radeonsi: move PS LLVM code into si_shader_llvm_ps.c

This is an attempt to clean up si_shader.c.

v2: don't move code that is not specific to LLVM

Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com> (v1)

src/gallium/drivers/radeonsi/Makefile.sources

@@ -37,6 +37,7 @@ C_SOURCES := \
 	si_shader_internal.h \
 	si_shader_llvm.c \
 	si_shader_llvm_build.c \
+	si_shader_llvm_ps.c \
 	si_shader_nir.c \
 	si_shaderlib_tgsi.c \
 	si_state.c \

src/gallium/drivers/radeonsi/meson.build

@@ -52,6 +52,7 @@ files_libradeonsi = files(
   'si_shader_internal.h',
   'si_shader_llvm.c',
   'si_shader_llvm_build.c',
+  'si_shader_llvm_ps.c',
   'si_shader_nir.c',
   'si_shaderlib_tgsi.c',
   'si_state.c',

src/gallium/drivers/radeonsi/si_shader_internal.h

@@ -35,6 +35,11 @@ struct pipe_debug_callback;
 
 #define RADEON_LLVM_MAX_INPUTS 32 * 4
 
+/* Ideally pass the sample mask input to the PS epilog as v14, which
+ * is its usual location, so that the shader doesn't have to add v_mov.
+ */
+#define PS_EPILOG_SAMPLEMASK_MIN_LOC 14
+
 struct si_shader_output_values {
 	LLVMValueRef values[4];
 	unsigned semantic_name;
@@ -235,9 +240,7 @@ LLVMValueRef si_nir_load_input_tes(struct ac_shader_abi *abi,
 				   bool is_patch,
 				   bool is_compact,
 				   bool load_input);
-LLVMValueRef si_nir_lookup_interp_param(struct ac_shader_abi *abi,
-					enum glsl_interp_mode interp,
-					unsigned location);
+bool si_is_merged_shader(struct si_shader_context *ctx);
 LLVMValueRef si_get_sample_id(struct si_shader_context *ctx);
 LLVMValueRef si_load_sampler_desc(struct si_shader_context *ctx,
 				  LLVMValueRef list, LLVMValueRef index,
@@ -246,7 +249,10 @@ LLVMValueRef si_load_image_desc(struct si_shader_context *ctx,
 				LLVMValueRef list, LLVMValueRef index,
 				enum ac_descriptor_type desc_type,
 				bool uses_store, bool bindless);
-LLVMValueRef si_nir_emit_fbfetch(struct ac_shader_abi *abi);
+LLVMValueRef si_buffer_load_const(struct si_shader_context *ctx,
+				  LLVMValueRef resource, LLVMValueRef offset);
+void si_llvm_build_ret(struct si_shader_context *ctx, LLVMValueRef ret);
+LLVMValueRef si_prolog_get_rw_buffers(struct si_shader_context *ctx);
 void si_declare_compute_memory(struct si_shader_context *ctx);
 LLVMValueRef si_get_primitive_id(struct si_shader_context *ctx,
 				 unsigned swizzle);
@@ -258,6 +264,11 @@ void si_emit_streamout_output(struct si_shader_context *ctx,
 			      LLVMValueRef const *so_write_offsets,
 			      struct pipe_stream_output *stream_out,
 			      struct si_shader_output_values *shader_out);
+void si_add_arg_checked(struct ac_shader_args *args,
+			enum ac_arg_regfile file,
+			unsigned registers, enum ac_arg_type type,
+			struct ac_arg *arg,
+			unsigned idx);
 
 void si_llvm_load_input_vs(
 	struct si_shader_context *ctx,
@@ -271,6 +282,15 @@ LLVMValueRef si_unpack_param(struct si_shader_context *ctx,
 			     unsigned bitwidth);
 LLVMValueRef si_is_es_thread(struct si_shader_context *ctx);
 LLVMValueRef si_is_gs_thread(struct si_shader_context *ctx);
+void si_build_wrapper_function(struct si_shader_context *ctx, LLVMValueRef *parts,
+			       unsigned num_parts, unsigned main_part,
+			       unsigned next_shader_first_part);
+bool si_need_ps_prolog(const union si_shader_part_key *key);
+void si_get_ps_prolog_key(struct si_shader *shader,
+			  union si_shader_part_key *key,
+			  bool separate_prolog);
+void si_get_ps_epilog_key(struct si_shader *shader,
+			  union si_shader_part_key *key);
 
 void gfx10_emit_ngg_epilogue(struct ac_shader_abi *abi,
 			     unsigned max_outputs,
@@ -282,4 +302,13 @@ void gfx10_ngg_gs_emit_prologue(struct si_shader_context *ctx);
 void gfx10_ngg_gs_emit_epilogue(struct si_shader_context *ctx);
 void gfx10_ngg_calculate_subgroup_info(struct si_shader *shader);
 
+/* si_shader_llvm_ps.c */
+void si_llvm_build_ps_prolog(struct si_shader_context *ctx,
+			     union si_shader_part_key *key);
+void si_llvm_build_ps_epilog(struct si_shader_context *ctx,
+			     union si_shader_part_key *key);
+void si_llvm_build_monolithic_ps(struct si_shader_context *ctx,
+				 struct si_shader *shader);
+void si_llvm_init_ps_callbacks(struct si_shader_context *ctx);
+
 #endif

src/gallium/drivers/radeonsi/si_shader_llvm_build.c

@@ -159,61 +159,31 @@ LLVMValueRef si_load_sampler_desc(struct si_shader_context *ctx,
 	return ac_build_load_to_sgpr(&ctx->ac, list, index);
 }
 
-LLVMValueRef si_nir_emit_fbfetch(struct ac_shader_abi *abi)
+/**
+ * Load a dword from a constant buffer.
+ */
+LLVMValueRef si_buffer_load_const(struct si_shader_context *ctx,
+				  LLVMValueRef resource, LLVMValueRef offset)
 {
-	struct si_shader_context *ctx = si_shader_context_from_abi(abi);
-	struct ac_image_args args = {};
-	LLVMValueRef ptr, image, fmask;
-
-	/* Ignore src0, because KHR_blend_func_extended disallows multiple render
-	 * targets.
-	 */
-
-	/* Load the image descriptor. */
-	STATIC_ASSERT(SI_PS_IMAGE_COLORBUF0 % 2 == 0);
-	ptr = ac_get_arg(&ctx->ac, ctx->rw_buffers);
-	ptr = LLVMBuildPointerCast(ctx->ac.builder, ptr,
-				   ac_array_in_const32_addr_space(ctx->v8i32), "");
-	image = ac_build_load_to_sgpr(&ctx->ac, ptr,
-			LLVMConstInt(ctx->i32, SI_PS_IMAGE_COLORBUF0 / 2, 0));
-
-	unsigned chan = 0;
-
-	args.coords[chan++] = si_unpack_param(ctx, ctx->pos_fixed_pt, 0, 16);
-
-	if (!ctx->shader->key.mono.u.ps.fbfetch_is_1D)
-		args.coords[chan++] = si_unpack_param(ctx, ctx->pos_fixed_pt, 16, 16);
-
-	/* Get the current render target layer index. */
-	if (ctx->shader->key.mono.u.ps.fbfetch_layered)
-		args.coords[chan++] = si_unpack_param(ctx, ctx->args.ancillary, 16, 11);
-
-	if (ctx->shader->key.mono.u.ps.fbfetch_msaa)
-		args.coords[chan++] = si_get_sample_id(ctx);
-
-	if (ctx->shader->key.mono.u.ps.fbfetch_msaa &&
-	    !(ctx->screen->debug_flags & DBG(NO_FMASK))) {
-		fmask = ac_build_load_to_sgpr(&ctx->ac, ptr,
-			LLVMConstInt(ctx->i32, SI_PS_IMAGE_COLORBUF0_FMASK / 2, 0));
-
-		ac_apply_fmask_to_sample(&ctx->ac, fmask, args.coords,
-					 ctx->shader->key.mono.u.ps.fbfetch_layered);
-	}
-
-	args.opcode = ac_image_load;
-	args.resource = image;
-	args.dmask = 0xf;
-	args.attributes = AC_FUNC_ATTR_READNONE;
-
-	if (ctx->shader->key.mono.u.ps.fbfetch_msaa)
-		args.dim = ctx->shader->key.mono.u.ps.fbfetch_layered ?
-			ac_image_2darraymsaa : ac_image_2dmsaa;
-	else if (ctx->shader->key.mono.u.ps.fbfetch_is_1D)
-		args.dim = ctx->shader->key.mono.u.ps.fbfetch_layered ?
-			ac_image_1darray : ac_image_1d;
-	else
-		args.dim = ctx->shader->key.mono.u.ps.fbfetch_layered ?
-			ac_image_2darray : ac_image_2d;
-
-	return ac_build_image_opcode(&ctx->ac, &args);
+	return ac_build_buffer_load(&ctx->ac, resource, 1, NULL, offset, NULL,
+				    0, 0, true, true);
+}
+
+void si_llvm_build_ret(struct si_shader_context *ctx, LLVMValueRef ret)
+{
+	if (LLVMGetTypeKind(LLVMTypeOf(ret)) == LLVMVoidTypeKind)
+		LLVMBuildRetVoid(ctx->ac.builder);
+	else
+		LLVMBuildRet(ctx->ac.builder, ret);
+}
+
+LLVMValueRef si_prolog_get_rw_buffers(struct si_shader_context *ctx)
+{
+	LLVMValueRef ptr[2], list;
+	bool merged_shader = si_is_merged_shader(ctx);
+
+	ptr[0] = LLVMGetParam(ctx->main_fn, (merged_shader ? 8 : 0) + SI_SGPR_RW_BUFFERS);
+	list = LLVMBuildIntToPtr(ctx->ac.builder, ptr[0],
+				 ac_array_in_const32_addr_space(ctx->v4i32), "");
+	return list;
 }

src/gallium/drivers/radeonsi/si_shader_nir.c

@@ -1005,38 +1005,6 @@ static void declare_nir_input_vs(struct si_shader_context *ctx,
 	si_llvm_load_input_vs(ctx, input_index, out);
 }
 
-LLVMValueRef
-si_nir_lookup_interp_param(struct ac_shader_abi *abi,
-			   enum glsl_interp_mode interp, unsigned location)
-{
-	struct si_shader_context *ctx = si_shader_context_from_abi(abi);
-
-	switch (interp) {
-	case INTERP_MODE_FLAT:
-		return NULL;
-	case INTERP_MODE_SMOOTH:
-	case INTERP_MODE_NONE:
-		if (location == INTERP_CENTER)
-			return ac_get_arg(&ctx->ac, ctx->args.persp_center);
-		else if (location == INTERP_CENTROID)
-			return ctx->abi.persp_centroid;
-		else if (location == INTERP_SAMPLE)
-			return ac_get_arg(&ctx->ac, ctx->args.persp_sample);
-		break;
-	case INTERP_MODE_NOPERSPECTIVE:
-		if (location == INTERP_CENTER)
-			return ac_get_arg(&ctx->ac, ctx->args.linear_center);
-		else if (location == INTERP_CENTROID)
-			return ac_get_arg(&ctx->ac, ctx->args.linear_centroid);
-		else if (location == INTERP_SAMPLE)
-			return ac_get_arg(&ctx->ac, ctx->args.linear_sample);
-		break;
-	default:
-		assert(!"Unhandled interpolation mode.");
-	}
-	return NULL;
-}
-
 static LLVMValueRef
 si_nir_load_sampler_desc(struct ac_shader_abi *abi,
 		         unsigned descriptor_set, unsigned base_index,