[clang] [clang] Use the VFS to get the OpenMP entry info (PR #160935)

[via cfe-commits]

carlobertolli wrote:

@jansvoboda11 this change breaks -save-temps builds.
Reproducer:
```
#include <stdio.h>
#include <omp.h>

int main()
{
  int N = 10;

  int a[N];
  int b[N];

  int i;

  for (i=0; i<N; i++)
    a[i]=0;

  for (i=0; i<N; i++)
    b[i]=i;

#pragma omp target parallel for
  {
    for (int j = 0; j< N; j++)
      a[j]=b[j];
  }

  int rc = 0;
  for (i=0; i<N; i++)
    if (a[i] != b[i] ) {
      rc++;
      printf ("Wrong varlue: a[%d]=%d\n", i, a[i]);
    }

  if (!rc)
    printf("Success\n");

  return rc;
}
```

Build without -save-temps and make llvm print out list of target regions it identified:
clang -mllvm  -amdgpu-dump-hsa-metadata  -O2    -fopenmp --offload-arch=gfx90a   -D__OFFLOAD_ARCH_gfx90a__ veccopy.c 

Excerpt from resulting diagnostic shows that there are multiple entry points:
  .name:           __omp_offloading_811_2ea3964_main_l19
    .private_segment_fixed_size: 0
    .sgpr_count:     14
    .sgpr_spill_count: 0
    .symbol:         __omp_offloading_811_2ea3964_main_l19.kd


Build with -save-temps and make llvm print out list of target regions it identified:
clang -mllvm  -amdgpu-dump-hsa-metadata  -O2 -save-temps    -fopenmp --offload-arch=gfx90a   -D__OFFLOAD_ARCH_gfx90a__ veccopy.c


Excerpt from resulting diagnostic shows that there are no kernels
AMDGPU HSA Metadata:
---
amdhsa.kernels:  []
amdhsa.target:   amdgcn-amd-amdhsa--gfx90a
amdhsa.version:
  - 1
  - 2
...

Please fix?

Debugging showed me that this line
if (CGM.getFileSystem()->exists(PLoc.getFilename())) { 
is now succeeding (it was not before and I presume that was incorrect, hence your fix).
When using -save-temps, the following statement
PLoc = SM.getPresumedLoc(BeginLoc, /*UseLineDirectives=*/false);
is now identifying the preprocessed <src_file>.i as the actual PLoc filename and not the input source file.

When not using -save-temps, the preprocessed <src_file>.i does not exist in the build folder and the source file remains the same.

For some reason, this has an effect on clang's ability to identify target regions when building for the host, which seems fragile at best.

https://github.com/llvm/llvm-project/pull/160935