Updated code generation so that for vertex shader output position is written at last...

[mesa.git] / src / libre-soc / vulkan / libresoc_llvm.c

#include "libresoc_llvm.h"
#include "libresoc_llvm_build.h"

void InitLLVM(struct libresoc_llvm *llvm_ref)
{
    // LLVMInitializeNativeTarget();
    // LLVMInitializeNativeAsmPrinter();
    LLVMInitializeAllAsmPrinters();
    LLVMInitializeAllTargets();
    LLVMInitializeAllTargetInfos();
    LLVMInitializeAllTargetMCs();
    LLVMInitializeAllDisassemblers();
    //LLVMLinkInMCJIT();
    char *def_triple = LLVMGetDefaultTargetTriple();   // E.g. "x86_64-linux-gnu"
    char *error;
    LLVMTargetRef target_ref;
    if (LLVMGetTargetFromTriple(def_triple, &target_ref, &error)) {
        // Fatal error
    }

    if (!LLVMTargetHasJIT(target_ref)) {
        // Fatal error, cannot do JIT on this platform
    }

    LLVMTargetMachineRef tm_ref =
        LLVMCreateTargetMachine(target_ref, def_triple, "", "",
                LLVMCodeGenLevelDefault,
                LLVMRelocDefault,
                LLVMCodeModelJITDefault);
    //assert(tm_ref);
    LLVMDisposeErrorMessage(def_triple);
    llvm_ref->orc_ref = LLVMOrcCreateInstance(tm_ref);
    llvm_ref->lc.context = LLVMContextCreate();
    llvm_ref->lc.builder = LLVMCreateBuilderInContext(llvm_ref->lc.context);
    llvm_ref->lc.voidt = LLVMVoidTypeInContext(llvm_ref->lc.context);
    llvm_ref->lc.i1 = LLVMInt1TypeInContext(llvm_ref->lc.context);
    llvm_ref->lc.i8 = LLVMInt8TypeInContext(llvm_ref->lc.context);
    llvm_ref->lc.i16 = LLVMIntTypeInContext(llvm_ref->lc.context, 16);
    llvm_ref->lc.i32 = LLVMIntTypeInContext(llvm_ref->lc.context, 32);
    llvm_ref->lc.i64 = LLVMIntTypeInContext(llvm_ref->lc.context, 64);
    llvm_ref->lc.i128 = LLVMIntTypeInContext(llvm_ref->lc.context, 128);
    llvm_ref->lc.intptr = llvm_ref->lc.i32;
    llvm_ref->lc.f16 = LLVMHalfTypeInContext(llvm_ref->lc.context);
    llvm_ref->lc.f32 = LLVMFloatTypeInContext(llvm_ref->lc.context);
    llvm_ref->lc.f64 = LLVMDoubleTypeInContext(llvm_ref->lc.context);
    llvm_ref->lc.v2i16 = LLVMVectorType(llvm_ref->lc.i16, 2);
    llvm_ref->lc.v4i16 = LLVMVectorType(llvm_ref->lc.i16, 4);
    llvm_ref->lc.v2f16 = LLVMVectorType(llvm_ref->lc.f16, 2);
    llvm_ref->lc.v4f16 = LLVMVectorType(llvm_ref->lc.f16, 4);
    llvm_ref->lc.v2i32 = LLVMVectorType(llvm_ref->lc.i32, 2);
    llvm_ref->lc.v3i32 = LLVMVectorType(llvm_ref->lc.i32, 3);
    llvm_ref->lc.v4i32 = LLVMVectorType(llvm_ref->lc.i32, 4);
    llvm_ref->lc.v2f32 = LLVMVectorType(llvm_ref->lc.f32, 2);
    llvm_ref->lc.v3f32 = LLVMVectorType(llvm_ref->lc.f32, 3);
    llvm_ref->lc.v4f32 = LLVMVectorType(llvm_ref->lc.f32, 4);
    llvm_ref->lc.v8i32 = LLVMVectorType(llvm_ref->lc.i32, 8);
    // llvm_ref->lc.iN_wavemask = LLVMIntTypeInContext(llvm_ref->lc.context, llvm_ref->lc.wave_size);
    // llvm_ref->lc.iN_ballotmask = LLVMIntTypeInContext(llvm_ref->lc.context, ballot_mask_bits);

    llvm_ref->lc.i8_0 = LLVMConstInt(llvm_ref->lc.i8, 0, false);
    llvm_ref->lc.i8_1 = LLVMConstInt(llvm_ref->lc.i8, 1, false);
    llvm_ref->lc.i16_0 = LLVMConstInt(llvm_ref->lc.i16, 0, false);
    llvm_ref->lc.i16_1 = LLVMConstInt(llvm_ref->lc.i16, 1, false);
    llvm_ref->lc.i32_0 = LLVMConstInt(llvm_ref->lc.i32, 0, false);
    llvm_ref->lc.i32_1 = LLVMConstInt(llvm_ref->lc.i32, 1, false);
    llvm_ref->lc.i64_0 = LLVMConstInt(llvm_ref->lc.i64, 0, false);
    llvm_ref->lc.i64_1 = LLVMConstInt(llvm_ref->lc.i64, 1, false);
    llvm_ref->lc.i128_0 = LLVMConstInt(llvm_ref->lc.i128, 0, false);
    llvm_ref->lc.i128_1 = LLVMConstInt(llvm_ref->lc.i128, 1, false);
    llvm_ref->lc.f16_0 = LLVMConstReal(llvm_ref->lc.f16, 0.0);
    llvm_ref->lc.f16_1 = LLVMConstReal(llvm_ref->lc.f16, 1.0);
    llvm_ref->lc.f32_0 = LLVMConstReal(llvm_ref->lc.f32, 0.0);
    llvm_ref->lc.f32_1 = LLVMConstReal(llvm_ref->lc.f32, 1.0);
    llvm_ref->lc.f64_0 = LLVMConstReal(llvm_ref->lc.f64, 0.0);
    llvm_ref->lc.f64_1 = LLVMConstReal(llvm_ref->lc.f64, 1.0);

    llvm_ref->lc.i1false = LLVMConstInt(llvm_ref->lc.i1, 0, false);
    llvm_ref->lc.i1true = LLVMConstInt(llvm_ref->lc.i1, 1, false);
    llvm_ref->lc.float_mode = 0; //TODO: default value, when required take this value as parameter
}

void DestroyLLVM(struct libresoc_llvm *llvm_ref)
{
    LLVMErrorRef error_ref = LLVMOrcDisposeInstance(llvm_ref->orc_ref);
}

static uint64_t orc_sym_resolver(const char *name, void *ctx)
{
    LLVMOrcJITStackRef orc_ref = (LLVMOrcJITStackRef) (ctx);
    LLVMOrcTargetAddress address;
    LLVMOrcGetSymbolAddress(orc_ref, &address, name);
    return (uint64_t)address;
}

void handle_shader_output_decl(struct libresoc_nir_tran_ctx *ctx,
                                  struct nir_shader *nir, struct nir_variable *variable,
                                  gl_shader_stage stage)
{
   unsigned output_loc = variable->data.driver_location / 4;
   unsigned attrib_count = glsl_count_attribute_slots(variable->type, false);

   /* tess ctrl has it's own load/store paths for outputs */
   if (stage == MESA_SHADER_TESS_CTRL)
      return;

   if (stage == MESA_SHADER_VERTEX || stage == MESA_SHADER_TESS_EVAL ||
       stage == MESA_SHADER_GEOMETRY) {
      int idx = variable->data.location + variable->data.index;
      if (idx == VARYING_SLOT_CLIP_DIST0) {
         int length = nir->info.clip_distance_array_size + nir->info.cull_distance_array_size;

         if (length > 4)
            attrib_count = 2;
         else
            attrib_count = 1;
      }
   }

   bool is_16bit = glsl_type_is_16bit(glsl_without_array(variable->type));
   LLVMTypeRef type = is_16bit ? ctx->lc.f16 : ctx->lc.f32;
   for (unsigned i = 0; i < attrib_count; ++i) {
      for (unsigned chan = 0; chan < 4; chan++) {
         ctx->outputs[llvm_reg_index_soa(output_loc + i, chan)] =
            build_alloca_undef(&ctx->lc, type, "");
      }
   }
}

LLVMValueRef extract_components(struct libresoc_llvm_context *ctx, LLVMValueRef value, unsigned start,
                                   unsigned channels)
{
   LLVMValueRef chan[channels];

   for (unsigned i = 0; i < channels; i++)
      chan[i] = llvm_extract_elem(ctx, value, i + start);

   return build_gather_values(ctx, chan, channels);
}

static void build_store_values_extended(struct libresoc_llvm_context *lc, LLVMValueRef *values,
                                        unsigned value_count, unsigned value_stride,
                                        LLVMValueRef vec)
{
   LLVMBuilderRef builder = lc->builder;
   unsigned i;

   for (i = 0; i < value_count; i++) {
      LLVMValueRef ptr = values[i * value_stride];
      LLVMValueRef index = LLVMConstInt(lc->i32, i, false);
      LLVMValueRef value = LLVMBuildExtractElement(builder, vec, index, "");
      LLVMBuildStore(builder, value, ptr);
   }
}

static LLVMTypeRef arg_llvm_type(enum arg_type type, unsigned size, struct libresoc_llvm_context *ctx)
{
   if (type == ARG_FLOAT) {
      return size == 1 ? ctx->f32 : LLVMVectorType(ctx->f32, size);
   } else if (type == ARG_INT) {
      return size == 1 ? ctx->i32 : LLVMVectorType(ctx->i32, size);
   } else {
      LLVMTypeRef ptr_type;
      switch (type) {
      case ARG_CONST_PTR:
         ptr_type = ctx->i8;
         break;
      case ARG_CONST_FLOAT_PTR:
         ptr_type = ctx->f32;
         break;
      case ARG_CONST_PTR_PTR:
         ptr_type = LLVMPointerType(ctx->i8, 0);
         break;
      case ARG_CONST_DESC_PTR:
         ptr_type = ctx->v4i32;
         break;
      case ARG_CONST_IMAGE_PTR:
         ptr_type = ctx->v8i32;
         break;
      default:
         unreachable("unknown arg type");
      }
      if (size == 1) {
         //return ac_array_in_const32_addr_space(ptr_type);
           return LLVMPointerType(ptr_type, 0); //address space may be wrong
      } else {
         assert(size == 2);
         return LLVMPointerType(ptr_type, 0);
      }
   }
}
static LLVMValueRef get_src(struct libresoc_nir_tran_ctx *ctx, nir_src src)
{
   assert(src.is_ssa);
   // printf("index %d\n", src.ssa->index);
   return ctx->ssa_defs[src.ssa->index];
}

static LLVMTypeRef get_def_type(struct libresoc_nir_tran_ctx *ctx, const nir_ssa_def *def)
{
   LLVMTypeRef type = LLVMIntTypeInContext(ctx->lc.context, def->bit_size);
   if (def->num_components > 1) {
      type = LLVMVectorType(type, def->num_components);
   }
   return type;
}

static LLVMValueRef get_memory_ptr(struct libresoc_nir_tran_ctx *ctx, nir_src src, unsigned bit_size)
{
   LLVMValueRef ptr = get_src(ctx, src);
   ptr = LLVMBuildGEP(ctx->lc.builder, ctx->lc.lds, &ptr, 1, "");
   int addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));

   LLVMTypeRef type = LLVMIntTypeInContext(ctx->lc.context, bit_size);

   return LLVMBuildBitCast(ctx->lc.builder, ptr, LLVMPointerType(type, addr_space), "");
}

static uint32_t widen_mask(uint32_t mask, unsigned multiplier)
{
   uint32_t new_mask = 0;
   for (unsigned i = 0; i < 32 && (1u << i) <= mask; ++i)
      if (mask & (1u << i))
         new_mask |= ((1u << multiplier) - 1u) << (i * multiplier);
   return new_mask;
}

static void get_deref_offset(struct libresoc_nir_tran_ctx *ctx, nir_deref_instr *instr, bool vs_in,
                             unsigned *vertex_index_out, LLVMValueRef *vertex_index_ref,
                             unsigned *const_out, LLVMValueRef *indir_out)
{
   nir_variable *var = nir_deref_instr_get_variable(instr);
   nir_deref_path path;
   unsigned idx_lvl = 1;

   nir_deref_path_init(&path, instr, NULL);

   if (vertex_index_out != NULL || vertex_index_ref != NULL) {
      if (vertex_index_ref) {
         *vertex_index_ref = get_src(ctx, path.path[idx_lvl]->arr.index);
         if (vertex_index_out)
            *vertex_index_out = 0;
      } else {
         *vertex_index_out = nir_src_as_uint(path.path[idx_lvl]->arr.index);
      }
      ++idx_lvl;
   }

   uint32_t const_offset = 0;
   LLVMValueRef offset = NULL;

   if (var->data.compact) {
      assert(instr->deref_type == nir_deref_type_array);
      const_offset = nir_src_as_uint(instr->arr.index);
      goto out;
   }

   for (; path.path[idx_lvl]; ++idx_lvl) {
      const struct glsl_type *parent_type = path.path[idx_lvl - 1]->type;
      if (path.path[idx_lvl]->deref_type == nir_deref_type_struct) {
         unsigned index = path.path[idx_lvl]->strct.index;

         for (unsigned i = 0; i < index; i++) {
            const struct glsl_type *ft = glsl_get_struct_field(parent_type, i);
            const_offset += glsl_count_attribute_slots(ft, vs_in);
         }
      } else if (path.path[idx_lvl]->deref_type == nir_deref_type_array) {
         unsigned size = glsl_count_attribute_slots(path.path[idx_lvl]->type, vs_in);
         if (nir_src_is_const(path.path[idx_lvl]->arr.index)) {
            const_offset += size * nir_src_as_uint(path.path[idx_lvl]->arr.index);
         } else {
            LLVMValueRef array_off =
               LLVMBuildMul(ctx->lc.builder, LLVMConstInt(ctx->lc.i32, size, 0),
                            get_src(ctx, path.path[idx_lvl]->arr.index), "");
            if (offset)
               offset = LLVMBuildAdd(ctx->lc.builder, offset, array_off, "");
            else
               offset = array_off;
         }
      } else
         unreachable("Uhandled deref type in get_deref_instr_offset");
   }

out:
   nir_deref_path_finish(&path);

   if (const_offset && offset)
      offset =
         LLVMBuildAdd(ctx->lc.builder, offset, LLVMConstInt(ctx->lc.i32, const_offset, 0), "");

   *const_out = const_offset;
   *indir_out = offset;
}

static unsigned type_scalar_size_bytes(const struct glsl_type *type)
{
   assert(glsl_type_is_vector_or_scalar(type) || glsl_type_is_matrix(type));
   return glsl_type_is_boolean(type) ? 4 : glsl_get_bit_size(type) / 8;
}


static LLVMValueRef emit_int_cmp(struct libresoc_llvm_context *lc, LLVMIntPredicate pred,
        LLVMValueRef src0, LLVMValueRef src1)
{
    LLVMTypeRef src0_type = LLVMTypeOf(src0);
    LLVMTypeRef src1_type = LLVMTypeOf(src1);

    if (LLVMGetTypeKind(src0_type) == LLVMPointerTypeKind &&
            LLVMGetTypeKind(src1_type) != LLVMPointerTypeKind) {
        src1 = LLVMBuildIntToPtr(lc->builder, src1, src0_type, "");
    } else if (LLVMGetTypeKind(src1_type) == LLVMPointerTypeKind &&
            LLVMGetTypeKind(src0_type) != LLVMPointerTypeKind) {
        src0 = LLVMBuildIntToPtr(lc->builder, src0, src1_type, "");
    }

    LLVMValueRef result = LLVMBuildICmp(lc->builder, pred, src0, src1, "");
    return LLVMBuildSelect(lc->builder, result, LLVMConstInt(lc->i32, 0xFFFFFFFF, false),
            lc->i32_0, "");
}

static LLVMValueRef emit_float_cmp(struct libresoc_llvm_context *lc, LLVMRealPredicate pred,
                                   LLVMValueRef src0, LLVMValueRef src1)
{
   LLVMValueRef result;
   src0 = to_float(lc, src0);
   src1 = to_float(lc, src1);
   result = LLVMBuildFCmp(lc->builder, pred, src0, src1, "");
   return LLVMBuildSelect(lc->builder, result, LLVMConstInt(lc->i32, 0xFFFFFFFF, false),
                          lc->i32_0, "");
}

static LLVMValueRef emit_intrin_1f_param(struct libresoc_llvm_context *lc, const char *intrin,
                                         LLVMTypeRef result_type, LLVMValueRef src0)
{
   char name[64], type[64];
   LLVMValueRef params[] = {
      to_float(lc, src0),
   };

   build_type_name_for_intr(LLVMTypeOf(params[0]), type, sizeof(type));
   ASSERTED const int length = snprintf(name, sizeof(name), "%s.%s", intrin, type);
   assert(length < sizeof(name));
   return build_intrinsic(lc, name, result_type, params, 1, FUNC_ATTR_READNONE);
}

static LLVMValueRef emit_intrin_1f_param_scalar(struct libresoc_llvm_context *lc, const char *intrin,
                                                LLVMTypeRef result_type, LLVMValueRef src0)
{
   if (LLVMGetTypeKind(result_type) != LLVMVectorTypeKind)
      return emit_intrin_1f_param(lc, intrin, result_type, src0);

   LLVMTypeRef elem_type = LLVMGetElementType(result_type);
   LLVMValueRef ret = LLVMGetUndef(result_type);

   /* Scalarize the intrinsic, because vectors are not supported. */
   for (unsigned i = 0; i < LLVMGetVectorSize(result_type); i++) {
      char name[64], type[64];
      LLVMValueRef params[] = {
         to_float(lc, llvm_extract_elem(lc, src0, i)),
      };

      build_type_name_for_intr(LLVMTypeOf(params[0]), type, sizeof(type));
      ASSERTED const int length = snprintf(name, sizeof(name), "%s.%s", intrin, type);
      assert(length < sizeof(name));
      ret = LLVMBuildInsertElement(
         lc->builder, ret,
         build_intrinsic(lc, name, elem_type, params, 1, FUNC_ATTR_READNONE),
         LLVMConstInt(lc->i32, i, 0), "");
   }
   return ret;
}

static LLVMValueRef emit_intrin_2f_param(struct libresoc_llvm_context *ctx, const char *intrin,
                                         LLVMTypeRef result_type, LLVMValueRef src0,
                                         LLVMValueRef src1)
{
   char name[64], type[64];
   LLVMValueRef params[] = {
      to_float(ctx, src0),
      to_float(ctx, src1),
   };

   build_type_name_for_intr(LLVMTypeOf(params[0]), type, sizeof(type));
   ASSERTED const int length = snprintf(name, sizeof(name), "%s.%s", intrin, type);
   assert(length < sizeof(name));
   return build_intrinsic(ctx, name, result_type, params, 2, FUNC_ATTR_READNONE);
}

static LLVMValueRef emit_intrin_3f_param(struct libresoc_llvm_context *ctx, const char *intrin,
                                         LLVMTypeRef result_type, LLVMValueRef src0,
                                         LLVMValueRef src1, LLVMValueRef src2)
{
   char name[64], type[64];
   LLVMValueRef params[] = {
      to_float(ctx, src0),
      to_float(ctx, src1),
      to_float(ctx, src2),
   };

   build_type_name_for_intr(LLVMTypeOf(params[0]), type, sizeof(type));
   ASSERTED const int length = snprintf(name, sizeof(name), "%s.%s", intrin, type);
   assert(length < sizeof(name));
   return build_intrinsic(ctx, name, result_type, params, 3, FUNC_ATTR_READNONE);
}

static LLVMValueRef emit_bcsel(struct libresoc_llvm_context *ctx, LLVMValueRef src0, LLVMValueRef src1,
                               LLVMValueRef src2)
{
   LLVMTypeRef src1_type = LLVMTypeOf(src1);
   LLVMTypeRef src2_type = LLVMTypeOf(src2);

   if (LLVMGetTypeKind(src1_type) == LLVMPointerTypeKind &&
       LLVMGetTypeKind(src2_type) != LLVMPointerTypeKind) {
      src2 = LLVMBuildIntToPtr(ctx->builder, src2, src1_type, "");
   } else if (LLVMGetTypeKind(src2_type) == LLVMPointerTypeKind &&
              LLVMGetTypeKind(src1_type) != LLVMPointerTypeKind) {
      src1 = LLVMBuildIntToPtr(ctx->builder, src1, src2_type, "");
   }

   LLVMValueRef v =
      LLVMBuildICmp(ctx->builder, LLVMIntNE, src0, LLVMConstNull(LLVMTypeOf(src0)), "");
   return LLVMBuildSelect(ctx->builder, v, to_integer_or_pointer(ctx, src1),
                          to_integer_or_pointer(ctx, src2), "");
}

static LLVMValueRef emit_iabs(struct libresoc_llvm_context *ctx, LLVMValueRef src0)
{
   return build_imax(ctx, src0, LLVMBuildNeg(ctx->builder, src0, ""));
}

static LLVMValueRef emit_uint_carry(struct libresoc_llvm_context *ctx, const char *intrin,
                                    LLVMValueRef src0, LLVMValueRef src1)
{
   LLVMTypeRef ret_type;
   LLVMTypeRef types[] = {ctx->i32, ctx->i1};
   LLVMValueRef res;
   LLVMValueRef params[] = {src0, src1};
   ret_type = LLVMStructTypeInContext(ctx->context, types, 2, true);

   res = build_intrinsic(ctx, intrin, ret_type, params, 2, FUNC_ATTR_READNONE);

   res = LLVMBuildExtractValue(ctx->builder, res, 1, "");
   res = LLVMBuildZExt(ctx->builder, res, ctx->i32, "");
   return res;
}

static LLVMValueRef emit_b2f(struct libresoc_llvm_context *ctx, LLVMValueRef src0, unsigned bitsize)
{
   assert(get_elem_bits(ctx, LLVMTypeOf(src0)) == 32);
   LLVMValueRef result =
      LLVMBuildAnd(ctx->builder, src0, const_uint_vec(ctx, LLVMTypeOf(src0), 0x3f800000), "");
   result = to_float(ctx, result);

   switch (bitsize) {
   case 16: {
      bool vec2 = LLVMGetTypeKind(LLVMTypeOf(result)) == LLVMVectorTypeKind;
      return LLVMBuildFPTrunc(ctx->builder, result, vec2 ? ctx->v2f16 : ctx->f16, "");
   }
   case 32:
      return result;
   case 64:
      return LLVMBuildFPExt(ctx->builder, result, ctx->f64, "");
   default:
      unreachable("Unsupported bit size.");
   }
}

static LLVMValueRef emit_f2b(struct libresoc_llvm_context *ctx, LLVMValueRef src0)
{
   src0 = to_float(ctx, src0);
   LLVMValueRef zero = LLVMConstNull(LLVMTypeOf(src0));
   return LLVMBuildSExt(ctx->builder, LLVMBuildFCmp(ctx->builder, LLVMRealUNE, src0, zero, ""),
                        ctx->i32, "");
}

static LLVMValueRef emit_b2i(struct libresoc_llvm_context *ctx, LLVMValueRef src0, unsigned bitsize)
{
   LLVMValueRef result = LLVMBuildAnd(ctx->builder, src0, ctx->i32_1, "");

   switch (bitsize) {
   case 8:
      return LLVMBuildTrunc(ctx->builder, result, ctx->i8, "");
   case 16:
      return LLVMBuildTrunc(ctx->builder, result, ctx->i16, "");
   case 32:
      return result;
   case 64:
      return LLVMBuildZExt(ctx->builder, result, ctx->i64, "");
   default:
      unreachable("Unsupported bit size.");
   }
}

static LLVMValueRef emit_i2b(struct libresoc_llvm_context *ctx, LLVMValueRef src0)
{
   LLVMValueRef zero = LLVMConstNull(LLVMTypeOf(src0));
   return LLVMBuildSExt(ctx->builder, LLVMBuildICmp(ctx->builder, LLVMIntNE, src0, zero, ""),
                        ctx->i32, "");
}

static LLVMValueRef emit_f2f16(struct libresoc_llvm_context *ctx, LLVMValueRef src0)
{
   LLVMValueRef result;
   LLVMValueRef cond = NULL;

   src0 = to_float(ctx, src0);
   result = LLVMBuildFPTrunc(ctx->builder, src0, ctx->f16, "");

   /* need to convert back up to f32 */
   result = LLVMBuildFPExt(ctx->builder, result, ctx->f32, "");
   return result;
}

static LLVMValueRef emit_umul_high(struct libresoc_llvm_context *ctx, LLVMValueRef src0,
                                   LLVMValueRef src1)
{
   LLVMValueRef dst64, result;
   src0 = LLVMBuildZExt(ctx->builder, src0, ctx->i64, "");
   src1 = LLVMBuildZExt(ctx->builder, src1, ctx->i64, "");

   dst64 = LLVMBuildMul(ctx->builder, src0, src1, "");
   dst64 = LLVMBuildLShr(ctx->builder, dst64, LLVMConstInt(ctx->i64, 32, false), "");
   result = LLVMBuildTrunc(ctx->builder, dst64, ctx->i32, "");
   return result;
}

static LLVMValueRef emit_imul_high(struct libresoc_llvm_context *ctx, LLVMValueRef src0,
                                   LLVMValueRef src1)
{
   LLVMValueRef dst64, result;
   src0 = LLVMBuildSExt(ctx->builder, src0, ctx->i64, "");
   src1 = LLVMBuildSExt(ctx->builder, src1, ctx->i64, "");

   dst64 = LLVMBuildMul(ctx->builder, src0, src1, "");
   dst64 = LLVMBuildAShr(ctx->builder, dst64, LLVMConstInt(ctx->i64, 32, false), "");
   result = LLVMBuildTrunc(ctx->builder, dst64, ctx->i32, "");
   return result;
}

static LLVMValueRef emit_bfm(struct libresoc_llvm_context *ctx, LLVMValueRef bits, LLVMValueRef offset)
{
   /* mask = ((1 << bits) - 1) << offset */
   return LLVMBuildShl(
      ctx->builder,
      LLVMBuildSub(ctx->builder, LLVMBuildShl(ctx->builder, ctx->i32_1, bits, ""), ctx->i32_1, ""),
      offset, "");
}

static LLVMValueRef emit_bitfield_select(struct libresoc_llvm_context *ctx, LLVMValueRef mask,
                                         LLVMValueRef insert, LLVMValueRef base)
{
   /* Calculate:
    *   (mask & insert) | (~mask & base) = base ^ (mask & (insert ^ base))
    * Use the right-hand side, which the LLVM backend can convert to V_BFI.
    */
   return LLVMBuildXor(
      ctx->builder, base,
      LLVMBuildAnd(ctx->builder, mask, LLVMBuildXor(ctx->builder, insert, base, ""), ""), "");
}

static LLVMValueRef emit_pack_2x16(struct libresoc_llvm_context *ctx, LLVMValueRef src0,
                                   LLVMValueRef (*pack)(struct libresoc_llvm_context *ctx,
                                                        LLVMValueRef args[2]))
{
   LLVMValueRef comp[2];

   src0 = to_float(ctx, src0);
   comp[0] = LLVMBuildExtractElement(ctx->builder, src0, ctx->i32_0, "");
   comp[1] = LLVMBuildExtractElement(ctx->builder, src0, ctx->i32_1, "");

   return LLVMBuildBitCast(ctx->builder, pack(ctx, comp), ctx->i32, "");
}

static LLVMValueRef emit_unpack_half_2x16(struct libresoc_llvm_context *ctx, LLVMValueRef src0)
{
   LLVMValueRef const16 = LLVMConstInt(ctx->i32, 16, false);
   LLVMValueRef temps[2], val;
   int i;

   for (i = 0; i < 2; i++) {
      val = i == 1 ? LLVMBuildLShr(ctx->builder, src0, const16, "") : src0;
      val = LLVMBuildTrunc(ctx->builder, val, ctx->i16, "");
      val = LLVMBuildBitCast(ctx->builder, val, ctx->f16, "");
      temps[i] = LLVMBuildFPExt(ctx->builder, val, ctx->f32, "");
   }
   return build_gather_values(ctx, temps, 2);
}

// TODO: enable this whn ac_builddxy() is added
// static LLVMValueRef emit_ddxy(struct libresoc_nir_context *ctx, nir_op op, LLVMValueRef src0)
// {
//    unsigned mask;
//    int idx;
//    LLVMValueRef result;

//    if (op == nir_op_fddx_fine)
//       mask = TID_MASK_LEFT;
//    else if (op == nir_op_fddy_fine)
//       mask = TID_MASK_TOP;
//    else
//       mask = TID_MASK_TOP_LEFT;

//    /* for DDX we want to next X pixel, DDY next Y pixel. */
//    if (op == nir_op_fddx_fine || op == nir_op_fddx_coarse || op == nir_op_fddx)
//       idx = 1;
//    else
//       idx = 2;

//    result = ac_build_ddxy(&ctx->ac, mask, idx, src0);
//    return result;
// }

static void setup_locals(struct libresoc_nir_tran_ctx *ctx, struct nir_function *func)
{
   int i, j;
   ctx->num_locals = 0;
   nir_foreach_function_temp_variable(variable, func->impl)
   {
      unsigned attrib_count = glsl_count_attribute_slots(variable->type, false);
      variable->data.driver_location = ctx->num_locals * 4;
      variable->data.location_frac = 0;
      ctx->num_locals += attrib_count;
   }
   ctx->locals = malloc(4 * ctx->num_locals * sizeof(LLVMValueRef));
   if (!ctx->locals)
      return;

   for (i = 0; i < ctx->num_locals; i++) {
      for (j = 0; j < 4; j++) {
         ctx->locals[i * 4 + j] = build_alloca_undef(&ctx->lc, ctx->lc.f32, "temp");
      }
   }
}

static void setup_scratch(struct libresoc_nir_tran_ctx *ctx, struct nir_shader *shader)
{
   if (shader->scratch_size == 0)
      return;

   ctx->scratch =
      build_alloca_undef(&ctx->lc, LLVMArrayType(ctx->lc.i8, shader->scratch_size), "scratch");
}

static void setup_constant_data(struct libresoc_nir_tran_ctx *ctx, struct nir_shader *shader)
{
    if (!shader->constant_data)
        return;

    LLVMValueRef data = LLVMConstStringInContext(ctx->lc.context, shader->constant_data,
            shader->constant_data_size, true);
    LLVMTypeRef type = LLVMArrayType(ctx->lc.i8, shader->constant_data_size);

    unsigned address_space = 0; //TODO: dummay value
    LLVMValueRef global =
        LLVMAddGlobalInAddressSpace(*(ctx->lc.module), type, "const_data", address_space);

    LLVMSetInitializer(global, data);
    LLVMSetGlobalConstant(global, true);
    LLVMSetVisibility(global, LLVMHiddenVisibility);
    ctx->constant_data = global;
}

static LLVMTypeRef glsl_base_to_llvm_type(struct libresoc_llvm_context *lc, enum glsl_base_type type)
{
   switch (type) {
   case GLSL_TYPE_INT:
   case GLSL_TYPE_UINT:
   case GLSL_TYPE_BOOL:
   case GLSL_TYPE_SUBROUTINE:
      return lc->i32;
   case GLSL_TYPE_INT8:
   case GLSL_TYPE_UINT8:
      return lc->i8;
   case GLSL_TYPE_INT16:
   case GLSL_TYPE_UINT16:
      return lc->i16;
   case GLSL_TYPE_FLOAT:
      return lc->f32;
   case GLSL_TYPE_FLOAT16:
      return lc->f16;
   case GLSL_TYPE_INT64:
   case GLSL_TYPE_UINT64:
      return lc->i64;
   case GLSL_TYPE_DOUBLE:
      return lc->f64;
   default:
      unreachable("unknown GLSL type");
   }
}

static LLVMTypeRef glsl_to_llvm_type(struct libresoc_llvm_context *lc, const struct glsl_type *type)
{
   if (glsl_type_is_scalar(type)) {
      return glsl_base_to_llvm_type(lc, glsl_get_base_type(type));
   }

   if (glsl_type_is_vector(type)) {
      return LLVMVectorType(glsl_base_to_llvm_type(lc, glsl_get_base_type(type)),
                            glsl_get_vector_elements(type));
   }

   if (glsl_type_is_matrix(type)) {
      return LLVMArrayType(glsl_to_llvm_type(lc, glsl_get_column_type(type)),
                           glsl_get_matrix_columns(type));
   }

   if (glsl_type_is_array(type)) {
      return LLVMArrayType(glsl_to_llvm_type(lc, glsl_get_array_element(type)),
                           glsl_get_length(type));
   }

   assert(glsl_type_is_struct_or_ifc(type));

   LLVMTypeRef member_types[glsl_get_length(type)];

   for (unsigned i = 0; i < glsl_get_length(type); i++) {
      member_types[i] = glsl_to_llvm_type(lc, glsl_get_struct_field(type, i));
   }

   return LLVMStructTypeInContext(lc->context, member_types, glsl_get_length(type), false);
}

// static LLVMValueRef visit_load(struct libresoc_nir_tran_ctx *ctx, nir_intrinsic_instr *instr,
//                                bool is_output)
// {
//    LLVMValueRef values[8];
//    LLVMTypeRef dest_type = get_def_type(ctx, &instr->dest.ssa);
//    LLVMTypeRef component_type;
//    unsigned base = nir_intrinsic_base(instr);
//    unsigned component = nir_intrinsic_component(instr);
//    unsigned count = instr->dest.ssa.num_components * (instr->dest.ssa.bit_size == 64 ? 2 : 1);
//    nir_src *vertex_index_src = nir_get_io_vertex_index_src(instr);
//    LLVMValueRef vertex_index = vertex_index_src ? get_src(ctx, *vertex_index_src) : NULL;
//    nir_src offset = *nir_get_io_offset_src(instr);
//    LLVMValueRef indir_index = NULL;

//    if (LLVMGetTypeKind(dest_type) == LLVMVectorTypeKind)
//       component_type = LLVMGetElementType(dest_type);
//    else
//       component_type = dest_type;

//    if (nir_src_is_const(offset))
//       assert(nir_src_as_uint(offset) == 0);
//    else
//       indir_index = get_src(ctx, offset);

//    if (ctx->stage == MESA_SHADER_TESS_CTRL || (ctx->stage == MESA_SHADER_TESS_EVAL && !is_output)) {
//       LLVMValueRef result = ctx->abi->load_tess_varyings(
//          ctx->abi, component_type, vertex_index, indir_index, 0, 0, base * 4, component,
//          instr->num_components, false, false, !is_output);
//       if (instr->dest.ssa.bit_size == 16) {
//          result = to_integer(&ctx->lc, result);
//          result = LLVMBuildTrunc(ctx->lc.builder, result, dest_type, "");
//       }
//       return LLVMBuildBitCast(ctx->lc.builder, result, dest_type, "");
//    }

//    /* No indirect indexing is allowed after this point. */
//    assert(!indir_index);

//    if (ctx->stage == MESA_SHADER_GEOMETRY) {
//       LLVMTypeRef type = LLVMIntTypeInContext(ctx->ac.context, instr->dest.ssa.bit_size);
//       assert(nir_src_is_const(*vertex_index_src));

//       return ctx->abi->load_inputs(ctx->abi, 0, base * 4, component, instr->num_components,
//                                    nir_src_as_uint(*vertex_index_src), 0, type);
//    }

//    if (ctx->stage == MESA_SHADER_FRAGMENT && is_output &&
//        nir_intrinsic_io_semantics(instr).fb_fetch_output)
//       return ctx->abi->emit_fbfetch(ctx->abi);

//    /* Other non-fragment cases have inputs and outputs in temporaries. */
//    if (ctx->stage != MESA_SHADER_FRAGMENT) {
//       for (unsigned chan = component; chan < count + component; chan++) {
//          if (is_output) {
//             values[chan] = LLVMBuildLoad(ctx->lc.builder, ctx->outputs[base * 4 + chan], "");
//          } else {
//             values[chan] = ctx->inputs[base * 4 + chan];
//             if (!values[chan])
//                values[chan] = LLVMGetUndef(ctx->lc.i32);
//          }
//       }
//       LLVMValueRef result = build_varying_gather_values(&ctx->lc, values, count, component);
//       return LLVMBuildBitCast(ctx->lc.builder, result, dest_type, "");
//    }

//    /* Fragment shader inputs. */
//    unsigned vertex_id = 2; /* P0 */

//    if (instr->intrinsic == nir_intrinsic_load_input_vertex) {
//       nir_const_value *src0 = nir_src_as_const_value(instr->src[0]);

//       switch (src0[0].i32) {
//       case 0:
//          vertex_id = 2;
//          break;
//       case 1:
//          vertex_id = 0;
//          break;
//       case 2:
//          vertex_id = 1;
//          break;
//       default:
//          unreachable("Invalid vertex index");
//       }
//    }

//    LLVMValueRef attr_number = LLVMConstInt(ctx->lc.i32, base, false);

//    for (unsigned chan = 0; chan < count; chan++) {
//       if (component + chan > 4)
//          attr_number = LLVMConstInt(ctx->lc.i32, base + 1, false);
//       LLVMValueRef llvm_chan = LLVMConstInt(ctx->lc.i32, (component + chan) % 4, false);
//       values[chan] =
//          build_fs_interp_mov(&ctx->lc, LLVMConstInt(ctx->lc.i32, vertex_id, false), llvm_chan,
//                                 attr_number, get_arg(&ctx->lc, ctx->args->prim_mask));
//       values[chan] = LLVMBuildBitCast(ctx->lc.builder, values[chan], ctx->lc.i32, "");
//       values[chan] =
//          LLVMBuildTruncOrBitCast(ctx->lc.builder, values[chan],
//                                  instr->dest.ssa.bit_size == 16 ? ctx->lc.i16 : ctx->lc.i32, "");
//    }

//    LLVMValueRef result = build_gather_values(&ctx->lc, values, count);
//    return LLVMBuildBitCast(ctx->lc.builder, result, dest_type, "");
// }

static LLVMValueRef visit_load_shared(struct libresoc_nir_tran_ctx *ctx, const nir_intrinsic_instr *instr)
{
   LLVMValueRef values[4], derived_ptr, index, ret;

   LLVMValueRef ptr = get_memory_ptr(ctx, instr->src[0], instr->dest.ssa.bit_size);

   for (int chan = 0; chan < instr->num_components; chan++) {
      index = LLVMConstInt(ctx->lc.i32, chan, 0);
      derived_ptr = LLVMBuildGEP(ctx->lc.builder, ptr, &index, 1, "");
      values[chan] = LLVMBuildLoad(ctx->lc.builder, derived_ptr, "");
   }

   ret = build_gather_values(&ctx->lc, values, instr->num_components);
   return LLVMBuildBitCast(ctx->lc.builder, ret, get_def_type(ctx, &instr->dest.ssa), "");
}

static void visit_store_shared(struct libresoc_nir_tran_ctx *ctx, const nir_intrinsic_instr *instr)
{
   LLVMValueRef derived_ptr, data, index;
   LLVMBuilderRef builder = ctx->lc.builder;

   LLVMValueRef ptr = get_memory_ptr(ctx, instr->src[1], instr->src[0].ssa->bit_size);
   LLVMValueRef src = get_src(ctx, instr->src[0]);

   int writemask = nir_intrinsic_write_mask(instr);
   for (int chan = 0; chan < 4; chan++) {
      if (!(writemask & (1 << chan))) {
         continue;
      }
      data = llvm_extract_elem(&ctx->lc, src, chan);
      index = LLVMConstInt(ctx->lc.i32, chan, 0);
      derived_ptr = LLVMBuildGEP(builder, ptr, &index, 1, "");
      LLVMBuildStore(builder, data, derived_ptr);
   }
}
static void visit_load_const(struct libresoc_nir_tran_ctx *ctx, const nir_load_const_instr *instr)
{
   LLVMValueRef values[4], value = NULL;
   LLVMTypeRef element_type = LLVMIntTypeInContext(ctx->lc.context, instr->def.bit_size);

   for (unsigned i = 0; i < instr->def.num_components; ++i) {
      switch (instr->def.bit_size) {
      case 8:
         values[i] = LLVMConstInt(element_type, instr->value[i].u8, false);
         break;
      case 16:
         values[i] = LLVMConstInt(element_type, instr->value[i].u16, false);
         break;
      case 32:
         values[i] = LLVMConstInt(element_type, instr->value[i].u32, false);
         break;
      case 64:
         values[i] = LLVMConstInt(element_type, instr->value[i].u64, false);
         break;
      default:
         fprintf(stderr, "unsupported nir load_const bit_size: %d\n", instr->def.bit_size);
         abort();
      }
   }
   if (instr->def.num_components > 1) {
      value = LLVMConstVector(values, instr->def.num_components);
   } else
      value = values[0];

   ctx->ssa_defs[instr->def.index] = value;
}

static void visit_store_output(struct libresoc_nir_tran_ctx *ctx, nir_intrinsic_instr *instr)
{
   // if (ctx->ac.postponed_kill) {
   //    LLVMValueRef cond = LLVMBuildLoad(ctx->ac.builder, ctx->ac.postponed_kill, "");
   //    ac_build_ifcc(&ctx->ac, cond, 7002);
   // }

   unsigned base = nir_intrinsic_base(instr);
   unsigned writemask = nir_intrinsic_write_mask(instr);
   unsigned component = nir_intrinsic_component(instr);
   LLVMValueRef src = to_float(&ctx->lc, get_src(ctx, instr->src[0]));
   nir_src offset = *nir_get_io_offset_src(instr);
   LLVMValueRef indir_index = NULL;

   if (nir_src_is_const(offset))
      assert(nir_src_as_uint(offset) == 0);
   else
      indir_index = get_src(ctx, offset);

   switch (get_elem_bits(&ctx->lc, LLVMTypeOf(src))) {
   case 32:
      break;
   case 64:
      writemask = widen_mask(writemask, 2);
      src = LLVMBuildBitCast(ctx->lc.builder, src,
                             LLVMVectorType(ctx->lc.f32, get_llvm_num_components(src) * 2), "");
      break;
   default:
      unreachable("unhandled store_output bit size");
      return;
   }

   writemask <<= component;

   // if (ctx->stage == MESA_SHADER_TESS_CTRL) {
   //    nir_src *vertex_index_src = nir_get_io_vertex_index_src(instr);
   //    LLVMValueRef vertex_index = vertex_index_src ? get_src(ctx, *vertex_index_src) : NULL;

   //    ctx->abi->store_tcs_outputs(ctx->abi, NULL, vertex_index, indir_index, 0, src, writemask,
   //                                component, base * 4);
   //    return;
   // }

   /* No indirect indexing is allowed after this point. */
   assert(!indir_index);

   for (unsigned chan = 0; chan < 8; chan++) {
      if (!(writemask & (1 << chan)))
         continue;

      LLVMValueRef value = llvm_extract_elem(&ctx->lc, src, chan - component);
      LLVMBuildStore(ctx->lc.builder, value, ctx->outputs[base * 4 + chan]);
   }

   // if (ctx->ac.postponed_kill)
   //    ac_build_endif(&ctx->ac, 7002);
}

static void visit_deref(struct libresoc_nir_tran_ctx *ctx, nir_deref_instr *instr)
{
   if (instr->mode != nir_var_mem_shared && instr->mode != nir_var_mem_global)
      return;

   LLVMValueRef result = NULL;
   switch (instr->deref_type) {
   case nir_deref_type_var: {
      struct hash_entry *entry = _mesa_hash_table_search(ctx->vars, instr->var);
      result = entry->data;
      break;
   }
   case nir_deref_type_struct:
      if (instr->mode == nir_var_mem_global) {
         nir_deref_instr *parent = nir_deref_instr_parent(instr);
         uint64_t offset = glsl_get_struct_field_offset(parent->type, instr->strct.index);
         result = build_gep_ptr(&ctx->lc, get_src(ctx, instr->parent),
                                   LLVMConstInt(ctx->lc.i32, offset, 0));
      } else {
         result = build_gep0(&ctx->lc, get_src(ctx, instr->parent),
                                LLVMConstInt(ctx->lc.i32, instr->strct.index, 0));
      }
      break;
   case nir_deref_type_array:
      if (instr->mode == nir_var_mem_global) {
         nir_deref_instr *parent = nir_deref_instr_parent(instr);
         unsigned stride = glsl_get_explicit_stride(parent->type);

         if ((glsl_type_is_matrix(parent->type) && glsl_matrix_type_is_row_major(parent->type)) ||
             (glsl_type_is_vector(parent->type) && stride == 0))
            stride = type_scalar_size_bytes(parent->type);

         assert(stride > 0);
         LLVMValueRef index = get_src(ctx, instr->arr.index);
         if (LLVMTypeOf(index) != ctx->lc.i64)
            index = LLVMBuildZExt(ctx->lc.builder, index, ctx->lc.i64, "");

         LLVMValueRef offset =
            LLVMBuildMul(ctx->lc.builder, index, LLVMConstInt(ctx->lc.i64, stride, 0), "");

         result = build_gep_ptr(&ctx->lc, get_src(ctx, instr->parent), offset);
      } else {
         result =
            build_gep0(&ctx->lc, get_src(ctx, instr->parent), get_src(ctx, instr->arr.index));
      }
      break;
   case nir_deref_type_ptr_as_array:
      if (instr->mode == nir_var_mem_global) {
         unsigned stride = nir_deref_instr_array_stride(instr);

         LLVMValueRef index = get_src(ctx, instr->arr.index);
         if (LLVMTypeOf(index) != ctx->lc.i64)
            index = LLVMBuildZExt(ctx->lc.builder, index, ctx->lc.i64, "");

         LLVMValueRef offset =
            LLVMBuildMul(ctx->lc.builder, index, LLVMConstInt(ctx->lc.i64, stride, 0), "");

         result = build_gep_ptr(&ctx->lc, get_src(ctx, instr->parent), offset);
      } else {
         result =
            build_gep_ptr(&ctx->lc, get_src(ctx, instr->parent), get_src(ctx, instr->arr.index));
      }
      break;
   case nir_deref_type_cast: {
      result = get_src(ctx, instr->parent);

      /* We can't use the structs from LLVM because the shader
       * specifies its own offsets. */
      LLVMTypeRef pointee_type = ctx->lc.i8;
      if (instr->mode == nir_var_mem_shared)
         pointee_type = glsl_to_llvm_type(&ctx->lc, instr->type);

      unsigned address_space;

      switch (instr->mode) {
      case nir_var_mem_shared:
         address_space = 1;
         break;
      case nir_var_mem_global:
         address_space = 0;
         break;
      default:
         unreachable("Unhandled address space");
      }

      LLVMTypeRef type = LLVMPointerType(pointee_type, address_space);

      if (LLVMTypeOf(result) != type) {
         if (LLVMGetTypeKind(LLVMTypeOf(result)) == LLVMVectorTypeKind) {
            result = LLVMBuildBitCast(ctx->lc.builder, result, type, "");
         } else {
            result = LLVMBuildIntToPtr(ctx->lc.builder, result, type, "");
         }
      }
      break;
   }
   default:
      unreachable("Unhandled deref_instr deref type");
   }

   ctx->ssa_defs[instr->dest.ssa.index] = result;
}

static void visit_phi(struct libresoc_nir_tran_ctx *ctx, nir_phi_instr *instr)
{
   LLVMTypeRef type = get_def_type(ctx, &instr->dest.ssa);
   LLVMValueRef result = LLVMBuildPhi(ctx->lc.builder, type, "");

   ctx->ssa_defs[instr->dest.ssa.index] = result;
   _mesa_hash_table_insert(ctx->phis, instr, result);
}

static bool is_def_used_in_an_export(const nir_ssa_def *def)
{
   nir_foreach_use (use_src, def) {
      if (use_src->parent_instr->type == nir_instr_type_intrinsic) {
         nir_intrinsic_instr *instr = nir_instr_as_intrinsic(use_src->parent_instr);
         if (instr->intrinsic == nir_intrinsic_store_deref)
            return true;
      } else if (use_src->parent_instr->type == nir_instr_type_alu) {
         nir_alu_instr *instr = nir_instr_as_alu(use_src->parent_instr);
         if (instr->op == nir_op_vec4 && is_def_used_in_an_export(&instr->dest.dest.ssa)) {
            return true;
         }
      }
   }
   return false;
}

static void visit_ssa_undef(struct libresoc_nir_tran_ctx *ctx, const nir_ssa_undef_instr *instr)
{
   unsigned num_components = instr->def.num_components;
   LLVMTypeRef type = LLVMIntTypeInContext(ctx->lc.context, instr->def.bit_size);

   if (/*!ctx->abi->convert_undef_to_zero ||*/ is_def_used_in_an_export(&instr->def)) {
      LLVMValueRef undef;

      if (num_components == 1)
         undef = LLVMGetUndef(type);
      else {
         undef = LLVMGetUndef(LLVMVectorType(type, num_components));
      }
      ctx->ssa_defs[instr->def.index] = undef;
   } else {
      LLVMValueRef zero = LLVMConstInt(type, 0, false);
      if (num_components > 1) {
         zero = build_gather_values_extended(&ctx->lc, &zero, 4, 0, false, false);
      }
      ctx->ssa_defs[instr->def.index] = zero;
   }
}

static void visit_jump(struct libresoc_llvm_context *lc, const nir_jump_instr *instr)
{
   switch (instr->type) {
   case nir_jump_break:
      build_break(lc);
      break;
   case nir_jump_continue:
      build_continue(lc);
      break;
   default:
      fprintf(stderr, "Unknown NIR jump instr: ");
      nir_print_instr(&instr->instr, stderr);
      fprintf(stderr, "\n");
      abort();
   }
}

static LLVMValueRef get_alu_src(struct libresoc_nir_tran_ctx *ctx, nir_alu_src src,
                                unsigned num_components)
{
   LLVMValueRef value = get_src(ctx, src.src);
   bool need_swizzle = false;

   assert(value);
   unsigned src_components = get_llvm_num_components(value);
   for (unsigned i = 0; i < num_components; ++i) {
      assert(src.swizzle[i] < src_components);
      if (src.swizzle[i] != i)
         need_swizzle = true;
   }

   if (need_swizzle || num_components != src_components) {
      LLVMValueRef masks[] = {LLVMConstInt(ctx->lc.i32, src.swizzle[0], false),
                              LLVMConstInt(ctx->lc.i32, src.swizzle[1], false),
                              LLVMConstInt(ctx->lc.i32, src.swizzle[2], false),
                              LLVMConstInt(ctx->lc.i32, src.swizzle[3], false)};

      if (src_components > 1 && num_components == 1) {
         value = LLVMBuildExtractElement(ctx->lc.builder, value, masks[0], "");
      } else if (src_components == 1 && num_components > 1) {
         LLVMValueRef values[] = {value, value, value, value};
         value = build_gather_values(&ctx->lc, values, num_components);
      } else {
         LLVMValueRef swizzle = LLVMConstVector(masks, num_components);
         value = LLVMBuildShuffleVector(ctx->lc.builder, value, value, swizzle, "");
      }
   }
   assert(!src.negate);
   assert(!src.abs);
   return value;
}

static void visit_alu(struct libresoc_nir_tran_ctx *ctx, const nir_alu_instr *instr)
{
   LLVMValueRef src[4], result = NULL;
   unsigned num_components = instr->dest.dest.ssa.num_components;
   unsigned src_components;
   LLVMTypeRef def_type = get_def_type(ctx, &instr->dest.dest.ssa);

   assert(nir_op_infos[instr->op].num_inputs <= ARRAY_SIZE(src));
   switch (instr->op) {
   case nir_op_vec2:
   case nir_op_vec3:
   case nir_op_vec4:
      src_components = 1;
      break;
   case nir_op_pack_half_2x16:
   case nir_op_pack_snorm_2x16:
   case nir_op_pack_unorm_2x16:
      src_components = 2;
      break;
   case nir_op_unpack_half_2x16:
      src_components = 1;
      break;
   case nir_op_cube_face_coord:
   case nir_op_cube_face_index:
      src_components = 3;
      break;
   default:
      src_components = num_components;
      break;
   }
   for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
      src[i] = get_alu_src(ctx, instr->src[i], src_components);

   switch (instr->op) {
   case nir_op_mov:
      result = src[0];
      break;
   case nir_op_fneg:
      src[0] = to_float(&ctx->lc, src[0]);
      result = LLVMBuildFNeg(ctx->lc.builder, src[0], "");
      if (ctx->lc.float_mode == FLOAT_MODE_DENORM_FLUSH_TO_ZERO) {
         /* fneg will be optimized by backend compiler with sign
          * bit removed via XOR. This is probably a LLVM bug.
          */
         result = build_canonicalize(&ctx->lc, result, instr->dest.dest.ssa.bit_size);
      }
      break;
   case nir_op_ineg:
      result = LLVMBuildNeg(ctx->lc.builder, src[0], "");
      break;
   case nir_op_inot:
      result = LLVMBuildNot(ctx->lc.builder, src[0], "");
      break;
   case nir_op_iadd:
      result = LLVMBuildAdd(ctx->lc.builder, src[0], src[1], "");
      break;
   case nir_op_fadd:
      src[0] = to_float(&ctx->lc, src[0]);
      src[1] = to_float(&ctx->lc, src[1]);
      result = LLVMBuildFAdd(ctx->lc.builder, src[0], src[1], "");
      break;
   case nir_op_fsub:
      src[0] = to_float(&ctx->lc, src[0]);
      src[1] = to_float(&ctx->lc, src[1]);
      result = LLVMBuildFSub(ctx->lc.builder, src[0], src[1], "");
      break;
   case nir_op_isub:
      result = LLVMBuildSub(ctx->lc.builder, src[0], src[1], "");
      break;
   case nir_op_imul:
      result = LLVMBuildMul(ctx->lc.builder, src[0], src[1], "");
      break;
   case nir_op_imod:
      result = LLVMBuildSRem(ctx->lc.builder, src[0], src[1], "");
      break;
   case nir_op_umod:
      result = LLVMBuildURem(ctx->lc.builder, src[0], src[1], "");
      break;
   case nir_op_irem:
      result = LLVMBuildSRem(ctx->lc.builder, src[0], src[1], "");
      break;
   case nir_op_idiv:
      result = LLVMBuildSDiv(ctx->lc.builder, src[0], src[1], "");
      break;
   case nir_op_udiv:
      result = LLVMBuildUDiv(ctx->lc.builder, src[0], src[1], "");
      break;
   case nir_op_fmul:
      src[0] = to_float(&ctx->lc, src[0]);
      src[1] = to_float(&ctx->lc, src[1]);
      result = LLVMBuildFMul(ctx->lc.builder, src[0], src[1], "");
      break;
   case nir_op_frcp:
      /* For doubles, we need precise division to pass GLCTS. */
      if (ctx->lc.float_mode == FLOAT_MODE_DEFAULT_OPENGL && get_type_size(def_type) == 8) {
         result = LLVMBuildFDiv(ctx->lc.builder, ctx->lc.f64_1, to_float(&ctx->lc, src[0]), "");
      } else {
         result = emit_intrin_1f_param_scalar(&ctx->lc, "llvm.amdgcn.rcp",
                                              to_float_type(&ctx->lc, def_type), src[0]);
      }
      // TODO: abi not supported
      // if (ctx->abi->clamp_div_by_zero)
      //    result = build_fmin(&ctx->lc, result,
      //                           LLVMConstReal(to_float_type(&ctx->lc, def_type), FLT_MAX));
      break;
   case nir_op_iand:
      result = LLVMBuildAnd(ctx->lc.builder, src[0], src[1], "");
      break;
   case nir_op_ior:
      result = LLVMBuildOr(ctx->lc.builder, src[0], src[1], "");
      break;
   case nir_op_ixor:
      result = LLVMBuildXor(ctx->lc.builder, src[0], src[1], "");
      break;
   case nir_op_ishl:
      if (get_elem_bits(&ctx->lc, LLVMTypeOf(src[1])) <
          get_elem_bits(&ctx->lc, LLVMTypeOf(src[0])))
         src[1] = LLVMBuildZExt(ctx->lc.builder, src[1], LLVMTypeOf(src[0]), "");
      else if (get_elem_bits(&ctx->lc, LLVMTypeOf(src[1])) >
               get_elem_bits(&ctx->lc, LLVMTypeOf(src[0])))
         src[1] = LLVMBuildTrunc(ctx->lc.builder, src[1], LLVMTypeOf(src[0]), "");
      result = LLVMBuildShl(ctx->lc.builder, src[0], src[1], "");
      break;
   case nir_op_ishr:
      if (get_elem_bits(&ctx->lc, LLVMTypeOf(src[1])) <
          get_elem_bits(&ctx->lc, LLVMTypeOf(src[0])))
         src[1] = LLVMBuildZExt(ctx->lc.builder, src[1], LLVMTypeOf(src[0]), "");
      else if (get_elem_bits(&ctx->lc, LLVMTypeOf(src[1])) >
               get_elem_bits(&ctx->lc, LLVMTypeOf(src[0])))
         src[1] = LLVMBuildTrunc(ctx->lc.builder, src[1], LLVMTypeOf(src[0]), "");
      result = LLVMBuildAShr(ctx->lc.builder, src[0], src[1], "");
      break;
   case nir_op_ushr:
      if (get_elem_bits(&ctx->lc, LLVMTypeOf(src[1])) <
          get_elem_bits(&ctx->lc, LLVMTypeOf(src[0])))
         src[1] = LLVMBuildZExt(ctx->lc.builder, src[1], LLVMTypeOf(src[0]), "");
      else if (get_elem_bits(&ctx->lc, LLVMTypeOf(src[1])) >
               get_elem_bits(&ctx->lc, LLVMTypeOf(src[0])))
         src[1] = LLVMBuildTrunc(ctx->lc.builder, src[1], LLVMTypeOf(src[0]), "");
      result = LLVMBuildLShr(ctx->lc.builder, src[0], src[1], "");
      break;
   case nir_op_ilt32:
      result = emit_int_cmp(&ctx->lc, LLVMIntSLT, src[0], src[1]);
      break;
   case nir_op_ine32:
      result = emit_int_cmp(&ctx->lc, LLVMIntNE, src[0], src[1]);
      break;
   case nir_op_ieq32:
      result = emit_int_cmp(&ctx->lc, LLVMIntEQ, src[0], src[1]);
      break;
   case nir_op_ige32:
      result = emit_int_cmp(&ctx->lc, LLVMIntSGE, src[0], src[1]);
      break;
   case nir_op_ult32:
      result = emit_int_cmp(&ctx->lc, LLVMIntULT, src[0], src[1]);
      break;
   case nir_op_uge32:
      result = emit_int_cmp(&ctx->lc, LLVMIntUGE, src[0], src[1]);
      break;
   case nir_op_feq32:
      result = emit_float_cmp(&ctx->lc, LLVMRealOEQ, src[0], src[1]);
      break;
   case nir_op_fneu32:
      result = emit_float_cmp(&ctx->lc, LLVMRealUNE, src[0], src[1]);
      break;
   case nir_op_flt32:
      result = emit_float_cmp(&ctx->lc, LLVMRealOLT, src[0], src[1]);
      break;
   case nir_op_fge32:
      result = emit_float_cmp(&ctx->lc, LLVMRealOGE, src[0], src[1]);
      break;
   case nir_op_fabs:
      result =
         emit_intrin_1f_param(&ctx->lc, "llvm.fabs", to_float_type(&ctx->lc, def_type), src[0]);
      if (ctx->lc.float_mode == FLOAT_MODE_DENORM_FLUSH_TO_ZERO) {
         /* fabs will be optimized by backend compiler with sign
          * bit removed via AND.
          */
         result = build_canonicalize(&ctx->lc, result, instr->dest.dest.ssa.bit_size);
      }
      break;
   case nir_op_iabs:
      result = emit_iabs(&ctx->lc, src[0]);
      break;
   case nir_op_imax:
      result = build_imax(&ctx->lc, src[0], src[1]);
      break;
   case nir_op_imin:
      result = build_imin(&ctx->lc, src[0], src[1]);
      break;
   case nir_op_umax:
      result = build_umax(&ctx->lc, src[0], src[1]);
      break;
   case nir_op_umin:
      result = build_umin(&ctx->lc, src[0], src[1]);
      break;
   case nir_op_isign:
      result = build_isign(&ctx->lc, src[0]);
      break;
   case nir_op_fsign:
      src[0] = to_float(&ctx->lc, src[0]);
      result = build_fsign(&ctx->lc, src[0]);
      break;
   case nir_op_ffloor:
      result =
         emit_intrin_1f_param(&ctx->lc, "llvm.floor", to_float_type(&ctx->lc, def_type), src[0]);
      break;
   case nir_op_ftrunc:
      result =
         emit_intrin_1f_param(&ctx->lc, "llvm.trunc", to_float_type(&ctx->lc, def_type), src[0]);
      break;
   case nir_op_fceil:
      result =
         emit_intrin_1f_param(&ctx->lc, "llvm.ceil", to_float_type(&ctx->lc, def_type), src[0]);
      break;
   case nir_op_fround_even:
      result =
         emit_intrin_1f_param(&ctx->lc, "llvm.rint", to_float_type(&ctx->lc, def_type), src[0]);
      break;
   case nir_op_ffract:
      result = emit_intrin_1f_param_scalar(&ctx->lc, "llvm.amdgcn.fract",
                                           to_float_type(&ctx->lc, def_type), src[0]);
      break;
   case nir_op_fsin:
      result =
         emit_intrin_1f_param(&ctx->lc, "llvm.sin", to_float_type(&ctx->lc, def_type), src[0]);
      break;
   case nir_op_fcos:
      result =
         emit_intrin_1f_param(&ctx->lc, "llvm.cos", to_float_type(&ctx->lc, def_type), src[0]);
      break;
   case nir_op_fsqrt:
      result =
         emit_intrin_1f_param(&ctx->lc, "llvm.sqrt", to_float_type(&ctx->lc, def_type), src[0]);
      break;
   case nir_op_fexp2:
      result =
         emit_intrin_1f_param(&ctx->lc, "llvm.exp2", to_float_type(&ctx->lc, def_type), src[0]);
      break;
   case nir_op_flog2:
      result =
         emit_intrin_1f_param(&ctx->lc, "llvm.log2", to_float_type(&ctx->lc, def_type), src[0]);
      break;
   case nir_op_frsq:
      result = emit_intrin_1f_param_scalar(&ctx->lc, "llvm.amdgcn.rsq",
                                           to_float_type(&ctx->lc, def_type), src[0]);
      // TODO: abi not enabled
      // if (ctx->abi->clamp_div_by_zero)
      //    result = build_fmin(&ctx->lc, result,
      //                           LLVMConstReal(to_float_type(&ctx->lc, def_type), FLT_MAX));
      break;
   case nir_op_frexp_exp:
      // TODO: enable this when ac_build_frexp_exp() is added
      // src[0] = to_float(&ctx->lc, src[0]);
      // result = ac_build_frexp_exp(&ctx->lc, src[0], get_elem_bits(&ctx->lc, LLVMTypeOf(src[0])));
      // if (get_elem_bits(&ctx->lc, LLVMTypeOf(src[0])) == 16)
      //    result = LLVMBuildSExt(ctx->lc.builder, result, ctx->lc.i32, "");
      break;
   case nir_op_frexp_sig:
      // TODO: enable this when ac_build_frexp_mant() is added
      // src[0] = to_float(&ctx->lc, src[0]);
      // result = ac_build_frexp_mant(&ctx->lc, src[0], instr->dest.dest.ssa.bit_size);
      // break;
   // case nir_op_fpow:
      // result = emit_intrin_2f_param(&ctx->lc, "llvm.pow", to_float_type(&ctx->lc, def_type),
      //                               src[0], src[1]);
      break;
   case nir_op_fmax:
      result = emit_intrin_2f_param(&ctx->lc, "llvm.maxnum", to_float_type(&ctx->lc, def_type),
                                    src[0], src[1]);
      break;
   case nir_op_fmin:
      result = emit_intrin_2f_param(&ctx->lc, "llvm.minnum", to_float_type(&ctx->lc, def_type),
                                    src[0], src[1]);
      break;
   case nir_op_ffma:
      result =
         emit_intrin_3f_param(&ctx->lc, "llvm.fmuladd",
                              to_float_type(&ctx->lc, def_type), src[0], src[1], src[2]);
      break;
   case nir_op_ldexp:
      src[0] = to_float(&ctx->lc, src[0]);
      if (get_elem_bits(&ctx->lc, def_type) == 32)
         result = build_intrinsic(&ctx->lc, "llvm.amdgcn.ldexp.f32", ctx->lc.f32, src, 2,
                                     FUNC_ATTR_READNONE);
      else if (get_elem_bits(&ctx->lc, def_type) == 16)
         result = build_intrinsic(&ctx->lc, "llvm.amdgcn.ldexp.f16", ctx->lc.f16, src, 2,
                                     FUNC_ATTR_READNONE);
      else
         result = build_intrinsic(&ctx->lc, "llvm.amdgcn.ldexp.f64", ctx->lc.f64, src, 2,
                                     FUNC_ATTR_READNONE);
      break;
   case nir_op_bfm:
      result = emit_bfm(&ctx->lc, src[0], src[1]);
      break;
   case nir_op_bitfield_select:
      result = emit_bitfield_select(&ctx->lc, src[0], src[1], src[2]);
      break;
   case nir_op_ubfe:
      result = build_bfe(&ctx->lc, src[0], src[1], src[2], false);
      break;
   case nir_op_ibfe:
      result = build_bfe(&ctx->lc, src[0], src[1], src[2], true);
      break;
   case nir_op_bitfield_reverse:
      result = build_bitfield_reverse(&ctx->lc, src[0]);
      break;
   case nir_op_bit_count:
      result = build_bit_count(&ctx->lc, src[0]);
      break;
   case nir_op_vec2:
   case nir_op_vec3:
   case nir_op_vec4:
      for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
         src[i] = to_integer(&ctx->lc, src[i]);
      result = build_gather_values(&ctx->lc, src, num_components);
      break;
   case nir_op_f2i8:
   case nir_op_f2i16:
   case nir_op_f2i32:
   case nir_op_f2i64:
      src[0] = to_float(&ctx->lc, src[0]);
      result = LLVMBuildFPToSI(ctx->lc.builder, src[0], def_type, "");
      break;
   case nir_op_f2u8:
   case nir_op_f2u16:
   case nir_op_f2u32:
   case nir_op_f2u64:
      src[0] = to_float(&ctx->lc, src[0]);
      result = LLVMBuildFPToUI(ctx->lc.builder, src[0], def_type, "");
      break;
   case nir_op_i2f16:
   case nir_op_i2f32:
   case nir_op_i2f64:
      result = LLVMBuildSIToFP(ctx->lc.builder, src[0], to_float_type(&ctx->lc, def_type), "");
      break;
   case nir_op_u2f16:
   case nir_op_u2f32:
   case nir_op_u2f64:
      result = LLVMBuildUIToFP(ctx->lc.builder, src[0], to_float_type(&ctx->lc, def_type), "");
      break;
   case nir_op_f2f16_rtz:
   case nir_op_f2f16:
   case nir_op_f2fmp:
      src[0] = to_float(&ctx->lc, src[0]);

      /* For OpenGL, we want fast packing with v_cvt_pkrtz_f16, but if we use it,
       * all f32->f16 conversions have to round towards zero, because both scalar
       * and vec2 down-conversions have to round equally.
       */
      if (ctx->lc.float_mode == FLOAT_MODE_DEFAULT_OPENGL || instr->op == nir_op_f2f16_rtz) {
         src[0] = to_float(&ctx->lc, src[0]);

         if (LLVMTypeOf(src[0]) == ctx->lc.f64)
            src[0] = LLVMBuildFPTrunc(ctx->lc.builder, src[0], ctx->lc.f32, "");

         /* Fast path conversion. This only works if NIR is vectorized
          * to vec2 16.
          */
         if (LLVMTypeOf(src[0]) == ctx->lc.v2f32) {
            LLVMValueRef args[] = {
               llvm_extract_elem(&ctx->lc, src[0], 0),
               llvm_extract_elem(&ctx->lc, src[0], 1),
            };
            result = build_cvt_pkrtz_f16(&ctx->lc, args);
            break;
         }

         assert(get_llvm_num_components(src[0]) == 1);
         LLVMValueRef param[2] = {src[0], LLVMGetUndef(ctx->lc.f32)};
         result = build_cvt_pkrtz_f16(&ctx->lc, param);
         result = LLVMBuildExtractElement(ctx->lc.builder, result, ctx->lc.i32_0, "");
      } else {
         if (get_elem_bits(&ctx->lc, LLVMTypeOf(src[0])) < get_elem_bits(&ctx->lc, def_type))
            result =
               LLVMBuildFPExt(ctx->lc.builder, src[0], to_float_type(&ctx->lc, def_type), "");
         else
            result =
               LLVMBuildFPTrunc(ctx->lc.builder, src[0], to_float_type(&ctx->lc, def_type), "");
      }
      break;
   case nir_op_f2f16_rtne:
   case nir_op_f2f32:
   case nir_op_f2f64:
      src[0] = to_float(&ctx->lc, src[0]);
      if (get_elem_bits(&ctx->lc, LLVMTypeOf(src[0])) < get_elem_bits(&ctx->lc, def_type))
         result = LLVMBuildFPExt(ctx->lc.builder, src[0], to_float_type(&ctx->lc, def_type), "");
      else
         result =
            LLVMBuildFPTrunc(ctx->lc.builder, src[0], to_float_type(&ctx->lc, def_type), "");
      break;
   case nir_op_u2u8:
   case nir_op_u2u16:
   case nir_op_u2ump:
   case nir_op_u2u32:
   case nir_op_u2u64:
      if (get_elem_bits(&ctx->lc, LLVMTypeOf(src[0])) < get_elem_bits(&ctx->lc, def_type))
         result = LLVMBuildZExt(ctx->lc.builder, src[0], def_type, "");
      else
         result = LLVMBuildTrunc(ctx->lc.builder, src[0], def_type, "");
      break;
   case nir_op_i2i8:
   case nir_op_i2i16:
   case nir_op_i2imp:
   case nir_op_i2i32:
   case nir_op_i2i64:
      if (get_elem_bits(&ctx->lc, LLVMTypeOf(src[0])) < get_elem_bits(&ctx->lc, def_type))
         result = LLVMBuildSExt(ctx->lc.builder, src[0], def_type, "");
      else
         result = LLVMBuildTrunc(ctx->lc.builder, src[0], def_type, "");
      break;
   case nir_op_b32csel:
      result = emit_bcsel(&ctx->lc, src[0], src[1], src[2]);
      break;
   case nir_op_find_lsb:
      result = find_lsb(&ctx->lc, ctx->lc.i32, src[0]);
      break;
   case nir_op_ufind_msb:
      result = build_umsb(&ctx->lc, src[0], ctx->lc.i32);
      break;
   case nir_op_ifind_msb:
      result = build_imsb(&ctx->lc, src[0], ctx->lc.i32);
      break;
   case nir_op_uadd_carry:
      result = emit_uint_carry(&ctx->lc, "llvm.uadd.with.overflow.i32", src[0], src[1]);
      break;
   case nir_op_usub_borrow:
      result = emit_uint_carry(&ctx->lc, "llvm.usub.with.overflow.i32", src[0], src[1]);
      break;
   case nir_op_b2f16:
   case nir_op_b2f32:
   case nir_op_b2f64:
      result = emit_b2f(&ctx->lc, src[0], instr->dest.dest.ssa.bit_size);
      break;
   case nir_op_f2b32:
      result = emit_f2b(&ctx->lc, src[0]);
      break;
   case nir_op_b2i8:
   case nir_op_b2i16:
   case nir_op_b2i32:
   case nir_op_b2i64:
      result = emit_b2i(&ctx->lc, src[0], instr->dest.dest.ssa.bit_size);
      break;
   case nir_op_i2b32:
      result = emit_i2b(&ctx->lc, src[0]);
      break;
   case nir_op_fquantize2f16:
      result = emit_f2f16(&ctx->lc, src[0]);
      break;
   case nir_op_umul_high:
      result = emit_umul_high(&ctx->lc, src[0], src[1]);
      break;
   case nir_op_imul_high:
      result = emit_imul_high(&ctx->lc, src[0], src[1]);
      break;
   case nir_op_pack_half_2x16:
      result = emit_pack_2x16(&ctx->lc, src[0], build_cvt_pkrtz_f16);
      break;
   case nir_op_pack_snorm_2x16:
      result = emit_pack_2x16(&ctx->lc, src[0], build_cvt_pknorm_i16);
      break;
   case nir_op_pack_unorm_2x16:
      result = emit_pack_2x16(&ctx->lc, src[0], build_cvt_pknorm_u16);
      break;
   case nir_op_unpack_half_2x16:
      result = emit_unpack_half_2x16(&ctx->lc, src[0]);
      break;
   case nir_op_fddx:
   case nir_op_fddy:
   case nir_op_fddx_fine:
   case nir_op_fddy_fine:
   case nir_op_fddx_coarse:
   case nir_op_fddy_coarse:
      // TODO: enable this when emit_ddxy() is added
      //result = emit_ddxy(ctx, instr->op, src[0]);
      break;

   case nir_op_unpack_64_2x32_split_x: {
      assert(get_llvm_num_components(src[0]) == 1);
      LLVMValueRef tmp = LLVMBuildBitCast(ctx->lc.builder, src[0], ctx->lc.v2i32, "");
      result = LLVMBuildExtractElement(ctx->lc.builder, tmp, ctx->lc.i32_0, "");
      break;
   }

   case nir_op_unpack_64_2x32_split_y: {
      assert(get_llvm_num_components(src[0]) == 1);
      LLVMValueRef tmp = LLVMBuildBitCast(ctx->lc.builder, src[0], ctx->lc.v2i32, "");
      result = LLVMBuildExtractElement(ctx->lc.builder, tmp, ctx->lc.i32_1, "");
      break;
   }

   case nir_op_pack_64_2x32_split: {
      LLVMValueRef tmp = build_gather_values(&ctx->lc, src, 2);
      result = LLVMBuildBitCast(ctx->lc.builder, tmp, ctx->lc.i64, "");
      break;
   }

   case nir_op_pack_32_2x16_split: {
      LLVMValueRef tmp = build_gather_values(&ctx->lc, src, 2);
      result = LLVMBuildBitCast(ctx->lc.builder, tmp, ctx->lc.i32, "");
      break;
   }

   case nir_op_unpack_32_2x16_split_x: {
      LLVMValueRef tmp = LLVMBuildBitCast(ctx->lc.builder, src[0], ctx->lc.v2i16, "");
      result = LLVMBuildExtractElement(ctx->lc.builder, tmp, ctx->lc.i32_0, "");
      break;
   }

   case nir_op_unpack_32_2x16_split_y: {
      LLVMValueRef tmp = LLVMBuildBitCast(ctx->lc.builder, src[0], ctx->lc.v2i16, "");
      result = LLVMBuildExtractElement(ctx->lc.builder, tmp, ctx->lc.i32_1, "");
      break;
   }

   case nir_op_cube_face_coord: {
      src[0] = to_float(&ctx->lc, src[0]);
      LLVMValueRef results[2];
      LLVMValueRef in[3];
      for (unsigned chan = 0; chan < 3; chan++)
         in[chan] = llvm_extract_elem(&ctx->lc, src[0], chan);
      results[0] = build_intrinsic(&ctx->lc, "llvm.amdgcn.cubesc", ctx->lc.f32, in, 3,
                                      FUNC_ATTR_READNONE);
      results[1] = build_intrinsic(&ctx->lc, "llvm.amdgcn.cubetc", ctx->lc.f32, in, 3,
                                      FUNC_ATTR_READNONE);
      LLVMValueRef ma = build_intrinsic(&ctx->lc, "llvm.amdgcn.cubema", ctx->lc.f32, in, 3,
                                           FUNC_ATTR_READNONE);
      results[0] = build_fdiv(&ctx->lc, results[0], ma);
      results[1] = build_fdiv(&ctx->lc, results[1], ma);
      LLVMValueRef offset = LLVMConstReal(ctx->lc.f32, 0.5);
      results[0] = LLVMBuildFAdd(ctx->lc.builder, results[0], offset, "");
      results[1] = LLVMBuildFAdd(ctx->lc.builder, results[1], offset, "");
      result = build_gather_values(&ctx->lc, results, 2);
      break;
   }

   case nir_op_cube_face_index: {
      src[0] = to_float(&ctx->lc, src[0]);
      LLVMValueRef in[3];
      for (unsigned chan = 0; chan < 3; chan++)
         in[chan] = llvm_extract_elem(&ctx->lc, src[0], chan);
      result = build_intrinsic(&ctx->lc, "llvm.amdgcn.cubeid", ctx->lc.f32, in, 3,
                                  FUNC_ATTR_READNONE);
      break;
   }

   default:
      fprintf(stderr, "Unknown NIR alu instr: ");
      nir_print_instr(&instr->instr, stderr);
      fprintf(stderr, "\n");
      abort();
   }

   if (result) {
      assert(instr->dest.dest.is_ssa);
      result = to_integer_or_pointer(&ctx->lc, result);
      ctx->ssa_defs[instr->dest.dest.ssa.index] = result;
   }
}

static LLVMValueRef visit_load_var(struct libresoc_nir_tran_ctx *ctx, nir_intrinsic_instr *instr)
{
   nir_deref_instr *deref = nir_instr_as_deref(instr->src[0].ssa->parent_instr);
   nir_variable *var = nir_deref_instr_get_variable(deref);

   LLVMValueRef values[8];
   int idx = 0;
   int ve = instr->dest.ssa.num_components;
   unsigned comp = 0;
   LLVMValueRef indir_index;
   LLVMValueRef ret;
   unsigned const_index;
   unsigned stride = 4;
   int mode = deref->mode;

   if (var) {
      bool vs_in = ctx->stage == MESA_SHADER_VERTEX && var->data.mode == nir_var_shader_in;
      idx = var->data.driver_location;
      comp = var->data.location_frac;
      mode = var->data.mode;

      get_deref_offset(ctx, deref, vs_in, NULL, NULL, &const_index, &indir_index);

      if (var->data.compact) {
         stride = 1;
         const_index += comp;
         comp = 0;
      }
   }

   if (instr->dest.ssa.bit_size == 64 &&
       (deref->mode == nir_var_shader_in || deref->mode == nir_var_shader_out ||
        deref->mode == nir_var_function_temp))
      ve *= 2;

   switch (mode) {
   case nir_var_shader_in:
      /* TODO: remove this after RADV switches to lowered IO */
      // if (ctx->stage == MESA_SHADER_TESS_CTRL || ctx->stage == MESA_SHADER_TESS_EVAL) {
      //    return load_tess_varyings(ctx, instr, true);
      // }

      // if (ctx->stage == MESA_SHADER_GEOMETRY) {
      //    LLVMTypeRef type = LLVMIntTypeInContext(ctx->lc.context, instr->dest.ssa.bit_size);
      //    LLVMValueRef indir_index;
      //    unsigned const_index, vertex_index;
      //    get_deref_offset(ctx, deref, false, &vertex_index, NULL, &const_index, &indir_index);
      //    assert(indir_index == NULL);

      //    return ctx->abi->load_inputs(ctx->abi, var->data.location, var->data.driver_location,
      //                                 var->data.location_frac, instr->num_components, vertex_index,
      //                                 const_index, type);
      // }

      for (unsigned chan = comp; chan < ve + comp; chan++) {
         if (indir_index) {
            unsigned count =
               glsl_count_attribute_slots(var->type, ctx->stage == MESA_SHADER_VERTEX);
            count -= chan / 4;
            LLVMValueRef tmp_vec = build_gather_values_extended(
               &ctx->lc, ctx->inputs + idx + chan, count, stride, false, true);

            values[chan] = LLVMBuildExtractElement(ctx->lc.builder, tmp_vec, indir_index, "");
         } else
            values[chan] = ctx->inputs[idx + chan + const_index * stride];
      }
      break;
   case nir_var_function_temp:
      for (unsigned chan = 0; chan < ve; chan++) {
         if (indir_index) {
            unsigned count = glsl_count_attribute_slots(var->type, false);
            count -= chan / 4;
            LLVMValueRef tmp_vec = build_gather_values_extended(
               &ctx->lc, ctx->locals + idx + chan, count, stride, true, true);

            values[chan] = LLVMBuildExtractElement(ctx->lc.builder, tmp_vec, indir_index, "");
         } else {
            values[chan] =
               LLVMBuildLoad(ctx->lc.builder, ctx->locals[idx + chan + const_index * stride], "");
         }
      }
      break;
   case nir_var_shader_out:
      /* TODO: remove this after RADV switches to lowered IO */
      // if (ctx->stage == MESA_SHADER_TESS_CTRL) {
      //    return load_tess_varyings(ctx, instr, false);
      // }

      // if (ctx->stage == MESA_SHADER_FRAGMENT && var->data.fb_fetch_output && ctx->abi->emit_fbfetch)
      //    return ctx->abi->emit_fbfetch(ctx->abi);

      for (unsigned chan = comp; chan < ve + comp; chan++) {
         if (indir_index) {
            unsigned count = glsl_count_attribute_slots(var->type, false);
            count -= chan / 4;
            LLVMValueRef tmp_vec = build_gather_values_extended(
               &ctx->lc, ctx->outputs + idx + chan, count, stride, true, true);

            values[chan] = LLVMBuildExtractElement(ctx->lc.builder, tmp_vec, indir_index, "");
         } else {
            values[chan] = LLVMBuildLoad(ctx->lc.builder,
                                         ctx->outputs[idx + chan + const_index * stride], "");
         }
      }
      break;
   case nir_var_mem_global: {
      LLVMValueRef address = get_src(ctx, instr->src[0]);
      LLVMTypeRef result_type = get_def_type(ctx, &instr->dest.ssa);
      unsigned explicit_stride = glsl_get_explicit_stride(deref->type);
      unsigned natural_stride = type_scalar_size_bytes(deref->type);
      unsigned stride = explicit_stride ? explicit_stride : natural_stride;
      int elem_size_bytes = get_elem_bits(&ctx->lc, result_type) / 8;
      bool split_loads = false;

      if (stride != natural_stride || split_loads) {
         if (LLVMGetTypeKind(result_type) == LLVMVectorTypeKind)
            result_type = LLVMGetElementType(result_type);

         LLVMTypeRef ptr_type =
            LLVMPointerType(result_type, LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
         address = LLVMBuildBitCast(ctx->lc.builder, address, ptr_type, "");

         for (unsigned i = 0; i < instr->dest.ssa.num_components; ++i) {
            LLVMValueRef offset = LLVMConstInt(ctx->lc.i32, i * stride / natural_stride, 0);
            values[i] =
               LLVMBuildLoad(ctx->lc.builder, build_gep_ptr(&ctx->lc, address, offset), "");

            if (nir_intrinsic_access(instr) & (ACCESS_COHERENT | ACCESS_VOLATILE))
               LLVMSetOrdering(values[i], LLVMAtomicOrderingMonotonic);
         }
         return build_gather_values(&ctx->lc, values, instr->dest.ssa.num_components);
      } else {
         LLVMTypeRef ptr_type =
            LLVMPointerType(result_type, LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
         address = LLVMBuildBitCast(ctx->lc.builder, address, ptr_type, "");
         LLVMValueRef val = LLVMBuildLoad(ctx->lc.builder, address, "");

         if (nir_intrinsic_access(instr) & (ACCESS_COHERENT | ACCESS_VOLATILE))
            LLVMSetOrdering(val, LLVMAtomicOrderingMonotonic);
         return val;
      }
   }
   default:
      unreachable("unhandle variable mode");
   }
   ret = build_varying_gather_values(&ctx->lc, values, ve, comp);
   return LLVMBuildBitCast(ctx->lc.builder, ret, get_def_type(ctx, &instr->dest.ssa), "");
}

static void visit_store_var(struct libresoc_nir_tran_ctx *ctx, nir_intrinsic_instr *instr)
{
   // if (ctx->lc.postponed_kill) {
   //    LLVMValueRef cond = LLVMBuildLoad(ctx->lc.builder, ctx->lc.postponed_kill, "");
   //    ac_build_ifcc(&ctx->lc, cond, 7002);
   // }

   nir_deref_instr *deref = nir_instr_as_deref(instr->src[0].ssa->parent_instr);
   nir_variable *var = nir_deref_instr_get_variable(deref);

   LLVMValueRef temp_ptr, value;
   int idx = 0;
   unsigned comp = 0;
   LLVMValueRef src = to_float(&ctx->lc, get_src(ctx, instr->src[1]));
   int writemask = instr->const_index[0];
   LLVMValueRef indir_index;
   unsigned const_index;

   if (var) {
      get_deref_offset(ctx, deref, false, NULL, NULL, &const_index, &indir_index);
      idx = var->data.driver_location;
      comp = var->data.location_frac;

      if (var->data.compact) {
         const_index += comp;
         comp = 0;
      }
   }

   if (get_elem_bits(&ctx->lc, LLVMTypeOf(src)) == 64 &&
       (deref->mode == nir_var_shader_out || deref->mode == nir_var_function_temp)) {

      src = LLVMBuildBitCast(ctx->lc.builder, src,
                             LLVMVectorType(ctx->lc.f32, get_llvm_num_components(src) * 2), "");

      writemask = widen_mask(writemask, 2);
   }

   writemask = writemask << comp;

   switch (deref->mode) {
   case nir_var_shader_out:
       if (ctx->stage == MESA_SHADER_VERTEX && var->data.location == VARYING_SLOT_POS) {

           LLVMValueRef vertexCache = LLVMGetParam(ctx->main_function, 2); 
           LLVMValueRef idx[3];

           idx[0] = LLVMConstInt(ctx->lc.i32, 0, false);
           idx[1] = LLVMConstInt(ctx->lc.i32, 0, false);
           LLVMValueRef outPos = LLVMConstInBoundsGEP(vertexCache,
                   idx,
                   2);
           LLVMBuildStore(ctx->lc.builder, src, outPos);
           break;
       }
       if (ctx->stage == MESA_SHADER_FRAGMENT && var->data.location == FRAG_RESULT_DATA0) {

           LLVMValueRef outColor = LLVMGetParam(ctx->main_function, 4); 
           // LLVMValueRef idx[3];

           // idx[0] = LLVMConstInt(ctx->lc.i32, 0, false);
           // idx[1] = LLVMConstInt(ctx->lc.i32, 0, false);
           // LLVMValueRef outPos = LLVMConstInBoundsGEP(vertexCache,
           //         idx,
           //         2);
           LLVMBuildStore(ctx->lc.builder, src, outColor);
           break;
       }
      /* TODO: remove this after RADV switches to lowered IO */
      // if (ctx->stage == MESA_SHADER_TESS_CTRL) {
      //    LLVMValueRef vertex_index = NULL;
      //    LLVMValueRef indir_index = NULL;
      //    unsigned const_index = 0;
      //    const bool is_patch = var->data.patch ||
      //                          var->data.location == VARYING_SLOT_TESS_LEVEL_INNER ||
      //                          var->data.location == VARYING_SLOT_TESS_LEVEL_OUTER;

      //    get_deref_offset(ctx, deref, false, NULL, is_patch ? NULL : &vertex_index, &const_index,
      //                     &indir_index);

      //    ctx->abi->store_tcs_outputs(ctx->abi, var, vertex_index, indir_index, const_index, src,
      //                                writemask, var->data.location_frac, var->data.driver_location);
      //    break;
      // }

      for (unsigned chan = 0; chan < 8; chan++) {
         int stride = 4;
         if (!(writemask & (1 << chan)))
            continue;

         value = llvm_extract_elem(&ctx->lc, src, chan - comp);

         if (var->data.compact)
            stride = 1;
         if (indir_index) {
            unsigned count = glsl_count_attribute_slots(var->type, false);
            count -= chan / 4;
            LLVMValueRef tmp_vec = build_gather_values_extended(
               &ctx->lc, ctx->outputs + idx + chan, count, stride, true, true);

            tmp_vec = LLVMBuildInsertElement(ctx->lc.builder, tmp_vec, value, indir_index, "");
            build_store_values_extended(&ctx->lc, ctx->outputs + idx + chan, count, stride,
                                        tmp_vec);

         } else {
            temp_ptr = ctx->outputs[idx + chan + const_index * stride];

            LLVMBuildStore(ctx->lc.builder, value, temp_ptr);
         }
      }
      break;
   case nir_var_function_temp:
      for (unsigned chan = 0; chan < 8; chan++) {
         if (!(writemask & (1 << chan)))
            continue;

         value = llvm_extract_elem(&ctx->lc, src, chan);
         if (indir_index) {
            unsigned count = glsl_count_attribute_slots(var->type, false);
            count -= chan / 4;
            LLVMValueRef tmp_vec = build_gather_values_extended(
               &ctx->lc, ctx->locals + idx + chan, count, 4, true, true);

            tmp_vec = LLVMBuildInsertElement(ctx->lc.builder, tmp_vec, value, indir_index, "");
            build_store_values_extended(&ctx->lc, ctx->locals + idx + chan, count, 4, tmp_vec);
         } else {
            temp_ptr = ctx->locals[idx + chan + const_index * 4];

            LLVMBuildStore(ctx->lc.builder, value, temp_ptr);
         }
      }
      break;

   case nir_var_mem_global: {
      int writemask = instr->const_index[0];
      LLVMValueRef address = get_src(ctx, instr->src[0]);
      LLVMValueRef val = get_src(ctx, instr->src[1]);

      unsigned explicit_stride = glsl_get_explicit_stride(deref->type);
      unsigned natural_stride = type_scalar_size_bytes(deref->type);
      unsigned stride = explicit_stride ? explicit_stride : natural_stride;
      int elem_size_bytes = get_elem_bits(&ctx->lc, LLVMTypeOf(val)) / 8;
      bool split_stores = false; 

      LLVMTypeRef ptr_type =
         LLVMPointerType(LLVMTypeOf(val), LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
      address = LLVMBuildBitCast(ctx->lc.builder, address, ptr_type, "");

      if (writemask == (1u << get_llvm_num_components(val)) - 1 && stride == natural_stride &&
          !split_stores) {
         LLVMTypeRef ptr_type =
            LLVMPointerType(LLVMTypeOf(val), LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
         address = LLVMBuildBitCast(ctx->lc.builder, address, ptr_type, "");

         val = LLVMBuildBitCast(ctx->lc.builder, val, LLVMGetElementType(LLVMTypeOf(address)), "");
         LLVMValueRef store = LLVMBuildStore(ctx->lc.builder, val, address);

         if (nir_intrinsic_access(instr) & (ACCESS_COHERENT | ACCESS_VOLATILE))
            LLVMSetOrdering(store, LLVMAtomicOrderingMonotonic);
      } else {
         LLVMTypeRef val_type = LLVMTypeOf(val);
         if (LLVMGetTypeKind(LLVMTypeOf(val)) == LLVMVectorTypeKind)
            val_type = LLVMGetElementType(val_type);

         LLVMTypeRef ptr_type =
            LLVMPointerType(val_type, LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
         address = LLVMBuildBitCast(ctx->lc.builder, address, ptr_type, "");
         for (unsigned chan = 0; chan < 4; chan++) {
            if (!(writemask & (1 << chan)))
               continue;

            LLVMValueRef offset = LLVMConstInt(ctx->lc.i32, chan * stride / natural_stride, 0);

            LLVMValueRef ptr = build_gep_ptr(&ctx->lc, address, offset);
            LLVMValueRef src = llvm_extract_elem(&ctx->lc, val, chan);
            src = LLVMBuildBitCast(ctx->lc.builder, src, LLVMGetElementType(LLVMTypeOf(ptr)), "");
            LLVMValueRef store = LLVMBuildStore(ctx->lc.builder, src, ptr);

            if (nir_intrinsic_access(instr) & (ACCESS_COHERENT | ACCESS_VOLATILE))
               LLVMSetOrdering(store, LLVMAtomicOrderingMonotonic);
         }
      }
      break;
   }
   default:
      abort();
      break;
   }

   // if (ctx->ac.postponed_kill)
   //    ac_build_endif(&ctx->ac, 7002);
}

static void visit_intrinsic(struct libresoc_nir_tran_ctx *ctx, nir_intrinsic_instr *instr)
{
   LLVMValueRef result = NULL;

   switch (instr->intrinsic) {
   case nir_intrinsic_ballot:
      // result = ac_build_ballot(&ctx->ac, get_src(ctx, instr->src[0]));
      // if (ctx->ac.ballot_mask_bits > ctx->ac.wave_size)
      //    result = LLVMBuildZExt(ctx->ac.builder, result, ctx->ac.iN_ballotmask, "");
      break;
   case nir_intrinsic_read_invocation:
      // result =
      //    ac_build_readlane(&ctx->ac, get_src(ctx, instr->src[0]), get_src(ctx, instr->src[1]));
      break;
   case nir_intrinsic_read_first_invocation:
      // result = ac_build_readlane(&ctx->ac, get_src(ctx, instr->src[0]), NULL);
      break;
   case nir_intrinsic_load_subgroup_invocation:
      // result = ac_get_thread_id(&ctx->ac);
      break;
   case nir_intrinsic_load_work_group_id: {
      // LLVMValueRef values[3];

      // for (int i = 0; i < 3; i++) {
      //    values[i] = ctx->args->workgroup_ids[i].used
      //                   ? ac_get_arg(&ctx->ac, ctx->args->workgroup_ids[i])
      //                   : ctx->ac.i32_0;
      // }

      // result = ac_build_gather_values(&ctx->ac, values, 3);
      break;
   }
   case nir_intrinsic_load_base_vertex:
   case nir_intrinsic_load_first_vertex:
      //result = ctx->abi->load_base_vertex(ctx->abi);
        result = LLVMGetParam(ctx->main_function, ctx->args.base_vertex.arg_index); 
      break;
   case nir_intrinsic_load_local_group_size:
      // result = ctx->abi->load_local_group_size(ctx->abi);
      break;
   case nir_intrinsic_load_vertex_id:
      result = LLVMGetParam(ctx->main_function, 1); 
      // result = LLVMBuildAdd(ctx->lc.builder, LLVMGetParam(ctx->main_function, ctx->args.vertex_id.arg_index),
      //                        LLVMGetParam(ctx->main_function, ctx->args.base_vertex.arg_index), "");
      break;
   case nir_intrinsic_load_vertex_id_zero_base: {
      // result = ctx->abi->vertex_id;
      result = LLVMGetParam(ctx->main_function, ctx->args.vertex_id.arg_index); 
      break;
   }
   case nir_intrinsic_load_local_invocation_id: {
      // result = ac_get_arg(&ctx->ac, ctx->args->local_invocation_ids);
      break;
   }
   case nir_intrinsic_load_base_instance:
      // result = ac_get_arg(&ctx->ac, ctx->args->start_instance);
      break;
   case nir_intrinsic_load_draw_id:
      // result = ac_get_arg(&ctx->ac, ctx->args->draw_id);
      break;
   case nir_intrinsic_load_view_index:
      // result = ac_get_arg(&ctx->ac, ctx->args->view_index);
      break;
   case nir_intrinsic_load_invocation_id:
      // if (ctx->stage == MESA_SHADER_TESS_CTRL) {
      //    result = ac_unpack_param(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->tcs_rel_ids), 8, 5);
      // } else {
      //    if (ctx->ac.chip_class >= GFX10) {
      //       result =
      //          LLVMBuildAnd(ctx->ac.builder, ac_get_arg(&ctx->ac, ctx->args->gs_invocation_id),
      //                       LLVMConstInt(ctx->ac.i32, 127, 0), "");
      //    } else {
      //       result = ac_get_arg(&ctx->ac, ctx->args->gs_invocation_id);
      //    }
      // }
      break;
   case nir_intrinsic_load_primitive_id:
      // if (ctx->stage == MESA_SHADER_GEOMETRY) {
      //    result = ac_get_arg(&ctx->ac, ctx->args->gs_prim_id);
      // } else if (ctx->stage == MESA_SHADER_TESS_CTRL) {
      //    result = ac_get_arg(&ctx->ac, ctx->args->tcs_patch_id);
      // } else if (ctx->stage == MESA_SHADER_TESS_EVAL) {
      //    result = ac_get_arg(&ctx->ac, ctx->args->tes_patch_id);
      // } else
      //    fprintf(stderr, "Unknown primitive id intrinsic: %d", ctx->stage);
      // break;
   // case nir_intrinsic_load_sample_id:
      // result = ac_unpack_param(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->ancillary), 8, 4);
      break;
   case nir_intrinsic_load_sample_pos:
      // result = load_sample_pos(ctx);
      break;
   case nir_intrinsic_load_sample_mask_in:
      // result = ctx->abi->load_sample_mask_in(ctx->abi);
      break;
   case nir_intrinsic_load_frag_coord: {
      // LLVMValueRef values[4] = {
      //    ac_get_arg(&ctx->ac, ctx->args->frag_pos[0]), ac_get_arg(&ctx->ac, ctx->args->frag_pos[1]),
      //    ac_get_arg(&ctx->ac, ctx->args->frag_pos[2]),
      //    ac_build_fdiv(&ctx->ac, ctx->ac.f32_1, ac_get_arg(&ctx->ac, ctx->args->frag_pos[3]))};
      // result = ac_to_integer(&ctx->ac, ac_build_gather_values(&ctx->ac, values, 4));
      break;
   }
   case nir_intrinsic_load_layer_id:
      // result = ctx->abi->inputs[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER, 0)];
      break;
   case nir_intrinsic_load_front_face:
      // result = ac_get_arg(&ctx->ac, ctx->args->front_face);
      break;
   case nir_intrinsic_load_helper_invocation:
      // result = ac_build_load_helper_invocation(&ctx->ac);
      break;
   case nir_intrinsic_is_helper_invocation:
      // result = ac_build_is_helper_invocation(&ctx->ac);
      break;
   case nir_intrinsic_load_color0:
      // result = ctx->abi->color0;
      break;
   case nir_intrinsic_load_color1:
      // result = ctx->abi->color1;
      break;
   case nir_intrinsic_load_user_data_amd:
      // assert(LLVMTypeOf(ctx->abi->user_data) == ctx->ac.v4i32);
      // result = ctx->abi->user_data;
      break;
   case nir_intrinsic_load_instance_id:
      // result = ctx->abi->instance_id;
      break;
   case nir_intrinsic_load_num_work_groups:
      // result = ac_get_arg(&ctx->ac, ctx->args->num_work_groups);
      break;
   case nir_intrinsic_load_local_invocation_index:
      // result = visit_load_local_invocation_index(ctx);
      break;
   case nir_intrinsic_load_subgroup_id:
      // result = visit_load_subgroup_id(ctx);
      break;
   case nir_intrinsic_load_num_subgroups:
      // result = visit_load_num_subgroups(ctx);
      break;
   case nir_intrinsic_first_invocation:
      // result = visit_first_invocation(ctx);
      break;
   case nir_intrinsic_load_push_constant:
      // result = visit_load_push_constant(ctx, instr);
      break;
   case nir_intrinsic_vulkan_resource_index: {
      // LLVMValueRef index = get_src(ctx, instr->src[0]);
      // unsigned desc_set = nir_intrinsic_desc_set(instr);
      // unsigned binding = nir_intrinsic_binding(instr);

      // result = ctx->abi->load_resource(ctx->abi, index, desc_set, binding);
      break;
   }
   case nir_intrinsic_vulkan_resource_reindex:
      // result = visit_vulkan_resource_reindex(ctx, instr);
      break;
   case nir_intrinsic_store_ssbo:
      // visit_store_ssbo(ctx, instr);
      break;
   case nir_intrinsic_load_ssbo:
      // result = visit_load_buffer(ctx, instr);
      break;
   case nir_intrinsic_ssbo_atomic_add:
   case nir_intrinsic_ssbo_atomic_imin:
   case nir_intrinsic_ssbo_atomic_umin:
   case nir_intrinsic_ssbo_atomic_imax:
   case nir_intrinsic_ssbo_atomic_umax:
   case nir_intrinsic_ssbo_atomic_and:
   case nir_intrinsic_ssbo_atomic_or:
   case nir_intrinsic_ssbo_atomic_xor:
   case nir_intrinsic_ssbo_atomic_exchange:
   case nir_intrinsic_ssbo_atomic_comp_swap:
      // result = visit_atomic_ssbo(ctx, instr);
      break;
   case nir_intrinsic_load_ubo:
      // result = visit_load_ubo_buffer(ctx, instr);
      break;
   case nir_intrinsic_get_buffer_size:
      // result = visit_get_buffer_size(ctx, instr);
      break;
   case nir_intrinsic_load_deref:
      result = visit_load_var(ctx, instr);
      break;
   case nir_intrinsic_store_deref:
      visit_store_var(ctx, instr);
      break;
   case nir_intrinsic_load_input:
   case nir_intrinsic_load_input_vertex:
   case nir_intrinsic_load_per_vertex_input:
      // result = visit_load(ctx, instr, false);
      break;
   case nir_intrinsic_load_output:
   case nir_intrinsic_load_per_vertex_output:
      // result = visit_load(ctx, instr, true);
      break;
   case nir_intrinsic_store_output:
   case nir_intrinsic_store_per_vertex_output:
      visit_store_output(ctx, instr);
      break;
   case nir_intrinsic_load_shared:
      result = visit_load_shared(ctx, instr);
      break;
   case nir_intrinsic_store_shared:
      visit_store_shared(ctx, instr);
      break;
   case nir_intrinsic_bindless_image_samples:
   case nir_intrinsic_image_deref_samples:
      // result = visit_image_samples(ctx, instr);
      break;
   case nir_intrinsic_bindless_image_load:
      // result = visit_image_load(ctx, instr, true);
      break;
   case nir_intrinsic_image_deref_load:
      // result = visit_image_load(ctx, instr, false);
      break;
   case nir_intrinsic_bindless_image_store:
      // visit_image_store(ctx, instr, true);
      break;
   case nir_intrinsic_image_deref_store:
      // visit_image_store(ctx, instr, false);
      break;
   case nir_intrinsic_bindless_image_atomic_add:
   case nir_intrinsic_bindless_image_atomic_imin:
   case nir_intrinsic_bindless_image_atomic_umin:
   case nir_intrinsic_bindless_image_atomic_imax:
   case nir_intrinsic_bindless_image_atomic_umax:
   case nir_intrinsic_bindless_image_atomic_and:
   case nir_intrinsic_bindless_image_atomic_or:
   case nir_intrinsic_bindless_image_atomic_xor:
   case nir_intrinsic_bindless_image_atomic_exchange:
   case nir_intrinsic_bindless_image_atomic_comp_swap:
   case nir_intrinsic_bindless_image_atomic_inc_wrap:
   case nir_intrinsic_bindless_image_atomic_dec_wrap:
      // result = visit_image_atomic(ctx, instr, true);
      break;
   case nir_intrinsic_image_deref_atomic_add:
   case nir_intrinsic_image_deref_atomic_imin:
   case nir_intrinsic_image_deref_atomic_umin:
   case nir_intrinsic_image_deref_atomic_imax:
   case nir_intrinsic_image_deref_atomic_umax:
   case nir_intrinsic_image_deref_atomic_and:
   case nir_intrinsic_image_deref_atomic_or:
   case nir_intrinsic_image_deref_atomic_xor:
   case nir_intrinsic_image_deref_atomic_exchange:
   case nir_intrinsic_image_deref_atomic_comp_swap:
   case nir_intrinsic_image_deref_atomic_inc_wrap:
   case nir_intrinsic_image_deref_atomic_dec_wrap:
      // result = visit_image_atomic(ctx, instr, false);
      break;
   case nir_intrinsic_bindless_image_size:
      // result = visit_image_size(ctx, instr, true);
      break;
   case nir_intrinsic_image_deref_size:
      // result = visit_image_size(ctx, instr, false);
      break;
   case nir_intrinsic_shader_clock:
      // result = ac_build_shader_clock(&ctx->ac, nir_intrinsic_memory_scope(instr));
      break;
   case nir_intrinsic_discard:
   case nir_intrinsic_discard_if:
      // emit_discard(ctx, instr);
      break;
   case nir_intrinsic_demote:
   case nir_intrinsic_demote_if:
      // emit_demote(ctx, instr);
      break;
   case nir_intrinsic_memory_barrier:
   case nir_intrinsic_group_memory_barrier:
   case nir_intrinsic_memory_barrier_buffer:
   case nir_intrinsic_memory_barrier_image:
   case nir_intrinsic_memory_barrier_shared:
      // emit_membar(&ctx->ac, instr);
      break;
   case nir_intrinsic_scoped_barrier: {
      // assert(!(nir_intrinsic_memory_semantics(instr) &
      //          (NIR_MEMORY_MAKE_AVAILABLE | NIR_MEMORY_MAKE_VISIBLE)));

      // nir_variable_mode modes = nir_intrinsic_memory_modes(instr);

      // unsigned wait_flags = 0;
      // if (modes & (nir_var_mem_global | nir_var_mem_ssbo))
      //    wait_flags |= AC_WAIT_VLOAD | AC_WAIT_VSTORE;
      // if (modes & nir_var_mem_shared)
      //    wait_flags |= AC_WAIT_LGKM;

      // if (wait_flags)
      //    ac_build_waitcnt(&ctx->ac, wait_flags);

      // if (nir_intrinsic_execution_scope(instr) == NIR_SCOPE_WORKGROUP)
      //    ac_emit_barrier(&ctx->ac, ctx->stage);
      break;
   }
   case nir_intrinsic_memory_barrier_tcs_patch:
      break;
   case nir_intrinsic_control_barrier:
      // ac_emit_barrier(&ctx->ac, ctx->stage);
      break;
   case nir_intrinsic_shared_atomic_add:
   case nir_intrinsic_shared_atomic_imin:
   case nir_intrinsic_shared_atomic_umin:
   case nir_intrinsic_shared_atomic_imax:
   case nir_intrinsic_shared_atomic_umax:
   case nir_intrinsic_shared_atomic_and:
   case nir_intrinsic_shared_atomic_or:
   case nir_intrinsic_shared_atomic_xor:
   case nir_intrinsic_shared_atomic_exchange:
   case nir_intrinsic_shared_atomic_comp_swap:
   case nir_intrinsic_shared_atomic_fadd: {
      // LLVMValueRef ptr = get_memory_ptr(ctx, instr->src[0], instr->src[1].ssa->bit_size);
      // result = visit_var_atomic(ctx, instr, ptr, 1);
      break;
   }
   case nir_intrinsic_deref_atomic_add:
   case nir_intrinsic_deref_atomic_imin:
   case nir_intrinsic_deref_atomic_umin:
   case nir_intrinsic_deref_atomic_imax:
   case nir_intrinsic_deref_atomic_umax:
   case nir_intrinsic_deref_atomic_and:
   case nir_intrinsic_deref_atomic_or:
   case nir_intrinsic_deref_atomic_xor:
   case nir_intrinsic_deref_atomic_exchange:
   case nir_intrinsic_deref_atomic_comp_swap:
   case nir_intrinsic_deref_atomic_fadd: {
      // LLVMValueRef ptr = get_src(ctx, instr->src[0]);
      // result = visit_var_atomic(ctx, instr, ptr, 1);
      break;
   }
   case nir_intrinsic_load_barycentric_pixel:
      // result = barycentric_center(ctx, nir_intrinsic_interp_mode(instr));
      break;
   case nir_intrinsic_load_barycentric_centroid:
      // result = barycentric_centroid(ctx, nir_intrinsic_interp_mode(instr));
      break;
   case nir_intrinsic_load_barycentric_sample:
      // result = barycentric_sample(ctx, nir_intrinsic_interp_mode(instr));
      break;
   case nir_intrinsic_load_barycentric_model:
      // result = barycentric_model(ctx);
      break;
   case nir_intrinsic_load_barycentric_at_offset: {
      // LLVMValueRef offset = ac_to_float(&ctx->ac, get_src(ctx, instr->src[0]));
      // result = barycentric_offset(ctx, nir_intrinsic_interp_mode(instr), offset);
      break;
   }
   case nir_intrinsic_load_barycentric_at_sample: {
      // LLVMValueRef sample_id = get_src(ctx, instr->src[0]);
      // result = barycentric_at_sample(ctx, nir_intrinsic_interp_mode(instr), sample_id);
      break;
   }
   case nir_intrinsic_load_interpolated_input: {
      /* We assume any indirect loads have been lowered away */
      // ASSERTED nir_const_value *offset = nir_src_as_const_value(instr->src[1]);
      // assert(offset);
      // assert(offset[0].i32 == 0);

      // LLVMValueRef interp_param = get_src(ctx, instr->src[0]);
      // unsigned index = nir_intrinsic_base(instr);
      // unsigned component = nir_intrinsic_component(instr);
      // result = load_interpolated_input(ctx, interp_param, index, component,
      //                                  instr->dest.ssa.num_components, instr->dest.ssa.bit_size);
      break;
   }
   case nir_intrinsic_emit_vertex:
      // ctx->abi->emit_vertex(ctx->abi, nir_intrinsic_stream_id(instr), ctx->abi->outputs);
      break;
   case nir_intrinsic_emit_vertex_with_counter: {
      // unsigned stream = nir_intrinsic_stream_id(instr);
      // LLVMValueRef next_vertex = get_src(ctx, instr->src[0]);
      // ctx->abi->emit_vertex_with_counter(ctx->abi, stream, next_vertex, ctx->abi->outputs);
      break;
   }
   case nir_intrinsic_end_primitive:
   case nir_intrinsic_end_primitive_with_counter:
      // ctx->abi->emit_primitive(ctx->abi, nir_intrinsic_stream_id(instr));
      break;
   case nir_intrinsic_load_tess_coord:
      // result = ctx->abi->load_tess_coord(ctx->abi);
      break;
   case nir_intrinsic_load_tess_level_outer:
      // result = ctx->abi->load_tess_level(ctx->abi, VARYING_SLOT_TESS_LEVEL_OUTER, false);
      break;
   case nir_intrinsic_load_tess_level_inner:
      // result = ctx->abi->load_tess_level(ctx->abi, VARYING_SLOT_TESS_LEVEL_INNER, false);
      break;
   case nir_intrinsic_load_tess_level_outer_default:
      // result = ctx->abi->load_tess_level(ctx->abi, VARYING_SLOT_TESS_LEVEL_OUTER, true);
      break;
   case nir_intrinsic_load_tess_level_inner_default:
      // result = ctx->abi->load_tess_level(ctx->abi, VARYING_SLOT_TESS_LEVEL_INNER, true);
      break;
   case nir_intrinsic_load_patch_vertices_in:
      // result = ctx->abi->load_patch_vertices_in(ctx->abi);
      break;
   case nir_intrinsic_vote_all: {
      // LLVMValueRef tmp = ac_build_vote_all(&ctx->ac, get_src(ctx, instr->src[0]));
      // result = LLVMBuildSExt(ctx->ac.builder, tmp, ctx->ac.i32, "");
      break;
   }
   case nir_intrinsic_vote_any: {
      // LLVMValueRef tmp = ac_build_vote_any(&ctx->ac, get_src(ctx, instr->src[0]));
      // result = LLVMBuildSExt(ctx->ac.builder, tmp, ctx->ac.i32, "");
      break;
   }
   case nir_intrinsic_shuffle:
      // if (ctx->ac.chip_class == GFX8 || ctx->ac.chip_class == GFX9 ||
      //     (ctx->ac.chip_class >= GFX10 && ctx->ac.wave_size == 32)) {
      //    result =
      //       ac_build_shuffle(&ctx->ac, get_src(ctx, instr->src[0]), get_src(ctx, instr->src[1]));
      // } else {
      //    LLVMValueRef src = get_src(ctx, instr->src[0]);
      //    LLVMValueRef index = get_src(ctx, instr->src[1]);
      //    LLVMTypeRef type = LLVMTypeOf(src);
      //    struct waterfall_context wctx;
      //    LLVMValueRef index_val;

      //    index_val = enter_waterfall(ctx, &wctx, index, true);

      //    src = LLVMBuildZExt(ctx->ac.builder, src, ctx->ac.i32, "");

      //    result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.readlane", ctx->ac.i32,
      //                                (LLVMValueRef[]){src, index_val}, 2,
      //                                AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);

      //    result = LLVMBuildTrunc(ctx->ac.builder, result, type, "");

      //    result = exit_waterfall(ctx, &wctx, result);
      // }
      break;
   case nir_intrinsic_reduce:
      // result = ac_build_reduce(&ctx->ac, get_src(ctx, instr->src[0]), instr->const_index[0],
      //                          instr->const_index[1]);
      break;
   case nir_intrinsic_inclusive_scan:
      // result =
      //    ac_build_inclusive_scan(&ctx->ac, get_src(ctx, instr->src[0]), instr->const_index[0]);
      break;
   case nir_intrinsic_exclusive_scan:
      // result =
      //    ac_build_exclusive_scan(&ctx->ac, get_src(ctx, instr->src[0]), instr->const_index[0]);
      break;
   case nir_intrinsic_quad_broadcast: {
      // unsigned lane = nir_src_as_uint(instr->src[1]);
      // result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), lane, lane, lane, lane);
      break;
   }
   case nir_intrinsic_quad_swap_horizontal:
      // result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), 1, 0, 3, 2);
      break;
   case nir_intrinsic_quad_swap_vertical:
      // result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), 2, 3, 0, 1);
      break;
   case nir_intrinsic_quad_swap_diagonal:
      // result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), 3, 2, 1, 0);
      break;
   case nir_intrinsic_quad_swizzle_amd: {
      // uint32_t mask = nir_intrinsic_swizzle_mask(instr);
      // result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), mask & 0x3,
      //                                (mask >> 2) & 0x3, (mask >> 4) & 0x3, (mask >> 6) & 0x3);
      break;
   }
   case nir_intrinsic_masked_swizzle_amd: {
      // uint32_t mask = nir_intrinsic_swizzle_mask(instr);
      // result = ac_build_ds_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), mask);
      break;
   }
   case nir_intrinsic_write_invocation_amd:
      // result = ac_build_writelane(&ctx->ac, get_src(ctx, instr->src[0]),
      //                             get_src(ctx, instr->src[1]), get_src(ctx, instr->src[2]));
      break;
   case nir_intrinsic_mbcnt_amd:
      // result = ac_build_mbcnt(&ctx->ac, get_src(ctx, instr->src[0]));
      break;
   case nir_intrinsic_load_scratch: {
      LLVMValueRef offset = get_src(ctx, instr->src[0]);
      LLVMValueRef ptr = build_gep0(&ctx->lc, ctx->scratch, offset);
      LLVMTypeRef comp_type = LLVMIntTypeInContext(ctx->lc.context, instr->dest.ssa.bit_size);
      LLVMTypeRef vec_type = instr->dest.ssa.num_components == 1
                                ? comp_type
                                : LLVMVectorType(comp_type, instr->dest.ssa.num_components);
      unsigned addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
      ptr = LLVMBuildBitCast(ctx->lc.builder, ptr, LLVMPointerType(vec_type, addr_space), "");
      result = LLVMBuildLoad(ctx->lc.builder, ptr, "");
      break;
   }
   case nir_intrinsic_store_scratch: {
      LLVMValueRef offset = get_src(ctx, instr->src[1]);
      LLVMValueRef ptr = build_gep0(&ctx->lc, ctx->scratch, offset);
      LLVMTypeRef comp_type = LLVMIntTypeInContext(ctx->lc.context, instr->src[0].ssa->bit_size);
      unsigned addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
      ptr = LLVMBuildBitCast(ctx->lc.builder, ptr, LLVMPointerType(comp_type, addr_space), "");
      LLVMValueRef src = get_src(ctx, instr->src[0]);
      unsigned wrmask = nir_intrinsic_write_mask(instr);
      while (wrmask) {
         int start, count;
         u_bit_scan_consecutive_range(&wrmask, &start, &count);

         LLVMValueRef offset = LLVMConstInt(ctx->lc.i32, start, false);
         LLVMValueRef offset_ptr = LLVMBuildGEP(ctx->lc.builder, ptr, &offset, 1, "");
         LLVMTypeRef vec_type = count == 1 ? comp_type : LLVMVectorType(comp_type, count);
         offset_ptr = LLVMBuildBitCast(ctx->lc.builder, offset_ptr,
                                       LLVMPointerType(vec_type, addr_space), "");
         LLVMValueRef offset_src = extract_components(&ctx->lc, src, start, count);
         LLVMBuildStore(ctx->lc.builder, offset_src, offset_ptr);
      }
      break;
   }
   case nir_intrinsic_load_constant: {
      unsigned base = nir_intrinsic_base(instr);
      unsigned range = nir_intrinsic_range(instr);

      LLVMValueRef offset = get_src(ctx, instr->src[0]);
      offset = LLVMBuildAdd(ctx->lc.builder, offset, LLVMConstInt(ctx->lc.i32, base, false), "");

      /* Clamp the offset to avoid out-of-bound access because global
       * instructions can't handle them.
       */
      LLVMValueRef size = LLVMConstInt(ctx->lc.i32, base + range, false);
      LLVMValueRef cond = LLVMBuildICmp(ctx->lc.builder, LLVMIntULT, offset, size, "");
      offset = LLVMBuildSelect(ctx->lc.builder, cond, offset, size, "");

      LLVMValueRef ptr = build_gep0(&ctx->lc, ctx->constant_data, offset);
      LLVMTypeRef comp_type = LLVMIntTypeInContext(ctx->lc.context, instr->dest.ssa.bit_size);
      LLVMTypeRef vec_type = instr->dest.ssa.num_components == 1
                                ? comp_type
                                : LLVMVectorType(comp_type, instr->dest.ssa.num_components);
      unsigned addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
      ptr = LLVMBuildBitCast(ctx->lc.builder, ptr, LLVMPointerType(vec_type, addr_space), "");
      result = LLVMBuildLoad(ctx->lc.builder, ptr, "");
      break;
   }
   default:
      fprintf(stderr, "Unknown intrinsic: ");
      nir_print_instr(&instr->instr, stderr);
      fprintf(stderr, "\n");
      break;
   }
   if (result) {
      ctx->ssa_defs[instr->dest.ssa.index] = result;
   }
}

static void visit_cf_list(struct libresoc_nir_tran_ctx *ctx, struct exec_list *list);

static void visit_block(struct libresoc_nir_tran_ctx *ctx, nir_block *block)
{
    nir_foreach_instr (instr, block) {
        switch (instr->type) {
            case nir_instr_type_alu:
                visit_alu(ctx, nir_instr_as_alu(instr));
                break;
            case nir_instr_type_load_const:
                visit_load_const(ctx, nir_instr_as_load_const(instr));
                break;
            case nir_instr_type_intrinsic:
                visit_intrinsic(ctx, nir_instr_as_intrinsic(instr));
                break;
            case nir_instr_type_tex:
                // visit_tex(ctx, nir_instr_as_tex(instr));
                break;
            case nir_instr_type_phi:
                visit_phi(ctx, nir_instr_as_phi(instr));
                break;
            case nir_instr_type_ssa_undef:
                visit_ssa_undef(ctx, nir_instr_as_ssa_undef(instr));
                break;
            case nir_instr_type_jump:
                visit_jump(&ctx->lc, nir_instr_as_jump(instr));
                break;
            case nir_instr_type_deref:
                visit_deref(ctx, nir_instr_as_deref(instr));
                break;
            default:
                fprintf(stderr, "Unknown NIR instr type: ");
                nir_print_instr(instr, stderr);
                fprintf(stderr, "\n");
                abort();
        }
    }
}

static void visit_if(struct libresoc_nir_tran_ctx *ctx, nir_if *if_stmt)
{
   LLVMValueRef value = get_src(ctx, if_stmt->condition);

   nir_block *then_block = (nir_block *)exec_list_get_head(&if_stmt->then_list);

   build_uif(&ctx->lc, value, then_block->index);

   visit_cf_list(ctx, &if_stmt->then_list);

   if (!exec_list_is_empty(&if_stmt->else_list)) {
      nir_block *else_block = (nir_block *)exec_list_get_head(&if_stmt->else_list);

      build_else(&ctx->lc, else_block->index);
      visit_cf_list(ctx, &if_stmt->else_list);
   }

   build_endif(&ctx->lc, then_block->index);
}

static void visit_loop(struct libresoc_nir_tran_ctx *ctx, nir_loop *loop)
{
   nir_block *first_loop_block = (nir_block *)exec_list_get_head(&loop->body);

   build_bgnloop(&ctx->lc, first_loop_block->index);

   visit_cf_list(ctx, &loop->body);

   build_endloop(&ctx->lc, first_loop_block->index);
}

static void visit_cf_list(struct libresoc_nir_tran_ctx *ctx, struct exec_list *list)
{
   foreach_list_typed(nir_cf_node, node, node, list)
   {
      switch (node->type) {
      case nir_cf_node_block:
         visit_block(ctx, nir_cf_node_as_block(node));
         break;

      case nir_cf_node_if:
         visit_if(ctx, nir_cf_node_as_if(node));
         break;

      case nir_cf_node_loop:
         visit_loop(ctx, nir_cf_node_as_loop(node));
         break;

      default:
         assert(0);
      }
   }
}

LLVMModuleRef libresoc_nir_translate(struct libresoc_llvm *llvm_ref, struct nir_shader *nir)
{
    struct libresoc_nir_tran_ctx ctx = {};
    struct nir_function *func;
    char shader_name[60];
    sprintf(shader_name, "libresoc-shader-%s", gl_shader_stage_name(nir->info.stage)); 
    LLVMModuleRef mod = LLVMModuleCreateWithNameInContext(shader_name, llvm_ref->lc.context);
    ctx.lc.module = &mod;
    ctx.lc = llvm_ref->lc;
    ctx.stage = nir->info.stage;
    ctx.info = &nir->info;
    
    if (ctx.stage == MESA_SHADER_VERTEX) { 
        add_arg(&ctx.args, ARG_SGPR, 1, ARG_INT, &ctx.args.base_vertex);
        add_arg(&ctx.args, ARG_SGPR, 1, ARG_INT, &ctx.args.start_instance);
        add_arg(&ctx.args, ARG_VGPR, 1, ARG_INT, &ctx.args.vertex_id);
    }
    LLVMTypeRef arg_types[32];
   LLVMTypeRef ret_type = LLVMVoidTypeInContext(ctx.lc.context);
   for (unsigned i = 0; i < ctx.args.arg_count; i++) {
      arg_types[i] = arg_llvm_type(ctx.args.args[i].type, ctx.args.args[i].size, &ctx.lc);
   }

    LLVMTypeRef main_function_type;
    if(nir->info.stage == MESA_SHADER_VERTEX) {
        LLVMTypeRef arg_types[32];
        LLVMTypeRef floatType = LLVMFloatTypeInContext(llvm_ref->lc.context);
        LLVMTypeRef float4 = LLVMVectorType(floatType, 4);
        LLVMTypeRef f4arrayType = LLVMArrayType(float4, 10);
        arg_types[0] = float4;
        arg_types[1] = f4arrayType;
        LLVMTypeRef vertexCacheType = LLVMStructTypeInContext(llvm_ref->lc.context, arg_types, 2, false);
        LLVMTypeRef gpuStatePtrType = LLVMPointerType(LLVMStructCreateNamed(llvm_ref->lc.context, "GPUState"), 0);
        arg_types[0] = gpuStatePtrType;
        arg_types[1] = LLVMIntTypeInContext(llvm_ref->lc.context, 32);
        arg_types[2] = LLVMPointerType(vertexCacheType, 0);
        main_function_type = LLVMFunctionType(ret_type, arg_types, 3, 0);
    } else {

        LLVMTypeRef floatType = LLVMFloatTypeInContext(llvm_ref->lc.context);
        LLVMTypeRef float4 = LLVMVectorType(floatType, 4);
        LLVMTypeRef f4arrayType = LLVMArrayType(float4, 10);
        arg_types[0] = float4;
        arg_types[1] = f4arrayType;
        LLVMTypeRef vertexCacheType = LLVMStructTypeInContext(llvm_ref->lc.context, arg_types, 2, false);
        arg_types[3] = LLVMPointerType(vertexCacheType, 0);
        LLVMTypeRef gpuStatePtrType = LLVMPointerType(LLVMStructCreateNamed(llvm_ref->lc.context, "GPUState"), 0);
        arg_types[0] = gpuStatePtrType;
        arg_types[1] = LLVMFloatTypeInContext(llvm_ref->lc.context);
        arg_types[2] = LLVMPointerType(float4, 0);
        arg_types[4] = LLVMPointerType(float4, 0);
        main_function_type = LLVMFunctionType(ret_type, arg_types, 3, 0);
        //TODO: this is zero argument function and returns void
        main_function_type = LLVMFunctionType(ret_type, arg_types, 5, 0);
    }
   LLVMValueRef main_function = LLVMAddFunction(mod, gl_shader_stage_name(nir->info.stage), main_function_type);
   LLVMBasicBlockRef main_function_body =
      LLVMAppendBasicBlockInContext(ctx.lc.context, main_function, "main_body");
   LLVMPositionBuilderAtEnd(ctx.lc.builder, main_function_body);
    ctx.main_function = main_function;

   if (!nir->info.io_lowered) {
      nir_foreach_shader_out_variable(variable, nir)
      {
         handle_shader_output_decl(&ctx, nir, variable, ctx.stage);
      }
   }
    ctx.defs = _mesa_hash_table_create(NULL, _mesa_hash_pointer, _mesa_key_pointer_equal);
    ctx.phis = _mesa_hash_table_create(NULL, _mesa_hash_pointer, _mesa_key_pointer_equal);
    ctx.vars = _mesa_hash_table_create(NULL, _mesa_hash_pointer, _mesa_key_pointer_equal);
    func = (struct nir_function *)exec_list_get_head(&nir->functions);

    nir_index_ssa_defs(func->impl);
    ctx.ssa_defs = calloc(func->impl->ssa_alloc, sizeof(LLVMValueRef));
    setup_locals(&ctx, func);
    setup_scratch(&ctx, nir);
    setup_constant_data(&ctx, nir);

    // if (gl_shader_stage_is_compute(nir->info.stage))
    //     setup_shared(&ctx, nir);
    visit_cf_list(&ctx, &func->impl->body);
    LLVMBuildRetVoid(ctx.lc.builder);
    char *error = NULL;
    LLVMVerifyModule(mod, LLVMPrintMessageAction, &error);
    LLVMDumpModule(mod);
    LLVMDisposeMessage(error);
    LLVMOrcModuleHandle mod_handle;
    LLVMErrorRef error_ref =  LLVMOrcAddEagerlyCompiledIR(llvm_ref->orc_ref,
            &mod_handle,
            mod,
            orc_sym_resolver,
            (void *)(llvm_ref->orc_ref));
    LLVMDumpModule(mod);
    char *def_triple = LLVMGetDefaultTargetTriple();   // E.g. "x86_64-linux-gnu"
    LLVMDisasmContextRef disasm = LLVMCreateDisasm(def_triple, NULL,
                                      0, NULL,
                                     NULL);
    if (disasm) {
        LLVMOrcTargetAddress MainAddr;
        LLVMOrcGetSymbolAddress(llvm_ref->orc_ref, &MainAddr ,gl_shader_stage_name(nir->info.stage));
        // if(nir->info.stage == MESA_SHADER_VERTEX)
        // {
        //     pipeline->vs = (VertexShader)MainAddr;
        // }
        // else if(nir->info.stage == MESA_SHADER_FRAGMENT)
        // {
        //     pipeline->fs = (FragmentShader)MainAddr;
        // }
        const uint8_t *bytes = (const uint8_t *)MainAddr; 
        char outline[1024];
        uint64_t pc;
        pc = 0;
        uint64_t extent = 200;
        while (pc < extent) {
            size_t Size;

            /*
             * Print address.  We use addresses relative to the start of the function,
             * so that between runs.
             */


            Size = LLVMDisasmInstruction(disasm, (uint8_t *)bytes + pc, extent - pc, 0, outline,
                    sizeof outline);

            /*
             * Print the instruction.
             */
            printf("\t%s \n", outline);


            /*
             * Stop disassembling on return statements, if there is no record of a
             * jump to a successive address.
             *
             * XXX: This currently assumes x86
             */

            if (Size == 1 && bytes[pc] == 0xc3) {
                break;
            }

            /*
             * Advance.
             */

            pc += Size;

            if (pc >= extent) {
                break;
            }
        }    
    }
    return mod;
    // LLVMModuleRef mod = LLVMModuleCreateWithName("libresoc_mod");
    // LLVMTypeRef param_types[] = { LLVMInt32Type(), LLVMInt32Type() };
    // LLVMTypeRef ret_type = LLVMFunctionType(LLVMInt32Type(), param_types, 2, 0);
    // LLVMValueRef sum = LLVMAddFunction(mod, "sum", ret_type);
    // LLVMBasicBlockRef entry = LLVMAppendBasicBlock(sum, "entry");
    // LLVMBuilderRef builder = LLVMCreateBuilder();
    // LLVMPositionBuilderAtEnd(builder, entry);
    // LLVMValueRef tmp = LLVMBuildAdd(builder, LLVMGetParam(sum, 0), LLVMGetParam(sum, 1), "tmp");
    // LLVMBuildRet(builder, tmp);
    // char *error = NULL;
    // LLVMVerifyModule(mod, LLVMAbortProcessAction, &error);
    // LLVMDumpModule(mod);
    // LLVMDisposeMessage(error);
    // LLVMOrcModuleHandle mod_handle;
    // LLVMErrorRef error_ref =  LLVMOrcAddEagerlyCompiledIR(llvm_ref->orc_ref,
    //         &mod_handle,
    //         mod,
    //         orc_sym_resolver,
    //         (void *)(llvm_ref->orc_ref));
}

Shader GetFuncPointer(struct libresoc_llvm *llvm_ref, const char *name) {
    LLVMOrcTargetAddress MainAddr;
    LLVMOrcGetSymbolAddress(llvm_ref->orc_ref, &MainAddr , name);
    return (Shader)MainAddr;
}