r600g/llvm: rework handling of the constants

Vincent Lejeune:
  - tgsi to llvm now emits pointers for constants

Tom Stellard:
  - Only use texture cache for vtx fetch with compute shaders
  - Change address space used for constant loads to match LLVM
    backend.

Reviewed-by: Tom Stellard <[email protected]>

3 files changed, 54 insertions, 16 deletions

diff --git a/src/gallium/drivers/r600/r600_llvm.c b/src/gallium/drivers/r600/r600_llvm.c
index 17d362c0aaf..9e31171c711 100644
--- a/src/gallium/drivers/r600/r600_llvm.c
+++ b/src/gallium/drivers/r600/r600_llvm.c
@@ -19,18 +19,29 @@
 
 #if defined R600_USE_LLVM || defined HAVE_OPENCL
 
+#define CONSTANT_BUFFER_0_ADDR_SPACE 9
+
 static LLVMValueRef llvm_fetch_const(
 	struct lp_build_tgsi_context * bld_base,
 	const struct tgsi_full_src_register *reg,
 	enum tgsi_opcode_type type,
 	unsigned swizzle)
 {
-	LLVMValueRef idx = lp_build_const_int32(bld_base->base.gallivm,
-			radeon_llvm_reg_index_soa(reg->Register.Index, swizzle));
-	LLVMValueRef cval = build_intrinsic(bld_base->base.gallivm->builder,
-		"llvm.AMDGPU.load.const", bld_base->base.elem_type,
-		&idx, 1, LLVMReadNoneAttribute);
-
+	LLVMValueRef offset[2] = {
+		LLVMConstInt(LLVMInt64TypeInContext(bld_base->base.gallivm->context), 0, false),
+		lp_build_const_int32(bld_base->base.gallivm, reg->Register.Index)
+	};
+	if (reg->Register.Indirect) {
+		struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
+		LLVMValueRef index = LLVMBuildLoad(bld_base->base.gallivm->builder, bld->addr[reg->Indirect.Index][reg->Indirect.SwizzleX], "");
+		offset[1] = LLVMBuildAdd(bld_base->base.gallivm->builder, offset[1], index, "");
+	}
+	LLVMTypeRef const_ptr_type = LLVMPointerType(LLVMArrayType(LLVMVectorType(bld_base->base.elem_type, 4), 1024),
+							CONSTANT_BUFFER_0_ADDR_SPACE);
+	LLVMValueRef const_ptr = LLVMBuildIntToPtr(bld_base->base.gallivm->builder, lp_build_const_int32(bld_base->base.gallivm, 0), const_ptr_type, "");
+	LLVMValueRef ptr = LLVMBuildGEP(bld_base->base.gallivm->builder, const_ptr, offset, 2, "");
+	LLVMValueRef cvecval = LLVMBuildLoad(bld_base->base.gallivm->builder, ptr, "");
+	LLVMValueRef cval = LLVMBuildExtractElement(bld_base->base.gallivm->builder, cvecval, lp_build_const_int32(bld_base->base.gallivm, swizzle), "");
 	return bitcast(bld_base, type, cval);
 }
 
diff --git a/src/gallium/drivers/r600/r600_shader.c b/src/gallium/drivers/r600/r600_shader.c
index 83077504385..410ffce28ba 100644
--- a/src/gallium/drivers/r600/r600_shader.c
+++ b/src/gallium/drivers/r600/r600_shader.c
@@ -309,15 +309,21 @@ static unsigned r600_src_from_byte_stream(unsigned char * bytes,
 static unsigned r600_alu_from_byte_stream(struct r600_shader_ctx *ctx,
 				unsigned char * bytes, unsigned bytes_read)
 {
-	unsigned src_idx;
+	unsigned src_idx, src_num;
 	struct r600_bytecode_alu alu;
-	unsigned src_const_reg[3];
+	unsigned src_use_sel[3];
+	unsigned src_sel[3] = {};
 	uint32_t word0, word1;
 
+	src_num = bytes[bytes_read++];
+
 	memset(&alu, 0, sizeof(alu));
-	for(src_idx = 0; src_idx < 3; src_idx++) {
+	for(src_idx = 0; src_idx < src_num; src_idx++) {
 		unsigned i;
-		src_const_reg[src_idx] = bytes[bytes_read++];
+		src_use_sel[src_idx] = bytes[bytes_read++];
+		for (i = 0; i < 4; i++) {
+			src_sel[src_idx] |= bytes[bytes_read++] << (i * 8);
+		}
 		for (i = 0; i < 4; i++) {
 			alu.src[src_idx].value |= bytes[bytes_read++] << (i * 8);
 		}
@@ -338,9 +344,22 @@ static unsigned r600_alu_from_byte_stream(struct r600_shader_ctx *ctx,
 		break;
 	}
 
-	for(src_idx = 0; src_idx < 3; src_idx++) {
-		if (src_const_reg[src_idx])
-			alu.src[src_idx].sel += 512;
+	for(src_idx = 0; src_idx < src_num; src_idx++) {
+		if (src_use_sel[src_idx]) {
+			unsigned sel = src_sel[src_idx];
+
+			alu.src[src_idx].chan = sel & 3;
+			sel >>= 2;
+
+			if (sel>=512) { /* constant */
+				sel -= 512;
+				alu.src[src_idx].kc_bank = sel >> 12;
+				alu.src[src_idx].sel = (sel & 4095) + 512;
+			}
+			else {
+				alu.src[src_idx].sel = sel;
+			}
+		}
 	}
 
 #if HAVE_LLVM < 0x0302
@@ -521,8 +540,12 @@ static int r600_vtx_from_byte_stream(struct r600_shader_ctx *ctx,
 	if (r600_bytecode_add_vtx(ctx->bc, &vtx)) {
 		fprintf(stderr, "Error adding vtx\n");
 	}
-	/* Use the Texture Cache */
-	ctx->bc->cf_last->inst = EG_V_SQ_CF_WORD1_SQ_CF_INST_TEX;
+
+	/* Use the Texture Cache for compute shaders*/
+	if (ctx->bc->chip_class >= EVERGREEN &&
+		ctx->bc->type == TGSI_PROCESSOR_COMPUTE) {
+		ctx->bc->cf_last->inst = EG_V_SQ_CF_WORD1_SQ_CF_INST_TEX;
+	}
 	return bytes_read;
 }
 
@@ -1268,7 +1291,7 @@ static int r600_shader_from_tgsi(struct r600_screen *rscreen,
 	}
 
 #ifdef R600_USE_LLVM
-	if (use_llvm && ctx.info.indirect_files) {
+	if (use_llvm && ctx.info.indirect_files && (ctx.info.indirect_files & (1 << TGSI_FILE_CONSTANT)) != ctx.info.indirect_files) {
 		fprintf(stderr, "Warning: R600 LLVM backend does not support "
 				"indirect adressing.  Falling back to TGSI "
 				"backend.\n");
diff --git a/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c b/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
index d390f927a62..258e796595a 100644
--- a/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
+++ b/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
@@ -324,6 +324,10 @@ emit_store(
 		}
 
 		switch(reg->Register.File) {
+		case TGSI_FILE_ADDRESS:
+			temp_ptr = bld->addr[reg->Register.Index][chan_index];
+			LLVMBuildStore(builder, value, temp_ptr);
+			continue;
 		case TGSI_FILE_OUTPUT:
 			temp_ptr = bld->outputs[reg->Register.Index][chan_index];
 			break;