[llvm] AMDGPU: Relax shouldCoalesce to allow more register tuple widening (PR #166475)

Valery Pykhtin via llvm-commits llvm-commits at lists.llvm.org
Wed Nov 5 05:22:24 PST 2025


================
@@ -6,9 +6,9 @@ define amdgpu_kernel void @foo() {
 ; CHECK:       ; %bb.0: ; %entry
 ; CHECK-NEXT:    s_mov_b64 s[0:1], src_shared_base
 ; CHECK-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; CHECK-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s1
-; CHECK-NEXT:    v_dual_mov_b32 v2, v0 :: v_dual_mov_b32 v3, v0
-; CHECK-NEXT:    flat_store_b64 v[0:1], v[2:3]
+; CHECK-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s1
+; CHECK-NEXT:    v_mov_b32_e32 v0, v1
+; CHECK-NEXT:    flat_store_b64 v[1:2], v[0:1]
----------------
vpykhtin wrote:

good, less instructions and registers used!

https://github.com/llvm/llvm-project/pull/166475


More information about the llvm-commits mailing list