[AMDGPU] Add a regression test for -mattr=dumpcode

[jayfoad]

No description provided.

[llvmbot]

@llvm/pr-subscribers-backend-amdgpu

Author: Jay Foad (jayfoad)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/116982.diff

1 Files Affected:

• (added) llvm/test/CodeGen/AMDGPU/dumpcode.ll (+30)

diff --git a/llvm/test/CodeGen/AMDGPU/dumpcode.ll b/llvm/test/CodeGen/AMDGPU/dumpcode.ll
new file mode 100644
index 00000000000000..1acec2997aa8ff
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/dumpcode.ll
@@ -0,0 +1,30 @@
+; RUN: llc -mtriple=amdgcn -mcpu=gfx1010 -mattr=dumpcode -filetype=obj < %s | llvm-objcopy --dump-section .AMDGPU.disasm=- - /dev/null | FileCheck %s -check-prefix=GFX10
+
+; GFX10: f:
+; GFX10-NEXT: BB0_0:
+; GFX10-NEXT:   s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; BF8C0000
+; GFX10-NEXT:   v_mov_b32_e32 v3, 0xde                  ; 7E0602FF 000000DE
+; GFX10-NEXT:   v_add_nc_u32_e32 v2, 1, v4              ; 4A040881
+; GFX10-NEXT:   s_mov_b32 s4, 0                         ; BE840380
+; GFX10-NEXT:   global_store_dword v[0:1], v3, off      ; DC708000 007D0300
+; GFX10-NEXT: BB0_1:
+; GFX10-NEXT:   v_add_nc_u32_e32 v2, -1, v2             ; 4A0404C1
+; GFX10-NEXT:   v_cmp_eq_u32_e32 vcc_lo, 0, v2          ; 7D840480
+; GFX10-NEXT:   s_or_b32 s4, vcc_lo, s4                 ; 8804046A
+; GFX10-NEXT:   s_andn2_b32 exec_lo, exec_lo, s4        ; 8A7E047E
+; GFX10-NEXT:   s_cbranch_execnz ""                     ; BF890000
+; GFX10-NEXT:   s_or_b32 exec_lo, exec_lo, s4           ; 887E047E
+; GFX10-NEXT:   s_setpc_b64 s[30:31]                    ; BE80201E
+
+define void @f(ptr addrspace(1) %out, ptr addrspace(1) %in, i32 %val) {
+entry:
+  br label %body
+body:
+  %i = phi i32 [0, %entry], [%inc, %body]
+  store i32 222, ptr addrspace(1) %out
+  %cmp = icmp ne i32 %i, %val
+  %inc = add i32 %i, 1
+  br i1 %cmp, label %body, label %end
+end:
+  ret void
+}

[arsenm]

Can we please be rid of this hack at last?

[jayfoad]

Ping. If the feature exists, it should be tested.

Can we please be rid of this hack at last?

We could start a separate discussion on that. It's a little awkward at the moment as we still have downstream uses of it.

[arsenm]

Does this need a +

[jayfoad]

No, it's optional and defaults to +.

[jayfoad]

This test failed on a big-endian host: https://lab.llvm.org/buildbot/#/builders/42/builds/3253

#127064 works around this.