[PATCH] D135462: [SelectionDAG] Do not second-guess alignment for alloca

[Andrew Savonichev via Phabricator via llvm-commits]

asavonic added a comment.

I checked OpenMP tests, and the difference is pretty much what was expected. This is what we have as an input (the patch does not change the IR):

  // x86_64-pc-linux-gnu :: mapping/ompx_hold/struct.c
  struct S {
    int i;
    int j;
  } s;
  // CHECK: presence of s, s.i, s.j: 0, 0, 0
  CHECK_PRESENCE(s, s.i, s.j);

IR:

  %s = alloca %struct.S, align 4
  %call = call i32 @omp_get_default_device()
  %call1 = call i32 @omp_target_is_present(ptr noundef %s, i32 noundef %call)

The assembly is different: we keep alignment of 4 and do not increase it to the preferred value of 8.
Before the patch:

  leaq    -16(%rbp), %rdi
  callq   omp_target_is_present at PLT

After the patch:

  leaq    -12(%rbp), %rdi
  callq   omp_target_is_present at PLT
  
  Libomptarget message: explicit extension not allowed: host address specified is 0x00007fff03109220 (12 bytes), but device allocation maps to host at 0x00007fff03109224 (8 bytes)

If I explicitly set `__attribute__((aligned(8)))` to the declaration of `s`, assembly before and after the patch becomes equivalent, so the problem should go away. Apparently there is some discrepancy in alignment that the frontend (or the OpenMP runtime) expects, but it is not obvious where we need to fix this.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D135462/new/

https://reviews.llvm.org/D135462