[PATCH] math: Fix x86_64 build for -Os (BZ 33367)
H.J. Lu
hjl.tools@gmail.com
Mon Sep 8 19:34:18 GMT 2025
On Mon, Sep 8, 2025 at 12:09 PM Adhemerval Zanella Netto
<adhemerval.zanella@linaro.org> wrote:
>
>
>
> On 08/09/25 15:59, H.J. Lu wrote:
> > On Mon, Sep 8, 2025 at 9:30 AM H.J. Lu <hjl.tools@gmail.com> wrote:
> >>
> >> On Mon, Sep 8, 2025 at 9:04 AM Adhemerval Zanella Netto
> >> <adhemerval.zanella@linaro.org> wrote:
> >>>
> >>>
> >>>
> >>> On 08/09/25 12:56, H.J. Lu wrote:
> >>>> On Mon, Sep 8, 2025 at 8:27 AM Adhemerval Zanella Netto
> >>>> <adhemerval.zanella@linaro.org> wrote:
> >>>>>
> >>>>>
> >>>>>
> >>>>> On 08/09/25 12:22, H.J. Lu wrote:
> >>>>>> On Mon, Sep 8, 2025 at 7:26 AM H.J. Lu <hjl.tools@gmail.com> wrote:
> >>>>>>>
> >>>>>>> On Sat, Sep 6, 2025 at 3:22 PM Adhemerval Zanella
> >>>>>>> <adhemerval.zanella@linaro.org> wrote:
> >>>>>>>>
> >>>>>>>> The compiler might not inline the trunc function call for
> >>>>>>>> USE_TRUNC_BUILTIN, which requires adding the required objects.
> >>>>>>>>
> >>>>>>>> Route trunc calls for the optimized modf variants to thei
> >>>>>>>> ABI-compatible version. This avoids invoking the ifunc variant
> >>>>>>>> for x86_64-v1 and requires fewer objects to be added.
> >>>>>>>>
> >>>>>>>> Checked on x86_64, x86_64-v2, x86_64-v3, and x86_64-v4. Used -O2 and
> >>>>>>>> -Os options. Performed a full make check on x86_64 with both
> >>>>>>>> optimizations.
> >>>>>>>> ---
> >>>>>>>> sysdeps/x86_64/fpu/multiarch/Makefile | 19 ++++++++++++-------
> >>>>>>>> sysdeps/x86_64/fpu/multiarch/s_modf-avx.c | 5 +++++
> >>>>>>>> sysdeps/x86_64/fpu/multiarch/s_modf-sse4_1.c | 5 +++++
> >>>>>>>> sysdeps/x86_64/fpu/multiarch/s_modf.c | 1 -
> >>>>>>>> sysdeps/x86_64/fpu/multiarch/s_modff-avx.c | 5 +++++
> >>>>>>>> sysdeps/x86_64/fpu/multiarch/s_modff-sse4_1.c | 5 +++++
> >>>>>>>> sysdeps/x86_64/fpu/multiarch/s_modff.c | 1 -
> >>>>>>>> 7 files changed, 32 insertions(+), 9 deletions(-)
> >>>>>>>>
> >>>>>>>> diff --git a/sysdeps/x86_64/fpu/multiarch/Makefile b/sysdeps/x86_64/fpu/multiarch/Makefile
> >>>>>>>> index 708b14297f..baefe8dd07 100644
> >>>>>>>> --- a/sysdeps/x86_64/fpu/multiarch/Makefile
> >>>>>>>> +++ b/sysdeps/x86_64/fpu/multiarch/Makefile
> >>>>>>>> @@ -26,6 +26,11 @@ CFLAGS-s_sinf-fma.c = -mfma -mavx2
> >>>>>>>> CFLAGS-s_cosf-fma.c = -mfma -mavx2
> >>>>>>>> CFLAGS-s_sincosf-fma.c = -mfma -mavx2
> >>>>>>>>
> >>>>>>>> +sysdep_calls += \
> >>>>>>>> + s_trunc-sse4_1 \
> >>>>>>>> + s_truncf-sse4_1 \
> >>>>>>>> +# sysdep_calls
> >>>>>>>> +
> >>>>>>>> # Check if ISA level is 2 or above.
> >>>>>>>> ifeq (,$(filter $(have-x86-isa-level),$(x86-isa-level-2-or-above)))
> >>>>>>>> sysdep_calls += \
> >>>>>>>> @@ -36,6 +41,10 @@ endif
> >>>>>>>>
> >>>>>>>> # Check if ISA level is 3 or above.
> >>>>>>>> ifneq (,$(filter $(have-x86-isa-level),$(x86-isa-level-3-or-above)))
> >>>>>>>> +sysdep_calls += \
> >>>>>>>> + s_trunc-avx \
> >>>>>>>> + s_truncf-avx \
> >>>>>>>> +# sysdep_calls
> >>>>>>>> libm-sysdep_routines += \
> >>>>>>>> s_ceil-avx \
> >>>>>>>> s_ceilf-avx \
> >>>>>>>> @@ -47,8 +56,6 @@ libm-sysdep_routines += \
> >>>>>>>> s_rintf-avx \
> >>>>>>>> s_roundeven-avx \
> >>>>>>>> s_roundevenf-avx \
> >>>>>>>> - s_trunc-avx \
> >>>>>>>> - s_truncf-avx \
> >>>>>>>> # libm-sysdep_routines
> >>>>>>>> else
> >>>>>>>> sysdep_calls += \
> >>>>>>>> @@ -114,8 +121,6 @@ libm-sysdep_routines += \
> >>>>>>>> s_tan-avx \
> >>>>>>>> s_tan-fma \
> >>>>>>>> s_tanh-fma \
> >>>>>>>> - s_trunc-sse4_1 \
> >>>>>>>> - s_truncf-sse4_1 \
> >>>>>>>> # libm-sysdep_routines
> >>>>>>>> ifeq ($(have-x86-isa-level),baseline)
> >>>>>>>> libm-sysdep_routines += \
> >>>>>>>> @@ -132,13 +137,13 @@ libm-sysdep_routines += \
> >>>>>>>> s_trunc-c \
> >>>>>>>> s_truncf-c \
> >>>>>>>> # libm-sysdep_routines
> >>>>>>>> -endif
> >>>>>>>> +endif # ($(have-x86-isa-level),baseline)
> >>>>>>>> +endif # (,$(filter $(have-x86-isa-level),$(x86-isa-level-3-or-above)))
> >>>>>>>>
> >>>>>>>> # $(sysdep_calls) functions are built both for libc and libm. While the
> >>>>>>>> # libc objects have the prefix s_, the libm ones are prefixed with m_.
> >>>>>>>> -sysdep_routines += $(sysdep_calls)
> >>>>>>>> +sysdep_routines += $(sysdep_calls)
> >>>>>>>> libm-sysdep_routines += $(sysdep_calls:s_%=m_%)
> >>>>>>>> -endif
> >>>>>>>>
> >>>>>>>> CFLAGS-e_asin-fma4.c = -mfma4
> >>>>>>>> CFLAGS-e_atan2-fma4.c = -mfma4
> >>>>>>>> diff --git a/sysdeps/x86_64/fpu/multiarch/s_modf-avx.c b/sysdeps/x86_64/fpu/multiarch/s_modf-avx.c
> >>>>>>>> index ab4f03db0e..25617c4fdc 100644
> >>>>>>>> --- a/sysdeps/x86_64/fpu/multiarch/s_modf-avx.c
> >>>>>>>> +++ b/sysdeps/x86_64/fpu/multiarch/s_modf-avx.c
> >>>>>>>> @@ -1,3 +1,8 @@
> >>>>>>>> +#include <sysdeps/x86/isa-level.h>
> >>>>>>>> +#if MINIMUM_X86_ISA_LEVEL < SSE4_1_X86_ISA_LEVEL
> >>>>>>>> +asm ("__trunc = __trunc_sse41");
> >>>>>>>
> >>>>>>> It looks odd for __modf_avx to call __trunc_sse41. But this is a
> >>>>>>> separate issue.
> >>>>>>> We do have s_truncf-avx.S, but it isn't used by default. I think we should
> >>>>>>> make it available for s_modf.c and s_modf-avx.c first.
> >>>>>>
> >>>>>> Can we use inline functions?
> >>>>>
> >>>>> It might an option, and I think it simplified the Makefile requirement.
> >>>>> I will update the patch with this.
> >>>>>
> >>>>> Do we have a way to force compiler always expand the builin
> >>>>> regardless of the optimization level?
> >>>>
> >>>> #include <math.h>
> >>>> #include <x86intrin.h>
> >>>>
> >>>> double
> >>>> __trunc (double x)
> >>>> {
> >>>> #ifdef __AVX__
> >>>> asm ("vroundsd $11, %1, %1, %0" : "=v" (x) : "v" (x));
> >>>> #elif defined __SSE4_1__
> >>>> asm ("roundsd $11, %1, %0" : "=x" (x) : "x" (x));
> >>>> #else
> >>>> x = trunc (x);
> >>>> #endif
> >>>> return x;
> >>>> }
> >>>>
> >>> Yeah, that was my idea to change it. But my question was if there is an extra
> >>> compiler option to avoid the libcall generation for -Os if the builtin lowering
> >>> does allow to use an instruction.
> >>
> >> I don't know.
> >>
> >>> As a side note, I think it is a missed optimization that gcc is creating a
> >>> libcall for x86_64-v2 and forward.
> >>
> >> Can you open a GCC bug and CC me?
> >>
> >
> > I opened:
> >
> > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121861
> >
>
> Thanks, using inline resulted in a simpler patch. I will send a new version.
Will adding -fno-trapping-math with -mavx/-msse4.1 break these 2 functions?
If not, we can add -fno-trapping-math with -mavx/-msse4.1.
--
H.J.
More information about the Libc-alpha
mailing list