[Mesa-dev] [PATCH 6/6] ac: use llvm.amdgcn.s.buffer.load

[Marek Olšák]

From: Marek Olšák <marek.olsak at amd.com>

---
 src/amd/common/ac_llvm_build.c                 | 18 +++++++++++++-----
 .../drivers/radeonsi/si_shader_tgsi_mem.c      |  4 ++--
 2 files changed, 15 insertions(+), 7 deletions(-)

diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c
index 76047148a6a..c0d90ada2be 100644
--- a/src/amd/common/ac_llvm_build.c
+++ b/src/amd/common/ac_llvm_build.c
@@ -1234,25 +1234,33 @@ ac_build_buffer_load(struct ac_llvm_context *ctx,
 	if (allow_smem && !glc && !slc) {
 		assert(vindex == NULL);
 
 		LLVMValueRef result[8];
 
 		for (int i = 0; i < num_channels; i++) {
 			if (i) {
 				offset = LLVMBuildAdd(ctx->builder, offset,
 						      LLVMConstInt(ctx->i32, 4, 0), "");
 			}
-			LLVMValueRef args[2] = {rsrc, offset};
-			result[i] = ac_build_intrinsic(ctx, "llvm.SI.load.const.v4i32",
-						       ctx->f32, args, 2,
-						       AC_FUNC_ATTR_READNONE |
-						       AC_FUNC_ATTR_LEGACY);
+
+			if (HAVE_LLVM >= 0x0800) {
+				LLVMValueRef args[3] = {rsrc, offset, ctx->i32_0};
+				result[i] = ac_build_intrinsic(ctx, "llvm.amdgcn.s.buffer.load.i32",
+							       ctx->f32, args, 3,
+							       AC_FUNC_ATTR_READNONE);
+			} else {
+				LLVMValueRef args[2] = {rsrc, offset};
+				result[i] = ac_build_intrinsic(ctx, "llvm.SI.load.const.v4i32",
+							       ctx->f32, args, 2,
+							       AC_FUNC_ATTR_READNONE |
+							       AC_FUNC_ATTR_LEGACY);
+			}
 		}
 		if (num_channels == 1)
 			return result[0];
 
 		if (num_channels == 3)
 			result[num_channels++] = LLVMGetUndef(ctx->f32);
 		return ac_build_gather_values(ctx, result, num_channels);
 	}
 
 	return ac_build_buffer_load_common(ctx, rsrc, vindex, offset,
diff --git a/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c b/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c
index 727def56f65..2f49685c642 100644
--- a/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c
+++ b/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c
@@ -533,24 +533,24 @@ static void load_emit(
 						info->images_store |
 						info->images_atomic,
 						info->uses_bindless_buffer_store |
 						info->uses_bindless_buffer_atomic,
 						info->uses_bindless_image_store |
 						info->uses_bindless_image_atomic);
 	args.cache_policy = get_cache_policy(ctx, inst, false, false, false);
 
 	if (inst->Src[0].Register.File == TGSI_FILE_BUFFER) {
 		/* Don't use SMEM for shader buffer loads, because LLVM doesn't
-		 * select SMEM for SI.load.const with a non-constant offset, and
+		 * select SMEM for amdgcn.s.buffer.load with a non-constant offset, and
 		 * constant offsets practically don't exist with shader buffers.
 		 *
-		 * Also, SI.load.const doesn't use inst_offset when it's lowered
+		 * Also, amdgcn.s.buffer.load doesn't use inst_offset when it's lowered
 		 * to VMEM, so we just end up with more VALU instructions in the end
 		 * and no benefit.
 		 *
 		 * TODO: Remove this line once LLVM can select SMEM with a non-constant
 		 *       offset, and can derive inst_offset when VMEM is selected.
 		 *       After that, si_memory_barrier should invalidate sL1 for shader
 		 *       buffers.
 		 */
 		emit_data->output[emit_data->chan] =
 			ac_build_buffer_load(&ctx->ac, args.resource,
-- 
2.17.1