On 2/10/19, H.J. Lu <hjl.to...@gmail.com> wrote: > In 64-bit mode, SSE2 can be used to emulate MMX instructions without > 3DNOW. We can use SSE2 to support 64-bit vectors. > > PR target/89021 > * config/i386/i386.c (ix86_set_reg_reg_cost): Also support > VALID_MMX_WITH_SSE_REG_MODE. > (ix86_vector_mode_supported_p): Likewise. > * config/i386/i386.h (TARGET_MMX_WITH_SSE): New. > (TARGET_MMX_WITH_SSE_P): Likewise. > (VALID_MMX_WITH_SSE_REG_MODE): Likewise. > --- > gcc/config/i386/i386.c | 3 +++ > gcc/config/i386/i386.h | 14 ++++++++++++++ > 2 files changed, 17 insertions(+) > > diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c > index 12bc7926f86..ba02c26c8b2 100644 > --- a/gcc/config/i386/i386.c > +++ b/gcc/config/i386/i386.c > @@ -40235,6 +40235,7 @@ ix86_set_reg_reg_cost (machine_mode mode) > || (TARGET_AVX && VALID_AVX256_REG_MODE (mode)) > || (TARGET_SSE2 && VALID_SSE2_REG_MODE (mode)) > || (TARGET_SSE && VALID_SSE_REG_MODE (mode)) > + || (TARGET_MMX_WITH_SSE && VALID_MMX_WITH_SSE_REG_MODE (mode)) > || (TARGET_MMX && VALID_MMX_REG_MODE (mode)))
With V2SFmode out of the way (see below) we can finaly use (TARGET_MMX || TARGET_MMX_WITH_SSE) && VALID_MMX_REG_MODE (mode). This is a cost function, and we do have DImode and SImode in SSE registers. > units = GET_MODE_SIZE (mode); > } > @@ -44057,6 +44058,8 @@ ix86_vector_mode_supported_p (machine_mode mode) > return true; > if (TARGET_SSE2 && VALID_SSE2_REG_MODE (mode)) > return true; > + if (TARGET_MMX_WITH_SSE && VALID_MMX_WITH_SSE_REG_MODE (mode)) > + return true; Assuming middle end won't ask for scalar modes, and with V2SFmode out of the way, we can also use (TARGET_MMX || TARGET_MMX_WITH_SSE) && VALID_MMX_REG_MODE (mode) here. > if (TARGET_AVX && VALID_AVX256_REG_MODE (mode)) > return true; > if (TARGET_AVX512F && VALID_AVX512F_REG_MODE (mode)) > diff --git a/gcc/config/i386/i386.h b/gcc/config/i386/i386.h > index 83b025e0cf5..3ae0900caa0 100644 > --- a/gcc/config/i386/i386.h > +++ b/gcc/config/i386/i386.h > @@ -201,6 +201,13 @@ see the files COPYING3 and COPYING.RUNTIME > respectively. If not, see > #define TARGET_16BIT TARGET_CODE16 > #define TARGET_16BIT_P(x) TARGET_CODE16_P(x) > > +/* In 64-bit mode, SSE2 can be used to emulate MMX instructions. > + FIXME: All 3DNOW patterns needs to be updated with SSE emulation. */ > +#define TARGET_MMX_WITH_SSE \ > + (TARGET_64BIT && TARGET_SSE2 && !TARGET_3DNOW) > +#define TARGET_MMX_WITH_SSE_P(x) \ > + (TARGET_64BIT_P (x) && TARGET_SSE2_P (x) && !TARGET_3DNOW_P (x)) The above is not acceptable, the choice of native MMX should not depend on -m3dnow flag. So, instead of FIXME, please leave partial conversion of V2SF mode out of the patchset, V2SF values should still live in MMX registers. Actually, -m3dnow is a dead end, deprecated insn set, so I see no reason to emulate V2SF at all. SSE doesn't have native V2SF instructions, and emulating reciprocals will trap due to 0.0 in the high two elements. There are also hard to emulate reciprocal step instructions. Also, the purpose of the patchset is to convert MMX builtins, since SSE builtins depend on them, so at the end we can avoid enabling MMX registers with -msse, and thus making -mmmx orthogonal to -msse. We don't wan to sneak in an autovectorization of V2SF with the patchset. > #include "config/vxworks-dummy.h" > > #include "config/i386/i386-opts.h" > @@ -1143,6 +1150,13 @@ extern const char *host_detect_local_cpu (int argc, > const char **argv); > || (MODE) == V4SImode || (MODE) == V4SFmode || (MODE) == V8HImode \ > || (MODE) == TFmode || (MODE) == V1TImode) > > +/* NB: Don't use VALID_MMX_REG_MODE with TARGET_MMX_WITH_SSE since we > + want to include only 8-byte vector modes, like V2SFmode, but not No, we don't want to include V2SF mode vectors. > + DImode nor SImode. */ > +#define VALID_MMX_WITH_SSE_REG_MODE(MODE) \ > + ((MODE) == V1DImode || (MODE) == V8QImode || (MODE) == V4HImode \ > + || (MODE) == V2SImode || (MODE) == V2SFmode) Without V2SFmode, the above definition is unneeded. Uros. > #define VALID_SSE2_REG_MODE(MODE) \ > ((MODE) == V16QImode || (MODE) == V8HImode || (MODE) == V2DFmode \ > || (MODE) == V2DImode || (MODE) == DFmode) > -- > 2.20.1 > >
