[clang] [llvm] [InstCombine] Strip leading zero indices from GEP (PR #155415)

[Jay Foad via llvm-commits]

jayfoad wrote:

This diff looks suspicious to me in function `_ZN22__llvm_libc_22_0_0_git8internal18CharacterConverter8pop_utf8Ev`:
```diff
--- 01-preopt-good.txt  2025-09-05 10:39:57.226224451 +0100
+++ 01-preopt-bad.txt   2025-09-05 10:39:56.149207707 +0100
@@ -11112,10 +11113,10 @@
 if.then4:                                         ; preds = %if.end
-  %sub7 = add nsw i64 %conv, -1
-  %4 = extractelement <4 x i8> <i8 0, i8 -64, i8 -32, i8 -16>, i64 %sub7
-  %5 = load i32, ptr %0, align 4, !tbaa !58
+  %4 = add nuw nsw i64 %conv, 4294967295
+  %5 = extractelement <4 x i8> <i8 0, i8 -64, i8 -32, i8 -16>, i64 %4
+  %6 = load i32, ptr %0, align 4, !tbaa !58
   %sh_prom = trunc nuw nsw i64 %mul to i32
-  %shr = lshr i32 %5, %sh_prom
+  %shr = lshr i32 %6, %sh_prom
   %shr.tr = trunc i32 %shr to i8
-  %or.narrow = or i8 %4, %shr.tr
-  %6 = zext i8 %or.narrow to i32
+  %or.narrow = or i8 %5, %shr.tr
+  %7 = zext i8 %or.narrow to i32
   br label %if.end15
```
I am guessing that the `add nsw i64 %conv, -1` vs `add nuw nsw i64 %conv, 4294967295` are the result of transforming some `getelementptr` instructions, and it seems suspicious that the 64-bit offset is small and negative in the good case but large and positive in the bad case.

https://github.com/llvm/llvm-project/pull/155415