*/
#include "si_shader_internal.h"
-#include "radeon/radeon_elf_util.h"
+#include "si_pipe.h"
#include "gallivm/lp_bld_const.h"
#include "gallivm/lp_bld_gather.h"
#include <stdio.h>
#include <llvm-c/Transforms/IPO.h>
#include <llvm-c/Transforms/Scalar.h>
+#include <llvm-c/Support.h>
/* Data for if/else/endif and bgnloop/endloop control flow structures.
*/
LLVMBasicBlockRef loop_entry_block;
};
-#define CPU_STRING_LEN 30
-#define FS_STRING_LEN 30
-#define TRIPLE_STRING_LEN 7
-
-/**
- * Shader types for the LLVM backend.
- */
-enum si_llvm_shader_type {
- RADEON_LLVM_SHADER_PS = 0,
- RADEON_LLVM_SHADER_VS = 1,
- RADEON_LLVM_SHADER_GS = 2,
- RADEON_LLVM_SHADER_CS = 3,
-};
-
enum si_llvm_calling_convention {
RADEON_LLVM_AMDGPU_VS = 87,
RADEON_LLVM_AMDGPU_GS = 88,
LLVMAddTargetDependentFunctionAttr(F, name, str);
}
-/**
- * Set the shader type we want to compile
- *
- * @param type shader type to set
- */
-void si_llvm_shader_type(LLVMValueRef F, unsigned type)
-{
- enum si_llvm_shader_type llvm_type;
- enum si_llvm_calling_convention calling_conv;
-
- switch (type) {
- case PIPE_SHADER_VERTEX:
- case PIPE_SHADER_TESS_CTRL:
- case PIPE_SHADER_TESS_EVAL:
- llvm_type = RADEON_LLVM_SHADER_VS;
- calling_conv = RADEON_LLVM_AMDGPU_VS;
- break;
- case PIPE_SHADER_GEOMETRY:
- llvm_type = RADEON_LLVM_SHADER_GS;
- calling_conv = RADEON_LLVM_AMDGPU_GS;
- break;
- case PIPE_SHADER_FRAGMENT:
- llvm_type = RADEON_LLVM_SHADER_PS;
- calling_conv = RADEON_LLVM_AMDGPU_PS;
- break;
- case PIPE_SHADER_COMPUTE:
- llvm_type = RADEON_LLVM_SHADER_CS;
- calling_conv = RADEON_LLVM_AMDGPU_CS;
- break;
- default:
- unreachable("Unhandle shader type");
- }
-
- if (HAVE_LLVM >= 0x309)
- LLVMSetFunctionCallConv(F, calling_conv);
- else
- si_llvm_add_attribute(F, "ShaderType", llvm_type);
-}
-
static void init_amdgpu_target()
{
gallivm_init_llvm_targets();
-#if HAVE_LLVM < 0x0307
- LLVMInitializeR600TargetInfo();
- LLVMInitializeR600Target();
- LLVMInitializeR600TargetMC();
- LLVMInitializeR600AsmPrinter();
-#else
LLVMInitializeAMDGPUTargetInfo();
LLVMInitializeAMDGPUTarget();
LLVMInitializeAMDGPUTargetMC();
LLVMInitializeAMDGPUAsmPrinter();
-#endif
+ /* For inline assembly. */
+ LLVMInitializeAMDGPUAsmParser();
+
+ if (HAVE_LLVM >= 0x0400) {
+ /*
+ * Workaround for bug in llvm 4.0 that causes image intrinsics
+ * to disappear.
+ * https://reviews.llvm.org/D26348
+ */
+ const char *argv[2] = {"mesa", "-simplifycfg-sink-common=false"};
+ LLVMParseCommandLineOptions(2, argv, NULL);
+ }
}
static once_flag init_amdgpu_target_once_flag = ONCE_FLAG_INIT;
*
* @returns 0 for success, 1 for failure
*/
-unsigned si_llvm_compile(LLVMModuleRef M, struct radeon_shader_binary *binary,
+unsigned si_llvm_compile(LLVMModuleRef M, struct ac_shader_binary *binary,
LLVMTargetMachineRef tm,
struct pipe_debug_callback *debug)
{
buffer_size = LLVMGetBufferSize(out_buffer);
buffer_data = LLVMGetBufferStart(out_buffer);
- radeon_elf_read(buffer_data, buffer_size, binary);
+ ac_elf_read(buffer_data, buffer_size, binary);
/* Clean up */
LLVMDisposeMemoryBuffer(out_buffer);
{
struct gallivm_state *gallivm = &ctx->gallivm;
LLVMBuilderRef builder = gallivm->builder;
- LLVMValueRef c_max = lp_build_const_int32(gallivm, num - 1);
+ LLVMValueRef c_max = LLVMConstInt(ctx->i32, num - 1, 0);
LLVMValueRef cc;
if (util_is_power_of_two(num)) {
const struct tgsi_ind_register *reg)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- unsigned num_arrays = ctx->soa.bld_base.info->array_max[TGSI_FILE_TEMPORARY];
+ unsigned num_arrays = ctx->bld_base.info->array_max[TGSI_FILE_TEMPORARY];
unsigned i;
if (reg && reg->ArrayID > 0 && reg->ArrayID <= num_arrays)
}
static LLVMValueRef
-emit_array_index(struct lp_build_tgsi_soa_context *bld,
+emit_array_index(struct si_shader_context *ctx,
const struct tgsi_ind_register *reg,
unsigned offset)
{
- struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
if (!reg) {
- return lp_build_const_int32(gallivm, offset);
+ return LLVMConstInt(ctx->i32, offset, 0);
}
- LLVMValueRef addr = LLVMBuildLoad(gallivm->builder, bld->addr[reg->Index][reg->Swizzle], "");
- return LLVMBuildAdd(gallivm->builder, addr, lp_build_const_int32(gallivm, offset), "");
+ LLVMValueRef addr = LLVMBuildLoad(gallivm->builder, ctx->addrs[reg->Index][reg->Swizzle], "");
+ return LLVMBuildAdd(gallivm->builder, addr, LLVMConstInt(ctx->i32, offset, 0), "");
}
/**
{
unsigned array_id;
struct tgsi_array_info *array;
- struct gallivm_state *gallivm = ctx->soa.bld_base.base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef idxs[2];
LLVMValueRef index;
if (file != TGSI_FILE_TEMPORARY)
return NULL;
- array_id = get_temp_array_id(&ctx->soa.bld_base, reg_index, reg_indirect);
+ array_id = get_temp_array_id(&ctx->bld_base, reg_index, reg_indirect);
if (!array_id)
return NULL;
if (!(array->writemask & (1 << swizzle)))
return ctx->undef_alloca;
- index = emit_array_index(&ctx->soa, reg_indirect,
+ index = emit_array_index(ctx, reg_indirect,
reg_index - ctx->temp_arrays[array_id - 1].range.First);
/* Ensure that the index is within a valid range, to guard against
index = LLVMBuildMul(
builder, index,
- lp_build_const_int32(gallivm, util_bitcount(array->writemask)),
+ LLVMConstInt(ctx->i32, util_bitcount(array->writemask), 0),
"");
index = LLVMBuildAdd(
builder, index,
- lp_build_const_int32(
- gallivm,
- util_bitcount(array->writemask & ((1 << swizzle) - 1))),
+ LLVMConstInt(ctx->i32,
+ util_bitcount(array->writemask & ((1 << swizzle) - 1)), 0),
"");
- idxs[0] = ctx->soa.bld_base.uint_bld.zero;
+ idxs[0] = ctx->i32_0;
idxs[1] = index;
return LLVMBuildGEP(builder, alloca, idxs, 2, "");
}
struct tgsi_declaration_range range,
unsigned swizzle)
{
- struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
- struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
- LLVMBuilderRef builder = bld_base->base.gallivm->builder;
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+
+ LLVMBuilderRef builder = ctx->gallivm.builder;
unsigned i, size = range.Last - range.First + 1;
LLVMTypeRef vec = LLVMVectorType(tgsi2llvmtype(bld_base, type), size);
tmp_reg.Register.Index = i + range.First;
LLVMValueRef temp = si_llvm_emit_fetch(bld_base, &tmp_reg, type, swizzle);
result = LLVMBuildInsertElement(builder, result, temp,
- lp_build_const_int32(gallivm, i), "array_vector");
+ LLVMConstInt(ctx->i32, i, 0), "array_vector");
}
return result;
}
const struct tgsi_ind_register *reg_indirect)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef ptr;
LLVMValueRef val = LLVMBuildLoad(builder, ptr, "");
if (tgsi_type_is_64bit(type)) {
LLVMValueRef ptr_hi, val_hi;
- ptr_hi = LLVMBuildGEP(builder, ptr, &bld_base->uint_bld.one, 1, "");
+ ptr_hi = LLVMBuildGEP(builder, ptr, &ctx->i32_1, 1, "");
val_hi = LLVMBuildLoad(builder, ptr_hi, "");
val = si_llvm_emit_fetch_64bit(bld_base, type, val, val_hi);
}
struct tgsi_declaration_range range =
get_array_range(bld_base, file, reg_index, reg_indirect);
LLVMValueRef index =
- emit_array_index(bld, reg_indirect, reg_index - range.First);
+ emit_array_index(ctx, reg_indirect, reg_index - range.First);
LLVMValueRef array =
emit_array_fetch(bld_base, file, type, range, swizzle);
return LLVMBuildExtractElement(builder, array, index, "");
const struct tgsi_ind_register *reg_indirect)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef ptr;
} else {
unsigned i, size;
struct tgsi_declaration_range range = get_array_range(bld_base, file, reg_index, reg_indirect);
- LLVMValueRef index = emit_array_index(bld, reg_indirect, reg_index - range.First);
+ LLVMValueRef index = emit_array_index(ctx, reg_indirect, reg_index - range.First);
LLVMValueRef array =
emit_array_fetch(bld_base, file, TGSI_TYPE_FLOAT, range, chan_index);
LLVMValueRef temp_ptr;
for (i = 0; i < size; ++i) {
switch(file) {
case TGSI_FILE_OUTPUT:
- temp_ptr = bld->outputs[i + range.First][chan_index];
+ temp_ptr = ctx->outputs[i + range.First][chan_index];
break;
case TGSI_FILE_TEMPORARY:
continue;
}
value = LLVMBuildExtractElement(builder, array,
- lp_build_const_int32(gallivm, i), "");
+ LLVMConstInt(ctx->i32, i, 0), "");
LLVMBuildStore(builder, value, temp_ptr);
}
}
}
+/* If this is true, preload FS inputs at the beginning of shaders. Otherwise,
+ * reload them at each use. This must be true if the shader is using
+ * derivatives and KILL, because KILL can leave the WQM and then a lazy
+ * input load isn't in the WQM anymore.
+ */
+static bool si_preload_fs_inputs(struct si_shader_context *ctx)
+{
+ struct si_shader_selector *sel = ctx->shader->selector;
+
+ return sel->info.uses_derivatives &&
+ sel->info.uses_kill;
+}
+
+static LLVMValueRef
+get_output_ptr(struct lp_build_tgsi_context *bld_base, unsigned index,
+ unsigned chan)
+{
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+
+ assert(index <= ctx->bld_base.info->file_max[TGSI_FILE_OUTPUT]);
+ return ctx->outputs[index][chan];
+}
+
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)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
- LLVMBuilderRef builder = bld_base->base.gallivm->builder;
+ LLVMBuilderRef builder = ctx->gallivm.builder;
LLVMValueRef result = NULL, ptr, ptr2;
if (swizzle == ~0) {
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
values[chan] = si_llvm_emit_fetch(bld_base, reg, type, chan);
}
- return lp_build_gather_values(bld_base->base.gallivm, values,
+ return lp_build_gather_values(&ctx->gallivm, values,
TGSI_NUM_CHANNELS);
}
case TGSI_FILE_IMMEDIATE: {
LLVMTypeRef ctype = tgsi2llvmtype(bld_base, type);
if (tgsi_type_is_64bit(type)) {
- result = LLVMGetUndef(LLVMVectorType(LLVMIntTypeInContext(bld_base->base.gallivm->context, 32), bld_base->base.type.length * 2));
+ result = LLVMGetUndef(LLVMVectorType(ctx->i32, bld_base->base.type.length * 2));
result = LLVMConstInsertElement(result,
- bld->immediates[reg->Register.Index][swizzle],
- bld_base->int_bld.zero);
+ ctx->imms[reg->Register.Index * TGSI_NUM_CHANNELS + swizzle],
+ ctx->i32_0);
result = LLVMConstInsertElement(result,
- bld->immediates[reg->Register.Index][swizzle + 1],
- bld_base->int_bld.one);
+ ctx->imms[reg->Register.Index * TGSI_NUM_CHANNELS + swizzle + 1],
+ ctx->i32_1);
return LLVMConstBitCast(result, ctype);
} else {
- return LLVMConstBitCast(bld->immediates[reg->Register.Index][swizzle], ctype);
+ return LLVMConstBitCast(ctx->imms[reg->Register.Index * TGSI_NUM_CHANNELS + swizzle], ctype);
}
}
* only once. Fragment shaders don't care much, because
* v_interp instructions are much cheaper than VMEM loads.
*/
- if (ctx->soa.bld_base.info->processor == PIPE_SHADER_FRAGMENT)
+ if (!si_preload_fs_inputs(ctx) &&
+ ctx->bld_base.info->processor == PIPE_SHADER_FRAGMENT)
ctx->load_input(ctx, index, &ctx->input_decls[index], input);
else
memcpy(input, &ctx->inputs[index * 4], sizeof(input));
break;
case TGSI_FILE_OUTPUT:
- ptr = lp_get_output_ptr(bld, reg->Register.Index, swizzle);
+ ptr = get_output_ptr(bld_base, reg->Register.Index, swizzle);
if (tgsi_type_is_64bit(type)) {
- ptr2 = lp_get_output_ptr(bld, reg->Register.Index, swizzle + 1);
+ ptr2 = get_output_ptr(bld_base, reg->Register.Index, swizzle + 1);
return si_llvm_emit_fetch_64bit(bld_base, type,
LLVMBuildLoad(builder, ptr, ""),
LLVMBuildLoad(builder, ptr2, ""));
unsigned swizzle)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
-
+ LLVMBuilderRef builder = ctx->gallivm.builder;
LLVMValueRef cval = ctx->system_values[reg->Register.Index];
+
+ if (tgsi_type_is_64bit(type)) {
+ LLVMValueRef lo, hi;
+
+ assert(swizzle == 0 || swizzle == 2);
+
+ lo = LLVMBuildExtractElement(
+ builder, cval, LLVMConstInt(ctx->i32, swizzle, 0), "");
+ hi = LLVMBuildExtractElement(
+ builder, cval, LLVMConstInt(ctx->i32, swizzle + 1, 0), "");
+
+ return si_llvm_emit_fetch_64bit(bld_base, type, lo, hi);
+ }
+
if (LLVMGetTypeKind(LLVMTypeOf(cval)) == LLVMVectorTypeKind) {
- cval = LLVMBuildExtractElement(gallivm->builder, cval,
- lp_build_const_int32(gallivm, swizzle), "");
+ cval = LLVMBuildExtractElement(
+ builder, cval, LLVMConstInt(ctx->i32, swizzle, 0), "");
+ } else {
+ assert(swizzle == 0);
}
+
return bitcast(bld_base, type, cval);
}
const struct tgsi_full_declaration *decl)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- LLVMBuilderRef builder = bld_base->base.gallivm->builder;
+ LLVMBuilderRef builder = ctx->gallivm.builder;
unsigned first, last, i;
switch(decl->Declaration.File) {
case TGSI_FILE_ADDRESS:
for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
unsigned chan;
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
- ctx->soa.addr[idx][chan] = lp_build_alloca_undef(
+ ctx->addrs[idx][chan] = lp_build_alloca_undef(
&ctx->gallivm,
- ctx->soa.bld_base.uint_bld.elem_type, "");
+ ctx->i32, "");
}
}
break;
* FIXME: We shouldn't need to have the non-alloca
* code path for arrays. LLVM should be smart enough to
* promote allocas into registers when profitable.
- *
- * LLVM 3.8 crashes with this.
*/
- if (HAVE_LLVM >= 0x0309 && array_size > 16) {
+ if (array_size > 16 ||
+ /* TODO: VGPR indexing is buggy on GFX9. */
+ ctx->screen->b.chip_class == GFX9) {
array_alloca = LLVMBuildAlloca(builder,
- LLVMArrayType(bld_base->base.vec_type,
+ LLVMArrayType(ctx->f32,
array_size), "array");
ctx->temp_array_allocas[id] = array_alloca;
}
first + i / 4, "xyzw"[i % 4]);
#endif
ctx->temps[first * TGSI_NUM_CHANNELS + i] =
- lp_build_alloca_undef(bld_base->base.gallivm,
- bld_base->base.vec_type,
+ lp_build_alloca_undef(&ctx->gallivm,
+ ctx->f32,
name);
}
} else {
LLVMValueRef idxs[2] = {
- bld_base->uint_bld.zero,
+ ctx->i32_0,
NULL
};
unsigned j = 0;
* it never writes to.
*/
ctx->undef_alloca = lp_build_alloca_undef(
- bld_base->base.gallivm,
- bld_base->base.vec_type, "undef");
+ &ctx->gallivm,
+ ctx->f32, "undef");
}
for (i = 0; i < decl_size; ++i) {
snprintf(name, sizeof(name), "TEMP%d.%c",
first + i / 4, "xyzw"[i % 4]);
#endif
- idxs[1] = lp_build_const_int32(bld_base->base.gallivm, j);
+ idxs[1] = LLVMConstInt(ctx->i32, j, 0);
ptr = LLVMBuildGEP(builder, array_alloca, idxs, 2, name);
j++;
} else {
if (ctx->load_input &&
ctx->input_decls[idx].Declaration.File != TGSI_FILE_INPUT) {
ctx->input_decls[idx] = *decl;
+ ctx->input_decls[idx].Range.First = idx;
+ ctx->input_decls[idx].Range.Last = idx;
+ ctx->input_decls[idx].Semantic.Index += idx - decl->Range.First;
- if (bld_base->info->processor != PIPE_SHADER_FRAGMENT)
- ctx->load_input(ctx, idx, decl,
+ if (si_preload_fs_inputs(ctx) ||
+ bld_base->info->processor != PIPE_SHADER_FRAGMENT)
+ ctx->load_input(ctx, idx, &ctx->input_decls[idx],
&ctx->inputs[idx * 4]);
}
}
for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
unsigned chan;
assert(idx < RADEON_LLVM_MAX_OUTPUTS);
- if (ctx->soa.outputs[idx][0])
+ if (ctx->outputs[idx][0])
continue;
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
#ifdef DEBUG
snprintf(name, sizeof(name), "OUT%d.%c",
idx, "xyzw"[chan % 4]);
#endif
- ctx->soa.outputs[idx][chan] = lp_build_alloca_undef(
+ ctx->outputs[idx][chan] = lp_build_alloca_undef(
&ctx->gallivm,
- ctx->soa.bld_base.base.elem_type, name);
+ ctx->f32, name);
}
}
break;
}
}
-LLVMValueRef si_llvm_saturate(struct lp_build_tgsi_context *bld_base,
- LLVMValueRef value)
-{
- struct lp_build_emit_data clamp_emit_data;
-
- memset(&clamp_emit_data, 0, sizeof(clamp_emit_data));
- clamp_emit_data.arg_count = 3;
- clamp_emit_data.args[0] = value;
- clamp_emit_data.args[2] = bld_base->base.one;
- clamp_emit_data.args[1] = bld_base->base.zero;
-
- return lp_build_emit_llvm(bld_base, TGSI_OPCODE_CLAMP,
- &clamp_emit_data);
-}
-
void si_llvm_emit_store(struct lp_build_tgsi_context *bld_base,
const struct tgsi_full_instruction *inst,
const struct tgsi_opcode_info *info,
LLVMValueRef dst[4])
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
- struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
const struct tgsi_full_dst_register *reg = &inst->Dst[0];
- LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
+ LLVMBuilderRef builder = ctx->gallivm.builder;
LLVMValueRef temp_ptr, temp_ptr2 = NULL;
unsigned chan, chan_index;
bool is_vec_store = false;
if (is_vec_store) {
LLVMValueRef values[4] = {};
TGSI_FOR_EACH_DST0_ENABLED_CHANNEL(inst, chan) {
- LLVMValueRef index = lp_build_const_int32(gallivm, chan);
+ LLVMValueRef index = LLVMConstInt(ctx->i32, chan, 0);
values[chan] = LLVMBuildExtractElement(gallivm->builder,
dst[0], index, "");
}
if (tgsi_type_is_64bit(dtype) && (chan_index == 1 || chan_index == 3))
continue;
if (inst->Instruction.Saturate)
- value = si_llvm_saturate(bld_base, value);
+ value = ac_build_clamp(&ctx->ac, value);
if (reg->Register.File == TGSI_FILE_ADDRESS) {
- temp_ptr = bld->addr[reg->Register.Index][chan_index];
+ temp_ptr = ctx->addrs[reg->Register.Index][chan_index];
LLVMBuildStore(builder, value, temp_ptr);
continue;
}
} else {
switch(reg->Register.File) {
case TGSI_FILE_OUTPUT:
- temp_ptr = bld->outputs[reg->Register.Index][chan_index];
+ temp_ptr = ctx->outputs[reg->Register.Index][chan_index];
if (tgsi_type_is_64bit(dtype))
- temp_ptr2 = bld->outputs[reg->Register.Index][chan_index + 1];
+ temp_ptr2 = ctx->outputs[reg->Register.Index][chan_index + 1];
break;
case TGSI_FILE_TEMPORARY:
LLVMBuildStore(builder, value, temp_ptr);
else {
LLVMValueRef ptr = LLVMBuildBitCast(builder, value,
- LLVMVectorType(LLVMIntTypeInContext(bld_base->base.gallivm->context, 32), 2), "");
+ LLVMVectorType(ctx->i32, 2), "");
LLVMValueRef val2;
value = LLVMBuildExtractElement(builder, ptr,
- bld_base->uint_bld.zero, "");
+ ctx->i32_0, "");
val2 = LLVMBuildExtractElement(builder, ptr,
- bld_base->uint_bld.one, "");
+ ctx->i32_1, "");
LLVMBuildStore(builder, bitcast(bld_base, TGSI_TYPE_FLOAT, value), temp_ptr);
LLVMBuildStore(builder, bitcast(bld_base, TGSI_TYPE_FLOAT, val2), temp_ptr2);
struct lp_build_emit_data *emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
struct si_llvm_flow *flow = push_flow(ctx);
flow->loop_entry_block = append_basic_block(ctx, "LOOP");
flow->next_block = append_basic_block(ctx, "ENDLOOP");
struct lp_build_emit_data *emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
struct si_llvm_flow *flow = get_innermost_loop(ctx);
LLVMBuildBr(gallivm->builder, flow->next_block);
struct lp_build_emit_data *emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
struct si_llvm_flow *flow = get_innermost_loop(ctx);
LLVMBuildBr(gallivm->builder, flow->loop_entry_block);
struct lp_build_emit_data *emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
struct si_llvm_flow *current_branch = get_current_flow(ctx);
LLVMBasicBlockRef endif_block;
struct lp_build_emit_data *emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
struct si_llvm_flow *current_branch = get_current_flow(ctx);
assert(!current_branch->loop_entry_block);
struct lp_build_emit_data *emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
struct si_llvm_flow *current_loop = get_current_flow(ctx);
assert(current_loop->loop_entry_block);
LLVMValueRef cond)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
struct si_llvm_flow *flow = push_flow(ctx);
LLVMBasicBlockRef if_block;
struct si_shader_context *ctx = si_shader_context(bld_base);
for (i = 0; i < 4; ++i) {
- ctx->soa.immediates[ctx->soa.num_immediates][i] =
- LLVMConstInt(bld_base->uint_bld.elem_type, imm->u[i].Uint, false );
+ ctx->imms[ctx->imms_num * TGSI_NUM_CHANNELS + i] =
+ LLVMConstInt(ctx->i32, imm->u[i].Uint, false );
}
- ctx->soa.num_immediates++;
+ ctx->imms_num++;
}
void si_llvm_context_init(struct si_shader_context *ctx,
struct si_screen *sscreen,
- struct si_shader *shader,
- LLVMTargetMachineRef tm,
- const struct tgsi_shader_info *info,
- const struct tgsi_token *tokens)
+ LLVMTargetMachineRef tm)
{
struct lp_type type;
* helper functions in the gallivm module.
*/
memset(ctx, 0, sizeof(*ctx));
- ctx->shader = shader;
ctx->screen = sscreen;
ctx->tm = tm;
- ctx->type = info ? info->processor : -1;
ctx->gallivm.context = LLVMContextCreate();
ctx->gallivm.module = LLVMModuleCreateWithNameInContext("tgsi",
ctx->gallivm.context);
LLVMSetTarget(ctx->gallivm.module, "amdgcn--");
- ctx->gallivm.builder = LLVMCreateBuilderInContext(ctx->gallivm.context);
- struct lp_build_tgsi_context *bld_base = &ctx->soa.bld_base;
+ 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);
- bld_base->info = info;
+ bool unsafe_fpmath = (sscreen->b.debug_flags & DBG_UNSAFE_MATH) != 0;
+ enum lp_float_mode float_mode =
+ unsafe_fpmath ? LP_FLOAT_MODE_UNSAFE_FP_MATH :
+ LP_FLOAT_MODE_NO_SIGNED_ZEROS_FP_MATH;
- if (info && info->array_max[TGSI_FILE_TEMPORARY] > 0) {
- int size = info->array_max[TGSI_FILE_TEMPORARY];
+ ctx->gallivm.builder = lp_create_builder(ctx->gallivm.context,
+ float_mode);
- ctx->temp_arrays = CALLOC(size, sizeof(ctx->temp_arrays[0]));
- ctx->temp_array_allocas = CALLOC(size, sizeof(ctx->temp_array_allocas[0]));
+ ac_llvm_context_init(&ctx->ac, ctx->gallivm.context);
+ ctx->ac.module = ctx->gallivm.module;
+ ctx->ac.builder = ctx->gallivm.builder;
- if (tokens)
- tgsi_scan_arrays(tokens, TGSI_FILE_TEMPORARY, size,
- ctx->temp_arrays);
- }
+ struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
type.floating = true;
type.fixed = false;
type.length = 1;
lp_build_context_init(&bld_base->base, &ctx->gallivm, type);
- lp_build_context_init(&ctx->soa.bld_base.uint_bld, &ctx->gallivm, lp_uint_type(type));
- lp_build_context_init(&ctx->soa.bld_base.int_bld, &ctx->gallivm, lp_int_type(type));
+ lp_build_context_init(&ctx->bld_base.uint_bld, &ctx->gallivm, lp_uint_type(type));
+ lp_build_context_init(&ctx->bld_base.int_bld, &ctx->gallivm, lp_int_type(type));
type.width *= 2;
- lp_build_context_init(&ctx->soa.bld_base.dbl_bld, &ctx->gallivm, type);
- lp_build_context_init(&ctx->soa.bld_base.uint64_bld, &ctx->gallivm, lp_uint_type(type));
- lp_build_context_init(&ctx->soa.bld_base.int64_bld, &ctx->gallivm, lp_int_type(type));
+ lp_build_context_init(&ctx->bld_base.dbl_bld, &ctx->gallivm, type);
+ lp_build_context_init(&ctx->bld_base.uint64_bld, &ctx->gallivm, lp_uint_type(type));
+ lp_build_context_init(&ctx->bld_base.int64_bld, &ctx->gallivm, lp_int_type(type));
bld_base->soa = 1;
- bld_base->emit_store = si_llvm_emit_store;
bld_base->emit_swizzle = emit_swizzle;
bld_base->emit_declaration = emit_declaration;
bld_base->emit_immediate = emit_immediate;
- bld_base->emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = si_llvm_emit_fetch;
- bld_base->emit_fetch_funcs[TGSI_FILE_INPUT] = si_llvm_emit_fetch;
- bld_base->emit_fetch_funcs[TGSI_FILE_TEMPORARY] = si_llvm_emit_fetch;
- bld_base->emit_fetch_funcs[TGSI_FILE_OUTPUT] = si_llvm_emit_fetch;
- bld_base->emit_fetch_funcs[TGSI_FILE_SYSTEM_VALUE] = fetch_system_value;
-
/* metadata allowing 2.5 ULP */
ctx->fpmath_md_kind = LLVMGetMDKindIDInContext(ctx->gallivm.context,
"fpmath", 6);
ctx->fpmath_md_2p5_ulp = LLVMMDNodeInContext(ctx->gallivm.context,
&arg, 1);
- /* Allocate outputs */
- ctx->soa.outputs = ctx->outputs;
-
bld_base->op_actions[TGSI_OPCODE_BGNLOOP].emit = bgnloop_emit;
bld_base->op_actions[TGSI_OPCODE_BRK].emit = brk_emit;
bld_base->op_actions[TGSI_OPCODE_CONT].emit = cont_emit;
bld_base->op_actions[TGSI_OPCODE_ENDIF].emit = endif_emit;
bld_base->op_actions[TGSI_OPCODE_ENDLOOP].emit = endloop_emit;
- si_shader_context_init_alu(&ctx->soa.bld_base);
+ si_shader_context_init_alu(&ctx->bld_base);
ctx->voidt = LLVMVoidTypeInContext(ctx->gallivm.context);
ctx->i1 = LLVMInt1TypeInContext(ctx->gallivm.context);
ctx->i64 = LLVMInt64TypeInContext(ctx->gallivm.context);
ctx->i128 = LLVMIntTypeInContext(ctx->gallivm.context, 128);
ctx->f32 = LLVMFloatTypeInContext(ctx->gallivm.context);
- ctx->v16i8 = LLVMVectorType(ctx->i8, 16);
ctx->v2i32 = LLVMVectorType(ctx->i32, 2);
ctx->v4i32 = LLVMVectorType(ctx->i32, 4);
ctx->v4f32 = LLVMVectorType(ctx->f32, 4);
ctx->v8i32 = LLVMVectorType(ctx->i32, 8);
+
+ ctx->i32_0 = LLVMConstInt(ctx->i32, 0, 0);
+ ctx->i32_1 = LLVMConstInt(ctx->i32, 1, 0);
+}
+
+/* Set the context to a certain TGSI shader. Can be called repeatedly
+ * to change the shader. */
+void si_llvm_context_set_tgsi(struct si_shader_context *ctx,
+ struct si_shader *shader)
+{
+ const struct tgsi_shader_info *info = NULL;
+ const struct tgsi_token *tokens = NULL;
+
+ if (shader && shader->selector) {
+ info = &shader->selector->info;
+ tokens = shader->selector->tokens;
+ }
+
+ ctx->shader = shader;
+ ctx->type = info ? info->processor : -1;
+ ctx->bld_base.info = info;
+
+ /* Clean up the old contents. */
+ FREE(ctx->temp_arrays);
+ ctx->temp_arrays = NULL;
+ FREE(ctx->temp_array_allocas);
+ ctx->temp_array_allocas = NULL;
+
+ FREE(ctx->imms);
+ ctx->imms = NULL;
+ ctx->imms_num = 0;
+
+ FREE(ctx->temps);
+ ctx->temps = NULL;
+ ctx->temps_count = 0;
+
+ if (!info || !tokens)
+ return;
+
+ if (info->array_max[TGSI_FILE_TEMPORARY] > 0) {
+ int size = info->array_max[TGSI_FILE_TEMPORARY];
+
+ ctx->temp_arrays = CALLOC(size, sizeof(ctx->temp_arrays[0]));
+ ctx->temp_array_allocas = CALLOC(size, sizeof(ctx->temp_array_allocas[0]));
+
+ tgsi_scan_arrays(tokens, TGSI_FILE_TEMPORARY, size,
+ ctx->temp_arrays);
+ }
+ if (info->file_max[TGSI_FILE_IMMEDIATE] >= 0) {
+ int size = info->file_max[TGSI_FILE_IMMEDIATE] + 1;
+ ctx->imms = MALLOC(size * TGSI_NUM_CHANNELS * sizeof(LLVMValueRef));
+ }
+
+ /* Re-set these to start with a clean slate. */
+ ctx->bld_base.num_instructions = 0;
+ ctx->bld_base.pc = 0;
+ memset(ctx->outputs, 0, sizeof(ctx->outputs));
+
+ ctx->bld_base.emit_store = si_llvm_emit_store;
+ ctx->bld_base.emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = si_llvm_emit_fetch;
+ ctx->bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = si_llvm_emit_fetch;
+ ctx->bld_base.emit_fetch_funcs[TGSI_FILE_TEMPORARY] = si_llvm_emit_fetch;
+ ctx->bld_base.emit_fetch_funcs[TGSI_FILE_OUTPUT] = si_llvm_emit_fetch;
+ ctx->bld_base.emit_fetch_funcs[TGSI_FILE_SYSTEM_VALUE] = fetch_system_value;
}
void si_llvm_create_func(struct si_shader_context *ctx,
{
LLVMTypeRef main_fn_type, ret_type;
LLVMBasicBlockRef main_fn_body;
+ enum si_llvm_calling_convention call_conv;
if (num_return_elems)
ret_type = LLVMStructTypeInContext(ctx->gallivm.context,
main_fn_body = LLVMAppendBasicBlockInContext(ctx->gallivm.context,
ctx->main_fn, "main_body");
LLVMPositionBuilderAtEnd(ctx->gallivm.builder, main_fn_body);
+
+ switch (ctx->type) {
+ case PIPE_SHADER_VERTEX:
+ case PIPE_SHADER_TESS_CTRL:
+ case PIPE_SHADER_TESS_EVAL:
+ call_conv = RADEON_LLVM_AMDGPU_VS;
+ break;
+ case PIPE_SHADER_GEOMETRY:
+ call_conv = RADEON_LLVM_AMDGPU_GS;
+ break;
+ case PIPE_SHADER_FRAGMENT:
+ call_conv = RADEON_LLVM_AMDGPU_PS;
+ break;
+ case PIPE_SHADER_COMPUTE:
+ call_conv = RADEON_LLVM_AMDGPU_CS;
+ break;
+ default:
+ unreachable("Unhandle shader type");
+ }
+
+ LLVMSetFunctionCallConv(ctx->main_fn, call_conv);
}
-void si_llvm_finalize_module(struct si_shader_context *ctx,
- bool run_verifier)
+void si_llvm_optimize_module(struct si_shader_context *ctx)
{
- struct gallivm_state *gallivm = ctx->soa.bld_base.base.gallivm;
+ 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->b.debug_flags & DBG_PREOPT_IR &&
+ r600_can_dump_shader(&ctx->screen->b, 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 (run_verifier)
+ if (r600_extra_shader_checks(&ctx->screen->b, ctx->type))
LLVMAddVerifierPass(gallivm->passmgr);
LLVMAddAlwaysInlinerPass(gallivm->passmgr);
void si_llvm_dispose(struct si_shader_context *ctx)
{
- LLVMDisposeModule(ctx->soa.bld_base.base.gallivm->module);
- LLVMContextDispose(ctx->soa.bld_base.base.gallivm->context);
+ LLVMDisposeModule(ctx->gallivm.module);
+ LLVMContextDispose(ctx->gallivm.context);
FREE(ctx->temp_arrays);
ctx->temp_arrays = NULL;
FREE(ctx->temp_array_allocas);
FREE(ctx->temps);
ctx->temps = NULL;
ctx->temps_count = 0;
+ FREE(ctx->imms);
+ ctx->imms = NULL;
+ ctx->imms_num = 0;
FREE(ctx->flow);
ctx->flow = NULL;
ctx->flow_depth_max = 0;