Skip to content

Commit 5b1a158

Browse files
committed
Simplify and correct ll test
- Must use -mtriple to reproduce the bug on the unfixed branch - Function does not need to be a kernel
1 parent d68738f commit 5b1a158

File tree

1 file changed

+2
-7
lines changed

1 file changed

+2
-7
lines changed

llvm/test/CodeGen/AMDGPU/sdwa-peephole-cndmask-fail.ll

Lines changed: 2 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -1,17 +1,12 @@
1-
; RUN: llc -march=amdgcn -mcpu=gfx1030 < %s | FileCheck %s
1+
; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1030 < %s | FileCheck %s
22

33
; In this test, V_CNDMASK_B32_e64 gets converted to V_CNDMASK_B32_e32,
44
; but the expected conversion to SDWA does not occur. This led to a
55
; compilation error, because the use of $vcc in the resulting
66
; instruction must be fixed to $vcc_lo for wave32 which only happened
77
; after the full conversion to SDWA.
88

9-
10-
; CHECK-NOT: {{.*}}V_CNDMASK_B32_e32{{.*}}$vcc
11-
; CHECK-NOT: {{.*}}Bad machine code: Virtual register defs don't dominate all uses
12-
; CHECK: {{.*}}v_cndmask_b32_e32{{.*}}vcc_lo
13-
14-
define amdgpu_kernel void @quux(i32 %arg, i1 %arg1, i1 %arg2) {
9+
define void @quux(i32 %arg, i1 %arg1, i1 %arg2) {
1510
bb:
1611
br i1 %arg1, label %bb9, label %bb3
1712

0 commit comments

Comments
 (0)