#include "si_shader_internal.h"
#include "si_pipe.h"
-#include "radeon/radeon_elf_util.h"
#include "gallivm/lp_bld_const.h"
#include "gallivm/lp_bld_gather.h"
*
* @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);
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)
const struct tgsi_ind_register *reg,
unsigned offset)
{
- struct gallivm_state *gallivm = ctx->soa.bld_base.base.gallivm;
+ struct gallivm_state *gallivm = ctx->bld_base.base.gallivm;
if (!reg) {
return lp_build_const_int32(gallivm, offset);
{
unsigned array_id;
struct tgsi_array_info *array;
- struct gallivm_state *gallivm = ctx->soa.bld_base.base.gallivm;
+ struct gallivm_state *gallivm = ctx->bld_base.base.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;
gallivm,
util_bitcount(array->writemask & ((1 << swizzle) - 1))),
"");
- idxs[0] = ctx->soa.bld_base.uint_bld.zero;
+ idxs[0] = ctx->bld_base.uint_bld.zero;
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;
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ struct gallivm_state *gallivm = ctx->bld_base.base.gallivm;
+
LLVMBuilderRef builder = bld_base->base.gallivm->builder;
unsigned i, size = range.Last - range.First + 1;
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;
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef 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:
/* 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, because all inputs should be loaded in the WQM mode.
+ * 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)
{
- return ctx->shader->selector->info.uses_derivatives;
+ 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,
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;
LLVMValueRef result = NULL, ptr, ptr2;
* v_interp instructions are much cheaper than VMEM loads.
*/
if (!si_preload_fs_inputs(ctx) &&
- ctx->soa.bld_base.info->processor == PIPE_SHADER_FRAGMENT)
+ 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, ""));
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
ctx->addrs[idx][chan] = lp_build_alloca_undef(
&ctx->gallivm,
- ctx->soa.bld_base.uint_bld.elem_type, "");
+ ctx->bld_base.uint_bld.elem_type, "");
}
}
break;
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->bld_base.base.elem_type, 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->bld_base.base.gallivm;
const struct tgsi_full_dst_register *reg = &inst->Dst[0];
- LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
+ LLVMBuilderRef builder = ctx->bld_base.base.gallivm->builder;
LLVMValueRef temp_ptr, temp_ptr2 = NULL;
unsigned chan, chan_index;
bool is_vec_store = false;
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 = ctx->addrs[reg->Register.Index][chan_index];
} 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:
ctx->gallivm.context);
LLVMSetTarget(ctx->gallivm.module, "amdgcn--");
+#if HAVE_LLVM >= 0x0309
+ 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);
+#endif
+
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;
+
ctx->gallivm.builder = lp_create_builder(ctx->gallivm.context,
- unsafe_fpmath);
+ float_mode);
ac_llvm_context_init(&ctx->ac, ctx->gallivm.context);
ctx->ac.module = ctx->gallivm.module;
ctx->ac.builder = ctx->gallivm.builder;
- struct lp_build_tgsi_context *bld_base = &ctx->soa.bld_base;
+ struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
bld_base->info = info;
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;
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->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);
}
void si_llvm_create_func(struct si_shader_context *ctx,
void si_llvm_finalize_module(struct si_shader_context *ctx,
bool run_verifier)
{
- struct gallivm_state *gallivm = ctx->soa.bld_base.base.gallivm;
+ struct gallivm_state *gallivm = ctx->bld_base.base.gallivm;
const char *triple = LLVMGetTarget(gallivm->module);
LLVMTargetLibraryInfoRef target_library_info;
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->bld_base.base.gallivm->module);
+ LLVMContextDispose(ctx->bld_base.base.gallivm->context);
FREE(ctx->temp_arrays);
ctx->temp_arrays = NULL;
FREE(ctx->temp_array_allocas);
projects / mesa.git / blobdiff