#include "compiler/nir_types.h"
+static int
+type_size(const struct glsl_type *type)
+{
+ return glsl_count_attribute_slots(type, false);
+}
+
static void scan_instruction(struct tgsi_shader_info *info,
nir_instr *instr)
{
}
}
+/**
+ * Perform "lowering" operations on the NIR that are run once when the shader
+ * selector is created.
+ */
+void
+si_lower_nir(struct si_shader_selector* sel)
+{
+ /* Adjust the driver location of inputs and outputs. The state tracker
+ * interprets them as slots, while the ac/nir backend interprets them
+ * as individual components.
+ */
+ nir_foreach_variable(variable, &sel->nir->inputs)
+ variable->data.driver_location *= 4;
+
+ nir_foreach_variable(variable, &sel->nir->outputs) {
+ variable->data.driver_location *= 4;
+
+ if (sel->nir->stage == MESA_SHADER_FRAGMENT) {
+ if (variable->data.location == FRAG_RESULT_DEPTH)
+ variable->data.driver_location += 2;
+ else if (variable->data.location == FRAG_RESULT_STENCIL)
+ variable->data.driver_location += 1;
+ }
+ }
+
+ /* Perform lowerings (and optimizations) of code.
+ *
+ * Performance considerations aside, we must:
+ * - lower certain ALU operations
+ * - ensure constant offsets for texture instructions are folded
+ * and copy-propagated
+ */
+ NIR_PASS_V(sel->nir, nir_lower_io, nir_var_uniform, type_size,
+ (nir_lower_io_options)0);
+ NIR_PASS_V(sel->nir, nir_lower_uniforms_to_ubo);
+
+ NIR_PASS_V(sel->nir, nir_lower_returns);
+ NIR_PASS_V(sel->nir, nir_lower_vars_to_ssa);
+ NIR_PASS_V(sel->nir, nir_lower_alu_to_scalar);
+ NIR_PASS_V(sel->nir, nir_lower_phis_to_scalar);
+
+ static const struct nir_lower_tex_options lower_tex_options = {
+ .lower_txp = ~0u,
+ };
+ NIR_PASS_V(sel->nir, nir_lower_tex, &lower_tex_options);
+
+ bool progress;
+ do {
+ progress = false;
+
+ /* (Constant) copy propagation is needed for txf with offsets. */
+ NIR_PASS(progress, sel->nir, nir_copy_prop);
+ NIR_PASS(progress, sel->nir, nir_opt_remove_phis);
+ NIR_PASS(progress, sel->nir, nir_opt_dce);
+ if (nir_opt_trivial_continues(sel->nir)) {
+ progress = true;
+ NIR_PASS(progress, sel->nir, nir_copy_prop);
+ NIR_PASS(progress, sel->nir, nir_opt_dce);
+ }
+ NIR_PASS(progress, sel->nir, nir_opt_if);
+ NIR_PASS(progress, sel->nir, nir_opt_dead_cf);
+ NIR_PASS(progress, sel->nir, nir_opt_cse);
+ NIR_PASS(progress, sel->nir, nir_opt_peephole_select, 8);
+
+ /* Needed for algebraic lowering */
+ NIR_PASS(progress, sel->nir, nir_opt_algebraic);
+ NIR_PASS(progress, sel->nir, nir_opt_constant_folding);
+
+ NIR_PASS(progress, sel->nir, nir_opt_undef);
+ NIR_PASS(progress, sel->nir, nir_opt_conditional_discard);
+ if (sel->nir->options->max_unroll_iterations) {
+ NIR_PASS(progress, sel->nir, nir_opt_loop_unroll, 0);
+ }
+ } while (progress);
+}
+
static void declare_nir_input_vs(struct si_shader_context *ctx,
struct nir_variable *variable, unsigned rel,
LLVMValueRef out[4])
(*fs_attr_idx)++;
}
+static LLVMValueRef
+si_nir_load_sampler_desc(struct ac_shader_abi *abi,
+ unsigned descriptor_set, unsigned base_index,
+ unsigned constant_index, LLVMValueRef dynamic_index,
+ enum ac_descriptor_type desc_type, bool image,
+ bool write)
+{
+ struct si_shader_context *ctx = si_shader_context_from_abi(abi);
+ LLVMBuilderRef builder = ctx->ac.builder;
+ LLVMValueRef list = LLVMGetParam(ctx->main_fn, ctx->param_samplers_and_images);
+ LLVMValueRef index = dynamic_index;
+
+ assert(!descriptor_set);
+
+ if (!index)
+ index = ctx->ac.i32_0;
+
+ index = LLVMBuildAdd(builder, index,
+ LLVMConstInt(ctx->ac.i32, base_index + constant_index, false),
+ "");
+
+ if (image) {
+ assert(desc_type == AC_DESC_IMAGE || desc_type == AC_DESC_BUFFER);
+ assert(base_index + constant_index < ctx->num_images);
+
+ if (dynamic_index)
+ index = si_llvm_bound_index(ctx, index, ctx->num_images);
+
+ index = LLVMBuildSub(ctx->gallivm.builder,
+ LLVMConstInt(ctx->i32, SI_NUM_IMAGES - 1, 0),
+ index, "");
+
+ /* TODO: be smarter about when we use dcc_off */
+ return si_load_image_desc(ctx, list, index, desc_type, write);
+ }
+
+ assert(base_index + constant_index < ctx->num_samplers);
+
+ if (dynamic_index)
+ index = si_llvm_bound_index(ctx, index, ctx->num_samplers);
+
+ index = LLVMBuildAdd(ctx->gallivm.builder, index,
+ LLVMConstInt(ctx->i32, SI_NUM_IMAGES / 2, 0), "");
+
+ return si_load_sampler_desc(ctx, list, index, desc_type);
+}
+
bool si_nir_build_llvm(struct si_shader_context *ctx, struct nir_shader *nir)
{
+ struct tgsi_shader_info *info = &ctx->shader->selector->info;
+
unsigned fs_attr_idx = 0;
nir_foreach_variable(variable, &nir->inputs) {
unsigned attrib_count = glsl_count_attribute_slots(variable->type,
}
ctx->abi.inputs = &ctx->inputs[0];
+ ctx->abi.load_sampler_desc = si_nir_load_sampler_desc;
+
+ ctx->num_samplers = util_last_bit(info->samplers_declared);
+ ctx->num_images = util_last_bit(info->images_declared);
ac_nir_translate(&ctx->ac, &ctx->abi, nir, NULL);