| Message ID | Z+O5eOzRC1lFSzY0@tucnak |
|---|---|
| State | New |
| Headers | show |
| Series | i386: Fix up pr55583.c testcase [PR119465] | expand |
On Wed, Mar 26, 2025 at 9:23 AM Jakub Jelinek <jakub@redhat.com> wrote: > > Hi! > > In r15-4289 H.J. fixed up the pr55583.c testcase to use unsigned long long > or long long instead of unsigned long or long. That change looks correct to > me because the > void test64r () { b = ((u64)b >> n) | (a << (64 - n)); } > etc. functions otherwise aren't really 64-bit rotates, but something that > triggers UB all the time (at least one of the shifts is out of bounds). > I assume that change fixed the FAILs on -mx32, but it caused > FAIL: gcc.target/i386/pr55583.c scan-assembler-times (?n)shldl?[\\\\t ]*\\\\\$2 1 > FAIL: gcc.target/i386/pr55583.c scan-assembler-times (?n)shrdl?[\\\\t ]*\\\\\$2 2 > regression on i686-linux (but just for -m32 without defaulting to SSE2 or > what). The difference is that for say -m32 -march=x86-64 the stv pass > handles some of the rotates in SSE and so we get different sh[rl]dl > instruction counts from the case when SSE isn't enabled and stv pass isn't > done. > > The following patch fixes that by disabling SSE for ia32 and always testing > for the same number of instructions. Yes, in the testsuite we (mostly) assume that ia32 lacks MMX and SSE by default. > Bootstrapped/regtested on x86_64-linux and i686-linux and tested with all of > make check-gcc RUNTESTFLAGS='--target_board=unix\{-m32/-march=x86-64,-m32/-march=i686,-mx32,-m64\} i386.exp=pr55583.c' > ok for trunk? > > 2025-03-26 Jakub Jelinek <jakub@redhat.com> > > PR target/55583 > PR target/119465 > * gcc.target/i386/pr55583.c: Add -mno-sse -mno-mmx to > dg-additional-options. Expect 4 shrdl and 2 shldl instructions on > ia32. OK. Thanks, Uros. > --- gcc/testsuite/gcc.target/i386/pr55583.c.jj 2024-10-12 13:47:21.966350407 +0200 > +++ gcc/testsuite/gcc.target/i386/pr55583.c 2025-03-25 16:36:46.320986804 +0100 > @@ -1,9 +1,8 @@ > /* { dg-do compile } */ > /* { dg-options "-O2 -Wno-shift-count-overflow" } */ > -/* { dg-final { scan-assembler-times {(?n)shrd[ql]?[\t ]*\$2} 4 { target { ! ia32 } } } } */ > -/* { dg-final { scan-assembler-times {(?n)shrdl?[\t ]*\$2} 2 { target ia32 } } } */ > -/* { dg-final { scan-assembler-times {(?n)shldl?[\t ]*\$2} 1 { target ia32 } } } */ > -/* { dg-final { scan-assembler-times {(?n)shld[ql]?[\t ]*\$2} 2 { target { ! ia32 } } } } */ > +/* { dg-additional-options "-mno-sse -mno-mmx" { target ia32 } } */ > +/* { dg-final { scan-assembler-times {(?n)shrd[ql]?[\t ]*\$2} 4 } } */ > +/* { dg-final { scan-assembler-times {(?n)shld[ql]?[\t ]*\$2} 2 } } */ > > typedef unsigned long long u64; > typedef unsigned int u32; > > Jakub >
--- gcc/testsuite/gcc.target/i386/pr55583.c.jj 2024-10-12 13:47:21.966350407 +0200 +++ gcc/testsuite/gcc.target/i386/pr55583.c 2025-03-25 16:36:46.320986804 +0100 @@ -1,9 +1,8 @@ /* { dg-do compile } */ /* { dg-options "-O2 -Wno-shift-count-overflow" } */ -/* { dg-final { scan-assembler-times {(?n)shrd[ql]?[\t ]*\$2} 4 { target { ! ia32 } } } } */ -/* { dg-final { scan-assembler-times {(?n)shrdl?[\t ]*\$2} 2 { target ia32 } } } */ -/* { dg-final { scan-assembler-times {(?n)shldl?[\t ]*\$2} 1 { target ia32 } } } */ -/* { dg-final { scan-assembler-times {(?n)shld[ql]?[\t ]*\$2} 2 { target { ! ia32 } } } } */ +/* { dg-additional-options "-mno-sse -mno-mmx" { target ia32 } } */ +/* { dg-final { scan-assembler-times {(?n)shrd[ql]?[\t ]*\$2} 4 } } */ +/* { dg-final { scan-assembler-times {(?n)shld[ql]?[\t ]*\$2} 2 } } */ typedef unsigned long long u64; typedef unsigned int u32;
Hi! In r15-4289 H.J. fixed up the pr55583.c testcase to use unsigned long long or long long instead of unsigned long or long. That change looks correct to me because the void test64r () { b = ((u64)b >> n) | (a << (64 - n)); } etc. functions otherwise aren't really 64-bit rotates, but something that triggers UB all the time (at least one of the shifts is out of bounds). I assume that change fixed the FAILs on -mx32, but it caused FAIL: gcc.target/i386/pr55583.c scan-assembler-times (?n)shldl?[\\\\t ]*\\\\\$2 1 FAIL: gcc.target/i386/pr55583.c scan-assembler-times (?n)shrdl?[\\\\t ]*\\\\\$2 2 regression on i686-linux (but just for -m32 without defaulting to SSE2 or what). The difference is that for say -m32 -march=x86-64 the stv pass handles some of the rotates in SSE and so we get different sh[rl]dl instruction counts from the case when SSE isn't enabled and stv pass isn't done. The following patch fixes that by disabling SSE for ia32 and always testing for the same number of instructions. Bootstrapped/regtested on x86_64-linux and i686-linux and tested with all of make check-gcc RUNTESTFLAGS='--target_board=unix\{-m32/-march=x86-64,-m32/-march=i686,-mx32,-m64\} i386.exp=pr55583.c' ok for trunk? 2025-03-26 Jakub Jelinek <jakub@redhat.com> PR target/55583 PR target/119465 * gcc.target/i386/pr55583.c: Add -mno-sse -mno-mmx to dg-additional-options. Expect 4 shrdl and 2 shldl instructions on ia32. Jakub