[AMDGPU] promote i1 arg type for amdgpu_cs

[nickleus27]

fixes #68087
Not sure where to put regression tests for this pr? Also, should i1 args not in reg also be promoted?

[llvmbot]

@llvm/pr-subscribers-backend-amdgpu

Author: Nick Anderson (nickleus27)

Changes

fixes #68087
Not sure where to put regression tests for this pr? Also, should i1 args not in reg also be promoted?

---

Full diff: https://github.com/llvm/llvm-project/pull/82971.diff

1 Files Affected:

• (modified) llvm/lib/Target/AMDGPU/AMDGPUCallingConv.td (+1)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUCallingConv.td b/llvm/lib/Target/AMDGPU/AMDGPUCallingConv.td
index c5207228dc913f..a988140df3e89b 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUCallingConv.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPUCallingConv.td
@@ -66,6 +66,7 @@ def RetCC_SI_Gfx : CallingConv<[
 
 def CC_SI_SHADER : CallingConv<[
 
+  CCIfType<[i1], CCPromoteToType<i16>>,
   CCIfInReg<CCIfType<[f32, i32, f16, i16, v2i16, v2f16, bf16, v2bf16] , CCAssignToReg<[
     SGPR0, SGPR1, SGPR2, SGPR3, SGPR4, SGPR5, SGPR6, SGPR7,
     SGPR8, SGPR9, SGPR10, SGPR11, SGPR12, SGPR13, SGPR14, SGPR15,

[arsenm]

fixes #68087 Not sure where to put regression tests for this pr? Also, should i1 args not in reg also be promoted?

test/CodeGen/AMDGPU/calling-conventions.ll is probably the best bet right now (although I'm not happy with how these tests are organized)

Both inreg and not inreg should be promoted. The inreg just flips from VGPR to SGPR.

[rovka]

Hi, while you're at it, could you please make the same change for amdgpu_cs_chain and amdgpu_cs_chain_preserve?