/*
* Copyright 2016 Advanced Micro Devices, Inc.
+ * All Rights Reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
#include "si_shader_internal.h"
#include "si_pipe.h"
-
-#include "gallivm/lp_bld_const.h"
-#include "gallivm/lp_bld_gather.h"
-#include "gallivm/lp_bld_flow.h"
-#include "gallivm/lp_bld_init.h"
-#include "gallivm/lp_bld_intr.h"
-#include "gallivm/lp_bld_misc.h"
-#include "gallivm/lp_bld_swizzle.h"
-#include "tgsi/tgsi_info.h"
-#include "tgsi/tgsi_parse.h"
-#include "util/u_math.h"
+#include "ac_llvm_util.h"
#include "util/u_memory.h"
-#include "util/u_debug.h"
-
-#include <stdio.h>
-#include <llvm-c/Transforms/IPO.h>
-#include <llvm-c/Transforms/Scalar.h>
enum si_llvm_calling_convention {
RADEON_LLVM_AMDGPU_VS = 87,
*
* @returns 0 for success, 1 for failure
*/
-unsigned si_llvm_compile(LLVMModuleRef M, struct ac_shader_binary *binary,
- LLVMTargetMachineRef tm,
- struct pipe_debug_callback *debug)
+unsigned si_llvm_compile(LLVMModuleRef M, struct si_shader_binary *binary,
+ struct ac_llvm_compiler *compiler,
+ struct pipe_debug_callback *debug,
+ bool less_optimized)
{
+ struct ac_compiler_passes *passes =
+ less_optimized && compiler->low_opt_passes ?
+ compiler->low_opt_passes : compiler->passes;
struct si_llvm_diagnostics diag;
- char *err;
LLVMContextRef llvm_ctx;
- LLVMMemoryBufferRef out_buffer;
- unsigned buffer_size;
- const char *buffer_data;
- LLVMBool mem_err;
diag.debug = debug;
diag.retval = 0;
LLVMContextSetDiagnosticHandler(llvm_ctx, si_diagnostic_handler, &diag);
- /* Compile IR*/
- mem_err = LLVMTargetMachineEmitToMemoryBuffer(tm, M, LLVMObjectFile, &err,
- &out_buffer);
-
- /* Process Errors/Warnings */
- if (mem_err) {
- fprintf(stderr, "%s: %s", __FUNCTION__, err);
- pipe_debug_message(debug, SHADER_INFO,
- "LLVM emit error: %s", err);
- FREE(err);
+ /* Compile IR. */
+ if (!ac_compile_module_to_elf(passes, M, (char **)&binary->elf_buffer,
+ &binary->elf_size))
diag.retval = 1;
- goto out;
- }
- /* Extract Shader Code*/
- buffer_size = LLVMGetBufferSize(out_buffer);
- buffer_data = LLVMGetBufferStart(out_buffer);
-
- if (!ac_elf_read(buffer_data, buffer_size, binary)) {
- fprintf(stderr, "radeonsi: cannot read an ELF shader binary\n");
- diag.retval = 1;
- }
-
- /* Clean up */
- LLVMDisposeMemoryBuffer(out_buffer);
-
-out:
if (diag.retval != 0)
pipe_debug_message(debug, SHADER_INFO, "LLVM compile failed");
return diag.retval;
}
+void si_shader_binary_clean(struct si_shader_binary *binary)
+{
+ free((void *)binary->elf_buffer);
+ binary->elf_buffer = NULL;
+
+ free(binary->llvm_ir_string);
+ binary->llvm_ir_string = NULL;
+}
+
LLVMTypeRef tgsi2llvmtype(struct lp_build_tgsi_context *bld_base,
enum tgsi_opcode_type type)
{
LLVMValueRef c_max = LLVMConstInt(ctx->i32, num - 1, 0);
LLVMValueRef cc;
- if (util_is_power_of_two(num)) {
+ if (util_is_power_of_two_or_zero(num)) {
index = LLVMBuildAnd(builder, index, c_max, "");
} else {
/* In theory, this MAX pattern should result in code that is
{
unsigned array_id;
struct tgsi_array_info *array;
- LLVMBuilderRef builder = ctx->ac.builder;
LLVMValueRef idxs[2];
LLVMValueRef index;
LLVMValueRef alloca;
*/
index = si_llvm_bound_index(ctx, index, array->range.Last - array->range.First + 1);
- index = LLVMBuildMul(
- builder, index,
- LLVMConstInt(ctx->i32, util_bitcount(array->writemask), 0),
- "");
- index = LLVMBuildAdd(
- builder, index,
- LLVMConstInt(ctx->i32,
- util_bitcount(array->writemask & ((1 << swizzle) - 1)), 0),
- "");
+ index = ac_build_imad(&ctx->ac, index,
+ LLVMConstInt(ctx->i32, util_bitcount(array->writemask), 0),
+ LLVMConstInt(ctx->i32,
+ util_bitcount(array->writemask & ((1 << swizzle) - 1)), 0));
idxs[0] = ctx->i32_0;
idxs[1] = index;
return LLVMBuildGEP(ctx->ac.builder, alloca, idxs, 2, "");
LLVMValueRef ptr2)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- LLVMValueRef result;
-
- result = LLVMGetUndef(LLVMVectorType(ctx->i32, 2));
-
- result = LLVMBuildInsertElement(ctx->ac.builder,
- result,
- ac_to_integer(&ctx->ac, ptr),
- ctx->i32_0, "");
- result = LLVMBuildInsertElement(ctx->ac.builder,
- result,
- ac_to_integer(&ctx->ac, ptr2),
- ctx->i32_1, "");
+ LLVMValueRef values[2] = {
+ ac_to_integer(&ctx->ac, ptr),
+ ac_to_integer(&ctx->ac, ptr2),
+ };
+ LLVMValueRef result = ac_build_gather_values(&ctx->ac, values, 2);
return LLVMBuildBitCast(ctx->ac.builder, result, type, "");
}
emit_array_fetch(struct lp_build_tgsi_context *bld_base,
unsigned File, enum tgsi_opcode_type type,
struct tgsi_declaration_range range,
- unsigned swizzle)
+ unsigned swizzle_in)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
unsigned i, size = range.Last - range.First + 1;
LLVMTypeRef vec = LLVMVectorType(tgsi2llvmtype(bld_base, type), size);
LLVMValueRef result = LLVMGetUndef(vec);
-
+ unsigned swizzle = swizzle_in;
struct tgsi_full_src_register tmp_reg = {};
tmp_reg.Register.File = File;
+ if (tgsi_type_is_64bit(type))
+ swizzle |= (swizzle_in + 1) << 16;
for (i = 0; i < size; ++i) {
tmp_reg.Register.Index = i + range.First;
+
LLVMValueRef temp = si_llvm_emit_fetch(bld_base, &tmp_reg, type, swizzle);
result = LLVMBuildInsertElement(ctx->ac.builder, result, temp,
LLVMConstInt(ctx->i32, i, 0), "array_vector");
LLVMValueRef si_llvm_emit_fetch(struct lp_build_tgsi_context *bld_base,
const struct tgsi_full_src_register *reg,
enum tgsi_opcode_type type,
- unsigned swizzle)
+ unsigned swizzle_in)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
LLVMBuilderRef builder = ctx->ac.builder;
LLVMValueRef result = NULL, ptr, ptr2;
+ unsigned swizzle = swizzle_in & 0xffff;
- if (swizzle == ~0) {
+ if (swizzle_in == ~0) {
LLVMValueRef values[TGSI_NUM_CHANNELS];
unsigned chan;
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
values[chan] = si_llvm_emit_fetch(bld_base, reg, type, chan);
}
- return lp_build_gather_values(&ctx->gallivm, values,
+ return ac_build_gather_values(&ctx->ac, values,
TGSI_NUM_CHANNELS);
}
ctx->imms[reg->Register.Index * TGSI_NUM_CHANNELS + swizzle],
ctx->i32_0);
result = LLVMConstInsertElement(result,
- ctx->imms[reg->Register.Index * TGSI_NUM_CHANNELS + swizzle + 1],
+ ctx->imms[reg->Register.Index * TGSI_NUM_CHANNELS + (swizzle_in >> 16)],
ctx->i32_1);
return LLVMConstBitCast(result, ctype);
} else {
if (tgsi_type_is_64bit(type)) {
ptr = result;
- ptr2 = input[swizzle + 1];
+ ptr2 = input[swizzle_in >> 16];
return si_llvm_emit_fetch_64bit(bld_base, tgsi2llvmtype(bld_base, type),
ptr, ptr2);
}
return LLVMGetUndef(tgsi2llvmtype(bld_base, type));
ptr = ctx->temps[reg->Register.Index * TGSI_NUM_CHANNELS + swizzle];
if (tgsi_type_is_64bit(type)) {
- ptr2 = ctx->temps[reg->Register.Index * TGSI_NUM_CHANNELS + swizzle + 1];
+ ptr2 = ctx->temps[reg->Register.Index * TGSI_NUM_CHANNELS + (swizzle_in >> 16)];
return si_llvm_emit_fetch_64bit(bld_base, tgsi2llvmtype(bld_base, type),
LLVMBuildLoad(builder, ptr, ""),
LLVMBuildLoad(builder, ptr2, ""));
case TGSI_FILE_OUTPUT:
ptr = get_output_ptr(bld_base, reg->Register.Index, swizzle);
if (tgsi_type_is_64bit(type)) {
- ptr2 = get_output_ptr(bld_base, reg->Register.Index, swizzle + 1);
+ ptr2 = get_output_ptr(bld_base, reg->Register.Index, (swizzle_in >> 16));
return si_llvm_emit_fetch_64bit(bld_base, tgsi2llvmtype(bld_base, type),
LLVMBuildLoad(builder, ptr, ""),
LLVMBuildLoad(builder, ptr2, ""));
static LLVMValueRef fetch_system_value(struct lp_build_tgsi_context *bld_base,
const struct tgsi_full_src_register *reg,
enum tgsi_opcode_type type,
- unsigned swizzle)
+ unsigned swizzle_in)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
LLVMBuilderRef builder = ctx->ac.builder;
LLVMValueRef cval = ctx->system_values[reg->Register.Index];
+ unsigned swizzle = swizzle_in & 0xffff;
if (tgsi_type_is_64bit(type)) {
LLVMValueRef lo, hi;
lo = LLVMBuildExtractElement(
builder, cval, LLVMConstInt(ctx->i32, swizzle, 0), "");
hi = LLVMBuildExtractElement(
- builder, cval, LLVMConstInt(ctx->i32, swizzle + 1, 0), "");
+ builder, cval, LLVMConstInt(ctx->i32, (swizzle_in >> 16), 0), "");
return si_llvm_emit_fetch_64bit(bld_base, tgsi2llvmtype(bld_base, type),
lo, hi);
for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
unsigned chan;
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
- ctx->addrs[idx][chan] = lp_build_alloca_undef(
- &ctx->gallivm,
- ctx->i32, "");
+ ctx->addrs[idx][chan] = ac_build_alloca_undef(
+ &ctx->ac, ctx->i32, "");
}
}
break;
*/
if (array_size > 16 ||
!ctx->screen->llvm_has_working_vgpr_indexing) {
- array_alloca = lp_build_alloca_undef(&ctx->gallivm,
+ array_alloca = ac_build_alloca_undef(&ctx->ac,
LLVMArrayType(ctx->f32,
array_size), "array");
ctx->temp_array_allocas[id] = array_alloca;
}
if (!array_alloca) {
for (i = 0; i < decl_size; ++i) {
-#ifdef DEBUG
+#ifndef NDEBUG
snprintf(name, sizeof(name), "TEMP%d.%c",
first + i / 4, "xyzw"[i % 4]);
#endif
ctx->temps[first * TGSI_NUM_CHANNELS + i] =
- lp_build_alloca_undef(&ctx->gallivm,
+ ac_build_alloca_undef(&ctx->ac,
ctx->f32,
name);
}
* a shader ever reads from a channel that
* it never writes to.
*/
- ctx->undef_alloca = lp_build_alloca_undef(
- &ctx->gallivm,
- ctx->f32, "undef");
+ ctx->undef_alloca = ac_build_alloca_undef(
+ &ctx->ac, ctx->f32, "undef");
}
for (i = 0; i < decl_size; ++i) {
LLVMValueRef ptr;
if (writemask & (1 << (i % 4))) {
-#ifdef DEBUG
+#ifndef NDEBUG
snprintf(name, sizeof(name), "TEMP%d.%c",
first + i / 4, "xyzw"[i % 4]);
#endif
if (ctx->outputs[idx][0])
continue;
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
-#ifdef DEBUG
+#ifndef NDEBUG
snprintf(name, sizeof(name), "OUT%d.%c",
idx, "xyzw"[chan % 4]);
#endif
- ctx->outputs[idx][chan] = lp_build_alloca_undef(
- &ctx->gallivm,
- ctx->f32, name);
+ ctx->outputs[idx][chan] = ac_build_alloca_undef(
+ &ctx->ac, ctx->f32, name);
}
}
break;
void si_llvm_context_init(struct si_shader_context *ctx,
struct si_screen *sscreen,
- LLVMTargetMachineRef tm)
+ struct ac_llvm_compiler *compiler)
{
struct lp_type type;
*/
memset(ctx, 0, sizeof(*ctx));
ctx->screen = sscreen;
- ctx->tm = tm;
-
- ctx->gallivm.context = LLVMContextCreate();
- ctx->gallivm.module = LLVMModuleCreateWithNameInContext("tgsi",
- ctx->gallivm.context);
- LLVMSetTarget(ctx->gallivm.module, "amdgcn--");
+ ctx->compiler = compiler;
- LLVMTargetDataRef data_layout = LLVMCreateTargetDataLayout(tm);
- char *data_layout_str = LLVMCopyStringRepOfTargetData(data_layout);
- LLVMSetDataLayout(ctx->gallivm.module, data_layout_str);
- LLVMDisposeTargetData(data_layout);
- LLVMDisposeMessage(data_layout_str);
+ ac_llvm_context_init(&ctx->ac, sscreen->info.chip_class, sscreen->info.family, 64);
+ ctx->ac.module = ac_create_module(compiler->tm, ctx->ac.context);
- bool unsafe_fpmath = (sscreen->debug_flags & DBG(UNSAFE_MATH)) != 0;
enum ac_float_mode float_mode =
- unsafe_fpmath ? AC_FLOAT_MODE_UNSAFE_FP_MATH :
- AC_FLOAT_MODE_NO_SIGNED_ZEROS_FP_MATH;
+ sscreen->debug_flags & DBG(UNSAFE_MATH) ?
+ AC_FLOAT_MODE_UNSAFE_FP_MATH :
+ AC_FLOAT_MODE_NO_SIGNED_ZEROS_FP_MATH;
+ ctx->ac.builder = ac_create_builder(ctx->ac.context, float_mode);
- ctx->gallivm.builder = ac_create_builder(ctx->gallivm.context,
- float_mode);
-
- ac_llvm_context_init(&ctx->ac, ctx->gallivm.context,
- sscreen->info.chip_class, sscreen->info.family);
- ctx->ac.module = ctx->gallivm.module;
- ctx->ac.builder = ctx->gallivm.builder;
+ ctx->gallivm.context = ctx->ac.context;
+ ctx->gallivm.module = ctx->ac.module;
+ ctx->gallivm.builder = ctx->ac.builder;
struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
ctx->i32_0 = LLVMConstInt(ctx->i32, 0, 0);
ctx->i32_1 = LLVMConstInt(ctx->i32, 1, 0);
+ ctx->i1false = LLVMConstInt(ctx->i1, 0, 0);
+ ctx->i1true = LLVMConstInt(ctx->i1, 1, 0);
}
/* Set the context to a certain TGSI shader. Can be called repeatedly
LLVMTypeRef main_fn_type, ret_type;
LLVMBasicBlockRef main_fn_body;
enum si_llvm_calling_convention call_conv;
- unsigned real_shader_type;
+ enum pipe_shader_type real_shader_type;
if (num_return_elems)
ret_type = LLVMStructTypeInContext(ctx->ac.context,
if (ctx->screen->info.chip_class >= GFX9) {
if (ctx->shader->key.as_ls)
real_shader_type = PIPE_SHADER_TESS_CTRL;
- else if (ctx->shader->key.as_es)
+ else if (ctx->shader->key.as_es || ctx->shader->key.as_ngg)
real_shader_type = PIPE_SHADER_GEOMETRY;
}
call_conv = RADEON_LLVM_AMDGPU_VS;
break;
case PIPE_SHADER_TESS_CTRL:
- call_conv = HAVE_LLVM >= 0x0500 ? RADEON_LLVM_AMDGPU_HS :
- RADEON_LLVM_AMDGPU_VS;
+ call_conv = RADEON_LLVM_AMDGPU_HS;
break;
case PIPE_SHADER_GEOMETRY:
call_conv = RADEON_LLVM_AMDGPU_GS;
void si_llvm_optimize_module(struct si_shader_context *ctx)
{
- struct gallivm_state *gallivm = &ctx->gallivm;
- const char *triple = LLVMGetTarget(gallivm->module);
- LLVMTargetLibraryInfoRef target_library_info;
-
/* Dump LLVM IR before any optimization passes */
if (ctx->screen->debug_flags & DBG(PREOPT_IR) &&
si_can_dump_shader(ctx->screen, ctx->type))
LLVMDumpModule(ctx->gallivm.module);
- /* Create the pass manager */
- gallivm->passmgr = LLVMCreatePassManager();
-
- target_library_info = gallivm_create_target_library_info(triple);
- LLVMAddTargetLibraryInfo(target_library_info, gallivm->passmgr);
-
- if (si_extra_shader_checks(ctx->screen, ctx->type))
- LLVMAddVerifierPass(gallivm->passmgr);
-
- LLVMAddAlwaysInlinerPass(gallivm->passmgr);
-
- /* This pass should eliminate all the load and store instructions */
- LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
-
- /* Add some optimization passes */
- LLVMAddScalarReplAggregatesPass(gallivm->passmgr);
- LLVMAddLICMPass(gallivm->passmgr);
- LLVMAddAggressiveDCEPass(gallivm->passmgr);
- LLVMAddCFGSimplificationPass(gallivm->passmgr);
- /* This is recommended by the instruction combining pass. */
- LLVMAddEarlyCSEMemSSAPass(gallivm->passmgr);
- LLVMAddInstructionCombiningPass(gallivm->passmgr);
-
/* Run the pass */
- LLVMRunPassManager(gallivm->passmgr, ctx->gallivm.module);
-
+ LLVMRunPassManager(ctx->compiler->passmgr, ctx->gallivm.module);
LLVMDisposeBuilder(ctx->ac.builder);
- LLVMDisposePassManager(gallivm->passmgr);
- gallivm_dispose_target_library_info(target_library_info);
}
void si_llvm_dispose(struct si_shader_context *ctx)