lapack
lapack copied to clipboard
Reference-LAPACK
[v3.12.0] sblat1.f (SNRM2) test fail on i386 architectures
Description
The sblat1.f test fails on i386 architectures, including i386 (effectively i686), and hurd-i386:
https://buildd.debian.org/status/package.php?p=lapack&suite=experimental
https://buildd.debian.org/status/fetch.php?pkg=lapack&arch=i386&ver=3.12.0-1&stamp=1702247469&raw=0
https://buildd.debian.org/status/fetch.php?pkg=lapack&arch=hurd-i386&ver=3.12.0-1&stamp=1702271309&raw=0
The typical failure is the SNRM2 call, which look like:
Test of subprogram number 7 SNRM2
FAIL
SNRM2: N= 4, INCX= -2, IV= 9, IW= 9, test= 0.20971520E+07
FAIL
SNRM2: N= 4, INCX= 1, IV= 9, IW= 9, test= 0.20971520E+07
----- PASS -----
I'm confused about why this failure does not occur on other 32-bit architectures such as armhf and m68k. Any idea? It does not seem like a compiler issue neither. Both gfortran-12 and gfortran-13 lead to the same results.
The minimal example to reproduce this issue is to build this package on a clean Debian unstable (sid) chroot.
Checklist
- [x] I've included a minimal example to reproduce the issue
- [x] I'd be willing to make a PR to solve this issue
There is a typo in the test that on most systems apparently yields a zero, but not on the two systems above.
@cdluminate I sent you an email.
BTW, the following is the default build flags for Debian (distribution-wide):
CFLAGS=-g -O2 -ffile-prefix-map=/home/lumin=. -fstack-protector-strong -fstack-clash-protection -Wformat -Werror=format-security -fcf-protection
CPPFLAGS=-Wdate-time -D_FORTIFY_SOURCE=2
FFLAGS=-g -O2 -ffile-prefix-map=/home/lumin=. -fstack-protector-strong -fstack-clash-protection -fcf-protection
LDFLAGS=-Wl,-z,relro
I've not yet verified which argument triggered the bug. But I cannot reproduce the issue on i386 with the default flags in make.inc. The bug only appears when built using debian's flags and tools. There is a resemblance to https://github.com/Reference-LAPACK/lapack/pull/961 which is only triggered by coverage build.
Ok, after fixing the test implementation (https://github.com/Reference-LAPACK/lapack/pull/964), the ZDNRM2 test breaks again with Debian specific build flags. I think I should try to contribute a CI test with Debian build flags. Those stack protection flags can help you guys catch more bugs.
With the default flags in make.inc. The test passed. But with stack protections, you can reproduce the issue with (1) the following config, (2) Patch #964, and (3) the command make blas_testing on an amd64 machine.
21c21
< FFLAGS = -O2 -frecursive
---
> FFLAGS = -g -O2 -fstack-protector-strong -fstack-clash-protection -fcf-protection
27c27
< LDFLAGS =
---
> LDFLAGS = -Wl,-z,relro
The error looks like
Test of subprogram number 6 DZNRM2
FAIL
DZNRM2: N= 1, INCX= -2, IV= 9, IW= 1, test= NaN
DZNRM2: N= 1, INCX= -2, IV= 9, IW= 2, test= NaN
DZNRM2: N= 1, INCX= -2, IV= 9, IW= 3, test= NaN
DZNRM2: N= 1, INCX= -2, IV= 9, IW= 4, test= NaN
DZNRM2: N= 1, INCX= -2, IV= 9, IW= 5, test= NaN
DZNRM2: N= 1, INCX= -2, IV= 9, IW= 6, test= NaN
DZNRM2: N= 1, INCX= -2, IV= 9, IW= 7, test= NaN
DZNRM2: N= 1, INCX= -2, IV= 9, IW= 8, test= NaN
DZNRM2: N= 1, INCX= -2, IV= 9, IW= 9, test= NaN
DZNRM2: N= 1, INCX= -2, IV= 9, IW=10, test= NaN
FAIL
DZNRM2: N= 1, INCX= 1, IV= 9, IW= 1, test= NaN
DZNRM2: N= 1, INCX= 1, IV= 9, IW= 2, test= NaN
DZNRM2: N= 1, INCX= 1, IV= 9, IW= 3, test= NaN
DZNRM2: N= 1, INCX= 1, IV= 9, IW= 4, test= NaN
DZNRM2: N= 1, INCX= 1, IV= 9, IW= 5, test= NaN
DZNRM2: N= 1, INCX= 1, IV= 9, IW= 6, test= NaN
DZNRM2: N= 1, INCX= 1, IV= 9, IW= 7, test= NaN
DZNRM2: N= 1, INCX= 1, IV= 9, IW= 8, test= NaN
DZNRM2: N= 1, INCX= 1, IV= 9, IW= 9, test= NaN
DZNRM2: N= 1, INCX= 1, IV= 9, IW=10, test= NaN
FAIL
DZNRM2: N= 2, INCX= -2, IV= 1, IW= 9, test= NaN
DZNRM2: N= 2, INCX= -2, IV= 2, IW= 9, test= NaN
DZNRM2: N= 2, INCX= -2, IV= 3, IW= 9, test= NaN
DZNRM2: N= 2, INCX= -2, IV= 4, IW= 9, test= NaN
DZNRM2: N= 2, INCX= -2, IV= 5, IW= 9, test= NaN
DZNRM2: N= 2, INCX= -2, IV= 6, IW= 9, test= NaN
DZNRM2: N= 2, INCX= -2, IV= 7, IW= 9, test= NaN
DZNRM2: N= 2, INCX= -2, IV= 8, IW= 9, test= NaN
DZNRM2: N= 2, INCX= -2, IV= 9, IW= 1, test= NaN
DZNRM2: N= 2, INCX= -2, IV= 9, IW= 2, test= NaN
DZNRM2: N= 2, INCX= -2, IV= 9, IW= 3, test= NaN
DZNRM2: N= 2, INCX= -2, IV= 9, IW= 4, test= NaN
DZNRM2: N= 2, INCX= -2, IV= 9, IW= 5, test= NaN
DZNRM2: N= 2, INCX= -2, IV= 9, IW= 6, test= NaN
DZNRM2: N= 2, INCX= -2, IV= 9, IW= 7, test= NaN
DZNRM2: N= 2, INCX= -2, IV= 9, IW= 8, test= NaN
DZNRM2: N= 2, INCX= -2, IV= 9, IW= 9, test= NaN
FAIL
DZNRM2: N= 2, INCX= 1, IV= 1, IW= 9, test= NaN
DZNRM2: N= 2, INCX= 1, IV= 2, IW= 9, test= NaN
DZNRM2: N= 2, INCX= 1, IV= 3, IW= 9, test= NaN
[...truncated...]
Update: cblas1.f fails just like zcblas1.f.
@cdluminate Do you confirm that this issue is indeed fixed by #964? My understanding was that this PR was not enough.
