[llvm] [InstCombine] Simplify nonnull pointers (PR #128111)

[via llvm-commits]

asmok-g wrote:

This is causing a probable miscompile.

I'm working on a  repro. Bisecting for the exact file and exact function that causes the miscompile when optimized; the IR diff looks like, before: 
```
; Function Attrs: mustprogress nounwind uwtable
define linkonce_odr dso_local void @_ZN3gvr21WST12GetTransformEv(ptr dead_on_unwind noalias writable sret(%"M") align 4 %0, ptr noundef nonnull align 8 dereferenceable(32) %1) unnamed_addr #0 comdat align 2 {
  %3 = getelementptr inbounds nuw i8, ptr %1, i64 16
  %4 = getelementptr inbounds nuw i8, ptr %1, i64 24
  %5 = load ptr, ptr %4, align 8, !tbaa !51, !nonnull !50, !noundef !50
  %6 = tail call noundef ptr @_ZNSt3__u19__shared_weak_count4lockEv(ptr noundef nonnull align 8 dereferenceable(24) %5) #12
  %7 = icmp eq ptr %6, null
  %8 = load ptr, ptr %3, align 8
  %9 = select i1 %7, ptr null, ptr %8
  %10 = load ptr, ptr %9, align 8, !tbaa !3
  %11 = getelementptr inbounds nuw i8, ptr %10, i64 24
  %12 = load ptr, ptr %11, align 8
  tail call void %12(ptr dead_on_unwind writable sret(%"M") align 4 %0, ptr noundef nonnull align 8 dereferenceable(112) %9) #12
  br i1 %7, label %21, label %13

13:                                               ; preds = %2
  %14 = getelementptr inbounds nuw i8, ptr %6, i64 8
  %15 = atomicrmw add ptr %14, i64 -1 acq_rel, align 8
  %16 = icmp eq i64 %15, 0
  br i1 %16, label %17, label %21

17:                                               ; preds = %13
  %18 = load ptr, ptr %6, align 8, !tbaa !3
  %19 = getelementptr inbounds nuw i8, ptr %18, i64 16
  %20 = load ptr, ptr %19, align 8
  tail call void %20(ptr noundef nonnull align 8 dereferenceable(24) %6) #12
  tail call void @_ZNSt3__u19__shared_weak_count14__release_weakEv(ptr noundef nonnull align 8 dereferenceable(24) %6) #12
  br label %21

21:                                               ; preds = %2, %13, %17
  ret void
}
```
after:
```
; Function Attrs: mustprogress nounwind uwtable
define linkonce_odr dso_local void @_ZN3gvr21WST12GetTransformEv(ptr dead_on_unwind noalias writable sret(%"M") align 4 %0, ptr noundef nonnull align 8 dereferenceable(32) %1) unnamed_addr #0 comdat align 2 {
  %3 = getelementptr inbounds nuw i8, ptr %1, i64 16
  %4 = getelementptr inbounds nuw i8, ptr %1, i64 24
  %5 = load ptr, ptr %4, align 8, !tbaa !51, !nonnull !50, !noundef !50
  %6 = tail call noundef ptr @_ZNSt3__u19__shared_weak_count4lockEv(ptr noundef nonnull align 8 dereferenceable(24) %5) #12
  %7 = load ptr, ptr %3, align 8
  %8 = load ptr, ptr %7, align 8, !tbaa !3
  %9 = getelementptr inbounds nuw i8, ptr %8, i64 24
  %10 = load ptr, ptr %9, align 8
  tail call void %10(ptr dead_on_unwind writable sret(%"M") align 4 %0, ptr noundef nonnull align 8 dereferenceable(112) %7) #12
  %11 = icmp eq ptr %6, null
  br i1 %11, label %20, label %12

12:                                               ; preds = %2
  %13 = getelementptr inbounds nuw i8, ptr %6, i64 8
  %14 = atomicrmw add ptr %13, i64 -1 acq_rel, align 8
  %15 = icmp eq i64 %14, 0
  br i1 %15, label %16, label %20

16:                                               ; preds = %12
  %17 = load ptr, ptr %6, align 8, !tbaa !3
  %18 = getelementptr inbounds nuw i8, ptr %17, i64 16
  %19 = load ptr, ptr %18, align 8
  tail call void %19(ptr noundef nonnull align 8 dereferenceable(24) %6) #12
  tail call void @_ZNSt3__u19__shared_weak_count14__release_weakEv(ptr noundef nonnull align 8 dereferenceable(24) %6) #12
  br label %20

20:                                               ; preds = %2, %12, %16
  ret void
}
```

The clang invocation to reproduce is `clang  -fno-exceptions   -O3  '-std=gnu++20' pre.ii -emit-llvm -S  -o case.ll`

The entire preprocessed file that i'm reducing has other changes, but i thought i'd show the diff for the specific function will we have a completer reduced case.

https://github.com/llvm/llvm-project/pull/128111