On Wed, 18 Aug 2021, Hongtao Liu wrote:
> On Wed, Aug 18, 2021 at 11:24 AM Hongtao Liu <[email protected]> wrote:
> >
> > On Tue, Aug 17, 2021 at 10:43 PM Richard Biener via Gcc-patches
> > <[email protected]> wrote:
> > >
> > > On Tue, Aug 17, 2021 at 3:29 PM Richard Biener via Gcc-patches
> > > <[email protected]> wrote:
> > > >
> > > > This is an attempt to start moving the x86 backend to use
> > > > standard pattern names for [mask_]gather_load and [mask_]scatter_store
> > > > rather than using the builtin_{gather,scatter} target hooks.
> > > >
> > > > I've started with AVX2 gathers and given x86 only supports masked
> > > > gather I only implemented mask_gather_load. Note while for
> > > > the builtin_gather case the vectorizer will provide an all-true
> > > > mask operand for non-masked gathers this capability does not
> > > > exist for the IFN path yet, so only testcases with actual masked
> > > > gathers will work.
> > > >
> > > > If this looks reasonable on the backend side I'll see to first
> > > > complete the vectorizer part, ripping out the target hook and
> > > > arranging for the missing pieces. Another one is the support
> > > > for SImode indices with DFmode data which requires unpacking
> > > > the index vector and actually recognizing the IFN.
> > > >
> > > > 2021-08-17 Richard Biener <[email protected]>
> > > >
> > > > * tree-vect-data-refs.c (vect_check_gather_scatter):
> > > > Always use internal functions.
> > > > * config/i386/sse.md
> > > > (mask_gather_load<mode><vec_gather_idxsi>): New expander.
> > > > (mask_gather_load<mode><vec_gather_idxdi>): Likewise.
> > > > ---
> > > > gcc/config/i386/sse.md | 56 +++++++++++++++++++++++++++++++++++++++
> > > > gcc/tree-vect-data-refs.c | 4 +--
> > > > 2 files changed, 57 insertions(+), 3 deletions(-)
> > > >
> > > > diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md
> > > > index 3957c86c3df..40bec98d9f7 100644
> > > > --- a/gcc/config/i386/sse.md
> > > > +++ b/gcc/config/i386/sse.md
> > > > @@ -23232,12 +23232,22 @@
> > > > (V2DF "V4SI") (V4DF "V4SI") (V8DF "V8SI")
> > > > (V4SI "V4SI") (V8SI "V8SI") (V16SI "V16SI")
> > > > (V4SF "V4SI") (V8SF "V8SI") (V16SF "V16SI")])
> > > > +(define_mode_attr vec_gather_idxsi
> > > > + [(V2DI "v4si") (V4DI "v4si") (V8DI "v8si")
> > > > + (V2DF "v4si") (V4DF "v4si") (V8DF "v8si")
> > > > + (V4SI "v4si") (V8SI "v8si") (V16SI "v16si")
> > > > + (V4SF "v4si") (V8SF "v8si") (V16SF "v16si")])
> > > >
> > > > (define_mode_attr VEC_GATHER_IDXDI
> > > > [(V2DI "V2DI") (V4DI "V4DI") (V8DI "V8DI")
> > > > (V2DF "V2DI") (V4DF "V4DI") (V8DF "V8DI")
> > > > (V4SI "V2DI") (V8SI "V4DI") (V16SI "V8DI")
> > > > (V4SF "V2DI") (V8SF "V4DI") (V16SF "V8DI")])
> > > > +(define_mode_attr vec_gather_idxdi
> > > > + [(V2DI "v2di") (V4DI "v4di") (V8DI "v8di")
> > > > + (V2DF "v2di") (V4DF "v4di") (V8DF "v8di")
> > > > + (V4SI "v2di") (V8SI "v4di") (V16SI "v8di")
> > > > + (V4SF "v2di") (V8SF "v4di") (V16SF "v8di")])
> > > >
> > > > (define_mode_attr VEC_GATHER_SRCDI
> > > > [(V2DI "V2DI") (V4DI "V4DI") (V8DI "V8DI")
> > > > @@ -23245,6 +23255,29 @@
> > > > (V4SI "V4SI") (V8SI "V4SI") (V16SI "V8SI")
> > > > (V4SF "V4SF") (V8SF "V4SF") (V16SF "V8SF")])
> > > >
> > > > +(define_expand "mask_gather_load<mode><vec_gather_idxsi>"
> > > > + [(parallel [(set (match_operand:VEC_GATHER_MODE 0 "register_operand")
> > > > + (unspec:VEC_GATHER_MODE
> > > > + [(pc)
> > > > + (mem:<ssescalarmode>
> > > > + (match_par_dup 6
> > > > + [(match_operand 1 "vsib_address_operand")
> > > > + (match_operand:<VEC_GATHER_IDXSI>
> > > > + 2 "register_operand")
> > > > + (match_operand:SI 4 "const1248_operand")
> > > > + (match_operand:SI 3 "const0_operand")]))
> > > > + (mem:BLK (scratch))
> > > > + (match_operand:<sseintvecmode> 5
> > > > "register_operand")]
> > >
> > > One problem of these is that when AVX512[VL] is enabled we get a AVX512
> > > mask
> > > mode here and while the internal function expansion check succeeds (it
> > > never checks
> > > the mask operand!), RTL expansion fails unexpectedly because of this
> > > mismatch.
> > >
> > > I suppose a more complicated define_mode_attr might do the trick or do I
> > I don't think define_mode_attr supports conditional selection based on
> > target feature.
> > > need to add && !TARGET_AVX512F to these expanders?
> > There'll be a duplicated definition of mask_loadv8sfv8si for avx512
> > and no-avx512 versions.
> Note gather_loadmn can be shared by avx512 and non-avx512 version, we
> can handle mask mode in the preparation statements based on target
> feature.
> > >
> > The best way is like maskloadmn to accept different mask modes in its
> > name, but that may create too much complexity in the middle-end to
> > choose between avx2 mask_gather_load and avx512 mask_gather_load.
So when we have -mavx512f only then we can gather V8DF using a QImode
maks with AVX512 gather but we have to use AVX2 for V4DF gather
using a V4DImode mask. With -mavx512vl we use a QImode mask for that
case as well.
Note the
operands[5] = gen_lowpart_SUBREG (<VEC_GATHER_SRCDI>mode, operands[5]);
is because the define_insns for the gathers somehow expect odd mask
modes (the vectorizer generates an integer vector mode matching the
layout of the result vector). I suppose we could somehow fix that
but I wanted to change as little as possible for now.
The following seems to "work" in generating AVX512 variants for
-mavx512f -mavx512vl -mprefer-vector-width=512 and AVX2 variants
when I drop the -mavx512vl arg:
(define_expand "mask_gather_load<mode><vec_gather_idxdi>"
[(match_operand:VEC_GATHER_MODE 0 "register_operand")
(match_operand 1 "vsib_address_operand")
(match_operand:<VEC_GATHER_IDXDI> 2 "register_operand")
(match_operand:SI 3 "const0_operand")
(match_operand:SI 4 "const1248_operand")
(match_operand 5 "register_operand")]
"TARGET_AVX2 && TARGET_USE_GATHER"
{
if (TARGET_AVX512VL)
{
rtx scratch = gen_reg_rtx (<MODE>mode);
emit_move_insn (scratch, CONST0_RTX(<MODE>mode));
emit_insn (gen_avx512vl_gatherdi<mode>
(operands[0], scratch,
operands[1], operands[2], operands[5], operands[4]));
}
else
{
operands[5] = gen_lowpart_SUBREG (<VEC_GATHER_SRCDI>mode,
operands[5]);
rtx scratch = gen_reg_rtx (<MODE>mode);
emit_move_insn (scratch, CONST0_RTX(<MODE>mode));
emit_insn (gen_avx2_gatherdi<mode>
(operands[0], scratch,
operands[1], operands[2], operands[5], operands[4]));
}
DONE;
})
But I see quite horrible code for an always-true mask. From GIMPLE
# ivtmp.17_34 = PHI <ivtmp.17_19(3), 0(2)>
vect__4.5_24 = MEM <vector(8) int> [(int *)idx_12(D) + ivtmp.17_34 * 1];
vect_patt_28.6_23 = [vec_unpack_lo_expr] vect__4.5_24;
vect_patt_28.6_22 = [vec_unpack_hi_expr] vect__4.5_24;
vect_patt_27.7_17 = .MASK_GATHER_LOAD (_18, vect_patt_28.6_23, 8, { 0.0,
0.0, 0.0, 0.0 }, { -1, -1, -1, -1 });
vect_patt_27.8_16 = .MASK_GATHER_LOAD (_18, vect_patt_28.6_22, 8, { 0.0,
0.0, 0.0, 0.0 }, { -1, -1, -1, -1 });
MEM <vector(4) double> [(double *)&a + ivtmp.17_34 * 2] =
vect_patt_27.7_17;
MEM <vector(4) double> [(double *)&a + 32B + ivtmp.17_34 * 2] =
vect_patt_27.8_16;
ivtmp.17_19 = ivtmp.17_34 + 32;
if (ivtmp.17_19 != 4096)
we end up with
movl $-1, %edx
xorl %eax, %eax
vxorpd %xmm1, %xmm1, %xmm1
kmovw %edx, %k1
.p2align 4,,10
.p2align 3
.L2:
vpmovsxdq (%rsi,%rax), %ymm0
vmovdqu (%rsi,%rax), %ymm4
vmovapd %ymm1, %ymm3
kmovw %k1, %k2
vmovapd %ymm1, %ymm2
kmovw %k1, %k3
vgatherqpd (%rdi,%ymm0,8), %ymm3{%k2}
vextracti128 $0x1, %ymm4, %xmm0
vpmovsxdq %xmm0, %ymm0
vgatherqpd (%rdi,%ymm0,8), %ymm2{%k3}
vmovapd %ymm3, a(%rax,%rax)
vmovapd %ymm2, a+32(%rax,%rax)
addq $32, %rax
cmpq $4096, %rax
jne .L2
I wonder why we have those extra kmovw here. I see there's
extra (pc) patterns for always-true masks(?) like in the AVX2
case. But in the end we have to use {%kN} instructions, no?
We can always provide non-masked gather_load expanders if that's
better here.
Richard.
