[PATCH] D146629: [SimplifyCFG][LICM] Preserve nonnull, range and align metadata when speculating

[Nikita Popov via Phabricator via llvm-commits]

nikic added a comment.

Reduced to a miscompile in GVN:

  declare void @use(i64)
  
  define void @test(ptr %p) {
    %val = load ptr, ptr %p, align 8, !nonnull !{}
    %val.int = ptrtoint ptr %val to i64
    %val2 = load i64, ptr %p, align 8
    call void @use(i64 %val.int)
    call void @use(i64 %val2)
    ret void
  }
  
  ; RUN: opt -passes=gvn < %s
  
  define void @test(ptr %p) {
    %val = load ptr, ptr %p, align 8, !nonnull !0
    %val.int = ptrtoint ptr %val to i64
    call void @use(i64 %val.int)
    call void @use(i64 %val.int)
    ret void
  }

I believe the problem here is that this goes through VNCoercion and the replacement we do is not with a plain load, but a ptrtoint cast of the load. As such, we end up not performing the combineMetadataForCSE adjustment.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D146629/new/

https://reviews.llvm.org/D146629