[llvm-branch-commits] [clang] [llvm] [AMDGPU] Disable generic DAG combines at -O0 to preserve debuggability. (PR #176304)

[Daniil Fukalov via llvm-branch-commits]

================
@@ -872,6 +872,13 @@ bool AMDGPUTargetLowering::shouldReduceLoadWidth(
   return (OldSize < 32);
 }
 
+bool AMDGPUTargetLowering::shouldExpandPowerOf2DivRem(EVT VT) const {
+  // At -O0 we disable DAG combines, so power-of-2 div/rem won't be
----------------
dfukalov wrote:

@arsenm  I'm also not sure which can be better solution. I can imagine these approaches, but I'm not sure I prefer one of them:
- Modify ExpandIRInsts to always expand power-of-2 div/rem
- Handle power-of-2 expansion in AMDGPUISelLowering::LowerDIVREM
- Adjust MaxLegalDivRemBitWidth for AMDGPU at -O0

https://github.com/llvm/llvm-project/pull/176304