radeonsi: import all TGSI->LLVM code from gallium/radeon
authorMarek Olšák <marek.olsak@amd.com>
Mon, 17 Oct 2016 10:30:42 +0000 (12:30 +0200)
committerMarek Olšák <marek.olsak@amd.com>
Tue, 18 Oct 2016 16:41:04 +0000 (18:41 +0200)
Acked-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
Acked-by: Edward O'Callaghan <funfunctor@folklore1984.net>
12 files changed:
src/gallium/drivers/r600/evergreen_compute_internal.h
src/gallium/drivers/radeon/Makefile.sources
src/gallium/drivers/radeon/radeon_llvm.h [deleted file]
src/gallium/drivers/radeon/radeon_llvm_emit.c [deleted file]
src/gallium/drivers/radeon/radeon_llvm_emit.h [deleted file]
src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c [deleted file]
src/gallium/drivers/radeonsi/Makefile.sources
src/gallium/drivers/radeonsi/si_pipe.c
src/gallium/drivers/radeonsi/si_shader.c
src/gallium/drivers/radeonsi/si_shader_internal.h
src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c
src/gallium/drivers/radeonsi/si_shader_tgsi_setup.c [new file with mode: 0644]

index e6ff7609aea2f39e75e7006899e74e16b1ae1be9..34d96f6d23922f921ea70954586af359d0b261c6 100644 (file)
@@ -27,7 +27,6 @@
 
 #include "r600_asm.h"
 #ifdef HAVE_OPENCL
-#include "radeon/radeon_llvm.h"
 #include <llvm-c/Core.h>
 #endif
 
index 049da60c134f779fb60a65c680dabd33fc28546d..3e13dae3cbb6ff1766bd9cade18dad1ec9d3273a 100644 (file)
@@ -25,8 +25,4 @@ C_SOURCES := \
 
 LLVM_C_FILES := \
        radeon_elf_util.c \
-       radeon_elf_util.h \
-       radeon_llvm_emit.c \
-       radeon_llvm_emit.h \
-       radeon_llvm.h \
-       radeon_setup_tgsi_llvm.c
+       radeon_elf_util.h
diff --git a/src/gallium/drivers/radeon/radeon_llvm.h b/src/gallium/drivers/radeon/radeon_llvm.h
deleted file mode 100644 (file)
index b4b968b..0000000
+++ /dev/null
@@ -1,151 +0,0 @@
-/*
- * Copyright 2011 Advanced Micro Devices, Inc.
- *
- * Permission is hereby granted, free of charge, to any person obtaining a
- * copy of this software and associated documentation files (the "Software"),
- * to deal in the Software without restriction, including without limitation
- * the rights to use, copy, modify, merge, publish, distribute, sublicense,
- * and/or sell copies of the Software, and to permit persons to whom the
- * Software is furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice (including the next
- * paragraph) shall be included in all copies or substantial portions of the
- * Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
- * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- *
- * Authors: Tom Stellard <thomas.stellard@amd.com>
- *
- */
-
-#ifndef RADEON_LLVM_H
-#define RADEON_LLVM_H
-
-#include <llvm-c/Core.h>
-#include "gallivm/lp_bld_init.h"
-#include "gallivm/lp_bld_tgsi.h"
-#include "tgsi/tgsi_parse.h"
-
-#define RADEON_LLVM_MAX_INPUT_SLOTS 32
-#define RADEON_LLVM_MAX_INPUTS 32 * 4
-#define RADEON_LLVM_MAX_OUTPUTS 32 * 4
-
-#define RADEON_LLVM_INITIAL_CF_DEPTH 4
-
-#define RADEON_LLVM_MAX_SYSTEM_VALUES 4
-
-struct radeon_llvm_flow;
-
-struct radeon_llvm_context {
-       struct lp_build_tgsi_soa_context soa;
-
-       /*=== Front end configuration ===*/
-
-       /* Instructions that are not described by any of the TGSI opcodes. */
-
-       /** This function is responsible for initilizing the inputs array and will be
-         * called once for each input declared in the TGSI shader.
-         */
-       void (*load_input)(struct radeon_llvm_context *,
-                          unsigned input_index,
-                          const struct tgsi_full_declaration *decl,
-                          LLVMValueRef out[4]);
-
-       void (*load_system_value)(struct radeon_llvm_context *,
-                                 unsigned index,
-                                 const struct tgsi_full_declaration *decl);
-
-       void (*declare_memory_region)(struct radeon_llvm_context *,
-                                     const struct tgsi_full_declaration *decl);
-
-       /** This array contains the input values for the shader.  Typically these
-         * values will be in the form of a target intrinsic that will inform the
-         * backend how to load the actual inputs to the shader. 
-         */
-       struct tgsi_full_declaration input_decls[RADEON_LLVM_MAX_INPUT_SLOTS];
-       LLVMValueRef inputs[RADEON_LLVM_MAX_INPUTS];
-       LLVMValueRef outputs[RADEON_LLVM_MAX_OUTPUTS][TGSI_NUM_CHANNELS];
-
-       /** This pointer is used to contain the temporary values.
-         * The amount of temporary used in tgsi can't be bound to a max value and
-         * thus we must allocate this array at runtime.
-         */
-       LLVMValueRef *temps;
-       unsigned temps_count;
-       LLVMValueRef system_values[RADEON_LLVM_MAX_SYSTEM_VALUES];
-
-       /*=== Private Members ===*/
-
-       struct radeon_llvm_flow *flow;
-       unsigned flow_depth;
-       unsigned flow_depth_max;
-
-       struct tgsi_array_info *temp_arrays;
-       LLVMValueRef *temp_array_allocas;
-
-       LLVMValueRef undef_alloca;
-
-       LLVMValueRef main_fn;
-       LLVMTypeRef return_type;
-
-       unsigned fpmath_md_kind;
-       LLVMValueRef fpmath_md_2p5_ulp;
-
-       struct gallivm_state gallivm;
-};
-
-LLVMTypeRef tgsi2llvmtype(struct lp_build_tgsi_context *bld_base,
-                         enum tgsi_opcode_type type);
-
-LLVMValueRef bitcast(struct lp_build_tgsi_context *bld_base,
-                    enum tgsi_opcode_type type, LLVMValueRef value);
-
-LLVMValueRef radeon_llvm_bound_index(struct radeon_llvm_context *ctx,
-                                    LLVMValueRef index,
-                                    unsigned num);
-
-void radeon_llvm_context_init(struct radeon_llvm_context *ctx,
-                              const char *triple,
-                             const struct tgsi_shader_info *info,
-                             const struct tgsi_token *tokens);
-
-void radeon_llvm_create_func(struct radeon_llvm_context *ctx,
-                            LLVMTypeRef *return_types, unsigned num_return_elems,
-                            LLVMTypeRef *ParamTypes, unsigned ParamCount);
-
-void radeon_llvm_dispose(struct radeon_llvm_context *ctx);
-
-void radeon_llvm_finalize_module(struct radeon_llvm_context *ctx,
-                                bool run_verifier);
-
-LLVMValueRef radeon_llvm_emit_fetch_64bit(struct lp_build_tgsi_context *bld_base,
-                                         enum tgsi_opcode_type type,
-                                         LLVMValueRef ptr,
-                                         LLVMValueRef ptr2);
-
-LLVMValueRef radeon_llvm_saturate(struct lp_build_tgsi_context *bld_base,
-                                  LLVMValueRef value);
-
-LLVMValueRef radeon_llvm_emit_fetch(struct lp_build_tgsi_context *bld_base,
-                                   const struct tgsi_full_src_register *reg,
-                                   enum tgsi_opcode_type type,
-                                   unsigned swizzle);
-
-void radeon_llvm_emit_store(struct lp_build_tgsi_context *bld_base,
-                           const struct tgsi_full_instruction *inst,
-                           const struct tgsi_opcode_info *info,
-                           LLVMValueRef dst[4]);
-
-static inline struct radeon_llvm_context *
-radeon_llvm_context(struct lp_build_tgsi_context *bld_base)
-{
-       return (struct radeon_llvm_context*)bld_base;
-}
-
-#endif /* RADEON_LLVM_H */
diff --git a/src/gallium/drivers/radeon/radeon_llvm_emit.c b/src/gallium/drivers/radeon/radeon_llvm_emit.c
deleted file mode 100644 (file)
index 9611f15..0000000
+++ /dev/null
@@ -1,241 +0,0 @@
-/*
- * Copyright 2011 Advanced Micro Devices, Inc.
- *
- * Permission is hereby granted, free of charge, to any person obtaining a
- * copy of this software and associated documentation files (the "Software"),
- * to deal in the Software without restriction, including without limitation
- * the rights to use, copy, modify, merge, publish, distribute, sublicense,
- * and/or sell copies of the Software, and to permit persons to whom the
- * Software is furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice (including the next
- * paragraph) shall be included in all copies or substantial portions of the
- * Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
- * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- *
- * Authors: Tom Stellard <thomas.stellard@amd.com>
- *
- */
-
-#include "radeon_llvm_emit.h"
-#include "radeon_elf_util.h"
-#include "c11/threads.h"
-#include "gallivm/lp_bld_misc.h"
-#include "util/u_debug.h"
-#include "util/u_memory.h"
-#include "pipe/p_shader_tokens.h"
-#include "pipe/p_state.h"
-
-#include <llvm-c/Target.h>
-#include <llvm-c/TargetMachine.h>
-#include <llvm-c/Core.h>
-
-#include <stdlib.h>
-#include <stdio.h>
-#include <string.h>
-
-#define CPU_STRING_LEN 30
-#define FS_STRING_LEN 30
-#define TRIPLE_STRING_LEN 7
-
-/**
- * Shader types for the LLVM backend.
- */
-enum radeon_llvm_shader_type {
-       RADEON_LLVM_SHADER_PS = 0,
-       RADEON_LLVM_SHADER_VS = 1,
-       RADEON_LLVM_SHADER_GS = 2,
-       RADEON_LLVM_SHADER_CS = 3,
-};
-
-enum radeon_llvm_calling_convention {
-       RADEON_LLVM_AMDGPU_VS = 87,
-       RADEON_LLVM_AMDGPU_GS = 88,
-       RADEON_LLVM_AMDGPU_PS = 89,
-       RADEON_LLVM_AMDGPU_CS = 90,
-};
-
-void radeon_llvm_add_attribute(LLVMValueRef F, const char *name, int value)
-{
-       char str[16];
-
-       snprintf(str, sizeof(str), "%i", value);
-       LLVMAddTargetDependentFunctionAttr(F, name, str);
-}
-
-/**
- * Set the shader type we want to compile
- *
- * @param type shader type to set
- */
-void radeon_llvm_shader_type(LLVMValueRef F, unsigned type)
-{
-       enum radeon_llvm_shader_type llvm_type;
-       enum radeon_llvm_calling_convention calling_conv;
-
-       switch (type) {
-       case PIPE_SHADER_VERTEX:
-       case PIPE_SHADER_TESS_CTRL:
-       case PIPE_SHADER_TESS_EVAL:
-               llvm_type = RADEON_LLVM_SHADER_VS;
-               calling_conv = RADEON_LLVM_AMDGPU_VS;
-               break;
-       case PIPE_SHADER_GEOMETRY:
-               llvm_type = RADEON_LLVM_SHADER_GS;
-               calling_conv = RADEON_LLVM_AMDGPU_GS;
-               break;
-       case PIPE_SHADER_FRAGMENT:
-               llvm_type = RADEON_LLVM_SHADER_PS;
-               calling_conv = RADEON_LLVM_AMDGPU_PS;
-               break;
-       case PIPE_SHADER_COMPUTE:
-               llvm_type = RADEON_LLVM_SHADER_CS;
-               calling_conv = RADEON_LLVM_AMDGPU_CS;
-               break;
-       default:
-               unreachable("Unhandle shader type");
-       }
-
-       if (HAVE_LLVM >= 0x309)
-               LLVMSetFunctionCallConv(F, calling_conv);
-       else
-               radeon_llvm_add_attribute(F, "ShaderType", llvm_type);
-}
-
-static void init_r600_target()
-{
-       gallivm_init_llvm_targets();
-#if HAVE_LLVM < 0x0307
-       LLVMInitializeR600TargetInfo();
-       LLVMInitializeR600Target();
-       LLVMInitializeR600TargetMC();
-       LLVMInitializeR600AsmPrinter();
-#else
-       LLVMInitializeAMDGPUTargetInfo();
-       LLVMInitializeAMDGPUTarget();
-       LLVMInitializeAMDGPUTargetMC();
-       LLVMInitializeAMDGPUAsmPrinter();
-
-#endif
-}
-
-static once_flag init_r600_target_once_flag = ONCE_FLAG_INIT;
-
-LLVMTargetRef radeon_llvm_get_r600_target(const char *triple)
-{
-       LLVMTargetRef target = NULL;
-       char *err_message = NULL;
-
-       call_once(&init_r600_target_once_flag, init_r600_target);
-
-       if (LLVMGetTargetFromTriple(triple, &target, &err_message)) {
-               fprintf(stderr, "Cannot find target for triple %s ", triple);
-               if (err_message) {
-                       fprintf(stderr, "%s\n", err_message);
-               }
-               LLVMDisposeMessage(err_message);
-               return NULL;
-       }
-       return target;
-}
-
-struct radeon_llvm_diagnostics {
-       struct pipe_debug_callback *debug;
-       unsigned retval;
-};
-
-static void radeonDiagnosticHandler(LLVMDiagnosticInfoRef di, void *context)
-{
-       struct radeon_llvm_diagnostics *diag = (struct radeon_llvm_diagnostics *)context;
-       LLVMDiagnosticSeverity severity = LLVMGetDiagInfoSeverity(di);
-       char *description = LLVMGetDiagInfoDescription(di);
-       const char *severity_str = NULL;
-
-       switch (severity) {
-       case LLVMDSError:
-               severity_str = "error";
-               break;
-       case LLVMDSWarning:
-               severity_str = "warning";
-               break;
-       case LLVMDSRemark:
-               severity_str = "remark";
-               break;
-       case LLVMDSNote:
-               severity_str = "note";
-               break;
-       default:
-               severity_str = "unknown";
-       }
-
-       pipe_debug_message(diag->debug, SHADER_INFO,
-                          "LLVM diagnostic (%s): %s", severity_str, description);
-
-       if (severity == LLVMDSError) {
-               diag->retval = 1;
-               fprintf(stderr,"LLVM triggered Diagnostic Handler: %s\n", description);
-       }
-
-       LLVMDisposeMessage(description);
-}
-
-/**
- * Compile an LLVM module to machine code.
- *
- * @returns 0 for success, 1 for failure
- */
-unsigned radeon_llvm_compile(LLVMModuleRef M, struct radeon_shader_binary *binary,
-                            LLVMTargetMachineRef tm,
-                            struct pipe_debug_callback *debug)
-{
-       struct radeon_llvm_diagnostics diag;
-       char *err;
-       LLVMContextRef llvm_ctx;
-       LLVMMemoryBufferRef out_buffer;
-       unsigned buffer_size;
-       const char *buffer_data;
-       LLVMBool mem_err;
-
-       diag.debug = debug;
-       diag.retval = 0;
-
-       /* Setup Diagnostic Handler*/
-       llvm_ctx = LLVMGetModuleContext(M);
-
-       LLVMContextSetDiagnosticHandler(llvm_ctx, radeonDiagnosticHandler, &diag);
-
-       /* Compile IR*/
-       mem_err = LLVMTargetMachineEmitToMemoryBuffer(tm, M, LLVMObjectFile, &err,
-                                                                &out_buffer);
-
-       /* Process Errors/Warnings */
-       if (mem_err) {
-               fprintf(stderr, "%s: %s", __FUNCTION__, err);
-               pipe_debug_message(debug, SHADER_INFO,
-                                  "LLVM emit error: %s", err);
-               FREE(err);
-               diag.retval = 1;
-               goto out;
-       }
-
-       /* Extract Shader Code*/
-       buffer_size = LLVMGetBufferSize(out_buffer);
-       buffer_data = LLVMGetBufferStart(out_buffer);
-
-       radeon_elf_read(buffer_data, buffer_size, binary);
-
-       /* Clean up */
-       LLVMDisposeMemoryBuffer(out_buffer);
-
-out:
-       if (diag.retval != 0)
-               pipe_debug_message(debug, SHADER_INFO, "LLVM compile failed");
-       return diag.retval;
-}
diff --git a/src/gallium/drivers/radeon/radeon_llvm_emit.h b/src/gallium/drivers/radeon/radeon_llvm_emit.h
deleted file mode 100644 (file)
index 677978b..0000000
+++ /dev/null
@@ -1,46 +0,0 @@
-/*
- * Copyright 2012 Advanced Micro Devices, Inc.
- *
- * Permission is hereby granted, free of charge, to any person obtaining a
- * copy of this software and associated documentation files (the "Software"),
- * to deal in the Software without restriction, including without limitation
- * the rights to use, copy, modify, merge, publish, distribute, sublicense,
- * and/or sell copies of the Software, and to permit persons to whom the
- * Software is furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice (including the next
- * paragraph) shall be included in all copies or substantial portions of the
- * Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
- * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- *
- * Authors: Tom Stellard <thomas.stellard@amd.com>
- *
- */
-
-#ifndef RADEON_LLVM_EMIT_H
-#define RADEON_LLVM_EMIT_H
-
-#include <llvm-c/Core.h>
-#include <llvm-c/TargetMachine.h>
-#include <stdbool.h>
-
-struct pipe_debug_callback;
-struct radeon_shader_binary;
-
-void radeon_llvm_add_attribute(LLVMValueRef F, const char *name, int value);
-void radeon_llvm_shader_type(LLVMValueRef F, unsigned type);
-
-LLVMTargetRef radeon_llvm_get_r600_target(const char *triple);
-
-unsigned radeon_llvm_compile(LLVMModuleRef M, struct radeon_shader_binary *binary,
-                            LLVMTargetMachineRef tm,
-                            struct pipe_debug_callback *debug);
-
-#endif /* RADEON_LLVM_EMIT_H */
diff --git a/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c b/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
deleted file mode 100644 (file)
index c06eb3e..0000000
+++ /dev/null
@@ -1,1182 +0,0 @@
-/*
- * Copyright 2011 Advanced Micro Devices, Inc.
- *
- * Permission is hereby granted, free of charge, to any person obtaining a
- * copy of this software and associated documentation files (the "Software"),
- * to deal in the Software without restriction, including without limitation
- * the rights to use, copy, modify, merge, publish, distribute, sublicense,
- * and/or sell copies of the Software, and to permit persons to whom the
- * Software is furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice (including the next
- * paragraph) shall be included in all copies or substantial portions of the
- * Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
- * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- *
- * Authors: Tom Stellard <thomas.stellard@amd.com>
- *
- */
-#include "radeon_llvm.h"
-
-#include "gallivm/lp_bld_const.h"
-#include "gallivm/lp_bld_gather.h"
-#include "gallivm/lp_bld_flow.h"
-#include "gallivm/lp_bld_init.h"
-#include "gallivm/lp_bld_intr.h"
-#include "gallivm/lp_bld_misc.h"
-#include "gallivm/lp_bld_swizzle.h"
-#include "tgsi/tgsi_info.h"
-#include "tgsi/tgsi_parse.h"
-#include "util/u_math.h"
-#include "util/u_memory.h"
-#include "util/u_debug.h"
-
-#include <stdio.h>
-#include <llvm-c/Core.h>
-#include <llvm-c/Transforms/Scalar.h>
-
-/* Data for if/else/endif and bgnloop/endloop control flow structures.
- */
-struct radeon_llvm_flow {
-       /* Loop exit or next part of if/else/endif. */
-       LLVMBasicBlockRef next_block;
-       LLVMBasicBlockRef loop_entry_block;
-};
-
-LLVMTypeRef tgsi2llvmtype(struct lp_build_tgsi_context *bld_base,
-                         enum tgsi_opcode_type type)
-{
-       LLVMContextRef ctx = bld_base->base.gallivm->context;
-
-       switch (type) {
-       case TGSI_TYPE_UNSIGNED:
-       case TGSI_TYPE_SIGNED:
-               return LLVMInt32TypeInContext(ctx);
-       case TGSI_TYPE_UNSIGNED64:
-       case TGSI_TYPE_SIGNED64:
-               return LLVMInt64TypeInContext(ctx);
-       case TGSI_TYPE_DOUBLE:
-               return LLVMDoubleTypeInContext(ctx);
-       case TGSI_TYPE_UNTYPED:
-       case TGSI_TYPE_FLOAT:
-               return LLVMFloatTypeInContext(ctx);
-       default: break;
-       }
-       return 0;
-}
-
-LLVMValueRef bitcast(struct lp_build_tgsi_context *bld_base,
-                    enum tgsi_opcode_type type, LLVMValueRef value)
-{
-       LLVMBuilderRef builder = bld_base->base.gallivm->builder;
-       LLVMTypeRef dst_type = tgsi2llvmtype(bld_base, type);
-
-       if (dst_type)
-               return LLVMBuildBitCast(builder, value, dst_type, "");
-       else
-               return value;
-}
-
-/**
- * Return a value that is equal to the given i32 \p index if it lies in [0,num)
- * or an undefined value in the same interval otherwise.
- */
-LLVMValueRef radeon_llvm_bound_index(struct radeon_llvm_context *ctx,
-                                    LLVMValueRef index,
-                                    unsigned num)
-{
-       struct gallivm_state *gallivm = &ctx->gallivm;
-       LLVMBuilderRef builder = gallivm->builder;
-       LLVMValueRef c_max = lp_build_const_int32(gallivm, num - 1);
-       LLVMValueRef cc;
-
-       if (util_is_power_of_two(num)) {
-               index = LLVMBuildAnd(builder, index, c_max, "");
-       } else {
-               /* In theory, this MAX pattern should result in code that is
-                * as good as the bit-wise AND above.
-                *
-                * In practice, LLVM generates worse code (at the time of
-                * writing), because its value tracking is not strong enough.
-                */
-               cc = LLVMBuildICmp(builder, LLVMIntULE, index, c_max, "");
-               index = LLVMBuildSelect(builder, cc, index, c_max, "");
-       }
-
-       return index;
-}
-
-static struct radeon_llvm_flow *
-get_current_flow(struct radeon_llvm_context *ctx)
-{
-       if (ctx->flow_depth > 0)
-               return &ctx->flow[ctx->flow_depth - 1];
-       return NULL;
-}
-
-static struct radeon_llvm_flow *
-get_innermost_loop(struct radeon_llvm_context *ctx)
-{
-       for (unsigned i = ctx->flow_depth; i > 0; --i) {
-               if (ctx->flow[i - 1].loop_entry_block)
-                       return &ctx->flow[i - 1];
-       }
-       return NULL;
-}
-
-static struct radeon_llvm_flow *
-push_flow(struct radeon_llvm_context *ctx)
-{
-       struct radeon_llvm_flow *flow;
-
-       if (ctx->flow_depth >= ctx->flow_depth_max) {
-               unsigned new_max = MAX2(ctx->flow_depth << 1, RADEON_LLVM_INITIAL_CF_DEPTH);
-               ctx->flow = REALLOC(ctx->flow,
-                                   ctx->flow_depth_max * sizeof(*ctx->flow),
-                                   new_max * sizeof(*ctx->flow));
-               ctx->flow_depth_max = new_max;
-       }
-
-       flow = &ctx->flow[ctx->flow_depth];
-       ctx->flow_depth++;
-
-       flow->next_block = NULL;
-       flow->loop_entry_block = NULL;
-       return flow;
-}
-
-static LLVMValueRef emit_swizzle(struct lp_build_tgsi_context *bld_base,
-                                LLVMValueRef value,
-                                unsigned swizzle_x,
-                                unsigned swizzle_y,
-                                unsigned swizzle_z,
-                                unsigned swizzle_w)
-{
-       LLVMValueRef swizzles[4];
-       LLVMTypeRef i32t =
-               LLVMInt32TypeInContext(bld_base->base.gallivm->context);
-
-       swizzles[0] = LLVMConstInt(i32t, swizzle_x, 0);
-       swizzles[1] = LLVMConstInt(i32t, swizzle_y, 0);
-       swizzles[2] = LLVMConstInt(i32t, swizzle_z, 0);
-       swizzles[3] = LLVMConstInt(i32t, swizzle_w, 0);
-
-       return LLVMBuildShuffleVector(bld_base->base.gallivm->builder,
-                                     value,
-                                     LLVMGetUndef(LLVMTypeOf(value)),
-                                     LLVMConstVector(swizzles, 4), "");
-}
-
-/**
- * Return the description of the array covering the given temporary register
- * index.
- */
-static unsigned
-get_temp_array_id(struct lp_build_tgsi_context *bld_base,
-                 unsigned reg_index,
-                 const struct tgsi_ind_register *reg)
-{
-       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
-       unsigned num_arrays = ctx->soa.bld_base.info->array_max[TGSI_FILE_TEMPORARY];
-       unsigned i;
-
-       if (reg && reg->ArrayID > 0 && reg->ArrayID <= num_arrays)
-               return reg->ArrayID;
-
-       for (i = 0; i < num_arrays; i++) {
-               const struct tgsi_array_info *array = &ctx->temp_arrays[i];
-
-               if (reg_index >= array->range.First && reg_index <= array->range.Last)
-                       return i + 1;
-       }
-
-       return 0;
-}
-
-static struct tgsi_declaration_range
-get_array_range(struct lp_build_tgsi_context *bld_base,
-               unsigned File, unsigned reg_index,
-               const struct tgsi_ind_register *reg)
-{
-       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
-       struct tgsi_declaration_range range;
-
-       if (File == TGSI_FILE_TEMPORARY) {
-               unsigned array_id = get_temp_array_id(bld_base, reg_index, reg);
-               if (array_id)
-                       return ctx->temp_arrays[array_id - 1].range;
-       }
-
-       range.First = 0;
-       range.Last = bld_base->info->file_max[File];
-       return range;
-}
-
-static LLVMValueRef
-emit_array_index(struct lp_build_tgsi_soa_context *bld,
-                const struct tgsi_ind_register *reg,
-                unsigned offset)
-{
-       struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
-
-       if (!reg) {
-               return lp_build_const_int32(gallivm, offset);
-       }
-       LLVMValueRef addr = LLVMBuildLoad(gallivm->builder, bld->addr[reg->Index][reg->Swizzle], "");
-       return LLVMBuildAdd(gallivm->builder, addr, lp_build_const_int32(gallivm, offset), "");
-}
-
-/**
- * For indirect registers, construct a pointer directly to the requested
- * element using getelementptr if possible.
- *
- * Returns NULL if the insertelement/extractelement fallback for array access
- * must be used.
- */
-static LLVMValueRef
-get_pointer_into_array(struct radeon_llvm_context *ctx,
-                      unsigned file,
-                      unsigned swizzle,
-                      unsigned reg_index,
-                      const struct tgsi_ind_register *reg_indirect)
-{
-       unsigned array_id;
-       struct tgsi_array_info *array;
-       struct gallivm_state *gallivm = ctx->soa.bld_base.base.gallivm;
-       LLVMBuilderRef builder = gallivm->builder;
-       LLVMValueRef idxs[2];
-       LLVMValueRef index;
-       LLVMValueRef alloca;
-
-       if (file != TGSI_FILE_TEMPORARY)
-               return NULL;
-
-       array_id = get_temp_array_id(&ctx->soa.bld_base, reg_index, reg_indirect);
-       if (!array_id)
-               return NULL;
-
-       alloca = ctx->temp_array_allocas[array_id - 1];
-       if (!alloca)
-               return NULL;
-
-       array = &ctx->temp_arrays[array_id - 1];
-
-       if (!(array->writemask & (1 << swizzle)))
-               return ctx->undef_alloca;
-
-       index = emit_array_index(&ctx->soa, reg_indirect,
-                                reg_index - ctx->temp_arrays[array_id - 1].range.First);
-
-       /* Ensure that the index is within a valid range, to guard against
-        * VM faults and overwriting critical data (e.g. spilled resource
-        * descriptors).
-        *
-        * TODO It should be possible to avoid the additional instructions
-        * if LLVM is changed so that it guarantuees:
-        * 1. the scratch space descriptor isolates the current wave (this
-        *    could even save the scratch offset SGPR at the cost of an
-        *    additional SALU instruction)
-        * 2. the memory for allocas must be allocated at the _end_ of the
-        *    scratch space (after spilled registers)
-        */
-       index = radeon_llvm_bound_index(ctx, index, array->range.Last - array->range.First + 1);
-
-       index = LLVMBuildMul(
-               builder, index,
-               lp_build_const_int32(gallivm, util_bitcount(array->writemask)),
-               "");
-       index = LLVMBuildAdd(
-               builder, index,
-               lp_build_const_int32(
-                       gallivm,
-                       util_bitcount(array->writemask & ((1 << swizzle) - 1))),
-               "");
-       idxs[0] = ctx->soa.bld_base.uint_bld.zero;
-       idxs[1] = index;
-       return LLVMBuildGEP(builder, alloca, idxs, 2, "");
-}
-
-LLVMValueRef
-radeon_llvm_emit_fetch_64bit(struct lp_build_tgsi_context *bld_base,
-                            enum tgsi_opcode_type type,
-                            LLVMValueRef ptr,
-                            LLVMValueRef ptr2)
-{
-       LLVMBuilderRef builder = bld_base->base.gallivm->builder;
-       LLVMValueRef result;
-
-       result = LLVMGetUndef(LLVMVectorType(LLVMIntTypeInContext(bld_base->base.gallivm->context, 32), bld_base->base.type.length * 2));
-
-       result = LLVMBuildInsertElement(builder,
-                                       result,
-                                       bitcast(bld_base, TGSI_TYPE_UNSIGNED, ptr),
-                                       bld_base->int_bld.zero, "");
-       result = LLVMBuildInsertElement(builder,
-                                       result,
-                                       bitcast(bld_base, TGSI_TYPE_UNSIGNED, ptr2),
-                                       bld_base->int_bld.one, "");
-       return bitcast(bld_base, type, result);
-}
-
-static LLVMValueRef
-emit_array_fetch(struct lp_build_tgsi_context *bld_base,
-                unsigned File, enum tgsi_opcode_type type,
-                struct tgsi_declaration_range range,
-                unsigned swizzle)
-{
-       struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
-       struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
-       LLVMBuilderRef builder = bld_base->base.gallivm->builder;
-
-       unsigned i, size = range.Last - range.First + 1;
-       LLVMTypeRef vec = LLVMVectorType(tgsi2llvmtype(bld_base, type), size);
-       LLVMValueRef result = LLVMGetUndef(vec);
-
-       struct tgsi_full_src_register tmp_reg = {};
-       tmp_reg.Register.File = File;
-
-       for (i = 0; i < size; ++i) {
-               tmp_reg.Register.Index = i + range.First;
-               LLVMValueRef temp = radeon_llvm_emit_fetch(bld_base, &tmp_reg, type, swizzle);
-               result = LLVMBuildInsertElement(builder, result, temp,
-                       lp_build_const_int32(gallivm, i), "array_vector");
-       }
-       return result;
-}
-
-static LLVMValueRef
-load_value_from_array(struct lp_build_tgsi_context *bld_base,
-                     unsigned file,
-                     enum tgsi_opcode_type type,
-                     unsigned swizzle,
-                     unsigned reg_index,
-                     const struct tgsi_ind_register *reg_indirect)
-{
-       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
-       struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
-       struct gallivm_state *gallivm = bld_base->base.gallivm;
-       LLVMBuilderRef builder = gallivm->builder;
-       LLVMValueRef ptr;
-
-       ptr = get_pointer_into_array(ctx, file, swizzle, reg_index, reg_indirect);
-       if (ptr) {
-               LLVMValueRef val = LLVMBuildLoad(builder, ptr, "");
-               if (tgsi_type_is_64bit(type)) {
-                       LLVMValueRef ptr_hi, val_hi;
-                       ptr_hi = LLVMBuildGEP(builder, ptr, &bld_base->uint_bld.one, 1, "");
-                       val_hi = LLVMBuildLoad(builder, ptr_hi, "");
-                       val = radeon_llvm_emit_fetch_64bit(bld_base, type, val, val_hi);
-               }
-
-               return val;
-       } else {
-               struct tgsi_declaration_range range =
-                       get_array_range(bld_base, file, reg_index, reg_indirect);
-               LLVMValueRef index =
-                       emit_array_index(bld, reg_indirect, reg_index - range.First);
-               LLVMValueRef array =
-                       emit_array_fetch(bld_base, file, type, range, swizzle);
-               return LLVMBuildExtractElement(builder, array, index, "");
-       }
-}
-
-static void
-store_value_to_array(struct lp_build_tgsi_context *bld_base,
-                    LLVMValueRef value,
-                    unsigned file,
-                    unsigned chan_index,
-                    unsigned reg_index,
-                    const struct tgsi_ind_register *reg_indirect)
-{
-       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
-       struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
-       struct gallivm_state *gallivm = bld_base->base.gallivm;
-       LLVMBuilderRef builder = gallivm->builder;
-       LLVMValueRef ptr;
-
-       ptr = get_pointer_into_array(ctx, file, chan_index, reg_index, reg_indirect);
-       if (ptr) {
-               LLVMBuildStore(builder, value, ptr);
-       } else {
-               unsigned i, size;
-               struct tgsi_declaration_range range = get_array_range(bld_base, file, reg_index, reg_indirect);
-               LLVMValueRef index = emit_array_index(bld, reg_indirect, reg_index - range.First);
-               LLVMValueRef array =
-                       emit_array_fetch(bld_base, file, TGSI_TYPE_FLOAT, range, chan_index);
-               LLVMValueRef temp_ptr;
-
-               array = LLVMBuildInsertElement(builder, array, value, index, "");
-
-               size = range.Last - range.First + 1;
-               for (i = 0; i < size; ++i) {
-                       switch(file) {
-                       case TGSI_FILE_OUTPUT:
-                               temp_ptr = bld->outputs[i + range.First][chan_index];
-                               break;
-
-                       case TGSI_FILE_TEMPORARY:
-                               if (range.First + i >= ctx->temps_count)
-                                       continue;
-                               temp_ptr = ctx->temps[(i + range.First) * TGSI_NUM_CHANNELS + chan_index];
-                               break;
-
-                       default:
-                               continue;
-                       }
-                       value = LLVMBuildExtractElement(builder, array,
-                               lp_build_const_int32(gallivm, i), "");
-                       LLVMBuildStore(builder, value, temp_ptr);
-               }
-       }
-}
-
-LLVMValueRef radeon_llvm_emit_fetch(struct lp_build_tgsi_context *bld_base,
-                                   const struct tgsi_full_src_register *reg,
-                                   enum tgsi_opcode_type type,
-                                   unsigned swizzle)
-{
-       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
-       struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
-       LLVMBuilderRef builder = bld_base->base.gallivm->builder;
-       LLVMValueRef result = NULL, ptr, ptr2;
-
-       if (swizzle == ~0) {
-               LLVMValueRef values[TGSI_NUM_CHANNELS];
-               unsigned chan;
-               for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
-                       values[chan] = radeon_llvm_emit_fetch(bld_base, reg, type, chan);
-               }
-               return lp_build_gather_values(bld_base->base.gallivm, values,
-                                             TGSI_NUM_CHANNELS);
-       }
-
-       if (reg->Register.Indirect) {
-               LLVMValueRef load = load_value_from_array(bld_base, reg->Register.File, type,
-                               swizzle, reg->Register.Index, &reg->Indirect);
-               return bitcast(bld_base, type, load);
-       }
-
-       switch(reg->Register.File) {
-       case TGSI_FILE_IMMEDIATE: {
-               LLVMTypeRef ctype = tgsi2llvmtype(bld_base, type);
-               if (tgsi_type_is_64bit(type)) {
-                       result = LLVMGetUndef(LLVMVectorType(LLVMIntTypeInContext(bld_base->base.gallivm->context, 32), bld_base->base.type.length * 2));
-                       result = LLVMConstInsertElement(result,
-                                                       bld->immediates[reg->Register.Index][swizzle],
-                                                       bld_base->int_bld.zero);
-                       result = LLVMConstInsertElement(result,
-                                                       bld->immediates[reg->Register.Index][swizzle + 1],
-                                                       bld_base->int_bld.one);
-                       return LLVMConstBitCast(result, ctype);
-               } else {
-                       return LLVMConstBitCast(bld->immediates[reg->Register.Index][swizzle], ctype);
-               }
-       }
-
-       case TGSI_FILE_INPUT: {
-               unsigned index = reg->Register.Index;
-               LLVMValueRef input[4];
-
-               /* I don't think doing this for vertex shaders is beneficial.
-                * For those, we want to make sure the VMEM loads are executed
-                * only once. Fragment shaders don't care much, because
-                * v_interp instructions are much cheaper than VMEM loads.
-                */
-               if (ctx->soa.bld_base.info->processor == PIPE_SHADER_FRAGMENT)
-                       ctx->load_input(ctx, index, &ctx->input_decls[index], input);
-               else
-                       memcpy(input, &ctx->inputs[index * 4], sizeof(input));
-
-               result = input[swizzle];
-
-               if (tgsi_type_is_64bit(type)) {
-                       ptr = result;
-                       ptr2 = input[swizzle + 1];
-                       return radeon_llvm_emit_fetch_64bit(bld_base, type, ptr, ptr2);
-               }
-               break;
-       }
-
-       case TGSI_FILE_TEMPORARY:
-               if (reg->Register.Index >= ctx->temps_count)
-                       return LLVMGetUndef(tgsi2llvmtype(bld_base, type));
-               ptr = ctx->temps[reg->Register.Index * TGSI_NUM_CHANNELS + swizzle];
-               if (tgsi_type_is_64bit(type)) {
-                       ptr2 = ctx->temps[reg->Register.Index * TGSI_NUM_CHANNELS + swizzle + 1];
-                       return radeon_llvm_emit_fetch_64bit(bld_base, type,
-                                                LLVMBuildLoad(builder, ptr, ""),
-                                                LLVMBuildLoad(builder, ptr2, ""));
-               }
-               result = LLVMBuildLoad(builder, ptr, "");
-               break;
-
-       case TGSI_FILE_OUTPUT:
-               ptr = lp_get_output_ptr(bld, reg->Register.Index, swizzle);
-               if (tgsi_type_is_64bit(type)) {
-                       ptr2 = lp_get_output_ptr(bld, reg->Register.Index, swizzle + 1);
-                       return radeon_llvm_emit_fetch_64bit(bld_base, type,
-                                                LLVMBuildLoad(builder, ptr, ""),
-                                                LLVMBuildLoad(builder, ptr2, ""));
-               }
-               result = LLVMBuildLoad(builder, ptr, "");
-               break;
-
-       default:
-               return LLVMGetUndef(tgsi2llvmtype(bld_base, type));
-       }
-
-       return bitcast(bld_base, type, result);
-}
-
-static LLVMValueRef fetch_system_value(struct lp_build_tgsi_context *bld_base,
-                                      const struct tgsi_full_src_register *reg,
-                                      enum tgsi_opcode_type type,
-                                      unsigned swizzle)
-{
-       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
-       struct gallivm_state *gallivm = bld_base->base.gallivm;
-
-       LLVMValueRef cval = ctx->system_values[reg->Register.Index];
-       if (LLVMGetTypeKind(LLVMTypeOf(cval)) == LLVMVectorTypeKind) {
-               cval = LLVMBuildExtractElement(gallivm->builder, cval,
-                                              lp_build_const_int32(gallivm, swizzle), "");
-       }
-       return bitcast(bld_base, type, cval);
-}
-
-static void emit_declaration(struct lp_build_tgsi_context *bld_base,
-                            const struct tgsi_full_declaration *decl)
-{
-       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
-       LLVMBuilderRef builder = bld_base->base.gallivm->builder;
-       unsigned first, last, i;
-       switch(decl->Declaration.File) {
-       case TGSI_FILE_ADDRESS:
-       {
-                unsigned idx;
-               for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
-                       unsigned chan;
-                       for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
-                                ctx->soa.addr[idx][chan] = lp_build_alloca_undef(
-                                       &ctx->gallivm,
-                                       ctx->soa.bld_base.uint_bld.elem_type, "");
-                       }
-               }
-               break;
-       }
-
-       case TGSI_FILE_TEMPORARY:
-       {
-               char name[16] = "";
-               LLVMValueRef array_alloca = NULL;
-               unsigned decl_size;
-               unsigned writemask = decl->Declaration.UsageMask;
-               first = decl->Range.First;
-               last = decl->Range.Last;
-               decl_size = 4 * ((last - first) + 1);
-
-               if (decl->Declaration.Array) {
-                       unsigned id = decl->Array.ArrayID - 1;
-                       unsigned array_size;
-
-                       writemask &= ctx->temp_arrays[id].writemask;
-                       ctx->temp_arrays[id].writemask = writemask;
-                       array_size = ((last - first) + 1) * util_bitcount(writemask);
-
-                       /* If the array has more than 16 elements, store it
-                        * in memory using an alloca that spans the entire
-                        * array.
-                        *
-                        * Otherwise, store each array element individually.
-                        * We will then generate vectors (per-channel, up to
-                        * <16 x float> if the usagemask is a single bit) for
-                        * indirect addressing.
-                        *
-                        * Note that 16 is the number of vector elements that
-                        * LLVM will store in a register, so theoretically an
-                        * array with up to 4 * 16 = 64 elements could be
-                        * handled this way, but whether that's a good idea
-                        * depends on VGPR register pressure elsewhere.
-                        *
-                        * FIXME: We shouldn't need to have the non-alloca
-                        * code path for arrays. LLVM should be smart enough to
-                        * promote allocas into registers when profitable.
-                        *
-                        * LLVM 3.8 crashes with this.
-                        */
-                       if (HAVE_LLVM >= 0x0309 && array_size > 16) {
-                               array_alloca = LLVMBuildAlloca(builder,
-                                       LLVMArrayType(bld_base->base.vec_type,
-                                                     array_size), "array");
-                               ctx->temp_array_allocas[id] = array_alloca;
-                       }
-               }
-
-               if (!ctx->temps_count) {
-                       ctx->temps_count = bld_base->info->file_max[TGSI_FILE_TEMPORARY] + 1;
-                       ctx->temps = MALLOC(TGSI_NUM_CHANNELS * ctx->temps_count * sizeof(LLVMValueRef));
-               }
-               if (!array_alloca) {
-                       for (i = 0; i < decl_size; ++i) {
-#ifdef DEBUG
-                               snprintf(name, sizeof(name), "TEMP%d.%c",
-                                        first + i / 4, "xyzw"[i % 4]);
-#endif
-                               ctx->temps[first * TGSI_NUM_CHANNELS + i] =
-                                       lp_build_alloca_undef(bld_base->base.gallivm,
-                                                             bld_base->base.vec_type,
-                                                             name);
-                       }
-               } else {
-                       LLVMValueRef idxs[2] = {
-                               bld_base->uint_bld.zero,
-                               NULL
-                       };
-                       unsigned j = 0;
-
-                       if (writemask != TGSI_WRITEMASK_XYZW &&
-                           !ctx->undef_alloca) {
-                               /* Create a dummy alloca. We use it so that we
-                                * have a pointer that is safe to load from if
-                                * a shader ever reads from a channel that
-                                * it never writes to.
-                                */
-                               ctx->undef_alloca = lp_build_alloca_undef(
-                                       bld_base->base.gallivm,
-                                       bld_base->base.vec_type, "undef");
-                       }
-
-                       for (i = 0; i < decl_size; ++i) {
-                               LLVMValueRef ptr;
-                               if (writemask & (1 << (i % 4))) {
-#ifdef DEBUG
-                                       snprintf(name, sizeof(name), "TEMP%d.%c",
-                                                first + i / 4, "xyzw"[i % 4]);
-#endif
-                                       idxs[1] = lp_build_const_int32(bld_base->base.gallivm, j);
-                                       ptr = LLVMBuildGEP(builder, array_alloca, idxs, 2, name);
-                                       j++;
-                               } else {
-                                       ptr = ctx->undef_alloca;
-                               }
-                               ctx->temps[first * TGSI_NUM_CHANNELS + i] = ptr;
-                       }
-               }
-               break;
-       }
-       case TGSI_FILE_INPUT:
-       {
-               unsigned idx;
-               for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
-                       if (ctx->load_input &&
-                           ctx->input_decls[idx].Declaration.File != TGSI_FILE_INPUT) {
-                               ctx->input_decls[idx] = *decl;
-
-                               if (bld_base->info->processor != PIPE_SHADER_FRAGMENT)
-                                       ctx->load_input(ctx, idx, decl,
-                                                       &ctx->inputs[idx * 4]);
-                       }
-               }
-       }
-       break;
-
-       case TGSI_FILE_SYSTEM_VALUE:
-       {
-               unsigned idx;
-               for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
-                       ctx->load_system_value(ctx, idx, decl);
-               }
-       }
-       break;
-
-       case TGSI_FILE_OUTPUT:
-       {
-               char name[16] = "";
-               unsigned idx;
-               for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
-                       unsigned chan;
-                       assert(idx < RADEON_LLVM_MAX_OUTPUTS);
-                       if (ctx->soa.outputs[idx][0])
-                               continue;
-                       for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
-#ifdef DEBUG
-                               snprintf(name, sizeof(name), "OUT%d.%c",
-                                        idx, "xyzw"[chan % 4]);
-#endif
-                               ctx->soa.outputs[idx][chan] = lp_build_alloca_undef(
-                                       &ctx->gallivm,
-                                       ctx->soa.bld_base.base.elem_type, name);
-                       }
-               }
-               break;
-       }
-
-       case TGSI_FILE_MEMORY:
-               ctx->declare_memory_region(ctx, decl);
-               break;
-
-       default:
-               break;
-       }
-}
-
-LLVMValueRef radeon_llvm_saturate(struct lp_build_tgsi_context *bld_base,
-                                  LLVMValueRef value)
-{
-       struct lp_build_emit_data clamp_emit_data;
-
-       memset(&clamp_emit_data, 0, sizeof(clamp_emit_data));
-       clamp_emit_data.arg_count = 3;
-       clamp_emit_data.args[0] = value;
-       clamp_emit_data.args[2] = bld_base->base.one;
-       clamp_emit_data.args[1] = bld_base->base.zero;
-
-       return lp_build_emit_llvm(bld_base, TGSI_OPCODE_CLAMP,
-                                 &clamp_emit_data);
-}
-
-void radeon_llvm_emit_store(struct lp_build_tgsi_context *bld_base,
-                           const struct tgsi_full_instruction *inst,
-                           const struct tgsi_opcode_info *info,
-                           LLVMValueRef dst[4])
-{
-       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
-       struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
-       struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
-       const struct tgsi_full_dst_register *reg = &inst->Dst[0];
-       LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
-       LLVMValueRef temp_ptr, temp_ptr2 = NULL;
-       unsigned chan, chan_index;
-       bool is_vec_store = false;
-       enum tgsi_opcode_type dtype = tgsi_opcode_infer_dst_type(inst->Instruction.Opcode);
-
-       if (dst[0]) {
-               LLVMTypeKind k = LLVMGetTypeKind(LLVMTypeOf(dst[0]));
-               is_vec_store = (k == LLVMVectorTypeKind);
-       }
-
-       if (is_vec_store) {
-               LLVMValueRef values[4] = {};
-               TGSI_FOR_EACH_DST0_ENABLED_CHANNEL(inst, chan) {
-                       LLVMValueRef index = lp_build_const_int32(gallivm, chan);
-                       values[chan]  = LLVMBuildExtractElement(gallivm->builder,
-                                                       dst[0], index, "");
-               }
-               bld_base->emit_store(bld_base, inst, info, values);
-               return;
-       }
-
-       TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
-               LLVMValueRef value = dst[chan_index];
-
-               if (tgsi_type_is_64bit(dtype) && (chan_index == 1 || chan_index == 3))
-                       continue;
-               if (inst->Instruction.Saturate)
-                       value = radeon_llvm_saturate(bld_base, value);
-
-               if (reg->Register.File == TGSI_FILE_ADDRESS) {
-                       temp_ptr = bld->addr[reg->Register.Index][chan_index];
-                       LLVMBuildStore(builder, value, temp_ptr);
-                       continue;
-               }
-
-               if (!tgsi_type_is_64bit(dtype))
-                       value = bitcast(bld_base, TGSI_TYPE_FLOAT, value);
-
-               if (reg->Register.Indirect) {
-                       unsigned file = reg->Register.File;
-                       unsigned reg_index = reg->Register.Index;
-                       store_value_to_array(bld_base, value, file, chan_index,
-                                            reg_index, &reg->Indirect);
-               } else {
-                       switch(reg->Register.File) {
-                       case TGSI_FILE_OUTPUT:
-                               temp_ptr = bld->outputs[reg->Register.Index][chan_index];
-                               if (tgsi_type_is_64bit(dtype))
-                                       temp_ptr2 = bld->outputs[reg->Register.Index][chan_index + 1];
-                               break;
-
-                       case TGSI_FILE_TEMPORARY:
-                       {
-                               if (reg->Register.Index >= ctx->temps_count)
-                                       continue;
-
-                               temp_ptr = ctx->temps[ TGSI_NUM_CHANNELS * reg->Register.Index + chan_index];
-                               if (tgsi_type_is_64bit(dtype))
-                                       temp_ptr2 = ctx->temps[ TGSI_NUM_CHANNELS * reg->Register.Index + chan_index + 1];
-
-                               break;
-                       }
-                       default:
-                               return;
-                       }
-                       if (!tgsi_type_is_64bit(dtype))
-                               LLVMBuildStore(builder, value, temp_ptr);
-                       else {
-                               LLVMValueRef ptr = LLVMBuildBitCast(builder, value,
-                                                                   LLVMVectorType(LLVMIntTypeInContext(bld_base->base.gallivm->context, 32), 2), "");
-                               LLVMValueRef val2;
-                               value = LLVMBuildExtractElement(builder, ptr,
-                                                               bld_base->uint_bld.zero, "");
-                               val2 = LLVMBuildExtractElement(builder, ptr,
-                                                               bld_base->uint_bld.one, "");
-
-                               LLVMBuildStore(builder, bitcast(bld_base, TGSI_TYPE_FLOAT, value), temp_ptr);
-                               LLVMBuildStore(builder, bitcast(bld_base, TGSI_TYPE_FLOAT, val2), temp_ptr2);
-                       }
-               }
-       }
-}
-
-static void set_basicblock_name(LLVMBasicBlockRef bb, const char *base, int pc)
-{
-       char buf[32];
-       /* Subtract 1 so that the number shown is that of the corresponding
-        * opcode in the TGSI dump, e.g. an if block has the same suffix as
-        * the instruction number of the corresponding TGSI IF.
-        */
-       snprintf(buf, sizeof(buf), "%s%d", base, pc - 1);
-       LLVMSetValueName(LLVMBasicBlockAsValue(bb), buf);
-}
-
-/* Append a basic block at the level of the parent flow.
- */
-static LLVMBasicBlockRef append_basic_block(struct radeon_llvm_context *ctx,
-                                           const char *name)
-{
-       struct gallivm_state *gallivm = &ctx->gallivm;
-
-       assert(ctx->flow_depth >= 1);
-
-       if (ctx->flow_depth >= 2) {
-               struct radeon_llvm_flow *flow = &ctx->flow[ctx->flow_depth - 2];
-
-               return LLVMInsertBasicBlockInContext(gallivm->context,
-                                                    flow->next_block, name);
-       }
-
-       return LLVMAppendBasicBlockInContext(gallivm->context, ctx->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)
-{
-       if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(builder)))
-                LLVMBuildBr(builder, target);
-}
-
-static void bgnloop_emit(const struct lp_build_tgsi_action *action,
-                        struct lp_build_tgsi_context *bld_base,
-                        struct lp_build_emit_data *emit_data)
-{
-       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
-       struct gallivm_state *gallivm = bld_base->base.gallivm;
-       struct radeon_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", bld_base->pc);
-       LLVMBuildBr(gallivm->builder, flow->loop_entry_block);
-       LLVMPositionBuilderAtEnd(gallivm->builder, flow->loop_entry_block);
-}
-
-static void brk_emit(const struct lp_build_tgsi_action *action,
-                    struct lp_build_tgsi_context *bld_base,
-                    struct lp_build_emit_data *emit_data)
-{
-       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
-       struct gallivm_state *gallivm = bld_base->base.gallivm;
-       struct radeon_llvm_flow *flow = get_innermost_loop(ctx);
-
-       LLVMBuildBr(gallivm->builder, flow->next_block);
-}
-
-static void cont_emit(const struct lp_build_tgsi_action *action,
-                     struct lp_build_tgsi_context *bld_base,
-                     struct lp_build_emit_data *emit_data)
-{
-       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
-       struct gallivm_state *gallivm = bld_base->base.gallivm;
-       struct radeon_llvm_flow *flow = get_innermost_loop(ctx);
-
-       LLVMBuildBr(gallivm->builder, flow->loop_entry_block);
-}
-
-static void else_emit(const struct lp_build_tgsi_action *action,
-                     struct lp_build_tgsi_context *bld_base,
-                     struct lp_build_emit_data *emit_data)
-{
-       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
-       struct gallivm_state *gallivm = bld_base->base.gallivm;
-       struct radeon_llvm_flow *current_branch = get_current_flow(ctx);
-       LLVMBasicBlockRef endif_block;
-
-       assert(!current_branch->loop_entry_block);
-
-       endif_block = append_basic_block(ctx, "ENDIF");
-       emit_default_branch(gallivm->builder, endif_block);
-
-       LLVMPositionBuilderAtEnd(gallivm->builder, current_branch->next_block);
-       set_basicblock_name(current_branch->next_block, "else", bld_base->pc);
-
-       current_branch->next_block = endif_block;
-}
-
-static void endif_emit(const struct lp_build_tgsi_action *action,
-                      struct lp_build_tgsi_context *bld_base,
-                      struct lp_build_emit_data *emit_data)
-{
-       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
-       struct gallivm_state *gallivm = bld_base->base.gallivm;
-       struct radeon_llvm_flow *current_branch = get_current_flow(ctx);
-
-       assert(!current_branch->loop_entry_block);
-
-       emit_default_branch(gallivm->builder, current_branch->next_block);
-       LLVMPositionBuilderAtEnd(gallivm->builder, current_branch->next_block);
-       set_basicblock_name(current_branch->next_block, "endif", bld_base->pc);
-
-       ctx->flow_depth--;
-}
-
-static void endloop_emit(const struct lp_build_tgsi_action *action,
-                        struct lp_build_tgsi_context *bld_base,
-                        struct lp_build_emit_data *emit_data)
-{
-       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
-       struct gallivm_state *gallivm = bld_base->base.gallivm;
-       struct radeon_llvm_flow *current_loop = get_current_flow(ctx);
-
-       assert(current_loop->loop_entry_block);
-
-       emit_default_branch(gallivm->builder, current_loop->loop_entry_block);
-
-       LLVMPositionBuilderAtEnd(gallivm->builder, current_loop->next_block);
-       set_basicblock_name(current_loop->next_block, "endloop", bld_base->pc);
-       ctx->flow_depth--;
-}
-
-static void if_cond_emit(const struct lp_build_tgsi_action *action,
-                        struct lp_build_tgsi_context *bld_base,
-                        struct lp_build_emit_data *emit_data,
-                        LLVMValueRef cond)
-{
-       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
-       struct gallivm_state *gallivm = bld_base->base.gallivm;
-       struct radeon_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", bld_base->pc);
-       LLVMBuildCondBr(gallivm->builder, cond, if_block, flow->next_block);
-       LLVMPositionBuilderAtEnd(gallivm->builder, if_block);
-}
-
-static void if_emit(const struct lp_build_tgsi_action *action,
-                   struct lp_build_tgsi_context *bld_base,
-                   struct lp_build_emit_data *emit_data)
-{
-       struct gallivm_state *gallivm = bld_base->base.gallivm;
-       LLVMValueRef cond;
-
-       cond = LLVMBuildFCmp(gallivm->builder, LLVMRealUNE,
-                       emit_data->args[0],
-                       bld_base->base.zero, "");
-
-       if_cond_emit(action, bld_base, emit_data, cond);
-}
-
-static void uif_emit(const struct lp_build_tgsi_action *action,
-                    struct lp_build_tgsi_context *bld_base,
-                    struct lp_build_emit_data *emit_data)
-{
-       struct gallivm_state *gallivm = bld_base->base.gallivm;
-       LLVMValueRef cond;
-
-       cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE,
-               bitcast(bld_base, TGSI_TYPE_UNSIGNED, emit_data->args[0]),
-                       bld_base->int_bld.zero, "");
-
-       if_cond_emit(action, bld_base, emit_data, cond);
-}
-
-static void emit_immediate(struct lp_build_tgsi_context *bld_base,
-                          const struct tgsi_full_immediate *imm)
-{
-       unsigned i;
-       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
-
-       for (i = 0; i < 4; ++i) {
-               ctx->soa.immediates[ctx->soa.num_immediates][i] =
-                               LLVMConstInt(bld_base->uint_bld.elem_type, imm->u[i].Uint, false   );
-       }
-
-       ctx->soa.num_immediates++;
-}
-
-void radeon_llvm_context_init(struct radeon_llvm_context *ctx, const char *triple,
-                             const struct tgsi_shader_info *info,
-                             const struct tgsi_token *tokens)
-{
-       struct lp_type type;
-
-       /* Initialize the gallivm object:
-        * We are only using the module, context, and builder fields of this struct.
-        * This should be enough for us to be able to pass our gallivm struct to the
-        * helper functions in the gallivm module.
-        */
-       memset(&ctx->gallivm, 0, sizeof (ctx->gallivm));
-       memset(&ctx->soa, 0, sizeof(ctx->soa));
-       ctx->gallivm.context = LLVMContextCreate();
-       ctx->gallivm.module = LLVMModuleCreateWithNameInContext("tgsi",
-                                               ctx->gallivm.context);
-       LLVMSetTarget(ctx->gallivm.module, triple);
-       ctx->gallivm.builder = LLVMCreateBuilderInContext(ctx->gallivm.context);
-
-       struct lp_build_tgsi_context *bld_base = &ctx->soa.bld_base;
-
-       bld_base->info = info;
-
-       if (info && info->array_max[TGSI_FILE_TEMPORARY] > 0) {
-               int size = info->array_max[TGSI_FILE_TEMPORARY];
-
-               ctx->temp_arrays = CALLOC(size, sizeof(ctx->temp_arrays[0]));
-               ctx->temp_array_allocas = CALLOC(size, sizeof(ctx->temp_array_allocas[0]));
-
-               if (tokens)
-                       tgsi_scan_arrays(tokens, TGSI_FILE_TEMPORARY, size,
-                                        ctx->temp_arrays);
-       }
-
-       type.floating = true;
-       type.fixed = false;
-       type.sign = true;
-       type.norm = false;
-       type.width = 32;
-       type.length = 1;
-
-       lp_build_context_init(&bld_base->base, &ctx->gallivm, type);
-       lp_build_context_init(&ctx->soa.bld_base.uint_bld, &ctx->gallivm, lp_uint_type(type));
-       lp_build_context_init(&ctx->soa.bld_base.int_bld, &ctx->gallivm, lp_int_type(type));
-       type.width *= 2;
-       lp_build_context_init(&ctx->soa.bld_base.dbl_bld, &ctx->gallivm, type);
-       lp_build_context_init(&ctx->soa.bld_base.uint64_bld, &ctx->gallivm, lp_uint_type(type));
-       lp_build_context_init(&ctx->soa.bld_base.int64_bld, &ctx->gallivm, lp_int_type(type));
-
-       bld_base->soa = 1;
-       bld_base->emit_store = radeon_llvm_emit_store;
-       bld_base->emit_swizzle = emit_swizzle;
-       bld_base->emit_declaration = emit_declaration;
-       bld_base->emit_immediate = emit_immediate;
-
-       bld_base->emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = radeon_llvm_emit_fetch;
-       bld_base->emit_fetch_funcs[TGSI_FILE_INPUT] = radeon_llvm_emit_fetch;
-       bld_base->emit_fetch_funcs[TGSI_FILE_TEMPORARY] = radeon_llvm_emit_fetch;
-       bld_base->emit_fetch_funcs[TGSI_FILE_OUTPUT] = radeon_llvm_emit_fetch;
-       bld_base->emit_fetch_funcs[TGSI_FILE_SYSTEM_VALUE] = fetch_system_value;
-
-       /* metadata allowing 2.5 ULP */
-       ctx->fpmath_md_kind = LLVMGetMDKindIDInContext(ctx->gallivm.context,
-                                                      "fpmath", 6);
-       LLVMValueRef arg = lp_build_const_float(&ctx->gallivm, 2.5);
-       ctx->fpmath_md_2p5_ulp = LLVMMDNodeInContext(ctx->gallivm.context,
-                                                    &arg, 1);
-
-       /* Allocate outputs */
-       ctx->soa.outputs = ctx->outputs;
-
-       bld_base->op_actions[TGSI_OPCODE_BGNLOOP].emit = bgnloop_emit;
-       bld_base->op_actions[TGSI_OPCODE_BRK].emit = brk_emit;
-       bld_base->op_actions[TGSI_OPCODE_CONT].emit = cont_emit;
-       bld_base->op_actions[TGSI_OPCODE_IF].emit = if_emit;
-       bld_base->op_actions[TGSI_OPCODE_UIF].emit = uif_emit;
-       bld_base->op_actions[TGSI_OPCODE_ELSE].emit = else_emit;
-       bld_base->op_actions[TGSI_OPCODE_ENDIF].emit = endif_emit;
-       bld_base->op_actions[TGSI_OPCODE_ENDLOOP].emit = endloop_emit;
-}
-
-void radeon_llvm_create_func(struct radeon_llvm_context *ctx,
-                            LLVMTypeRef *return_types, unsigned num_return_elems,
-                            LLVMTypeRef *ParamTypes, unsigned ParamCount)
-{
-       LLVMTypeRef main_fn_type, ret_type;
-       LLVMBasicBlockRef main_fn_body;
-
-       if (num_return_elems)
-               ret_type = LLVMStructTypeInContext(ctx->gallivm.context,
-                                                  return_types,
-                                                  num_return_elems, true);
-       else
-               ret_type = LLVMVoidTypeInContext(ctx->gallivm.context);
-
-       /* Setup the function */
-       ctx->return_type = ret_type;
-       main_fn_type = LLVMFunctionType(ret_type, ParamTypes, ParamCount, 0);
-       ctx->main_fn = LLVMAddFunction(ctx->gallivm.module, "main", main_fn_type);
-       main_fn_body = LLVMAppendBasicBlockInContext(ctx->gallivm.context,
-                       ctx->main_fn, "main_body");
-       LLVMPositionBuilderAtEnd(ctx->gallivm.builder, main_fn_body);
-}
-
-void radeon_llvm_finalize_module(struct radeon_llvm_context *ctx,
-                                bool run_verifier)
-{
-       struct gallivm_state *gallivm = ctx->soa.bld_base.base.gallivm;
-       const char *triple = LLVMGetTarget(gallivm->module);
-       LLVMTargetLibraryInfoRef target_library_info;
-
-       /* Create the pass manager */
-       gallivm->passmgr = LLVMCreateFunctionPassManagerForModule(
-                                                       gallivm->module);
-
-       target_library_info = gallivm_create_target_library_info(triple);
-       LLVMAddTargetLibraryInfo(target_library_info, gallivm->passmgr);
-
-       if (run_verifier)
-               LLVMAddVerifierPass(gallivm->passmgr);
-
-       /* This pass should eliminate all the load and store instructions */
-       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
-
-       /* Add some optimization passes */
-       LLVMAddScalarReplAggregatesPass(gallivm->passmgr);
-       LLVMAddLICMPass(gallivm->passmgr);
-       LLVMAddAggressiveDCEPass(gallivm->passmgr);
-       LLVMAddCFGSimplificationPass(gallivm->passmgr);
-       LLVMAddInstructionCombiningPass(gallivm->passmgr);
-
-       /* Run the pass */
-       LLVMInitializeFunctionPassManager(gallivm->passmgr);
-       LLVMRunFunctionPassManager(gallivm->passmgr, ctx->main_fn);
-       LLVMFinalizeFunctionPassManager(gallivm->passmgr);
-
-       LLVMDisposeBuilder(gallivm->builder);
-       LLVMDisposePassManager(gallivm->passmgr);
-       gallivm_dispose_target_library_info(target_library_info);
-}
-
-void radeon_llvm_dispose(struct radeon_llvm_context *ctx)
-{
-       LLVMDisposeModule(ctx->soa.bld_base.base.gallivm->module);
-       LLVMContextDispose(ctx->soa.bld_base.base.gallivm->context);
-       FREE(ctx->temp_arrays);
-       ctx->temp_arrays = NULL;
-       FREE(ctx->temp_array_allocas);
-       ctx->temp_array_allocas = NULL;
-       FREE(ctx->temps);
-       ctx->temps = NULL;
-       ctx->temps_count = 0;
-       FREE(ctx->flow);
-       ctx->flow = NULL;
-       ctx->flow_depth_max = 0;
-}
index a16f19c7ffb10f77524a5ad01b9ea0a71daf0880..727a9cccfc20024f24977bf2577599a88b6bab2f 100644 (file)
@@ -17,6 +17,7 @@ C_SOURCES := \
        si_shader.h \
        si_shader_internal.h \
        si_shader_tgsi_alu.c \
+       si_shader_tgsi_setup.c \
        si_state.c \
        si_state_draw.c \
        si_state_shaders.c \
index 5a3f1014cf2e558550eba2493c2f72179c8d0405..7924375364727d1981845e5dc64af6e3ebf2f65f 100644 (file)
@@ -23,9 +23,9 @@
 
 #include "si_pipe.h"
 #include "si_public.h"
+#include "si_shader_internal.h"
 #include "sid.h"
 
-#include "radeon/radeon_llvm_emit.h"
 #include "radeon/radeon_uvd.h"
 #include "util/u_memory.h"
 #include "util/u_suballoc.h"
index 37a812a2840049a16a6a65940483ada450959576..194da3194f8a1a7329fab798f2cdab5f6b2f4e80 100644 (file)
@@ -33,9 +33,7 @@
 #include "gallivm/lp_bld_arit.h"
 #include "gallivm/lp_bld_flow.h"
 #include "gallivm/lp_bld_misc.h"
-#include "radeon/radeon_llvm.h"
 #include "radeon/radeon_elf_util.h"
-#include "radeon/radeon_llvm_emit.h"
 #include "util/u_memory.h"
 #include "util/u_string.h"
 #include "tgsi/tgsi_build.h"
index 44dd5fdc3b590158453ce542e3d731fd3b8c2921..b46cc1c379d5f909bf0fcd56f0b24de0f10b55c3 100644 (file)
 #include "si_shader.h"
 #include "gallivm/lp_bld_init.h"
 #include "gallivm/lp_bld_tgsi.h"
+#include "tgsi/tgsi_parse.h"
+
+#include <llvm-c/Core.h>
+#include <llvm-c/TargetMachine.h>
+
+struct pipe_debug_callback;
+struct radeon_shader_binary;
+
+#define RADEON_LLVM_MAX_INPUT_SLOTS 32
+#define RADEON_LLVM_MAX_INPUTS 32 * 4
+#define RADEON_LLVM_MAX_OUTPUTS 32 * 4
+
+#define RADEON_LLVM_INITIAL_CF_DEPTH 4
+
+#define RADEON_LLVM_MAX_SYSTEM_VALUES 4
+
+struct radeon_llvm_flow;
+
+struct radeon_llvm_context {
+       struct lp_build_tgsi_soa_context soa;
+
+       /*=== Front end configuration ===*/
+
+       /* Instructions that are not described by any of the TGSI opcodes. */
+
+       /** This function is responsible for initilizing the inputs array and will be
+         * called once for each input declared in the TGSI shader.
+         */
+       void (*load_input)(struct radeon_llvm_context *,
+                          unsigned input_index,
+                          const struct tgsi_full_declaration *decl,
+                          LLVMValueRef out[4]);
+
+       void (*load_system_value)(struct radeon_llvm_context *,
+                                 unsigned index,
+                                 const struct tgsi_full_declaration *decl);
+
+       void (*declare_memory_region)(struct radeon_llvm_context *,
+                                     const struct tgsi_full_declaration *decl);
+
+       /** This array contains the input values for the shader.  Typically these
+         * values will be in the form of a target intrinsic that will inform the
+         * backend how to load the actual inputs to the shader.
+         */
+       struct tgsi_full_declaration input_decls[RADEON_LLVM_MAX_INPUT_SLOTS];
+       LLVMValueRef inputs[RADEON_LLVM_MAX_INPUTS];
+       LLVMValueRef outputs[RADEON_LLVM_MAX_OUTPUTS][TGSI_NUM_CHANNELS];
+
+       /** This pointer is used to contain the temporary values.
+         * The amount of temporary used in tgsi can't be bound to a max value and
+         * thus we must allocate this array at runtime.
+         */
+       LLVMValueRef *temps;
+       unsigned temps_count;
+       LLVMValueRef system_values[RADEON_LLVM_MAX_SYSTEM_VALUES];
+
+       /*=== Private Members ===*/
+
+       struct radeon_llvm_flow *flow;
+       unsigned flow_depth;
+       unsigned flow_depth_max;
+
+       struct tgsi_array_info *temp_arrays;
+       LLVMValueRef *temp_array_allocas;
+
+       LLVMValueRef undef_alloca;
+
+       LLVMValueRef main_fn;
+       LLVMTypeRef return_type;
+
+       unsigned fpmath_md_kind;
+       LLVMValueRef fpmath_md_2p5_ulp;
+
+       struct gallivm_state gallivm;
+};
+
+static inline struct radeon_llvm_context *
+radeon_llvm_context(struct lp_build_tgsi_context *bld_base)
+{
+       return (struct radeon_llvm_context*)bld_base;
+}
+
+void radeon_llvm_add_attribute(LLVMValueRef F, const char *name, int value);
+void radeon_llvm_shader_type(LLVMValueRef F, unsigned type);
+
+LLVMTargetRef radeon_llvm_get_r600_target(const char *triple);
+
+unsigned radeon_llvm_compile(LLVMModuleRef M, struct radeon_shader_binary *binary,
+                            LLVMTargetMachineRef tm,
+                            struct pipe_debug_callback *debug);
+
+LLVMTypeRef tgsi2llvmtype(struct lp_build_tgsi_context *bld_base,
+                         enum tgsi_opcode_type type);
+
+LLVMValueRef bitcast(struct lp_build_tgsi_context *bld_base,
+                    enum tgsi_opcode_type type, LLVMValueRef value);
+
+LLVMValueRef radeon_llvm_bound_index(struct radeon_llvm_context *ctx,
+                                    LLVMValueRef index,
+                                    unsigned num);
+
+void radeon_llvm_context_init(struct radeon_llvm_context *ctx,
+                              const char *triple,
+                             const struct tgsi_shader_info *info,
+                             const struct tgsi_token *tokens);
+
+void radeon_llvm_create_func(struct radeon_llvm_context *ctx,
+                            LLVMTypeRef *return_types, unsigned num_return_elems,
+                            LLVMTypeRef *ParamTypes, unsigned ParamCount);
+
+void radeon_llvm_dispose(struct radeon_llvm_context *ctx);
+
+void radeon_llvm_finalize_module(struct radeon_llvm_context *ctx,
+                                bool run_verifier);
+
+LLVMValueRef radeon_llvm_emit_fetch_64bit(struct lp_build_tgsi_context *bld_base,
+                                         enum tgsi_opcode_type type,
+                                         LLVMValueRef ptr,
+                                         LLVMValueRef ptr2);
+
+LLVMValueRef radeon_llvm_saturate(struct lp_build_tgsi_context *bld_base,
+                                  LLVMValueRef value);
+
+LLVMValueRef radeon_llvm_emit_fetch(struct lp_build_tgsi_context *bld_base,
+                                   const struct tgsi_full_src_register *reg,
+                                   enum tgsi_opcode_type type,
+                                   unsigned swizzle);
+
+void radeon_llvm_emit_store(struct lp_build_tgsi_context *bld_base,
+                           const struct tgsi_full_instruction *inst,
+                           const struct tgsi_opcode_info *info,
+                           LLVMValueRef dst[4]);
 
 void si_shader_context_init_alu(struct lp_build_tgsi_context *bld_base);
 void si_prepare_cube_coords(struct lp_build_tgsi_context *bld_base,
index 192a2cbf0aef678382fd8ee44e2626428d51e482..75ceb577b9c6044454aeef438baf87807098c439 100644 (file)
@@ -26,7 +26,6 @@
 #include "gallivm/lp_bld_intr.h"
 #include "gallivm/lp_bld_gather.h"
 #include "tgsi/tgsi_parse.h"
-#include "radeon/radeon_llvm.h"
 
 static void kill_if_fetch_args(struct lp_build_tgsi_context *bld_base,
                               struct lp_build_emit_data *emit_data)
diff --git a/src/gallium/drivers/radeonsi/si_shader_tgsi_setup.c b/src/gallium/drivers/radeonsi/si_shader_tgsi_setup.c
new file mode 100644 (file)
index 0000000..e1534c7
--- /dev/null
@@ -0,0 +1,1379 @@
+/*
+ * Copyright 2016 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * on the rights to use, copy, modify, merge, publish, distribute, sub
+ * license, and/or sell copies of the Software, and to permit persons to whom
+ * the Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
+ * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
+ * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
+ * USE OR OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+#include "si_shader_internal.h"
+#include "radeon/radeon_elf_util.h"
+
+#include "gallivm/lp_bld_const.h"
+#include "gallivm/lp_bld_gather.h"
+#include "gallivm/lp_bld_flow.h"
+#include "gallivm/lp_bld_init.h"
+#include "gallivm/lp_bld_intr.h"
+#include "gallivm/lp_bld_misc.h"
+#include "gallivm/lp_bld_swizzle.h"
+#include "tgsi/tgsi_info.h"
+#include "tgsi/tgsi_parse.h"
+#include "util/u_math.h"
+#include "util/u_memory.h"
+#include "util/u_debug.h"
+
+#include <stdio.h>
+#include <llvm-c/Transforms/Scalar.h>
+
+/* Data for if/else/endif and bgnloop/endloop control flow structures.
+ */
+struct radeon_llvm_flow {
+       /* Loop exit or next part of if/else/endif. */
+       LLVMBasicBlockRef next_block;
+       LLVMBasicBlockRef loop_entry_block;
+};
+
+#define CPU_STRING_LEN 30
+#define FS_STRING_LEN 30
+#define TRIPLE_STRING_LEN 7
+
+/**
+ * Shader types for the LLVM backend.
+ */
+enum radeon_llvm_shader_type {
+       RADEON_LLVM_SHADER_PS = 0,
+       RADEON_LLVM_SHADER_VS = 1,
+       RADEON_LLVM_SHADER_GS = 2,
+       RADEON_LLVM_SHADER_CS = 3,
+};
+
+enum radeon_llvm_calling_convention {
+       RADEON_LLVM_AMDGPU_VS = 87,
+       RADEON_LLVM_AMDGPU_GS = 88,
+       RADEON_LLVM_AMDGPU_PS = 89,
+       RADEON_LLVM_AMDGPU_CS = 90,
+};
+
+void radeon_llvm_add_attribute(LLVMValueRef F, const char *name, int value)
+{
+       char str[16];
+
+       snprintf(str, sizeof(str), "%i", value);
+       LLVMAddTargetDependentFunctionAttr(F, name, str);
+}
+
+/**
+ * Set the shader type we want to compile
+ *
+ * @param type shader type to set
+ */
+void radeon_llvm_shader_type(LLVMValueRef F, unsigned type)
+{
+       enum radeon_llvm_shader_type llvm_type;
+       enum radeon_llvm_calling_convention calling_conv;
+
+       switch (type) {
+       case PIPE_SHADER_VERTEX:
+       case PIPE_SHADER_TESS_CTRL:
+       case PIPE_SHADER_TESS_EVAL:
+               llvm_type = RADEON_LLVM_SHADER_VS;
+               calling_conv = RADEON_LLVM_AMDGPU_VS;
+               break;
+       case PIPE_SHADER_GEOMETRY:
+               llvm_type = RADEON_LLVM_SHADER_GS;
+               calling_conv = RADEON_LLVM_AMDGPU_GS;
+               break;
+       case PIPE_SHADER_FRAGMENT:
+               llvm_type = RADEON_LLVM_SHADER_PS;
+               calling_conv = RADEON_LLVM_AMDGPU_PS;
+               break;
+       case PIPE_SHADER_COMPUTE:
+               llvm_type = RADEON_LLVM_SHADER_CS;
+               calling_conv = RADEON_LLVM_AMDGPU_CS;
+               break;
+       default:
+               unreachable("Unhandle shader type");
+       }
+
+       if (HAVE_LLVM >= 0x309)
+               LLVMSetFunctionCallConv(F, calling_conv);
+       else
+               radeon_llvm_add_attribute(F, "ShaderType", llvm_type);
+}
+
+static void init_r600_target()
+{
+       gallivm_init_llvm_targets();
+#if HAVE_LLVM < 0x0307
+       LLVMInitializeR600TargetInfo();
+       LLVMInitializeR600Target();
+       LLVMInitializeR600TargetMC();
+       LLVMInitializeR600AsmPrinter();
+#else
+       LLVMInitializeAMDGPUTargetInfo();
+       LLVMInitializeAMDGPUTarget();
+       LLVMInitializeAMDGPUTargetMC();
+       LLVMInitializeAMDGPUAsmPrinter();
+
+#endif
+}
+
+static once_flag init_r600_target_once_flag = ONCE_FLAG_INIT;
+
+LLVMTargetRef radeon_llvm_get_r600_target(const char *triple)
+{
+       LLVMTargetRef target = NULL;
+       char *err_message = NULL;
+
+       call_once(&init_r600_target_once_flag, init_r600_target);
+
+       if (LLVMGetTargetFromTriple(triple, &target, &err_message)) {
+               fprintf(stderr, "Cannot find target for triple %s ", triple);
+               if (err_message) {
+                       fprintf(stderr, "%s\n", err_message);
+               }
+               LLVMDisposeMessage(err_message);
+               return NULL;
+       }
+       return target;
+}
+
+struct radeon_llvm_diagnostics {
+       struct pipe_debug_callback *debug;
+       unsigned retval;
+};
+
+static void radeonDiagnosticHandler(LLVMDiagnosticInfoRef di, void *context)
+{
+       struct radeon_llvm_diagnostics *diag = (struct radeon_llvm_diagnostics *)context;
+       LLVMDiagnosticSeverity severity = LLVMGetDiagInfoSeverity(di);
+       char *description = LLVMGetDiagInfoDescription(di);
+       const char *severity_str = NULL;
+
+       switch (severity) {
+       case LLVMDSError:
+               severity_str = "error";
+               break;
+       case LLVMDSWarning:
+               severity_str = "warning";
+               break;
+       case LLVMDSRemark:
+               severity_str = "remark";
+               break;
+       case LLVMDSNote:
+               severity_str = "note";
+               break;
+       default:
+               severity_str = "unknown";
+       }
+
+       pipe_debug_message(diag->debug, SHADER_INFO,
+                          "LLVM diagnostic (%s): %s", severity_str, description);
+
+       if (severity == LLVMDSError) {
+               diag->retval = 1;
+               fprintf(stderr,"LLVM triggered Diagnostic Handler: %s\n", description);
+       }
+
+       LLVMDisposeMessage(description);
+}
+
+/**
+ * Compile an LLVM module to machine code.
+ *
+ * @returns 0 for success, 1 for failure
+ */
+unsigned radeon_llvm_compile(LLVMModuleRef M, struct radeon_shader_binary *binary,
+                            LLVMTargetMachineRef tm,
+                            struct pipe_debug_callback *debug)
+{
+       struct radeon_llvm_diagnostics diag;
+       char *err;
+       LLVMContextRef llvm_ctx;
+       LLVMMemoryBufferRef out_buffer;
+       unsigned buffer_size;
+       const char *buffer_data;
+       LLVMBool mem_err;
+
+       diag.debug = debug;
+       diag.retval = 0;
+
+       /* Setup Diagnostic Handler*/
+       llvm_ctx = LLVMGetModuleContext(M);
+
+       LLVMContextSetDiagnosticHandler(llvm_ctx, radeonDiagnosticHandler, &diag);
+
+       /* Compile IR*/
+       mem_err = LLVMTargetMachineEmitToMemoryBuffer(tm, M, LLVMObjectFile, &err,
+                                                                &out_buffer);
+
+       /* Process Errors/Warnings */
+       if (mem_err) {
+               fprintf(stderr, "%s: %s", __FUNCTION__, err);
+               pipe_debug_message(debug, SHADER_INFO,
+                                  "LLVM emit error: %s", err);
+               FREE(err);
+               diag.retval = 1;
+               goto out;
+       }
+
+       /* Extract Shader Code*/
+       buffer_size = LLVMGetBufferSize(out_buffer);
+       buffer_data = LLVMGetBufferStart(out_buffer);
+
+       radeon_elf_read(buffer_data, buffer_size, binary);
+
+       /* Clean up */
+       LLVMDisposeMemoryBuffer(out_buffer);
+
+out:
+       if (diag.retval != 0)
+               pipe_debug_message(debug, SHADER_INFO, "LLVM compile failed");
+       return diag.retval;
+}
+
+LLVMTypeRef tgsi2llvmtype(struct lp_build_tgsi_context *bld_base,
+                         enum tgsi_opcode_type type)
+{
+       LLVMContextRef ctx = bld_base->base.gallivm->context;
+
+       switch (type) {
+       case TGSI_TYPE_UNSIGNED:
+       case TGSI_TYPE_SIGNED:
+               return LLVMInt32TypeInContext(ctx);
+       case TGSI_TYPE_UNSIGNED64:
+       case TGSI_TYPE_SIGNED64:
+               return LLVMInt64TypeInContext(ctx);
+       case TGSI_TYPE_DOUBLE:
+               return LLVMDoubleTypeInContext(ctx);
+       case TGSI_TYPE_UNTYPED:
+       case TGSI_TYPE_FLOAT:
+               return LLVMFloatTypeInContext(ctx);
+       default: break;
+       }
+       return 0;
+}
+
+LLVMValueRef bitcast(struct lp_build_tgsi_context *bld_base,
+                    enum tgsi_opcode_type type, LLVMValueRef value)
+{
+       LLVMBuilderRef builder = bld_base->base.gallivm->builder;
+       LLVMTypeRef dst_type = tgsi2llvmtype(bld_base, type);
+
+       if (dst_type)
+               return LLVMBuildBitCast(builder, value, dst_type, "");
+       else
+               return value;
+}
+
+/**
+ * Return a value that is equal to the given i32 \p index if it lies in [0,num)
+ * or an undefined value in the same interval otherwise.
+ */
+LLVMValueRef radeon_llvm_bound_index(struct radeon_llvm_context *ctx,
+                                    LLVMValueRef index,
+                                    unsigned num)
+{
+       struct gallivm_state *gallivm = &ctx->gallivm;
+       LLVMBuilderRef builder = gallivm->builder;
+       LLVMValueRef c_max = lp_build_const_int32(gallivm, num - 1);
+       LLVMValueRef cc;
+
+       if (util_is_power_of_two(num)) {
+               index = LLVMBuildAnd(builder, index, c_max, "");
+       } else {
+               /* In theory, this MAX pattern should result in code that is
+                * as good as the bit-wise AND above.
+                *
+                * In practice, LLVM generates worse code (at the time of
+                * writing), because its value tracking is not strong enough.
+                */
+               cc = LLVMBuildICmp(builder, LLVMIntULE, index, c_max, "");
+               index = LLVMBuildSelect(builder, cc, index, c_max, "");
+       }
+
+       return index;
+}
+
+static struct radeon_llvm_flow *
+get_current_flow(struct radeon_llvm_context *ctx)
+{
+       if (ctx->flow_depth > 0)
+               return &ctx->flow[ctx->flow_depth - 1];
+       return NULL;
+}
+
+static struct radeon_llvm_flow *
+get_innermost_loop(struct radeon_llvm_context *ctx)
+{
+       for (unsigned i = ctx->flow_depth; i > 0; --i) {
+               if (ctx->flow[i - 1].loop_entry_block)
+                       return &ctx->flow[i - 1];
+       }
+       return NULL;
+}
+
+static struct radeon_llvm_flow *
+push_flow(struct radeon_llvm_context *ctx)
+{
+       struct radeon_llvm_flow *flow;
+
+       if (ctx->flow_depth >= ctx->flow_depth_max) {
+               unsigned new_max = MAX2(ctx->flow_depth << 1, RADEON_LLVM_INITIAL_CF_DEPTH);
+               ctx->flow = REALLOC(ctx->flow,
+                                   ctx->flow_depth_max * sizeof(*ctx->flow),
+                                   new_max * sizeof(*ctx->flow));
+               ctx->flow_depth_max = new_max;
+       }
+
+       flow = &ctx->flow[ctx->flow_depth];
+       ctx->flow_depth++;
+
+       flow->next_block = NULL;
+       flow->loop_entry_block = NULL;
+       return flow;
+}
+
+static LLVMValueRef emit_swizzle(struct lp_build_tgsi_context *bld_base,
+                                LLVMValueRef value,
+                                unsigned swizzle_x,
+                                unsigned swizzle_y,
+                                unsigned swizzle_z,
+                                unsigned swizzle_w)
+{
+       LLVMValueRef swizzles[4];
+       LLVMTypeRef i32t =
+               LLVMInt32TypeInContext(bld_base->base.gallivm->context);
+
+       swizzles[0] = LLVMConstInt(i32t, swizzle_x, 0);
+       swizzles[1] = LLVMConstInt(i32t, swizzle_y, 0);
+       swizzles[2] = LLVMConstInt(i32t, swizzle_z, 0);
+       swizzles[3] = LLVMConstInt(i32t, swizzle_w, 0);
+
+       return LLVMBuildShuffleVector(bld_base->base.gallivm->builder,
+                                     value,
+                                     LLVMGetUndef(LLVMTypeOf(value)),
+                                     LLVMConstVector(swizzles, 4), "");
+}
+
+/**
+ * Return the description of the array covering the given temporary register
+ * index.
+ */
+static unsigned
+get_temp_array_id(struct lp_build_tgsi_context *bld_base,
+                 unsigned reg_index,
+                 const struct tgsi_ind_register *reg)
+{
+       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
+       unsigned num_arrays = ctx->soa.bld_base.info->array_max[TGSI_FILE_TEMPORARY];
+       unsigned i;
+
+       if (reg && reg->ArrayID > 0 && reg->ArrayID <= num_arrays)
+               return reg->ArrayID;
+
+       for (i = 0; i < num_arrays; i++) {
+               const struct tgsi_array_info *array = &ctx->temp_arrays[i];
+
+               if (reg_index >= array->range.First && reg_index <= array->range.Last)
+                       return i + 1;
+       }
+
+       return 0;
+}
+
+static struct tgsi_declaration_range
+get_array_range(struct lp_build_tgsi_context *bld_base,
+               unsigned File, unsigned reg_index,
+               const struct tgsi_ind_register *reg)
+{
+       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
+       struct tgsi_declaration_range range;
+
+       if (File == TGSI_FILE_TEMPORARY) {
+               unsigned array_id = get_temp_array_id(bld_base, reg_index, reg);
+               if (array_id)
+                       return ctx->temp_arrays[array_id - 1].range;
+       }
+
+       range.First = 0;
+       range.Last = bld_base->info->file_max[File];
+       return range;
+}
+
+static LLVMValueRef
+emit_array_index(struct lp_build_tgsi_soa_context *bld,
+                const struct tgsi_ind_register *reg,
+                unsigned offset)
+{
+       struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
+
+       if (!reg) {
+               return lp_build_const_int32(gallivm, offset);
+       }
+       LLVMValueRef addr = LLVMBuildLoad(gallivm->builder, bld->addr[reg->Index][reg->Swizzle], "");
+       return LLVMBuildAdd(gallivm->builder, addr, lp_build_const_int32(gallivm, offset), "");
+}
+
+/**
+ * For indirect registers, construct a pointer directly to the requested
+ * element using getelementptr if possible.
+ *
+ * Returns NULL if the insertelement/extractelement fallback for array access
+ * must be used.
+ */
+static LLVMValueRef
+get_pointer_into_array(struct radeon_llvm_context *ctx,
+                      unsigned file,
+                      unsigned swizzle,
+                      unsigned reg_index,
+                      const struct tgsi_ind_register *reg_indirect)
+{
+       unsigned array_id;
+       struct tgsi_array_info *array;
+       struct gallivm_state *gallivm = ctx->soa.bld_base.base.gallivm;
+       LLVMBuilderRef builder = gallivm->builder;
+       LLVMValueRef idxs[2];
+       LLVMValueRef index;
+       LLVMValueRef alloca;
+
+       if (file != TGSI_FILE_TEMPORARY)
+               return NULL;
+
+       array_id = get_temp_array_id(&ctx->soa.bld_base, reg_index, reg_indirect);
+       if (!array_id)
+               return NULL;
+
+       alloca = ctx->temp_array_allocas[array_id - 1];
+       if (!alloca)
+               return NULL;
+
+       array = &ctx->temp_arrays[array_id - 1];
+
+       if (!(array->writemask & (1 << swizzle)))
+               return ctx->undef_alloca;
+
+       index = emit_array_index(&ctx->soa, reg_indirect,
+                                reg_index - ctx->temp_arrays[array_id - 1].range.First);
+
+       /* Ensure that the index is within a valid range, to guard against
+        * VM faults and overwriting critical data (e.g. spilled resource
+        * descriptors).
+        *
+        * TODO It should be possible to avoid the additional instructions
+        * if LLVM is changed so that it guarantuees:
+        * 1. the scratch space descriptor isolates the current wave (this
+        *    could even save the scratch offset SGPR at the cost of an
+        *    additional SALU instruction)
+        * 2. the memory for allocas must be allocated at the _end_ of the
+        *    scratch space (after spilled registers)
+        */
+       index = radeon_llvm_bound_index(ctx, index, array->range.Last - array->range.First + 1);
+
+       index = LLVMBuildMul(
+               builder, index,
+               lp_build_const_int32(gallivm, util_bitcount(array->writemask)),
+               "");
+       index = LLVMBuildAdd(
+               builder, index,
+               lp_build_const_int32(
+                       gallivm,
+                       util_bitcount(array->writemask & ((1 << swizzle) - 1))),
+               "");
+       idxs[0] = ctx->soa.bld_base.uint_bld.zero;
+       idxs[1] = index;
+       return LLVMBuildGEP(builder, alloca, idxs, 2, "");
+}
+
+LLVMValueRef
+radeon_llvm_emit_fetch_64bit(struct lp_build_tgsi_context *bld_base,
+                            enum tgsi_opcode_type type,
+                            LLVMValueRef ptr,
+                            LLVMValueRef ptr2)
+{
+       LLVMBuilderRef builder = bld_base->base.gallivm->builder;
+       LLVMValueRef result;
+
+       result = LLVMGetUndef(LLVMVectorType(LLVMIntTypeInContext(bld_base->base.gallivm->context, 32), bld_base->base.type.length * 2));
+
+       result = LLVMBuildInsertElement(builder,
+                                       result,
+                                       bitcast(bld_base, TGSI_TYPE_UNSIGNED, ptr),
+                                       bld_base->int_bld.zero, "");
+       result = LLVMBuildInsertElement(builder,
+                                       result,
+                                       bitcast(bld_base, TGSI_TYPE_UNSIGNED, ptr2),
+                                       bld_base->int_bld.one, "");
+       return bitcast(bld_base, type, result);
+}
+
+static LLVMValueRef
+emit_array_fetch(struct lp_build_tgsi_context *bld_base,
+                unsigned File, enum tgsi_opcode_type type,
+                struct tgsi_declaration_range range,
+                unsigned swizzle)
+{
+       struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
+       struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
+       LLVMBuilderRef builder = bld_base->base.gallivm->builder;
+
+       unsigned i, size = range.Last - range.First + 1;
+       LLVMTypeRef vec = LLVMVectorType(tgsi2llvmtype(bld_base, type), size);
+       LLVMValueRef result = LLVMGetUndef(vec);
+
+       struct tgsi_full_src_register tmp_reg = {};
+       tmp_reg.Register.File = File;
+
+       for (i = 0; i < size; ++i) {
+               tmp_reg.Register.Index = i + range.First;
+               LLVMValueRef temp = radeon_llvm_emit_fetch(bld_base, &tmp_reg, type, swizzle);
+               result = LLVMBuildInsertElement(builder, result, temp,
+                       lp_build_const_int32(gallivm, i), "array_vector");
+       }
+       return result;
+}
+
+static LLVMValueRef
+load_value_from_array(struct lp_build_tgsi_context *bld_base,
+                     unsigned file,
+                     enum tgsi_opcode_type type,
+                     unsigned swizzle,
+                     unsigned reg_index,
+                     const struct tgsi_ind_register *reg_indirect)
+{
+       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
+       struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
+       struct gallivm_state *gallivm = bld_base->base.gallivm;
+       LLVMBuilderRef builder = gallivm->builder;
+       LLVMValueRef ptr;
+
+       ptr = get_pointer_into_array(ctx, file, swizzle, reg_index, reg_indirect);
+       if (ptr) {
+               LLVMValueRef val = LLVMBuildLoad(builder, ptr, "");
+               if (tgsi_type_is_64bit(type)) {
+                       LLVMValueRef ptr_hi, val_hi;
+                       ptr_hi = LLVMBuildGEP(builder, ptr, &bld_base->uint_bld.one, 1, "");
+                       val_hi = LLVMBuildLoad(builder, ptr_hi, "");
+                       val = radeon_llvm_emit_fetch_64bit(bld_base, type, val, val_hi);
+               }
+
+               return val;
+       } else {
+               struct tgsi_declaration_range range =
+                       get_array_range(bld_base, file, reg_index, reg_indirect);
+               LLVMValueRef index =
+                       emit_array_index(bld, reg_indirect, reg_index - range.First);
+               LLVMValueRef array =
+                       emit_array_fetch(bld_base, file, type, range, swizzle);
+               return LLVMBuildExtractElement(builder, array, index, "");
+       }
+}
+
+static void
+store_value_to_array(struct lp_build_tgsi_context *bld_base,
+                    LLVMValueRef value,
+                    unsigned file,
+                    unsigned chan_index,
+                    unsigned reg_index,
+                    const struct tgsi_ind_register *reg_indirect)
+{
+       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
+       struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
+       struct gallivm_state *gallivm = bld_base->base.gallivm;
+       LLVMBuilderRef builder = gallivm->builder;
+       LLVMValueRef ptr;
+
+       ptr = get_pointer_into_array(ctx, file, chan_index, reg_index, reg_indirect);
+       if (ptr) {
+               LLVMBuildStore(builder, value, ptr);
+       } else {
+               unsigned i, size;
+               struct tgsi_declaration_range range = get_array_range(bld_base, file, reg_index, reg_indirect);
+               LLVMValueRef index = emit_array_index(bld, reg_indirect, reg_index - range.First);
+               LLVMValueRef array =
+                       emit_array_fetch(bld_base, file, TGSI_TYPE_FLOAT, range, chan_index);
+               LLVMValueRef temp_ptr;
+
+               array = LLVMBuildInsertElement(builder, array, value, index, "");
+
+               size = range.Last - range.First + 1;
+               for (i = 0; i < size; ++i) {
+                       switch(file) {
+                       case TGSI_FILE_OUTPUT:
+                               temp_ptr = bld->outputs[i + range.First][chan_index];
+                               break;
+
+                       case TGSI_FILE_TEMPORARY:
+                               if (range.First + i >= ctx->temps_count)
+                                       continue;
+                               temp_ptr = ctx->temps[(i + range.First) * TGSI_NUM_CHANNELS + chan_index];
+                               break;
+
+                       default:
+                               continue;
+                       }
+                       value = LLVMBuildExtractElement(builder, array,
+                               lp_build_const_int32(gallivm, i), "");
+                       LLVMBuildStore(builder, value, temp_ptr);
+               }
+       }
+}
+
+LLVMValueRef radeon_llvm_emit_fetch(struct lp_build_tgsi_context *bld_base,
+                                   const struct tgsi_full_src_register *reg,
+                                   enum tgsi_opcode_type type,
+                                   unsigned swizzle)
+{
+       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
+       struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
+       LLVMBuilderRef builder = bld_base->base.gallivm->builder;
+       LLVMValueRef result = NULL, ptr, ptr2;
+
+       if (swizzle == ~0) {
+               LLVMValueRef values[TGSI_NUM_CHANNELS];
+               unsigned chan;
+               for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
+                       values[chan] = radeon_llvm_emit_fetch(bld_base, reg, type, chan);
+               }
+               return lp_build_gather_values(bld_base->base.gallivm, values,
+                                             TGSI_NUM_CHANNELS);
+       }
+
+       if (reg->Register.Indirect) {
+               LLVMValueRef load = load_value_from_array(bld_base, reg->Register.File, type,
+                               swizzle, reg->Register.Index, &reg->Indirect);
+               return bitcast(bld_base, type, load);
+       }
+
+       switch(reg->Register.File) {
+       case TGSI_FILE_IMMEDIATE: {
+               LLVMTypeRef ctype = tgsi2llvmtype(bld_base, type);
+               if (tgsi_type_is_64bit(type)) {
+                       result = LLVMGetUndef(LLVMVectorType(LLVMIntTypeInContext(bld_base->base.gallivm->context, 32), bld_base->base.type.length * 2));
+                       result = LLVMConstInsertElement(result,
+                                                       bld->immediates[reg->Register.Index][swizzle],
+                                                       bld_base->int_bld.zero);
+                       result = LLVMConstInsertElement(result,
+                                                       bld->immediates[reg->Register.Index][swizzle + 1],
+                                                       bld_base->int_bld.one);
+                       return LLVMConstBitCast(result, ctype);
+               } else {
+                       return LLVMConstBitCast(bld->immediates[reg->Register.Index][swizzle], ctype);
+               }
+       }
+
+       case TGSI_FILE_INPUT: {
+               unsigned index = reg->Register.Index;
+               LLVMValueRef input[4];
+
+               /* I don't think doing this for vertex shaders is beneficial.
+                * For those, we want to make sure the VMEM loads are executed
+                * only once. Fragment shaders don't care much, because
+                * v_interp instructions are much cheaper than VMEM loads.
+                */
+               if (ctx->soa.bld_base.info->processor == PIPE_SHADER_FRAGMENT)
+                       ctx->load_input(ctx, index, &ctx->input_decls[index], input);
+               else
+                       memcpy(input, &ctx->inputs[index * 4], sizeof(input));
+
+               result = input[swizzle];
+
+               if (tgsi_type_is_64bit(type)) {
+                       ptr = result;
+                       ptr2 = input[swizzle + 1];
+                       return radeon_llvm_emit_fetch_64bit(bld_base, type, ptr, ptr2);
+               }
+               break;
+       }
+
+       case TGSI_FILE_TEMPORARY:
+               if (reg->Register.Index >= ctx->temps_count)
+                       return LLVMGetUndef(tgsi2llvmtype(bld_base, type));
+               ptr = ctx->temps[reg->Register.Index * TGSI_NUM_CHANNELS + swizzle];
+               if (tgsi_type_is_64bit(type)) {
+                       ptr2 = ctx->temps[reg->Register.Index * TGSI_NUM_CHANNELS + swizzle + 1];
+                       return radeon_llvm_emit_fetch_64bit(bld_base, type,
+                                                LLVMBuildLoad(builder, ptr, ""),
+                                                LLVMBuildLoad(builder, ptr2, ""));
+               }
+               result = LLVMBuildLoad(builder, ptr, "");
+               break;
+
+       case TGSI_FILE_OUTPUT:
+               ptr = lp_get_output_ptr(bld, reg->Register.Index, swizzle);
+               if (tgsi_type_is_64bit(type)) {
+                       ptr2 = lp_get_output_ptr(bld, reg->Register.Index, swizzle + 1);
+                       return radeon_llvm_emit_fetch_64bit(bld_base, type,
+                                                LLVMBuildLoad(builder, ptr, ""),
+                                                LLVMBuildLoad(builder, ptr2, ""));
+               }
+               result = LLVMBuildLoad(builder, ptr, "");
+               break;
+
+       default:
+               return LLVMGetUndef(tgsi2llvmtype(bld_base, type));
+       }
+
+       return bitcast(bld_base, type, result);
+}
+
+static LLVMValueRef fetch_system_value(struct lp_build_tgsi_context *bld_base,
+                                      const struct tgsi_full_src_register *reg,
+                                      enum tgsi_opcode_type type,
+                                      unsigned swizzle)
+{
+       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
+       struct gallivm_state *gallivm = bld_base->base.gallivm;
+
+       LLVMValueRef cval = ctx->system_values[reg->Register.Index];
+       if (LLVMGetTypeKind(LLVMTypeOf(cval)) == LLVMVectorTypeKind) {
+               cval = LLVMBuildExtractElement(gallivm->builder, cval,
+                                              lp_build_const_int32(gallivm, swizzle), "");
+       }
+       return bitcast(bld_base, type, cval);
+}
+
+static void emit_declaration(struct lp_build_tgsi_context *bld_base,
+                            const struct tgsi_full_declaration *decl)
+{
+       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
+       LLVMBuilderRef builder = bld_base->base.gallivm->builder;
+       unsigned first, last, i;
+       switch(decl->Declaration.File) {
+       case TGSI_FILE_ADDRESS:
+       {
+                unsigned idx;
+               for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
+                       unsigned chan;
+                       for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
+                                ctx->soa.addr[idx][chan] = lp_build_alloca_undef(
+                                       &ctx->gallivm,
+                                       ctx->soa.bld_base.uint_bld.elem_type, "");
+                       }
+               }
+               break;
+       }
+
+       case TGSI_FILE_TEMPORARY:
+       {
+               char name[16] = "";
+               LLVMValueRef array_alloca = NULL;
+               unsigned decl_size;
+               unsigned writemask = decl->Declaration.UsageMask;
+               first = decl->Range.First;
+               last = decl->Range.Last;
+               decl_size = 4 * ((last - first) + 1);
+
+               if (decl->Declaration.Array) {
+                       unsigned id = decl->Array.ArrayID - 1;
+                       unsigned array_size;
+
+                       writemask &= ctx->temp_arrays[id].writemask;
+                       ctx->temp_arrays[id].writemask = writemask;
+                       array_size = ((last - first) + 1) * util_bitcount(writemask);
+
+                       /* If the array has more than 16 elements, store it
+                        * in memory using an alloca that spans the entire
+                        * array.
+                        *
+                        * Otherwise, store each array element individually.
+                        * We will then generate vectors (per-channel, up to
+                        * <16 x float> if the usagemask is a single bit) for
+                        * indirect addressing.
+                        *
+                        * Note that 16 is the number of vector elements that
+                        * LLVM will store in a register, so theoretically an
+                        * array with up to 4 * 16 = 64 elements could be
+                        * handled this way, but whether that's a good idea
+                        * depends on VGPR register pressure elsewhere.
+                        *
+                        * FIXME: We shouldn't need to have the non-alloca
+                        * code path for arrays. LLVM should be smart enough to
+                        * promote allocas into registers when profitable.
+                        *
+                        * LLVM 3.8 crashes with this.
+                        */
+                       if (HAVE_LLVM >= 0x0309 && array_size > 16) {
+                               array_alloca = LLVMBuildAlloca(builder,
+                                       LLVMArrayType(bld_base->base.vec_type,
+                                                     array_size), "array");
+                               ctx->temp_array_allocas[id] = array_alloca;
+                       }
+               }
+
+               if (!ctx->temps_count) {
+                       ctx->temps_count = bld_base->info->file_max[TGSI_FILE_TEMPORARY] + 1;
+                       ctx->temps = MALLOC(TGSI_NUM_CHANNELS * ctx->temps_count * sizeof(LLVMValueRef));
+               }
+               if (!array_alloca) {
+                       for (i = 0; i < decl_size; ++i) {
+#ifdef DEBUG
+                               snprintf(name, sizeof(name), "TEMP%d.%c",
+                                        first + i / 4, "xyzw"[i % 4]);
+#endif
+                               ctx->temps[first * TGSI_NUM_CHANNELS + i] =
+                                       lp_build_alloca_undef(bld_base->base.gallivm,
+                                                             bld_base->base.vec_type,
+                                                             name);
+                       }
+               } else {
+                       LLVMValueRef idxs[2] = {
+                               bld_base->uint_bld.zero,
+                               NULL
+                       };
+                       unsigned j = 0;
+
+                       if (writemask != TGSI_WRITEMASK_XYZW &&
+                           !ctx->undef_alloca) {
+                               /* Create a dummy alloca. We use it so that we
+                                * have a pointer that is safe to load from if
+                                * a shader ever reads from a channel that
+                                * it never writes to.
+                                */
+                               ctx->undef_alloca = lp_build_alloca_undef(
+                                       bld_base->base.gallivm,
+                                       bld_base->base.vec_type, "undef");
+                       }
+
+                       for (i = 0; i < decl_size; ++i) {
+                               LLVMValueRef ptr;
+                               if (writemask & (1 << (i % 4))) {
+#ifdef DEBUG
+                                       snprintf(name, sizeof(name), "TEMP%d.%c",
+                                                first + i / 4, "xyzw"[i % 4]);
+#endif
+                                       idxs[1] = lp_build_const_int32(bld_base->base.gallivm, j);
+                                       ptr = LLVMBuildGEP(builder, array_alloca, idxs, 2, name);
+                                       j++;
+                               } else {
+                                       ptr = ctx->undef_alloca;
+                               }
+                               ctx->temps[first * TGSI_NUM_CHANNELS + i] = ptr;
+                       }
+               }
+               break;
+       }
+       case TGSI_FILE_INPUT:
+       {
+               unsigned idx;
+               for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
+                       if (ctx->load_input &&
+                           ctx->input_decls[idx].Declaration.File != TGSI_FILE_INPUT) {
+                               ctx->input_decls[idx] = *decl;
+
+                               if (bld_base->info->processor != PIPE_SHADER_FRAGMENT)
+                                       ctx->load_input(ctx, idx, decl,
+                                                       &ctx->inputs[idx * 4]);
+                       }
+               }
+       }
+       break;
+
+       case TGSI_FILE_SYSTEM_VALUE:
+       {
+               unsigned idx;
+               for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
+                       ctx->load_system_value(ctx, idx, decl);
+               }
+       }
+       break;
+
+       case TGSI_FILE_OUTPUT:
+       {
+               char name[16] = "";
+               unsigned idx;
+               for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
+                       unsigned chan;
+                       assert(idx < RADEON_LLVM_MAX_OUTPUTS);
+                       if (ctx->soa.outputs[idx][0])
+                               continue;
+                       for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
+#ifdef DEBUG
+                               snprintf(name, sizeof(name), "OUT%d.%c",
+                                        idx, "xyzw"[chan % 4]);
+#endif
+                               ctx->soa.outputs[idx][chan] = lp_build_alloca_undef(
+                                       &ctx->gallivm,
+                                       ctx->soa.bld_base.base.elem_type, name);
+                       }
+               }
+               break;
+       }
+
+       case TGSI_FILE_MEMORY:
+               ctx->declare_memory_region(ctx, decl);
+               break;
+
+       default:
+               break;
+       }
+}
+
+LLVMValueRef radeon_llvm_saturate(struct lp_build_tgsi_context *bld_base,
+                                  LLVMValueRef value)
+{
+       struct lp_build_emit_data clamp_emit_data;
+
+       memset(&clamp_emit_data, 0, sizeof(clamp_emit_data));
+       clamp_emit_data.arg_count = 3;
+       clamp_emit_data.args[0] = value;
+       clamp_emit_data.args[2] = bld_base->base.one;
+       clamp_emit_data.args[1] = bld_base->base.zero;
+
+       return lp_build_emit_llvm(bld_base, TGSI_OPCODE_CLAMP,
+                                 &clamp_emit_data);
+}
+
+void radeon_llvm_emit_store(struct lp_build_tgsi_context *bld_base,
+                           const struct tgsi_full_instruction *inst,
+                           const struct tgsi_opcode_info *info,
+                           LLVMValueRef dst[4])
+{
+       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
+       struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
+       struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
+       const struct tgsi_full_dst_register *reg = &inst->Dst[0];
+       LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
+       LLVMValueRef temp_ptr, temp_ptr2 = NULL;
+       unsigned chan, chan_index;
+       bool is_vec_store = false;
+       enum tgsi_opcode_type dtype = tgsi_opcode_infer_dst_type(inst->Instruction.Opcode);
+
+       if (dst[0]) {
+               LLVMTypeKind k = LLVMGetTypeKind(LLVMTypeOf(dst[0]));
+               is_vec_store = (k == LLVMVectorTypeKind);
+       }
+
+       if (is_vec_store) {
+               LLVMValueRef values[4] = {};
+               TGSI_FOR_EACH_DST0_ENABLED_CHANNEL(inst, chan) {
+                       LLVMValueRef index = lp_build_const_int32(gallivm, chan);
+                       values[chan]  = LLVMBuildExtractElement(gallivm->builder,
+                                                       dst[0], index, "");
+               }
+               bld_base->emit_store(bld_base, inst, info, values);
+               return;
+       }
+
+       TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
+               LLVMValueRef value = dst[chan_index];
+
+               if (tgsi_type_is_64bit(dtype) && (chan_index == 1 || chan_index == 3))
+                       continue;
+               if (inst->Instruction.Saturate)
+                       value = radeon_llvm_saturate(bld_base, value);
+
+               if (reg->Register.File == TGSI_FILE_ADDRESS) {
+                       temp_ptr = bld->addr[reg->Register.Index][chan_index];
+                       LLVMBuildStore(builder, value, temp_ptr);
+                       continue;
+               }
+
+               if (!tgsi_type_is_64bit(dtype))
+                       value = bitcast(bld_base, TGSI_TYPE_FLOAT, value);
+
+               if (reg->Register.Indirect) {
+                       unsigned file = reg->Register.File;
+                       unsigned reg_index = reg->Register.Index;
+                       store_value_to_array(bld_base, value, file, chan_index,
+                                            reg_index, &reg->Indirect);
+               } else {
+                       switch(reg->Register.File) {
+                       case TGSI_FILE_OUTPUT:
+                               temp_ptr = bld->outputs[reg->Register.Index][chan_index];
+                               if (tgsi_type_is_64bit(dtype))
+                                       temp_ptr2 = bld->outputs[reg->Register.Index][chan_index + 1];
+                               break;
+
+                       case TGSI_FILE_TEMPORARY:
+                       {
+                               if (reg->Register.Index >= ctx->temps_count)
+                                       continue;
+
+                               temp_ptr = ctx->temps[ TGSI_NUM_CHANNELS * reg->Register.Index + chan_index];
+                               if (tgsi_type_is_64bit(dtype))
+                                       temp_ptr2 = ctx->temps[ TGSI_NUM_CHANNELS * reg->Register.Index + chan_index + 1];
+
+                               break;
+                       }
+                       default:
+                               return;
+                       }
+                       if (!tgsi_type_is_64bit(dtype))
+                               LLVMBuildStore(builder, value, temp_ptr);
+                       else {
+                               LLVMValueRef ptr = LLVMBuildBitCast(builder, value,
+                                                                   LLVMVectorType(LLVMIntTypeInContext(bld_base->base.gallivm->context, 32), 2), "");
+                               LLVMValueRef val2;
+                               value = LLVMBuildExtractElement(builder, ptr,
+                                                               bld_base->uint_bld.zero, "");
+                               val2 = LLVMBuildExtractElement(builder, ptr,
+                                                               bld_base->uint_bld.one, "");
+
+                               LLVMBuildStore(builder, bitcast(bld_base, TGSI_TYPE_FLOAT, value), temp_ptr);
+                               LLVMBuildStore(builder, bitcast(bld_base, TGSI_TYPE_FLOAT, val2), temp_ptr2);
+                       }
+               }
+       }
+}
+
+static void set_basicblock_name(LLVMBasicBlockRef bb, const char *base, int pc)
+{
+       char buf[32];
+       /* Subtract 1 so that the number shown is that of the corresponding
+        * opcode in the TGSI dump, e.g. an if block has the same suffix as
+        * the instruction number of the corresponding TGSI IF.
+        */
+       snprintf(buf, sizeof(buf), "%s%d", base, pc - 1);
+       LLVMSetValueName(LLVMBasicBlockAsValue(bb), buf);
+}
+
+/* Append a basic block at the level of the parent flow.
+ */
+static LLVMBasicBlockRef append_basic_block(struct radeon_llvm_context *ctx,
+                                           const char *name)
+{
+       struct gallivm_state *gallivm = &ctx->gallivm;
+
+       assert(ctx->flow_depth >= 1);
+
+       if (ctx->flow_depth >= 2) {
+               struct radeon_llvm_flow *flow = &ctx->flow[ctx->flow_depth - 2];
+
+               return LLVMInsertBasicBlockInContext(gallivm->context,
+                                                    flow->next_block, name);
+       }
+
+       return LLVMAppendBasicBlockInContext(gallivm->context, ctx->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)
+{
+       if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(builder)))
+                LLVMBuildBr(builder, target);
+}
+
+static void bgnloop_emit(const struct lp_build_tgsi_action *action,
+                        struct lp_build_tgsi_context *bld_base,
+                        struct lp_build_emit_data *emit_data)
+{
+       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
+       struct gallivm_state *gallivm = bld_base->base.gallivm;
+       struct radeon_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", bld_base->pc);
+       LLVMBuildBr(gallivm->builder, flow->loop_entry_block);
+       LLVMPositionBuilderAtEnd(gallivm->builder, flow->loop_entry_block);
+}
+
+static void brk_emit(const struct lp_build_tgsi_action *action,
+                    struct lp_build_tgsi_context *bld_base,
+                    struct lp_build_emit_data *emit_data)
+{
+       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
+       struct gallivm_state *gallivm = bld_base->base.gallivm;
+       struct radeon_llvm_flow *flow = get_innermost_loop(ctx);
+
+       LLVMBuildBr(gallivm->builder, flow->next_block);
+}
+
+static void cont_emit(const struct lp_build_tgsi_action *action,
+                     struct lp_build_tgsi_context *bld_base,
+                     struct lp_build_emit_data *emit_data)
+{
+       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
+       struct gallivm_state *gallivm = bld_base->base.gallivm;
+       struct radeon_llvm_flow *flow = get_innermost_loop(ctx);
+
+       LLVMBuildBr(gallivm->builder, flow->loop_entry_block);
+}
+
+static void else_emit(const struct lp_build_tgsi_action *action,
+                     struct lp_build_tgsi_context *bld_base,
+                     struct lp_build_emit_data *emit_data)
+{
+       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
+       struct gallivm_state *gallivm = bld_base->base.gallivm;
+       struct radeon_llvm_flow *current_branch = get_current_flow(ctx);
+       LLVMBasicBlockRef endif_block;
+
+       assert(!current_branch->loop_entry_block);
+
+       endif_block = append_basic_block(ctx, "ENDIF");
+       emit_default_branch(gallivm->builder, endif_block);
+
+       LLVMPositionBuilderAtEnd(gallivm->builder, current_branch->next_block);
+       set_basicblock_name(current_branch->next_block, "else", bld_base->pc);
+
+       current_branch->next_block = endif_block;
+}
+
+static void endif_emit(const struct lp_build_tgsi_action *action,
+                      struct lp_build_tgsi_context *bld_base,
+                      struct lp_build_emit_data *emit_data)
+{
+       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
+       struct gallivm_state *gallivm = bld_base->base.gallivm;
+       struct radeon_llvm_flow *current_branch = get_current_flow(ctx);
+
+       assert(!current_branch->loop_entry_block);
+
+       emit_default_branch(gallivm->builder, current_branch->next_block);
+       LLVMPositionBuilderAtEnd(gallivm->builder, current_branch->next_block);
+       set_basicblock_name(current_branch->next_block, "endif", bld_base->pc);
+
+       ctx->flow_depth--;
+}
+
+static void endloop_emit(const struct lp_build_tgsi_action *action,
+                        struct lp_build_tgsi_context *bld_base,
+                        struct lp_build_emit_data *emit_data)
+{
+       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
+       struct gallivm_state *gallivm = bld_base->base.gallivm;
+       struct radeon_llvm_flow *current_loop = get_current_flow(ctx);
+
+       assert(current_loop->loop_entry_block);
+
+       emit_default_branch(gallivm->builder, current_loop->loop_entry_block);
+
+       LLVMPositionBuilderAtEnd(gallivm->builder, current_loop->next_block);
+       set_basicblock_name(current_loop->next_block, "endloop", bld_base->pc);
+       ctx->flow_depth--;
+}
+
+static void if_cond_emit(const struct lp_build_tgsi_action *action,
+                        struct lp_build_tgsi_context *bld_base,
+                        struct lp_build_emit_data *emit_data,
+                        LLVMValueRef cond)
+{
+       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
+       struct gallivm_state *gallivm = bld_base->base.gallivm;
+       struct radeon_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", bld_base->pc);
+       LLVMBuildCondBr(gallivm->builder, cond, if_block, flow->next_block);
+       LLVMPositionBuilderAtEnd(gallivm->builder, if_block);
+}
+
+static void if_emit(const struct lp_build_tgsi_action *action,
+                   struct lp_build_tgsi_context *bld_base,
+                   struct lp_build_emit_data *emit_data)
+{
+       struct gallivm_state *gallivm = bld_base->base.gallivm;
+       LLVMValueRef cond;
+
+       cond = LLVMBuildFCmp(gallivm->builder, LLVMRealUNE,
+                       emit_data->args[0],
+                       bld_base->base.zero, "");
+
+       if_cond_emit(action, bld_base, emit_data, cond);
+}
+
+static void uif_emit(const struct lp_build_tgsi_action *action,
+                    struct lp_build_tgsi_context *bld_base,
+                    struct lp_build_emit_data *emit_data)
+{
+       struct gallivm_state *gallivm = bld_base->base.gallivm;
+       LLVMValueRef cond;
+
+       cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE,
+               bitcast(bld_base, TGSI_TYPE_UNSIGNED, emit_data->args[0]),
+                       bld_base->int_bld.zero, "");
+
+       if_cond_emit(action, bld_base, emit_data, cond);
+}
+
+static void emit_immediate(struct lp_build_tgsi_context *bld_base,
+                          const struct tgsi_full_immediate *imm)
+{
+       unsigned i;
+       struct radeon_llvm_context *ctx = radeon_llvm_context(bld_base);
+
+       for (i = 0; i < 4; ++i) {
+               ctx->soa.immediates[ctx->soa.num_immediates][i] =
+                               LLVMConstInt(bld_base->uint_bld.elem_type, imm->u[i].Uint, false   );
+       }
+
+       ctx->soa.num_immediates++;
+}
+
+void radeon_llvm_context_init(struct radeon_llvm_context *ctx, const char *triple,
+                             const struct tgsi_shader_info *info,
+                             const struct tgsi_token *tokens)
+{
+       struct lp_type type;
+
+       /* Initialize the gallivm object:
+        * We are only using the module, context, and builder fields of this struct.
+        * This should be enough for us to be able to pass our gallivm struct to the
+        * helper functions in the gallivm module.
+        */
+       memset(&ctx->gallivm, 0, sizeof (ctx->gallivm));
+       memset(&ctx->soa, 0, sizeof(ctx->soa));
+       ctx->gallivm.context = LLVMContextCreate();
+       ctx->gallivm.module = LLVMModuleCreateWithNameInContext("tgsi",
+                                               ctx->gallivm.context);
+       LLVMSetTarget(ctx->gallivm.module, triple);
+       ctx->gallivm.builder = LLVMCreateBuilderInContext(ctx->gallivm.context);
+
+       struct lp_build_tgsi_context *bld_base = &ctx->soa.bld_base;
+
+       bld_base->info = info;
+
+       if (info && info->array_max[TGSI_FILE_TEMPORARY] > 0) {
+               int size = info->array_max[TGSI_FILE_TEMPORARY];
+
+               ctx->temp_arrays = CALLOC(size, sizeof(ctx->temp_arrays[0]));
+               ctx->temp_array_allocas = CALLOC(size, sizeof(ctx->temp_array_allocas[0]));
+
+               if (tokens)
+                       tgsi_scan_arrays(tokens, TGSI_FILE_TEMPORARY, size,
+                                        ctx->temp_arrays);
+       }
+
+       type.floating = true;
+       type.fixed = false;
+       type.sign = true;
+       type.norm = false;
+       type.width = 32;
+       type.length = 1;
+
+       lp_build_context_init(&bld_base->base, &ctx->gallivm, type);
+       lp_build_context_init(&ctx->soa.bld_base.uint_bld, &ctx->gallivm, lp_uint_type(type));
+       lp_build_context_init(&ctx->soa.bld_base.int_bld, &ctx->gallivm, lp_int_type(type));
+       type.width *= 2;
+       lp_build_context_init(&ctx->soa.bld_base.dbl_bld, &ctx->gallivm, type);
+       lp_build_context_init(&ctx->soa.bld_base.uint64_bld, &ctx->gallivm, lp_uint_type(type));
+       lp_build_context_init(&ctx->soa.bld_base.int64_bld, &ctx->gallivm, lp_int_type(type));
+
+       bld_base->soa = 1;
+       bld_base->emit_store = radeon_llvm_emit_store;
+       bld_base->emit_swizzle = emit_swizzle;
+       bld_base->emit_declaration = emit_declaration;
+       bld_base->emit_immediate = emit_immediate;
+
+       bld_base->emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = radeon_llvm_emit_fetch;
+       bld_base->emit_fetch_funcs[TGSI_FILE_INPUT] = radeon_llvm_emit_fetch;
+       bld_base->emit_fetch_funcs[TGSI_FILE_TEMPORARY] = radeon_llvm_emit_fetch;
+       bld_base->emit_fetch_funcs[TGSI_FILE_OUTPUT] = radeon_llvm_emit_fetch;
+       bld_base->emit_fetch_funcs[TGSI_FILE_SYSTEM_VALUE] = fetch_system_value;
+
+       /* metadata allowing 2.5 ULP */
+       ctx->fpmath_md_kind = LLVMGetMDKindIDInContext(ctx->gallivm.context,
+                                                      "fpmath", 6);
+       LLVMValueRef arg = lp_build_const_float(&ctx->gallivm, 2.5);
+       ctx->fpmath_md_2p5_ulp = LLVMMDNodeInContext(ctx->gallivm.context,
+                                                    &arg, 1);
+
+       /* Allocate outputs */
+       ctx->soa.outputs = ctx->outputs;
+
+       bld_base->op_actions[TGSI_OPCODE_BGNLOOP].emit = bgnloop_emit;
+       bld_base->op_actions[TGSI_OPCODE_BRK].emit = brk_emit;
+       bld_base->op_actions[TGSI_OPCODE_CONT].emit = cont_emit;
+       bld_base->op_actions[TGSI_OPCODE_IF].emit = if_emit;
+       bld_base->op_actions[TGSI_OPCODE_UIF].emit = uif_emit;
+       bld_base->op_actions[TGSI_OPCODE_ELSE].emit = else_emit;
+       bld_base->op_actions[TGSI_OPCODE_ENDIF].emit = endif_emit;
+       bld_base->op_actions[TGSI_OPCODE_ENDLOOP].emit = endloop_emit;
+}
+
+void radeon_llvm_create_func(struct radeon_llvm_context *ctx,
+                            LLVMTypeRef *return_types, unsigned num_return_elems,
+                            LLVMTypeRef *ParamTypes, unsigned ParamCount)
+{
+       LLVMTypeRef main_fn_type, ret_type;
+       LLVMBasicBlockRef main_fn_body;
+
+       if (num_return_elems)
+               ret_type = LLVMStructTypeInContext(ctx->gallivm.context,
+                                                  return_types,
+                                                  num_return_elems, true);
+       else
+               ret_type = LLVMVoidTypeInContext(ctx->gallivm.context);
+
+       /* Setup the function */
+       ctx->return_type = ret_type;
+       main_fn_type = LLVMFunctionType(ret_type, ParamTypes, ParamCount, 0);
+       ctx->main_fn = LLVMAddFunction(ctx->gallivm.module, "main", main_fn_type);
+       main_fn_body = LLVMAppendBasicBlockInContext(ctx->gallivm.context,
+                       ctx->main_fn, "main_body");
+       LLVMPositionBuilderAtEnd(ctx->gallivm.builder, main_fn_body);
+}
+
+void radeon_llvm_finalize_module(struct radeon_llvm_context *ctx,
+                                bool run_verifier)
+{
+       struct gallivm_state *gallivm = ctx->soa.bld_base.base.gallivm;
+       const char *triple = LLVMGetTarget(gallivm->module);
+       LLVMTargetLibraryInfoRef target_library_info;
+
+       /* Create the pass manager */
+       gallivm->passmgr = LLVMCreateFunctionPassManagerForModule(
+                                                       gallivm->module);
+
+       target_library_info = gallivm_create_target_library_info(triple);
+       LLVMAddTargetLibraryInfo(target_library_info, gallivm->passmgr);
+
+       if (run_verifier)
+               LLVMAddVerifierPass(gallivm->passmgr);
+
+       /* This pass should eliminate all the load and store instructions */
+       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
+
+       /* Add some optimization passes */
+       LLVMAddScalarReplAggregatesPass(gallivm->passmgr);
+       LLVMAddLICMPass(gallivm->passmgr);
+       LLVMAddAggressiveDCEPass(gallivm->passmgr);
+       LLVMAddCFGSimplificationPass(gallivm->passmgr);
+       LLVMAddInstructionCombiningPass(gallivm->passmgr);
+
+       /* Run the pass */
+       LLVMInitializeFunctionPassManager(gallivm->passmgr);
+       LLVMRunFunctionPassManager(gallivm->passmgr, ctx->main_fn);
+       LLVMFinalizeFunctionPassManager(gallivm->passmgr);
+
+       LLVMDisposeBuilder(gallivm->builder);
+       LLVMDisposePassManager(gallivm->passmgr);
+       gallivm_dispose_target_library_info(target_library_info);
+}
+
+void radeon_llvm_dispose(struct radeon_llvm_context *ctx)
+{
+       LLVMDisposeModule(ctx->soa.bld_base.base.gallivm->module);
+       LLVMContextDispose(ctx->soa.bld_base.base.gallivm->context);
+       FREE(ctx->temp_arrays);
+       ctx->temp_arrays = NULL;
+       FREE(ctx->temp_array_allocas);
+       ctx->temp_array_allocas = NULL;
+       FREE(ctx->temps);
+       ctx->temps = NULL;
+       ctx->temps_count = 0;
+       FREE(ctx->flow);
+       ctx->flow = NULL;
+       ctx->flow_depth_max = 0;
+}