The implementation fails to handle this test case properly:
typedef double __attribute__((vector_size(32))) v4df;
void use1(double);
__attribute__((noipa)) double use(double)
{
register double x asm("f24") = 114.514;
__asm__("" : "+f" (x));
return x;
}
void test(void)
{
register v4df x asm("f24") = {1, 2, 3, 4};
__asm__("" : "+f" (x));
use(x[1]);
use1(x[3]);
}
Here use() attempts to save and restore f24, but it uses fst.d/fld.d,
clobbering the high 192 bits of xr24. Now test() passes a wrong value
of x[3] to use1().
Note that saving and restoring f24 with xvst/xvld in use() won't really
fix the issue because in real life use() can be in another translation
unit (or even a shared library) compiled with -mno-lsx. So it seems we
need to tell the compiler "a function call may clobber the high bits of
a vector register even if the corresponding floating-point register is
saved". I'm not sure how to accomplish this...
On Tue, 2023-08-15 at 09:05 +0800, Chenghui Pan wrote:
> This is an update of:
> https://gcc.gnu.org/pipermail/gcc-patches/2023-August/626194.html
>
> This version of patch set only introduces some small simplications of
> implementation. Because I missed the size limitation of mail size, the
> huge testsuite patches of v2 and v3 are not shown in the mail list.
> So,
> testsuite patches are splited from this patch set again and will be
> submitted
> independently in the future.
>
> Binutils-gdb introduced LSX/LASX support since 2.41 release:
> https://lists.gnu.org/archive/html/info-gnu/2023-07/msg00009.html
>
> Brief history of patch set version:
> v1 -> v2:
> - Reduce usage of "unspec" in RTL template.
> - Append Support of ADDR_REG_REG in LSX and LASX.
> - Constraint docs are appended in gcc/doc/md.texi and ccomment block.
> - Codes related to vecarg are removed.
> - Testsuite of LSX and LASX is added in v2. (Because of the size
> limitation of
> mail list, these patches are not shown)
> - Adjust the loongarch_expand_vector_init() function to reduce
> instruction
> output amount.
> - Some minor implementation changes of RTL templates.
>
> v2 -> v3:
> - Revert vabsd/xvabsd RTL templates to unspec impl.
> - Resolve warning in gcc/config/loongarch/loongarch.cc when
> bootstrapping
> with BOOT_CFLAGS="-O2 -ftree-vectorize -fno-vect-cost-model -mlasx".
> - Remove redundant definitions in lasxintrin.h.
> - Refine commit info.
>
> Lulu Cheng (6):
> LoongArch: Add Loongson SX vector directive compilation framework.
> LoongArch: Add Loongson SX base instruction support.
> LoongArch: Add Loongson SX directive builtin function support.
> LoongArch: Add Loongson ASX vector directive compilation framework.
> LoongArch: Add Loongson ASX base instruction support.
> LoongArch: Add Loongson ASX directive builtin function support.
>
> gcc/config.gcc | 2 +-
> gcc/config/loongarch/constraints.md | 131 +-
> .../loongarch/genopts/loongarch-strings | 4 +
> gcc/config/loongarch/genopts/loongarch.opt.in | 12 +-
> gcc/config/loongarch/lasx.md | 5122 ++++++++++++++++
> gcc/config/loongarch/lasxintrin.h | 5338
> +++++++++++++++++
> gcc/config/loongarch/loongarch-builtins.cc | 2686 ++++++++-
> gcc/config/loongarch/loongarch-c.cc | 18 +
> gcc/config/loongarch/loongarch-def.c | 6 +
> gcc/config/loongarch/loongarch-def.h | 9 +-
> gcc/config/loongarch/loongarch-driver.cc | 10 +
> gcc/config/loongarch/loongarch-driver.h | 2 +
> gcc/config/loongarch/loongarch-ftypes.def | 666 +-
> gcc/config/loongarch/loongarch-modes.def | 39 +
> gcc/config/loongarch/loongarch-opts.cc | 89 +-
> gcc/config/loongarch/loongarch-opts.h | 3 +
> gcc/config/loongarch/loongarch-protos.h | 35 +
> gcc/config/loongarch/loongarch-str.h | 3 +
> gcc/config/loongarch/loongarch.cc | 4586 +++++++++++++-
> gcc/config/loongarch/loongarch.h | 117 +-
> gcc/config/loongarch/loongarch.md | 56 +-
> gcc/config/loongarch/loongarch.opt | 12 +-
> gcc/config/loongarch/lsx.md | 4481 ++++++++++++++
> gcc/config/loongarch/lsxintrin.h | 5181 ++++++++++++++++
> gcc/config/loongarch/predicates.md | 333 +-
> gcc/doc/md.texi | 11 +
> 26 files changed, 28668 insertions(+), 284 deletions(-)
> create mode 100644 gcc/config/loongarch/lasx.md
> create mode 100644 gcc/config/loongarch/lasxintrin.h
> create mode 100644 gcc/config/loongarch/lsx.md
> create mode 100644 gcc/config/loongarch/lsxintrin.h
>
--
Xi Ruoyao <[email protected]>
School of Aerospace Science and Technology, Xidian University
