#include "libresoc_llvm.h"
-#include "libresoc_shader_args.h"
#include "libresoc_llvm_build.h"
-#include <llvm-c/OrcBindings.h>
-#include <llvm-c/Core.h>
-#include <llvm-c/Analysis.h>
-#include "nir/nir.h"
-#include "nir/nir_deref.h"
-#include <float.h>
-
-struct libresoc_nir_tran_ctx {
- struct libresoc_llvm_context lc;
- gl_shader_stage stage;
- shader_info *info;
-
- struct shader_args args;
- LLVMValueRef *ssa_defs;
-
- LLVMValueRef scratch;
- LLVMValueRef constant_data;
-
- struct hash_table *defs;
- struct hash_table *phis;
- struct hash_table *vars;
- struct hash_table *verified_interp;
-
- LLVMValueRef main_function;
- LLVMBasicBlockRef continue_block;
- LLVMBasicBlockRef break_block;
-
- int num_locals;
- LLVMValueRef *locals;
-};
void InitLLVM(struct libresoc_llvm *llvm_ref)
{
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, "");
+ }
+ }
+}
+
+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) {
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)
// 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->abi->inputs + idx + chan, count, stride, false, true);
-
- // values[chan] = LLVMBuildExtractElement(ctx->lc.builder, tmp_vec, indir_index, "");
- // } else
- // values[chan] = ctx->abi->inputs[idx + chan + const_index * stride];
- // }
+ 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 (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->abi->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->abi->outputs[idx + chan + const_index * stride], "");
- // }
- // }
+ 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]);
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->abi->outputs + idx + chan, count, stride, true, true);
+ 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->abi->outputs + idx + chan, count, stride,
- // tmp_vec);
+ 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->abi->outputs[idx + chan + const_index * stride];
+ } else {
+ temp_ptr = ctx->outputs[idx + chan + const_index * stride];
- // LLVMBuildStore(ctx->lc.builder, value, temp_ptr);
- // }
+ LLVMBuildStore(ctx->lc.builder, value, temp_ptr);
+ }
}
break;
case nir_var_function_temp:
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);
- // }
+ 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;
break;
case nir_intrinsic_store_output:
case nir_intrinsic_store_per_vertex_output:
- // visit_store_output(ctx, instr);
+ visit_store_output(ctx, instr);
break;
case nir_intrinsic_load_shared:
// result = visit_load_shared(ctx, instr);
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);
projects / mesa.git / commitdiff