radv/ac: eliminate unused vertex shader outputs. (v2)

This is ported from radeonsi, and I can see at least one
Talos shader drops an export due to this, and saves some
VGPR usage.

v2: use shared code.

Reviewed-by: Bas Nieuwenhuizen <[email protected]>
Signed-off-by: Dave Airlie <[email protected]>

3 files changed, 45 insertions, 21 deletions

diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
index 97cd981ec47..d9962c783f0 100644
--- a/src/amd/common/ac_nir_to_llvm.c
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -31,6 +31,8 @@
 #include "util/bitscan.h"
 #include <llvm-c/Transforms/Scalar.h>
 #include "ac_shader_info.h"
+#include "ac_exp_param.h"
+
 enum radeon_llvm_calling_convention {
 	RADEON_LLVM_AMDGPU_VS = 87,
 	RADEON_LLVM_AMDGPU_GS = 88,
@@ -5133,7 +5135,7 @@ handle_vs_outputs_post(struct nir_to_llvm_context *ctx,
 	LLVMValueRef psize_value = NULL, layer_value = NULL, viewport_index_value = NULL;
 	int i;
 
-	memset(outinfo->vs_output_param_offset, EXP_PARAM_UNDEFINED,
+	memset(outinfo->vs_output_param_offset, AC_EXP_PARAM_UNDEFINED,
 	       sizeof(outinfo->vs_output_param_offset));
 
 	if (ctx->output_mask & (1ull << VARYING_SLOT_CLIP_DIST0)) {
@@ -5758,6 +5760,37 @@ static void ac_llvm_finalize_module(struct nir_to_llvm_context * ctx)
 }
 
 static void
+ac_nir_eliminate_const_vs_outputs(struct nir_to_llvm_context *ctx)
+{
+	struct ac_vs_output_info *outinfo;
+
+	if (ctx->stage == MESA_SHADER_FRAGMENT ||
+	    ctx->stage == MESA_SHADER_COMPUTE ||
+	    ctx->stage == MESA_SHADER_TESS_CTRL ||
+	    ctx->stage == MESA_SHADER_GEOMETRY)
+		return;
+
+	if (ctx->stage == MESA_SHADER_VERTEX) {
+		if (ctx->options->key.vs.as_ls ||
+		    ctx->options->key.vs.as_es)
+			return;
+		outinfo = &ctx->shader_info->vs.outinfo;
+	}
+
+	if (ctx->stage == MESA_SHADER_TESS_EVAL) {
+		if (ctx->options->key.vs.as_es)
+			return;
+		outinfo = &ctx->shader_info->tes.outinfo;
+	}
+
+	ac_eliminate_const_vs_outputs(&ctx->ac,
+				      ctx->main_function,
+				      outinfo->vs_output_param_offset,
+				      VARYING_SLOT_MAX,
+				      &outinfo->param_exports);
+}
+
+static void
 ac_setup_rings(struct nir_to_llvm_context *ctx)
 {
 	if ((ctx->stage == MESA_SHADER_VERTEX && ctx->options->key.vs.as_es) ||
@@ -5894,6 +5927,8 @@ LLVMModuleRef ac_translate_nir_to_llvm(LLVMTargetMachineRef tm,
 	LLVMBuildRetVoid(ctx.builder);
 
 	ac_llvm_finalize_module(&ctx);
+
+	ac_nir_eliminate_const_vs_outputs(&ctx);
 	free(ctx.locals);
 	ralloc_free(ctx.defs);
 	ralloc_free(ctx.phis);
diff --git a/src/amd/common/ac_nir_to_llvm.h b/src/amd/common/ac_nir_to_llvm.h
index f77a9b8d2bb..9ea3156b255 100644
--- a/src/amd/common/ac_nir_to_llvm.h
+++ b/src/amd/common/ac_nir_to_llvm.h
@@ -120,27 +120,15 @@ struct ac_userdata_locations {
 	struct ac_userdata_info shader_data[AC_UD_MAX_UD];
 };
 
-enum {
-	/* SPI_PS_INPUT_CNTL_i.OFFSET[0:4] */
-	EXP_PARAM_OFFSET_0 = 0,
-	EXP_PARAM_OFFSET_31 = 31,
-	/* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL[0:1] */
-	EXP_PARAM_DEFAULT_VAL_0000 = 64,
-	EXP_PARAM_DEFAULT_VAL_0001,
-	EXP_PARAM_DEFAULT_VAL_1110,
-	EXP_PARAM_DEFAULT_VAL_1111,
-	EXP_PARAM_UNDEFINED = 255,
-};
-
 struct ac_vs_output_info {
 	uint8_t	vs_output_param_offset[VARYING_SLOT_MAX];
 	uint8_t clip_dist_mask;
 	uint8_t cull_dist_mask;
+	uint8_t param_exports;
 	bool writes_pointsize;
 	bool writes_layer;
 	bool writes_viewport_index;
 	uint32_t export_mask;
-	unsigned param_exports;
 	unsigned pos_exports;
 };
 
diff --git a/src/amd/vulkan/radv_pipeline.c b/src/amd/vulkan/radv_pipeline.c
index d6989137a55..ce18178a9df 100644
--- a/src/amd/vulkan/radv_pipeline.c
+++ b/src/amd/vulkan/radv_pipeline.c
@@ -41,6 +41,7 @@
 #include "ac_nir_to_llvm.h"
 #include "vk_format.h"
 #include "util/debug.h"
+#include "ac_exp_param.h"
 
 void radv_shader_variant_destroy(struct radv_device *device,
                                  struct radv_shader_variant *variant);
@@ -1874,13 +1875,13 @@ static void calculate_pa_cl_vs_out_cntl(struct radv_pipeline *pipeline)
 static uint32_t offset_to_ps_input(uint32_t offset, bool flat_shade)
 {
 	uint32_t ps_input_cntl;
-	if (offset <= EXP_PARAM_OFFSET_31)
+	if (offset <= AC_EXP_PARAM_OFFSET_31)
 		ps_input_cntl = S_028644_OFFSET(offset);
 	else {
 		/* The input is a DEFAULT_VAL constant. */
-		assert(offset >= EXP_PARAM_DEFAULT_VAL_0000 &&
-		       offset <= EXP_PARAM_DEFAULT_VAL_1111);
-		offset -= EXP_PARAM_DEFAULT_VAL_0000;
+		assert(offset >= AC_EXP_PARAM_DEFAULT_VAL_0000 &&
+		       offset <= AC_EXP_PARAM_DEFAULT_VAL_1111);
+		offset -= AC_EXP_PARAM_DEFAULT_VAL_0000;
 		ps_input_cntl = S_028644_OFFSET(0x20) |
 			S_028644_DEFAULT_VAL(offset);
 	}
@@ -1903,7 +1904,7 @@ static void calculate_ps_inputs(struct radv_pipeline *pipeline)
 
 	if (ps->info.fs.prim_id_input) {
 		unsigned vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_PRIMITIVE_ID];
-		if (vs_offset != EXP_PARAM_UNDEFINED) {
+		if (vs_offset != AC_EXP_PARAM_UNDEFINED) {
 			pipeline->graphics.ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, true);
 			++ps_offset;
 		}
@@ -1911,7 +1912,7 @@ static void calculate_ps_inputs(struct radv_pipeline *pipeline)
 
 	if (ps->info.fs.layer_input) {
 		unsigned vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_LAYER];
-		if (vs_offset != EXP_PARAM_UNDEFINED) {
+		if (vs_offset != AC_EXP_PARAM_UNDEFINED) {
 			pipeline->graphics.ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, true);
 			++ps_offset;
 		}
@@ -1931,7 +1932,7 @@ static void calculate_ps_inputs(struct radv_pipeline *pipeline)
 			continue;
 
 		vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_VAR0 + i];
-		if (vs_offset == EXP_PARAM_UNDEFINED) {
+		if (vs_offset == AC_EXP_PARAM_UNDEFINED) {
 			pipeline->graphics.ps_input_cntl[ps_offset] = S_028644_OFFSET(0x20);
 			++ps_offset;
 			continue;