[llvm] [SeparateConstOffsetFromGEP] Don't set unsound inbounds flag (PR #130616)

[Fabian Ritter via llvm-commits]

================
@@ -0,0 +1,438 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+
+; RUN: llc -mtriple=amdgcn -mcpu=gfx90a -mattr=-enable-flat-scratch < %s | FileCheck --check-prefixes=GFX90A,GFX90A-FSOFF %s
+; RUN: llc -mtriple=amdgcn -mcpu=gfx90a -mattr=+enable-flat-scratch < %s | FileCheck --check-prefixes=GFX90A,GFX90A-FSON %s
+; RUN: llc -mtriple=amdgcn -mcpu=gfx1030 -mattr=-enable-flat-scratch < %s | FileCheck --check-prefixes=GFX10,GFX10-FSOFF %s
+; RUN: llc -mtriple=amdgcn -mcpu=gfx1030 -mattr=+enable-flat-scratch < %s | FileCheck --check-prefixes=GFX10,GFX10-FSON %s
+; RUN: llc -mtriple=amdgcn -mcpu=gfx942 < %s | FileCheck --check-prefixes=GFX942 %s
+; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 < %s | FileCheck --check-prefixes=GFX11 %s
+; RUN: llc -mtriple=amdgcn -mcpu=gfx1200 < %s | FileCheck --check-prefixes=GFX12 %s
+
+; This test documents when memory addresses with constant offset components can
+; be folded into memory accesses with immediate offsets.
+
+define i32 @flat(ptr addrspace(0) %p, i32 %i) {
----------------
ritter-x2a wrote:

As it is, this PR only affects flat scratch (and apparently SMEM, as the const32 test suggests, not sure why it makes a difference there). Ideally, it should also affect generic flat instructions, since we currently miscompile those when the final GEP is not inbounds, but changing that requires ISel changes for which I haven't opened a PR yet.

In the new version, the add no longer is `nuw`, so that the flat case is indeed a miscompile. I also only kept the flat and scratch cases and renamed them.
Should this test also check the cases where `%p` and/or `%i` are uniform?


https://github.com/llvm/llvm-project/pull/130616