xf, false);
-       }
-
-       /* wavefront shift_right by 1 on SI/CI */
-       LLVMValueRef active, tmp1, tmp2;
-       LLVMValueRef tid = ac_get_thread_id(ctx);
-       tmp1 = ac_build_ds_swizzle(ctx, src, (1 << 15) | dpp_quad_perm(0, 0, 1, 2));
-       tmp2 = ac_build_ds_swizzle(ctx, src, ds_pattern_bitmode(0x18, 0x03, 0x00));
-       active = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
-                              LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 0x7, 0), ""),
-                              LLVMConstInt(ctx->i32, 0x4, 0), "");
-       tmp1 = LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
-       tmp2 = ac_build_ds_swizzle(ctx, src, ds_pattern_bitmode(0x10, 0x07, 0x00));
-       active = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
-                              LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 0xf, 0), ""),
-                              LLVMConstInt(ctx->i32, 0x8, 0), "");
-       tmp1 = LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
-       tmp2 = ac_build_ds_swizzle(ctx, src, ds_pattern_bitmode(0x00, 0x0f, 0x00));
-       active = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
-                              LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 0x1f, 0), ""),
-                              LLVMConstInt(ctx->i32, 0x10, 0), "");
-       tmp1 = LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
-       tmp2 = ac_build_readlane(ctx, src, LLVMConstInt(ctx->i32, 31, 0));
-       active = LLVMBuildICmp(ctx->builder, LLVMIntEQ, tid, LLVMConstInt(ctx->i32, 32, 0), "");
-       tmp1 = LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
-       active = LLVMBuildICmp(ctx->builder, LLVMIntEQ, tid, LLVMConstInt(ctx->i32, 0, 0), "");
-       return LLVMBuildSelect(ctx->builder, active, identity, tmp1, "");
+static LLVMValueRef ac_wavefront_shift_right_1(struct ac_llvm_context *ctx, LLVMValueRef src,
+                                               LLVMValueRef identity, unsigned maxprefix)
+{
+   if (ctx->chip_class >= GFX10) {
+      /* wavefront shift_right by 1 on GFX10 (emulate dpp_wf_sr1) */
+      LLVMValueRef active, tmp1, tmp2;
+      LLVMValueRef tid = ac_get_thread_id(ctx);
+
+      tmp1 = ac_build_dpp(ctx, identity, src, dpp_row_sr(1), 0xf, 0xf, false);
+
+      tmp2 = ac_build_permlane16(ctx, src, (uint64_t)~0, true, false);
+
+      if (maxprefix > 32) {
+         active =
+            LLVMBuildICmp(ctx->builder, LLVMIntEQ, tid, LLVMConstInt(ctx->i32, 32, false), "");
+
+         tmp2 = LLVMBuildSelect(ctx->builder, active,
+                                ac_build_readlane(ctx, src, LLVMConstInt(ctx->i32, 31, false)),
+                                tmp2, "");
+
+         active = LLVMBuildOr(
+            ctx->builder, active,
+            LLVMBuildICmp(ctx->builder, LLVMIntEQ,
+                          LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 0x1f, false), ""),
+                          LLVMConstInt(ctx->i32, 0x10, false), ""),
+            "");
+         return LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
+      } else if (maxprefix > 16) {
+         active =
+            LLVMBuildICmp(ctx->builder, LLVMIntEQ, tid, LLVMConstInt(ctx->i32, 16, false), "");
+
+         return LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
+      }
+   } else if (ctx->chip_class >= GFX8) {
+      return ac_build_dpp(ctx, identity, src, dpp_wf_sr1, 0xf, 0xf, false);
+   }
+
+   /* wavefront shift_right by 1 on SI/CI */
+   LLVMValueRef active, tmp1, tmp2;
+   LLVMValueRef tid = ac_get_thread_id(ctx);
+   tmp1 = ac_build_ds_swizzle(ctx, src, (1 << 15) | dpp_quad_perm(0, 0, 1, 2));
+   tmp2 = ac_build_ds_swizzle(ctx, src, ds_pattern_bitmode(0x18, 0x03, 0x00));
+   active = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
+                          LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 0x7, 0), ""),
+                          LLVMConstInt(ctx->i32, 0x4, 0), "");
+   tmp1 = LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
+   tmp2 = ac_build_ds_swizzle(ctx, src, ds_pattern_bitmode(0x10, 0x07, 0x00));
+   active = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
+                          LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 0xf, 0), ""),
+                          LLVMConstInt(ctx->i32, 0x8, 0), "");
+   tmp1 = LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
+   tmp2 = ac_build_ds_swizzle(ctx, src, ds_pattern_bitmode(0x00, 0x0f, 0x00));
+   active = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
+                          LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 0x1f, 0), ""),
+                          LLVMConstInt(ctx->i32, 0x10, 0), "");
+   tmp1 = LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
+   tmp2 = ac_build_readlane(ctx, src, LLVMConstInt(ctx->i32, 31, 0));
+   active = LLVMBuildICmp(ctx->builder, LLVMIntEQ, tid, LLVMConstInt(ctx->i32, 32, 0), "");
+   tmp1 = LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
+   active = LLVMBuildICmp(ctx->builder, LLVMIntEQ, tid, LLVMConstInt(ctx->i32, 0, 0), "");
+   return LLVMBuildSelect(ctx->builder, active, identity, tmp1, "");
 }
 
 /**
  * \param maxprefix specifies that the result only needs to be correct for a
  *     prefix of this many threads
  */
-static LLVMValueRef
-ac_build_scan(struct ac_llvm_context *ctx, nir_op op, LLVMValueRef src, LLVMValueRef identity,
-             unsigned maxprefix, bool inclusive)
-{
-       LLVMValueRef result, tmp;
-
-       if (!inclusive)
-               src = ac_wavefront_shift_right_1(ctx, src, identity, maxprefix);
-
-       result = src;
-
-       if (ctx->chip_class <= GFX7) {
-               assert(maxprefix == 64);
-               LLVMValueRef tid = ac_get_thread_id(ctx);
-               LLVMValueRef active;
-               tmp = ac_build_ds_swizzle(ctx, src, ds_pattern_bitmode(0x1e, 0x00, 0x00));
-               active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
-                                      LLVMBuildAnd(ctx->builder, tid, ctx->i32_1, ""),
-                                      ctx->i32_0, "");
-               tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
-               result = ac_build_alu_op(ctx, result, tmp, op);
-               tmp = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1c, 0x01, 0x00));
-               active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
-                                      LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 2, 0), ""),
-                                      ctx->i32_0, "");
-               tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
-               result = ac_build_alu_op(ctx, result, tmp, op);
-               tmp = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x18, 0x03, 0x00));
-               active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
-                                      LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 4, 0), ""),
-                                      ctx->i32_0, "");
-               tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
-               result = ac_build_alu_op(ctx, result, tmp, op);
-               tmp = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x10, 0x07, 0x00));
-               active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
-                                      LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 8, 0), ""),
-                                      ctx->i32_0, "");
-               tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
-               result = ac_build_alu_op(ctx, result, tmp, op);
-               tmp = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x00, 0x0f, 0x00));
-               active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
-                                      LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 16, 0), ""),
-                                      ctx->i32_0, "");
-               tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
-               result = ac_build_alu_op(ctx, result, tmp, op);
-               tmp = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 31, 0));
-               active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
-                                      LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 32, 0), ""),
-                                      ctx->i32_0, "");
-               tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
-               result = ac_build_alu_op(ctx, result, tmp, op);
-               return result;
-       }
-
-       if (maxprefix <= 1)
-               return result;
-       tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(1), 0xf, 0xf, false);
-       result = ac_build_alu_op(ctx, result, tmp, op);
-       if (maxprefix <= 2)
-               return result;
-       tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(2), 0xf, 0xf, false);
-       result = ac_build_alu_op(ctx, result, tmp, op);
-       if (maxprefix <= 3)
-               return result;
-       tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(3), 0xf, 0xf, false);
-       result = ac_build_alu_op(ctx, result, tmp, op);
-       if (maxprefix <= 4)
-               return result;
-       tmp = ac_build_dpp(ctx, identity, result, dpp_row_sr(4), 0xf, 0xe, false);
-       result = ac_build_alu_op(ctx, result, tmp, op);
-       if (maxprefix <= 8)
-               return result;
-       tmp = ac_build_dpp(ctx, identity, result, dpp_row_sr(8), 0xf, 0xc, false);
-       result = ac_build_alu_op(ctx, result, tmp, op);
-       if (maxprefix <= 16)
-               return result;
-
-       if (ctx->chip_class >= GFX10) {
-               LLVMValueRef tid = ac_get_thread_id(ctx);
-               LLVMValueRef active;
-
-               tmp = ac_build_permlane16(ctx, result, ~(uint64_t)0, true, false);
-
-               active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
-                                      LLVMBuildAnd(ctx->builder, tid,
-                                                   LLVMConstInt(ctx->i32, 16, false), ""),
-                                      ctx->i32_0, "");
-
-               tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
-
-               result = ac_build_alu_op(ctx, result, tmp, op);
-
-               if (maxprefix <= 32)
-                       return result;
-
-               tmp = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 31, false));
-
-               active = LLVMBuildICmp(ctx->builder, LLVMIntUGE, tid,
-                                      LLVMConstInt(ctx->i32, 32, false), "");
-
-               tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
-
-               result = ac_build_alu_op(ctx, result, tmp, op);
-               return result;
-       }
-
-       tmp = ac_build_dpp(ctx, identity, result, dpp_row_bcast15, 0xa, 0xf, false);
-       result = ac_build_alu_op(ctx, result, tmp, op);
-       if (maxprefix <= 32)
-               return result;
-       tmp = ac_build_dpp(ctx, identity, result, dpp_row_bcast31, 0xc, 0xf, false);
-       result = ac_build_alu_op(ctx, result, tmp, op);
-       return result;
-}
-
-LLVMValueRef
-ac_build_inclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op)
-{
-       LLVMValueRef result;
-
-       if (LLVMTypeOf(src) == ctx->i1 && op == nir_op_iadd) {
-               LLVMBuilderRef builder = ctx->builder;
-               src = LLVMBuildZExt(builder, src, ctx->i32, "");
-               result = ac_build_ballot(ctx, src);
-               result = ac_build_mbcnt(ctx, result);
-               result = LLVMBuildAdd(builder, result, src, "");
-               return result;
-       }
-
-       ac_build_optimization_barrier(ctx, &src);
-
-       LLVMValueRef identity =
-               get_reduction_identity(ctx, op, ac_get_type_size(LLVMTypeOf(src)));
-       result = LLVMBuildBitCast(ctx->builder, ac_build_set_inactive(ctx, src, identity),
-                                 LLVMTypeOf(identity), "");
-       result = ac_build_scan(ctx, op, result, identity, ctx->wave_size, true);
-
-       return ac_build_wwm(ctx, result);
-}
-
-LLVMValueRef
-ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op)
-{
-       LLVMValueRef result;
-
-       if (LLVMTypeOf(src) == ctx->i1 && op == nir_op_iadd) {
-               LLVMBuilderRef builder = ctx->builder;
-               src = LLVMBuildZExt(builder, src, ctx->i32, "");
-               result = ac_build_ballot(ctx, src);
-               result = ac_build_mbcnt(ctx, result);
-               return result;
-       }
-
-       ac_build_optimization_barrier(ctx, &src);
-
-       LLVMValueRef identity =
-               get_reduction_identity(ctx, op, ac_get_type_size(LLVMTypeOf(src)));
-       result = LLVMBuildBitCast(ctx->builder, ac_build_set_inactive(ctx, src, identity),
-                                 LLVMTypeOf(identity), "");
-       result = ac_build_scan(ctx, op, result, identity, ctx->wave_size, false);
-
-       return ac_build_wwm(ctx, result);
-}
-
-LLVMValueRef
-ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op, unsigned cluster_size)
-{
-       if (cluster_size == 1) return src;
-       ac_build_optimization_barrier(ctx, &src);
-       LLVMValueRef result, swap;
-       LLVMValueRef identity = get_reduction_identity(ctx, op,
-                                                               ac_get_type_size(LLVMTypeOf(src)));
-       result = LLVMBuildBitCast(ctx->builder,
-                                                               ac_build_set_inactive(ctx, src, identity),
-                                                               LLVMTypeOf(identity), "");
-       swap = ac_build_quad_swizzle(ctx, result, 1, 0, 3, 2);
-       result = ac_build_alu_op(ctx, result, swap, op);
-       if (cluster_size == 2) return ac_build_wwm(ctx, result);
-
-       swap = ac_build_quad_swizzle(ctx, result, 2, 3, 0, 1);
-       result = ac_build_alu_op(ctx, result, swap, op);
-       if (cluster_size == 4) return ac_build_wwm(ctx, result);
-
-       if (ctx->chip_class >= GFX8)
-               swap = ac_build_dpp(ctx, identity, result, dpp_row_half_mirror, 0xf, 0xf, false);
-       else
-               swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x04));
-       result = ac_build_alu_op(ctx, result, swap, op);
-       if (cluster_size == 8) return ac_build_wwm(ctx, result);
-
-       if (ctx->chip_class >= GFX8)
-               swap = ac_build_dpp(ctx, identity, result, dpp_row_mirror, 0xf, 0xf, false);
-       else
-               swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x08));
-       result = ac_build_alu_op(ctx, result, swap, op);
-       if (cluster_size == 16) return ac_build_wwm(ctx, result);
-
-       if (ctx->chip_class >= GFX10)
-               swap = ac_build_permlane16(ctx, result, 0, true, false);
-       else if (ctx->chip_class >= GFX8 && cluster_size != 32)
-               swap = ac_build_dpp(ctx, identity, result, dpp_row_bcast15, 0xa, 0xf, false);
-       else
-               swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x10));
-       result = ac_build_alu_op(ctx, result, swap, op);
-       if (cluster_size == 32) return ac_build_wwm(ctx, result);
-
-       if (ctx->chip_class >= GFX8) {
-               if (ctx->wave_size == 64) {
-                       if (ctx->chip_class >= GFX10)
-                               swap = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 31, false));
-                       else
-                               swap = ac_build_dpp(ctx, identity, result, dpp_row_bcast31, 0xc, 0xf, false);
-                       result = ac_build_alu_op(ctx, result, swap, op);
-                       result = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 63, 0));
-               }
-
-               return ac_build_wwm(ctx, result);
-       } else {
-               swap = ac_build_readlane(ctx, result, ctx->i32_0);
-               result = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 32, 0));
-               result = ac_build_alu_op(ctx, result, swap, op);
-               return ac_build_wwm(ctx, result);
-       }
+static LLVMValueRef ac_build_scan(struct ac_llvm_context *ctx, nir_op op, LLVMValueRef src,
+                                  LLVMValueRef identity, unsigned maxprefix, bool inclusive)
+{
+   LLVMValueRef result, tmp;
+
+   if (!inclusive)
+      src = ac_wavefront_shift_right_1(ctx, src, identity, maxprefix);
+
+   result = src;
+
+   if (ctx->chip_class <= GFX7) {
+      assert(maxprefix == 64);
+      LLVMValueRef tid = ac_get_thread_id(ctx);
+      LLVMValueRef active;
+      tmp = ac_build_ds_swizzle(ctx, src, ds_pattern_bitmode(0x1e, 0x00, 0x00));
+      active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
+                             LLVMBuildAnd(ctx->builder, tid, ctx->i32_1, ""), ctx->i32_0, "");
+      tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
+      result = ac_build_alu_op(ctx, result, tmp, op);
+      tmp = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1c, 0x01, 0x00));
+      active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
+                             LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 2, 0), ""),
+                             ctx->i32_0, "");
+      tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
+      result = ac_build_alu_op(ctx, result, tmp, op);
+      tmp = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x18, 0x03, 0x00));
+      active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
+                             LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 4, 0), ""),
+                             ctx->i32_0, "");
+      tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
+      result = ac_build_alu_op(ctx, result, tmp, op);
+      tmp = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x10, 0x07, 0x00));
+      active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
+                             LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 8, 0), ""),
+                             ctx->i32_0, "");
+      tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
+      result = ac_build_alu_op(ctx, result, tmp, op);
+      tmp = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x00, 0x0f, 0x00));
+      active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
+                             LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 16, 0), ""),
+                             ctx->i32_0, "");
+      tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
+      result = ac_build_alu_op(ctx, result, tmp, op);
+      tmp = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 31, 0));
+      active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
+                             LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 32, 0), ""),
+                             ctx->i32_0, "");
+      tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
+      result = ac_build_alu_op(ctx, result, tmp, op);
+      return result;
+   }
+
+   if (maxprefix <= 1)
+      return result;
+   tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(1), 0xf, 0xf, false);
+   result = ac_build_alu_op(ctx, result, tmp, op);
+   if (maxprefix <= 2)
+      return result;
+   tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(2), 0xf, 0xf, false);
+   result = ac_build_alu_op(ctx, result, tmp, op);
+   if (maxprefix <= 3)
+      return result;
+   tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(3), 0xf, 0xf, false);
+   result = ac_build_alu_op(ctx, result, tmp, op);
+   if (maxprefix <= 4)
+      return result;
+   tmp = ac_build_dpp(ctx, identity, result, dpp_row_sr(4), 0xf, 0xe, false);
+   result = ac_build_alu_op(ctx, result, tmp, op);
+   if (maxprefix <= 8)
+      return result;
+   tmp = ac_build_dpp(ctx, identity, result, dpp_row_sr(8), 0xf, 0xc, false);
+   result = ac_build_alu_op(ctx, result, tmp, op);
+   if (maxprefix <= 16)
+      return result;
+
+   if (ctx->chip_class >= GFX10) {
+      LLVMValueRef tid = ac_get_thread_id(ctx);
+      LLVMValueRef active;
+
+      tmp = ac_build_permlane16(ctx, result, ~(uint64_t)0, true, false);
+
+      active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
+                             LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 16, false), ""),
+                             ctx->i32_0, "");
+
+      tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
+
+      result = ac_build_alu_op(ctx, result, tmp, op);
+
+      if (maxprefix <= 32)
+         return result;
+
+      tmp = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 31, false));
+
+      active = LLVMBuildICmp(ctx->builder, LLVMIntUGE, tid, LLVMConstInt(ctx->i32, 32, false), "");
+
+      tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
+
+      result = ac_build_alu_op(ctx, result, tmp, op);
+      return result;
+   }
+
+   tmp = ac_build_dpp(ctx, identity, result, dpp_row_bcast15, 0xa, 0xf, false);
+   result = ac_build_alu_op(ctx, result, tmp, op);
+   if (maxprefix <= 32)
+      return result;
+   tmp = ac_build_dpp(ctx, identity, result, dpp_row_bcast31, 0xc, 0xf, false);
+   result = ac_build_alu_op(ctx, result, tmp, op);
+   return result;
+}
+
+LLVMValueRef ac_build_inclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op)
+{
+   LLVMValueRef result;
+
+   if (LLVMTypeOf(src) == ctx->i1 && op == nir_op_iadd) {
+      LLVMBuilderRef builder = ctx->builder;
+      src = LLVMBuildZExt(builder, src, ctx->i32, "");
+      result = ac_build_ballot(ctx, src);
+      result = ac_build_mbcnt(ctx, result);
+      result = LLVMBuildAdd(builder, result, src, "");
+      return result;
+   }
+
+   ac_build_optimization_barrier(ctx, &src);
+
+   LLVMValueRef identity = get_reduction_identity(ctx, op, ac_get_type_size(LLVMTypeOf(src)));
+   result = LLVMBuildBitCast(ctx->builder, ac_build_set_inactive(ctx, src, identity),
+                             LLVMTypeOf(identity), "");
+   result = ac_build_scan(ctx, op, result, identity, ctx->wave_size, true);
+
+   return ac_build_wwm(ctx, result);
+}
+
+LLVMValueRef ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op)
+{
+   LLVMValueRef result;
+
+   if (LLVMTypeOf(src) == ctx->i1 && op == nir_op_iadd) {
+      LLVMBuilderRef builder = ctx->builder;
+      src = LLVMBuildZExt(builder, src, ctx->i32, "");
+      result = ac_build_ballot(ctx, src);
+      result = ac_build_mbcnt(ctx, result);
+      return result;
+   }
+
+   ac_build_optimization_barrier(ctx, &src);
+
+   LLVMValueRef identity = get_reduction_identity(ctx, op, ac_get_type_size(LLVMTypeOf(src)));
+   result = LLVMBuildBitCast(ctx->builder, ac_build_set_inactive(ctx, src, identity),
+                             LLVMTypeOf(identity), "");
+   result = ac_build_scan(ctx, op, result, identity, ctx->wave_size, false);
+
+   return ac_build_wwm(ctx, result);
+}
+
+LLVMValueRef ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op,
+                             unsigned cluster_size)
+{
+   if (cluster_size == 1)
+      return src;
+   ac_build_optimization_barrier(ctx, &src);
+   LLVMValueRef result, swap;
+   LLVMValueRef identity = get_reduction_identity(ctx, op, ac_get_type_size(LLVMTypeOf(src)));
+   result = LLVMBuildBitCast(ctx->builder, ac_build_set_inactive(ctx, src, identity),
+                             LLVMTypeOf(identity), "");
+   swap = ac_build_quad_swizzle(ctx, result, 1, 0, 3, 2);
+   result = ac_build_alu_op(ctx, result, swap, op);
+   if (cluster_size == 2)
+      return ac_build_wwm(ctx, result);
+
+   swap = ac_build_quad_swizzle(ctx, result, 2, 3, 0, 1);
+   result = ac_build_alu_op(ctx, result, swap, op);
+   if (cluster_size == 4)
+      return ac_build_wwm(ctx, result);
+
+   if (ctx->chip_class >= GFX8)
+      swap = ac_build_dpp(ctx, identity, result, dpp_row_half_mirror, 0xf, 0xf, false);
+   else
+      swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x04));
+   result = ac_build_alu_op(ctx, result, swap, op);
+   if (cluster_size == 8)
+      return ac_build_wwm(ctx, result);
+
+   if (ctx->chip_class >= GFX8)
+      swap = ac_build_dpp(ctx, identity, result, dpp_row_mirror, 0xf, 0xf, false);
+   else
+      swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x08));
+   result = ac_build_alu_op(ctx, result, swap, op);
+   if (cluster_size == 16)
+      return ac_build_wwm(ctx, result);
+
+   if (ctx->chip_class >= GFX10)
+      swap = ac_build_permlane16(ctx, result, 0, true, false);
+   else if (ctx->chip_class >= GFX8 && cluster_size != 32)
+      swap = ac_build_dpp(ctx, identity, result, dpp_row_bcast15, 0xa, 0xf, false);
+   else
+      swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x10));
+   result = ac_build_alu_op(ctx, result, swap, op);
+   if (cluster_size == 32)
+      return ac_build_wwm(ctx, result);
+
+   if (ctx->chip_class >= GFX8) {
+      if (ctx->wave_size == 64) {
+         if (ctx->chip_class >= GFX10)
+            swap = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 31, false));
+         else
+            swap = ac_build_dpp(ctx, identity, result, dpp_row_bcast31, 0xc, 0xf, false);
+         result = ac_build_alu_op(ctx, result, swap, op);
+         result = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 63, 0));
+      }
+
+      return ac_build_wwm(ctx, result);
+   } else {
+      swap = ac_build_readlane(ctx, result, ctx->i32_0);
+      result = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 32, 0));
+      result = ac_build_alu_op(ctx, result, swap, op);
+      return ac_build_wwm(ctx, result);
+   }
 }
 
 /**
@@ -4408,21 +4018,20 @@ ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op, unsign
  * The source value must be present in the highest lane of the wave, and the
  * highest lane must be live.
  */
-void
-ac_build_wg_wavescan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
+void ac_build_wg_wavescan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
 {
-       if (ws->maxwaves <= 1)
-               return;
+   if (ws->maxwaves <= 1)
+      return;
 
-       const LLVMValueRef last_lane = LLVMConstInt(ctx->i32, ctx->wave_size - 1, false);
-       LLVMBuilderRef builder = ctx->builder;
-       LLVMValueRef tid = ac_get_thread_id(ctx);
-       LLVMValueRef tmp;
+   const LLVMValueRef last_lane = LLVMConstInt(ctx->i32, ctx->wave_size - 1, false);
+   LLVMBuilderRef builder = ctx->builder;
+   LLVMValueRef tid = ac_get_thread_id(ctx);
+   LLVMValueRef tmp;
 
-       tmp = LLVMBuildICmp(builder, LLVMIntEQ, tid, last_lane, "");
-       ac_build_ifcc(ctx, tmp, 1000);
-       LLVMBuildStore(builder, ws->src, LLVMBuildGEP(builder, ws->scratch, &ws->waveidx, 1, ""));
-       ac_build_endif(ctx, 1000);
+   tmp = LLVMBuildICmp(builder, LLVMIntEQ, tid, last_lane, "");
+   ac_build_ifcc(ctx, tmp, 1000);
+   LLVMBuildStore(builder, ws->src, LLVMBuildGEP(builder, ws->scratch, &ws->waveidx, 1, ""));
+   ac_build_endif(ctx, 1000);
 }
 
 /**
@@ -4431,61 +4040,59 @@ ac_build_wg_wavescan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
  *
  * The caller must place a barrier between the top and bottom halves.
  */
-void
-ac_build_wg_wavescan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
-{
-       const LLVMTypeRef type = LLVMTypeOf(ws->src);
-       const LLVMValueRef identity =
-               get_reduction_identity(ctx, ws->op, ac_get_type_size(type));
-
-       if (ws->maxwaves <= 1) {
-               ws->result_reduce = ws->src;
-               ws->result_inclusive = ws->src;
-               ws->result_exclusive = identity;
-               return;
-       }
-       assert(ws->maxwaves <= 32);
-
-       LLVMBuilderRef builder = ctx->builder;
-       LLVMValueRef tid = ac_get_thread_id(ctx);
-       LLVMBasicBlockRef bbs[2];
-       LLVMValueRef phivalues_scan[2];
-       LLVMValueRef tmp, tmp2;
-
-       bbs[0] = LLVMGetInsertBlock(builder);
-       phivalues_scan[0] = LLVMGetUndef(type);
-
-       if (ws->enable_reduce)
-               tmp = LLVMBuildICmp(builder