[llvm-dev] llvm is illegally vectorizing with a recurrence on skylake

[Finkel, Hal J. via llvm-dev]

Hi, Scott,

Thanks for reporting this problem. We should get a bug filed on this issue at bugs.llvm.org. If you're not able to do this, please let us know, and someone else can take care of it.

 -Hal

Hal Finkel
Lead, Compiler Technology and Programming Languages
Leadership Computing Facility
Argonne National Laboratory

________________________________
From: llvm-dev <llvm-dev-bounces at lists.llvm.org> on behalf of Scott Manley via llvm-dev <llvm-dev at lists.llvm.org>
Sent: Thursday, May 2, 2019 4:14 PM
To: llvm-dev
Subject: [llvm-dev] llvm is illegally vectorizing with a recurrence on skylake

Hi -- I have found a bug in an HPC code where llvm is vectorizing a loop on Skylake that has an obvious recurrence. I derived a small test case based on the original benchmark below:

/*****************************************************************/
static void  __attribute__ ((always_inline)) one(
  const int *restrict in, const int *const end,
  const unsigned shift, int *const restrict index,
  int *const restrict out)
{
  do {
    int a_idx = *in>>shift;
    int b_idx = index[a_idx];
    out[b_idx] = *in;                // <-- reccurence as index[a_idx] can be the
    index[a_idx]++;                 //      same and incremented within the vector
  } while(++in!=end);              //     which leads to incorrect results
}

#ifndef NO_TWO
static void  __attribute__ ((noinline)) two(
  const int *restrict in, const int *const end,
  const unsigned shift, int *const restrict index,
  int *const restrict out)
{
  do out[index[(*in>>shift)]++]=*in; while(++in!=end);
}
#endif

void parent(
  int digits, int n, int *restrict work, int * restrict idx,
  int *restrict shift, int **restrict indicies)
{
  int *in = work;
  int *dst = work+n;
//  int *indicies[1024];
//  int shift[1024];
  int d;
  for(d=1;d!=digits-1;++d) {
    int *t;
    one(in,in+n,shift[d],indicies[d],dst);
    t=in,in=dst,dst=t;
  }
#ifndef NO_TWO
  two(in,in+n,shift[d],indicies[d],idx);
#endif
}
/*****************************************************************/

clang -S -O2 -Rpass=loop-vectorize small.c  -march=skylake-avx512
small.c:6:3: remark: vectorized loop (vectorization width: 16, interleaved count: 1) [-Rpass=loop-vectorize]
  do {
  ^

I believe the problem to be a issue with dependency information getting destroyed because if you remove the two() function (or compile one() on its own, or prevent inlining of one()), it correctly prevents vectorization.

clang -S -O2 -Rpass=loop-vectorize -Rpass-missed=loop-vectorize small.c  -march=skylake-avx512 -DNO_TWO
small.c:6:3: remark: loop not vectorized [-Rpass-missed=loop-vectorize]
  do {

I did trace it down to possibly being something within DepChecker->areDepsSafe() as it returns true for the incorrect case.

Thanks,

Scott
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20190502/2172d96d/attachment.html>