PING^1 [PATCH v4 0/2] x86: Convert CONST_WIDE_INT/CONST_VECTOR to broadcast

H.J. Lu hjl.tools@gmail.com
Sat Jun 26 13:44:57 GMT 2021


On Thu, Jun 24, 2021 at 11:04 PM Hongtao Liu <crazylht@gmail.com> wrote:
>
> On Fri, Jun 25, 2021 at 2:01 PM Hongtao Liu <crazylht@gmail.com> wrote:
> >
> > I didn't receive
> > https://gcc.gnu.org/pipermail/gcc-patches/2021-June/572436.html in my
> > gmail account, does anyone know why?
> >
> >
> > >--- a/gcc/config/i386/i386-protos.h
> > >+++ b/gcc/config/i386/i386-protos.h
> > >@@ -260,6 +260,7 @@ extern void ix86_expand_mul_widen_hilo (rtx, rtx, rtx, bool, bool);
> > > extern void ix86_expand_sse2_mulv4si3 (rtx, rtx, rtx);
> > > extern void ix86_expand_sse2_mulvxdi3 (rtx, rtx, rtx);
> > > extern void ix86_expand_sse2_abs (rtx, rtx);
> > >+extern bool ix86_expand_integer_vec_duplicate (rtx *);
> > >
> > > /* In i386-c.c  */
> > > extern void ix86_target_macros (void);
> > >diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md
> > >index 2a34756be2a..f094e5a2586 100644
> > >--- a/gcc/config/i386/sse.md
> > >+++ b/gcc/config/i386/sse.md
> > >@@ -24570,3 +24570,24 @@
> > >   "TARGET_WIDEKL"
> > >   "aes<aeswideklvariant>\t{%0}"
> > >   [(set_attr "type" "other")])
> > >+
> > >+;; Modes handled by broadcast patterns.
> > >+(define_mode_iterator INT_BROADCAST_MODE
> > >+  [(V64QI "TARGET_AVX512F") (V32QI "TARGET_AVX") V16QI
> > >+   (V32HI "TARGET_AVX512F") (V16HI "TARGET_AVX") V8HI
> > >+   (V16SI "TARGET_AVX512F") (V8SI "TARGET_AVX") V4SI
> > >+   (V8DI "TARGET_AVX512F && TARGET_64BIT")
> > >+   (V4DI "TARGET_AVX && TARGET_64BIT") (V2DI "TARGET_64BIT")])
> > TARGET_AVX512BW is needed for V64QI and V32HImode.

I will fix them.

> > And consider the scenario, TARGET_AVX2 is needed for all other
> > 128/256-bit modes?
>
> Actually we can use vbroadcastss/d for SI/DImode which means
> V2DI/V4DI/V8SI/4SI only need TARGET_AVX.
>
> (define_insn "vec_dup<mode>"
>   [(set (match_operand:AVX_VEC_DUP_MODE 0 "register_operand" "=x,x,x,v,x")
>         (vec_duplicate:AVX_VEC_DUP_MODE
>           (match_operand:<ssescalarmode> 1 "nonimmediate_operand"
> "m,m,x,v,?x")))]
>   "TARGET_AVX"
>
> I saw in patch 1/2 TARGET_AVX2 is also needed for every modes.

I will allow AVX with SI/DI broadcast.

> > >+
> > >+;; Broadcast from an integer.  NB: Enable broadcast only if we can move
> > >+;; from GPR to SSE register directly.
> > >+(define_expand "vec_duplicate<mode>"
> > >+  [(set (match_operand:INT_BROADCAST_MODE 0 "register_operand")
> > >+ (vec_duplicate:INT_BROADCAST_MODE
> > >+   (match_operand:<ssescalarmode> 1 "general_operand")))]
> > >+  "TARGET_SSE2 && TARGET_INTER_UNIT_MOVES_TO_VEC"
> > Why not directly use TARGET_AVX2 here, not in ix86_expand_integer_vec_duplicate.

We allow SSE2 for variable broadcast.

> > Also define some predicate to restrict operand[1] to be const_int but
> > not const0_rtx and constm1_rtx.
> > w/ this, guess there's no need for FAIL? and you can directly
> > call ix86_expand_vector_init_duplicate.

general_operand is used to allow both variable and constant broadcast
so that I can use vector broadcast in memset with both variable and constant.
Changing the predicate won't help SSE nor AVX for constant broadcast since
the expander will generate a sequence of codes to satisfy the
predicate.  If it is
allowed to FAIL, the memset expander can use load from constant pool for
SSE2 and AVX.

> > >+{
> > >+  if (!ix86_expand_integer_vec_duplicate (operands))
> > >+    FAIL;
> > >+  DONE;
> > >
> >
> > On Fri, Jun 25, 2021 at 1:30 AM H.J. Lu via Gcc-patches
> > <gcc-patches@gcc.gnu.org> wrote:
> > >
> > > On Wed, Jun 9, 2021 at 4:39 PM H.J. Lu <hjl.tools@gmail.com> wrote:
> > > >
> > > > 1. Update move expanders to convert the CONST_WIDE_INT and CONST_VECTO
> > > > operands to vector broadcast from an integer with AVX2.
> > > > 2. Add ix86_gen_scratch_sse_rtx to return a scratch SSE register which
> > > > won't increase stack alignment requirement and blocks transformation by
> > > > the combine pass.
> > > >
> > > > A small benchmark:
> > > >
> > > > https://gitlab.com/x86-benchmarks/microbenchmark/-/tree/memset/broadcast
> > > >
> > > > shows that broadcast is a little bit faster on Intel Core i7-8559U:
> > > >
> > > > $ make
> > > > gcc -g -I. -O2   -c -o test.o test.c
> > > > gcc -g   -c -o memory.o memory.S
> > > > gcc -g   -c -o broadcast.o broadcast.S
> > > > gcc -g   -c -o vec_dup_sse2.o vec_dup_sse2.S
> > > > gcc -o test test.o memory.o broadcast.o vec_dup_sse2.o
> > > > ./test
> > > > memory      : 147215
> > > > broadcast   : 121213
> > > > vec_dup_sse2: 171366
> > > > $
> > > >
> > > > broadcast is also smaller:
> > > >
> > > > $ size memory.o broadcast.o
> > > >    text    data     bss     dec     hex filename
> > > >     132       0       0     132      84 memory.o
> > > >     122       0       0     122      7a broadcast.o
> > > > $
> > > >
> > > > 3. Update PR 87767 tests to expect integer broadcast instead of broadcast
> > > > from memory.
> > > > 4. Update avx512f_cond_move.c to expect integer broadcast.
> > > >
> > > > A small benchmark:
> > > >
> > > > https://gitlab.com/x86-benchmarks/microbenchmark/-/tree/vpaddd/broadcast
> > > >
> > > > shows that integer broadcast is faster than embedded memory broadcast:
> > > >
> > > > $ make
> > > > gcc -g -I. -O2 -march=skylake-avx512   -c -o test.o test.c
> > > > gcc -g   -c -o memory.o memory.S
> > > > gcc -g   -c -o broadcast.o broadcast.S
> > > > gcc -o test test.o memory.o broadcast.o
> > > > ./test
> > > > memory      : 425538
> > > > broadcast   : 375260
> > > > $
> > > >
> > > > 5. Update vec_duplicate to allow to fail so that backend can only allow
> > > > broadcasting an integer constant to a vector when broadcast instruction
> > > > is available.  This can be used by memset expander to avoid vec_duplicate
> > > > when loading from constant pool is more efficient.
> > > > 6. Add vec_duplicate<mode> expander and enable vec_duplicate from a
> > > > non-standard SSE constant integer only if vector broadcast is available.
> > > >
> > > > H.J. Lu (2):
> > > >   x86: Convert CONST_WIDE_INT/CONST_VECTOR to broadcast
> > > >   x86: Add vec_duplicate<mode> expander
> > > >
> > > >  gcc/config/i386/i386-expand.c                 | 208 +++++++++++++++++-
> > > >  gcc/config/i386/i386-protos.h                 |   3 +
> > > >  gcc/config/i386/i386.c                        |  13 ++
> > > >  gcc/config/i386/sse.md                        |  21 ++
> > > >  gcc/doc/md.texi                               |   2 -
> > > >  .../i386/avx512f-broadcast-pr87767-1.c        |   7 +-
> > > >  .../i386/avx512f-broadcast-pr87767-5.c        |   5 +-
> > > >  .../gcc.target/i386/avx512f_cond_move.c       |   4 +-
> > > >  .../i386/avx512vl-broadcast-pr87767-1.c       |  12 +-
> > > >  .../i386/avx512vl-broadcast-pr87767-5.c       |   9 +-
> > > >  gcc/testsuite/gcc.target/i386/pr100865-1.c    |  13 ++
> > > >  gcc/testsuite/gcc.target/i386/pr100865-10a.c  |  33 +++
> > > >  gcc/testsuite/gcc.target/i386/pr100865-10b.c  |   7 +
> > > >  gcc/testsuite/gcc.target/i386/pr100865-11a.c  |  23 ++
> > > >  gcc/testsuite/gcc.target/i386/pr100865-11b.c  |   8 +
> > > >  gcc/testsuite/gcc.target/i386/pr100865-12a.c  |  20 ++
> > > >  gcc/testsuite/gcc.target/i386/pr100865-12b.c  |   8 +
> > > >  gcc/testsuite/gcc.target/i386/pr100865-2.c    |  14 ++
> > > >  gcc/testsuite/gcc.target/i386/pr100865-3.c    |  15 ++
> > > >  gcc/testsuite/gcc.target/i386/pr100865-4a.c   |  16 ++
> > > >  gcc/testsuite/gcc.target/i386/pr100865-4b.c   |   9 +
> > > >  gcc/testsuite/gcc.target/i386/pr100865-5a.c   |  16 ++
> > > >  gcc/testsuite/gcc.target/i386/pr100865-5b.c   |   9 +
> > > >  gcc/testsuite/gcc.target/i386/pr100865-6a.c   |  16 ++
> > > >  gcc/testsuite/gcc.target/i386/pr100865-6b.c   |   9 +
> > > >  gcc/testsuite/gcc.target/i386/pr100865-7a.c   |  17 ++
> > > >  gcc/testsuite/gcc.target/i386/pr100865-7b.c   |   9 +
> > > >  gcc/testsuite/gcc.target/i386/pr100865-8a.c   |  24 ++
> > > >  gcc/testsuite/gcc.target/i386/pr100865-8b.c   |   7 +
> > > >  gcc/testsuite/gcc.target/i386/pr100865-9a.c   |  25 +++
> > > >  gcc/testsuite/gcc.target/i386/pr100865-9b.c   |   7 +
> > > >  31 files changed, 563 insertions(+), 26 deletions(-)
> > > >  create mode 100644 gcc/testsuite/gcc.target/i386/pr100865-1.c
> > > >  create mode 100644 gcc/testsuite/gcc.target/i386/pr100865-10a.c
> > > >  create mode 100644 gcc/testsuite/gcc.target/i386/pr100865-10b.c
> > > >  create mode 100644 gcc/testsuite/gcc.target/i386/pr100865-11a.c
> > > >  create mode 100644 gcc/testsuite/gcc.target/i386/pr100865-11b.c
> > > >  create mode 100644 gcc/testsuite/gcc.target/i386/pr100865-12a.c
> > > >  create mode 100644 gcc/testsuite/gcc.target/i386/pr100865-12b.c
> > > >  create mode 100644 gcc/testsuite/gcc.target/i386/pr100865-2.c
> > > >  create mode 100644 gcc/testsuite/gcc.target/i386/pr100865-3.c
> > > >  create mode 100644 gcc/testsuite/gcc.target/i386/pr100865-4a.c
> > > >  create mode 100644 gcc/testsuite/gcc.target/i386/pr100865-4b.c
> > > >  create mode 100644 gcc/testsuite/gcc.target/i386/pr100865-5a.c
> > > >  create mode 100644 gcc/testsuite/gcc.target/i386/pr100865-5b.c
> > > >  create mode 100644 gcc/testsuite/gcc.target/i386/pr100865-6a.c
> > > >  create mode 100644 gcc/testsuite/gcc.target/i386/pr100865-6b.c
> > > >  create mode 100644 gcc/testsuite/gcc.target/i386/pr100865-7a.c
> > > >  create mode 100644 gcc/testsuite/gcc.target/i386/pr100865-7b.c
> > > >  create mode 100644 gcc/testsuite/gcc.target/i386/pr100865-8a.c
> > > >  create mode 100644 gcc/testsuite/gcc.target/i386/pr100865-8b.c
> > > >  create mode 100644 gcc/testsuite/gcc.target/i386/pr100865-9a.c
> > > >  create mode 100644 gcc/testsuite/gcc.target/i386/pr100865-9b.c
> > > >
> > > > --
> > > > 2.31.1
> > > >
> > >
> > > PING^1.
> > >
> > > --
> > > H.J.
> >
> >
> >
> > --
> > BR,
> > Hongtao
>
>
>
> --
> BR,
> Hongtao



-- 
H.J.


More information about the Gcc-patches mailing list