[AMDGPU] fix the check of inst-order in a test

[cmc-rep]

test was not well written to consider potential all-correct instruction ordering

[llvmbot]

@llvm/pr-subscribers-backend-amdgpu

Author: Gang Chen (cmc-rep)

Changes

test was not well written to consider potential all-correct instruction ordering

---

Full diff: https://github.com/llvm/llvm-project/pull/90744.diff

1 Files Affected:

• (modified) llvm/test/CodeGen/AMDGPU/preserve-wwm-copy-dst-reg.ll (+18-18)

diff --git a/llvm/test/CodeGen/AMDGPU/preserve-wwm-copy-dst-reg.ll b/llvm/test/CodeGen/AMDGPU/preserve-wwm-copy-dst-reg.ll
index 1be041c8dc9b0b..160f5785dbb0dd 100644
--- a/llvm/test/CodeGen/AMDGPU/preserve-wwm-copy-dst-reg.ll
+++ b/llvm/test/CodeGen/AMDGPU/preserve-wwm-copy-dst-reg.ll
@@ -31,15 +31,15 @@ define void @preserve_wwm_copy_dstreg(ptr %parg0, ptr %parg1, ptr %parg2) #0 {
 ; GFX906-NEXT:    s_mov_b32 s24, s12
 ; GFX906-NEXT:    v_writelane_b32 v2, s24, 5
 ; GFX906-NEXT:    s_mov_b64 s[26:27], s[10:11]
-; GFX906-NEXT:    v_writelane_b32 v2, s26, 6
-; GFX906-NEXT:    v_writelane_b32 v41, s34, 2
-; GFX906-NEXT:    v_writelane_b32 v2, s27, 7
-; GFX906-NEXT:    v_writelane_b32 v41, s35, 3
-; GFX906-NEXT:    v_writelane_b32 v2, s8, 8
-; GFX906-NEXT:    v_writelane_b32 v41, s16, 4
-; GFX906-NEXT:    v_writelane_b32 v2, s9, 9
-; GFX906-NEXT:    v_writelane_b32 v41, s30, 0
-; GFX906-NEXT:    v_writelane_b32 v2, s4, 10
+; GFX906-DAG:     v_writelane_b32 v2, s26, 6
+; GFX906-DAG:     v_writelane_b32 v41, s34, 2
+; GFX906-DAG:     v_writelane_b32 v2, s27, 7
+; GFX906-DAG:     v_writelane_b32 v41, s35, 3
+; GFX906-DAG:     v_writelane_b32 v2, s8, 8
+; GFX906-DAG:     v_writelane_b32 v41, s16, 4
+; GFX906-DAG:     v_writelane_b32 v2, s9, 9
+; GFX906-DAG:     v_writelane_b32 v41, s30, 0
+; GFX906-DAG:     v_writelane_b32 v2, s4, 10
 ; GFX906-NEXT:    s_addk_i32 s32, 0x2800
 ; GFX906-NEXT:    v_writelane_b32 v41, s31, 1
 ; GFX906-NEXT:    v_mov_b32_e32 v32, v31
@@ -340,9 +340,9 @@ define void @preserve_wwm_copy_dstreg(ptr %parg0, ptr %parg1, ptr %parg2) #0 {
 ; GFX906-NEXT:    v_readlane_b32 s31, v41, 1
 ; GFX906-NEXT:    v_readlane_b32 s30, v41, 0
 ; GFX906-NEXT:    ; kill: killed $vgpr40
-; GFX906-NEXT:    v_readlane_b32 s34, v41, 2
-; GFX906-NEXT:    v_readlane_b32 s35, v41, 3
-; GFX906-NEXT:    v_readlane_b32 s4, v41, 4
+; GFX906-DAG:     v_readlane_b32 s34, v41, 2
+; GFX906-DAG:     v_readlane_b32 s35, v41, 3
+; GFX906-DAG:     v_readlane_b32 s4, v41, 4
 ; GFX906-NEXT:    s_waitcnt vmcnt(0)
 ; GFX906-NEXT:    flat_store_dwordx4 v[0:1], v[30:33] offset:112
 ; GFX906-NEXT:    s_waitcnt vmcnt(0)
@@ -383,12 +383,12 @@ define void @preserve_wwm_copy_dstreg(ptr %parg0, ptr %parg1, ptr %parg2) #0 {
 ; GFX908-NEXT:    s_mov_b64 exec, -1
 ; GFX908-NEXT:    buffer_store_dword v40, off, s[0:3], s33 offset:152 ; 4-byte Folded Spill
 ; GFX908-NEXT:    s_mov_b64 exec, s[18:19]
-; GFX908-NEXT:    v_mov_b32_e32 v3, s34
-; GFX908-NEXT:    buffer_store_dword v3, off, s[0:3], s33 offset:160 ; 4-byte Folded Spill
-; GFX908-NEXT:    v_mov_b32_e32 v3, s35
-; GFX908-NEXT:    buffer_store_dword v3, off, s[0:3], s33 offset:164 ; 4-byte Folded Spill
-; GFX908-NEXT:    v_mov_b32_e32 v3, s16
-; GFX908-NEXT:    buffer_store_dword v3, off, s[0:3], s33 offset:168 ; 4-byte Folded Spill
+; GFX908-DAG:     v_mov_b32_e32 v3, s34
+; GFX908-DAG:     buffer_store_dword v3, off, s[0:3], s33 offset:160 ; 4-byte Folded Spill
+; GFX908-DAG:     v_mov_b32_e32 v3, s35
+; GFX908-DAG:     buffer_store_dword v3, off, s[0:3], s33 offset:164 ; 4-byte Folded Spill
+; GFX908-DAG:     v_mov_b32_e32 v3, s16
+; GFX908-DAG:     buffer_store_dword v3, off, s[0:3], s33 offset:168 ; 4-byte Folded Spill
 ; GFX908-NEXT:    s_addk_i32 s32, 0x2c00
 ; GFX908-NEXT:    s_mov_b64 s[16:17], exec
 ; GFX908-NEXT:    s_mov_b64 exec, 1

[jayfoad]

All these checks have been autogenerated (see comment line 1 of the test) so you really should not edit them by hand. If you do, your edits will get lost the next time some re-auto-generates the checks.

[cmc-rep]

All these checks have been autogenerated (see comment line 1 of the test) so you really should not edit them by hand. If you do, your edits will get lost the next time some re-auto-generates the checks.

@jayfoad I didn't notice that. This is tricky. However, it feels like that there is no real good way to reduce merge difficulty in this case? I should close this PR?

[cmc-rep]

Not the right way

[jayfoad]

However, it feels like that there is no real good way to reduce merge difficulty in this case?

Can you understand why your downstream changes cause churn in these tests? Maybe there is some underlying problem that could be fixed?

I should close this PR?

Yes :)