[llvm-bugs] [Bug 34682] New: [SKX] Bus error after Loop Vectorization
via llvm-bugs
llvm-bugs at lists.llvm.org
Wed Sep 20 01:08:40 PDT 2017
https://bugs.llvm.org/show_bug.cgi?id=34682
Bug ID: 34682
Summary: [SKX] Bus error after Loop Vectorization
Product: new-bugs
Version: trunk
Hardware: PC
OS: Windows NT
Status: NEW
Severity: enhancement
Priority: P
Component: new bugs
Assignee: unassignedbugs at nondot.org
Reporter: ilia.taraban at intel.com
CC: llvm-bugs at lists.llvm.org
This test fails at run with "Bus error (core dumped)" after Loop Vectorization
on O2, skx:
================= nice.c ==============
int main ()
{
unsigned int j = 0, k = 0, i = 0;
unsigned int x [17] [17] = {0}, a [17] = {0};
for (j = 1; j < 10; ++j)
for (k = j; k < 17; ++k)
for (i = 1; i < j; ++i)
x[k][0] = a[i - 1];
printf("res = %u\n", x[0][0]);
return 0;
}
============================================
>>> clang -v
clang version 6.0.0 (trunk 313612)
Target: x86_64-unknown-linux-gnu
Thread model: posix
...
>>> clang -march=skylake-avx512 -O0 -o nice.exe nice.c
>>> sde -skx -- ./nice.exe
res = 0
>>> clang -march=skylake-avx512 -O2 -o nice.exe nice.c
>>> sde -skx -- ./nice.exe
Bus error (core dumped)
>>> clang -march=skylake-avx512 -O2 -o nice.exe nice.c -mllvm -opt-bisect-limit=83
...
BISECT: running pass (82) Rotate Loops on loop
BISECT: running pass (83) Loop Distribution on function (main)
BISECT: NOT running pass (84) Loop Vectorization on function (main)
BISECT: NOT running pass (85) Loop Load Elimination on function (main)
...
>>> sde -skx -- ./nice.exe
res = 0
>>> clang -march=skylake-avx512 -O2 -o nice.exe nice.c -mllvm -opt-bisect-limit=84
...
BISECT: running pass (83) Loop Distribution on function (main)
BISECT: running pass (84) Loop Vectorization on function (main)
BISECT: NOT running pass (85) Loop Load Elimination on function (main)
BISECT: NOT running pass (86) Combine redundant instructions on function (main)
...
>>> sde -skx -- ./nice.exe
Segmentation fault (core dumped)
If we look at generated asm, we'll see that we put -1 into esi, but then use
rsi for pointer operations:
================= nice.s ==================
...
movl $16, %r10d
subq %r11, %r10
leal -1(%r11), %esi # put -1 into $esi
movq %rax, %rbx
cmpq $16, %r10
...
vpaddq %zmm12, %zmm3, %zmm12
vpbroadcastd (%rsp,%rsi,4), %ymm13 # SIGBUS rsi =
0x00000000FFFFFFFF
kmovq %k1, %k2
vpscatterqd %ymm13, (,%zmm12) {%k2}
...
===========================================
So if we change "leal -1(%r11), %esi" to "leaq -1(%r11), %rsi" , our
program will finish correctly with "res = 0"
>>> clang nice-fixed.s -o nice-fixed.exe
>>> sde -skx -- ./nice-fixed.exe
res = 0
--
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/20170920/70b8e471/attachment.html>
More information about the llvm-bugs
mailing list