[llvm-bugs] [Bug 31554] New: NewGVN miscompile

[via llvm-bugs]

https://llvm.org/bugs/show_bug.cgi?id=31554

            Bug ID: 31554
           Summary: NewGVN miscompile
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Windows NT
            Status: NEW
          Severity: normal
          Priority: P
         Component: Scalar Optimizations
          Assignee: unassignedbugs at nondot.org
          Reporter: greg_bedwell at sn.scee.net
                CC: llvm-bugs at lists.llvm.org
            Blocks: 30995
    Classification: Unclassified

Found a bad runtime result with -enable-newgvn on one of our test suites.  I'm
currently out of the office until mid-next-week so haven't had a chance to look
in-depth at the codegen to see if it's a duplicate of an existing report
against NewGVN, but I figured it would be better to raise it now than in a
week's time as I have a nice self-contained test-case already.

$./build/bin/clang --version
clang version 4.0.0 (trunk 291148)
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: /home/greg/public_git/./build/bin

$./build/bin/clang -O2 1.cpp -o 1.elf && ./1.elf
loop A: CC[0] <- AA[0] (10)
loop B: CC[1] <- BB[0] (20)
loop B: CC[2] <- BB[1] (21)
loop B: CC[3] <- BB[2] (22)
loop B: CC[4] <- BB[3] (23)
loop B: CC[5] <- BB[4] (24)
loop B: CC[6] <- BB[5] (25)
loop B: CC[7] <- BB[6] (26)
10, 20, 21, 22, 23, 24, 25, 26

$./build/bin/clang -O2 -mllvm -enable-newgvn 1.cpp -o 1.elf && ./1.elf
loop A: CC[0] <- AA[0] (10)
loop B: CC[1] <- BB[0] (20)
loop B: CC[2] <- BB[1] (21)
loop B: CC[3] <- BB[2] (22)
loop B: CC[4] <- BB[3] (23)
loop B: CC[5] <- BB[4] (24)
loop B: CC[6] <- BB[5] (25)
loop B: CC[7] <- BB[6] (26)
0, 20, 21, 22, 23, 24, 25, 26

^~~~~ should be 10


// ============================================================================
extern "C" int printf(const char *, ...);
typedef char __attribute__((vector_size(8))) vec8;

vec8 foo(unsigned mask) {

  vec8 AA = (vec8){10, 11, 12, 13, 14, 15, 16, 17};
  vec8 BB = (vec8){20, 21, 22, 23, 24, 25, 26, 27};
  vec8 CC;

  if (mask == 0)
    return BB;

  int idxA = 0;
  for (; idxA < mask; ++idxA, --mask) {
    int a = idxA;
    int b = mask - 1;
    int c = AA[b];
    printf("loop A: CC[%d] <- AA[%d] (%d)\n", a, b, c);
    CC[a] = c;
  }

  for (int idxB = 0; idxA < 8; ++idxA, ++idxB) {
    int a = idxA;
    int b = idxB;
    int c = BB[b];
    printf("loop B: CC[%d] <- BB[%d] (%d)\n", a, b, c);
    CC[a] = c;
  }
  return CC;
}

int main() {
  vec8 r = foo(1);
  printf("%d, %d, %d, %d, %d, %d, %d, %d\n", r[0], r[1], r[2], r[3], r[4],
r[5],
         r[6], r[7]);
}
// ============================================================================

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20170105/2263b96b/attachment.html>