https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100778
Bug ID: 100778
Summary: Get SIGFPE on simple test with -fpe-trap=invalid and
SLP vectorization ON, with gfortran 11.1.0 on x86_64
Product: gcc
Version: 11.1.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: fortran
Assignee: unassigned at gcc dot gnu.org
Reporter: gabrielle.hugo at cern dot ch
Target Milestone: ---
Hello,
Updating gfortran from 10.3.0 to 11.1.0 leads to SIGFPE in a Fortran 77 project
(all things else equal).
The issue can actually be isolated and reproduced on x86_64 with the equivalent
simple test.f [1]: it runs smoothly when compiled with gfortran 10.3.0, and
leads to SIGFPE with gfortran 11.1.0.
To reproduce:
gfortran -O3 -Wall -Wextra -g -ffpe-trap=invalid test.f -o test
./test
The FPE is thrown from `x = x / norm` (while the branch should not have been
entered).
The issue disappears if additionally compiling with -fno-tree-slp-vectorize, or
if placing a printout in between `x = x / norm` and ` y = y / norm`.
The issue does not appear on ARM (tested on Apple M1).
Thanks in advance
[1] test.f:
program test
implicit none
double precision input(2)
input(1) = 0.D+00
input(2) = 0.D+00
call normalize(input)
end
subroutine normalize(input)
implicit none
double precision input(2)
double precision x, y, norm
common / coordinates / x, y
save / coordinates /
x = input(1)
y = input(2)
norm = sqrt(x**2 + y**2)
if (norm .gt. 1.D+00) then
x = x / norm
y = y / norm
write (*, *) 'Normalized x and y.'
end if
write (*, *) 'Ended as expected.'
end
[2] gcc versions:
gcc version 10.3.0 (GCC)
Configured with:
/build/dkonst/gcc-clang/build/contrib/gcc-10.3.0/src/gcc/10.3.0/configure
--prefix=/build/dkonst/gcc-clang/lcgcmake-install/gcc/10.3.0/x86_64-centos7
-with-system-zlib --disable-multilib --enable-languages=all --with-gnu-ld
--with-gnu-as
Thread model: posix
Supported LTO compression algorithms: zlib
gcc version 11.1.0 (GCC)
Target: x86_64-pc-linux-gnu
Configured with:
/build/dkonst/gcc-clang-2/build/contrib/gcc-11.1.0/src/gcc/11.1.0/configure
--prefix=/build/dkonst/gcc-clang-2/lcgcmake-install/gcc/11.1.0/x86_64-centos7
-with-system-zlib --disable-multilib --enable-languages=all --with-gnu-ld
--with-gnu-as
Thread model: posix
Supported LTO compression algorithms: zlib
Tested on Centos7 and Fedora33.
[3] objdump -S test
gfortran 10.3.0 case:
x = input(1)
4011d9: f2 0f 10 0f movsd (%rdi),%xmm1
y = input(2)
4011dd: f2 0f 10 57 08 movsd 0x8(%rdi),%xmm2
x = input(1)
4011e2: 66 0f 10 27 movupd (%rdi),%xmm4
norm = sqrt(x**2 + y**2)
4011e6: 66 0f 28 c1 movapd %xmm1,%xmm0
4011ea: 66 0f 28 da movapd %xmm2,%xmm3
4011ee: f2 0f 59 c1 mulsd %xmm1,%xmm0
x = input(1)
4011f2: 0f 29 25 77 2e 00 00 movaps %xmm4,0x2e77(%rip) #
404070 <coordinates_>
norm = sqrt(x**2 + y**2)
4011f9: f2 0f 59 da mulsd %xmm2,%xmm3
4011fd: f2 0f 58 c3 addsd %xmm3,%xmm0
401201: f2 0f 51 c0 sqrtsd %xmm0,%xmm0
if (norm .gt. 1.D+00) then
401205: 66 0f 2f 05 53 0e 00 comisd 0xe53(%rip),%xmm0 #
402060 <options.2.0+0x20>
40120c: 00
40120d: 77 51 ja 401260 <normalize_+0x90>
40120f: 48 8b 1d 52 0e 00 00 mov 0xe52(%rip),%rbx # 402068
<options.2.0+0x28>
401216: 48 89 e5 mov %rsp,%rbp
x = x / norm
y = y / norm
write (*, *) 'Normalized x and y.'
end if
gfortran 11.1.0 case:
x = input(1)
4011d9: 66 0f 10 0f movupd (%rdi),%xmm1
4011dd: 66 0f 28 d9 movapd %xmm1,%xmm3
4011e1: 66 0f 28 c1 movapd %xmm1,%xmm0
4011e5: 0f 29 0d 84 2e 00 00 movaps %xmm1,0x2e84(%rip) #
404070 <coordinates_>
y = input(2)
norm = sqrt(x**2 + y**2)
4011ec: f2 0f 59 c1 mulsd %xmm1,%xmm0
4011f0: 66 0f 15 db unpckhpd %xmm3,%xmm3
4011f4: 66 0f 28 d3 movapd %xmm3,%xmm2
4011f8: f2 0f 59 d3 mulsd %xmm3,%xmm2
4011fc: f2 0f 58 c2 addsd %xmm2,%xmm0
401200: f2 0f 51 c0 sqrtsd %xmm0,%xmm0
if (norm .gt. 1.D+00) then
401204: 66 0f 2f 05 54 0e 00 comisd 0xe54(%rip),%xmm0 #
402060 <options.2.0+0x20>
40120b: 00
x = x / norm
40120c: 66 0f 28 d0 movapd %xmm0,%xmm2
401210: 66 0f 14 d2 unpcklpd %xmm2,%xmm2
401214: 66 0f 5e ca divpd %xmm2,%xmm1
if (norm .gt. 1.D+00) then
401218: 77 4e ja 401268 <normalize_+0x98>
40121a: 48 8b 1d 47 0e 00 00 mov 0xe47(%rip),%rbx # 402068
<options.2.0+0x28>
401221: 48 89 e5 mov %rsp,%rbp
y = y / norm
write (*, *) 'Normalized x and y.'
end if
Here, divpd is executed before ja.
