[Mesa-dev] [PATCH 3/5] radeonsi: set the convergent attribute where needed

[Marek Olšák]

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

---
 src/amd/common/ac_llvm_build.c           | 8 ++++++--
 src/gallium/drivers/radeonsi/si_shader.c | 2 +-
 2 files changed, 7 insertions(+), 3 deletions(-)

diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c
index 20c6eb8..0039e63 100644
--- a/src/amd/common/ac_llvm_build.c
+++ b/src/amd/common/ac_llvm_build.c
@@ -854,27 +854,31 @@ ac_build_ddxy(struct ac_llvm_context *ctx,
 
 	trbl_tid = LLVMBuildAdd(ctx->builder, tl_tid,
 				LLVMConstInt(ctx->i32, idx, false), "");
 
 	if (has_ds_bpermute) {
 		args[0] = LLVMBuildMul(ctx->builder, tl_tid,
 				       LLVMConstInt(ctx->i32, 4, false), "");
 		args[1] = val;
 		tl = ac_build_intrinsic(ctx,
 					"llvm.amdgcn.ds.bpermute", ctx->i32,
-					args, 2, AC_FUNC_ATTR_READNONE);
+					args, 2,
+					AC_FUNC_ATTR_READNONE |
+					AC_FUNC_ATTR_CONVERGENT);
 
 		args[0] = LLVMBuildMul(ctx->builder, trbl_tid,
 				       LLVMConstInt(ctx->i32, 4, false), "");
 		trbl = ac_build_intrinsic(ctx,
 					  "llvm.amdgcn.ds.bpermute", ctx->i32,
-					  args, 2, AC_FUNC_ATTR_READNONE);
+					  args, 2,
+					  AC_FUNC_ATTR_READNONE |
+					  AC_FUNC_ATTR_CONVERGENT);
 	} else {
 		LLVMValueRef store_ptr, load_ptr0, load_ptr1;
 
 		store_ptr = ac_build_gep0(ctx, lds, thread_id);
 		load_ptr0 = ac_build_gep0(ctx, lds, tl_tid);
 		load_ptr1 = ac_build_gep0(ctx, lds, trbl_tid);
 
 		LLVMBuildStore(ctx->builder, val, store_ptr);
 		tl = LLVMBuildLoad(ctx->builder, load_ptr0, "");
 		trbl = LLVMBuildLoad(ctx->builder, load_ptr1, "");
diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index 4e133ec..edfa284 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -5200,21 +5200,21 @@ static void si_llvm_emit_barrier(const struct lp_build_tgsi_action *action,
 	if (HAVE_LLVM >= 0x0309 &&
 	    ctx->screen->b.chip_class == SI &&
 	    ctx->type == PIPE_SHADER_TESS_CTRL) {
 		emit_waitcnt(ctx, LGKM_CNT & VM_CNT);
 		return;
 	}
 
 	lp_build_intrinsic(gallivm->builder,
 			   HAVE_LLVM >= 0x0309 ? "llvm.amdgcn.s.barrier"
 					       : "llvm.AMDGPU.barrier.local",
-			   ctx->voidt, NULL, 0, 0);
+			   ctx->voidt, NULL, 0, LP_FUNC_ATTR_CONVERGENT);
 }
 
 static const struct lp_build_tgsi_action tex_action = {
 	.fetch_args = tex_fetch_args,
 	.emit = build_tex_intrinsic,
 };
 
 static const struct lp_build_tgsi_action interp_action = {
 	.fetch_args = interp_fetch_args,
 	.emit = build_interp_intrinsic,
-- 
2.7.4