On Mon, Jan 4, 2021 at 6:54 AM Hongtao Liu <crazy...@gmail.com> wrote: > > Hi: > The following patch adds define_insn_and_split to optimize > > vpmovmskb %xmm0, %eax > - movzwl %ax, %eax > notl %eax > > Bootstrapped/regtested on x86_64-linux-gnu {,-m32}. > Ok for trunk? > > gcc/ChangeLog > PR target/98461 > * config/i386/sse.md (*sse2_pmovskb_zexthisi): New > define_insn_and_split for zero_extend of subreg HI of pmovskb > result. > > gcc/testsuite/ChangeLog > * gcc.target/i386/sse-pr98461-2.c: New test. > --- > gcc/config/i386/sse.md | 11 +++++++++++ > gcc/testsuite/gcc.target/i386/sse2-pr98461-2.c | 13 +++++++++++++ > 2 files changed, 24 insertions(+) > create mode 100644 gcc/testsuite/gcc.target/i386/sse2-pr98461-2.c > > diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md > index d84103807ff..4ed6b9ae476 100644 > --- a/gcc/config/i386/sse.md > +++ b/gcc/config/i386/sse.md > @@ -16099,6 +16099,17 @@ (define_insn "*sse2_pmovmskb_ext" > (set_attr "prefix" "maybe_vex") > (set_attr "mode" "SI")]) > > +(define_insn_and_split "*sse2_pmovskb_zexthisi" > + [(set (match_operand:SI 0 "register_operand") > + (zero_extend:SI (subreg:HI (unspec:SI > + [(match_operand:V16QI 1 "register_operand")] > + UNSPEC_MOVMSK) 0)))] > + "TARGET_SSE2"
This needs ix86_pre_reload_split () in insn predicate. Uros. > + "#" > + "&& 1" > + [(set (match_dup 0) > + (unspec:SI [(match_dup 1)] UNSPEC_MOVMSK))]) > + > (define_split > [(set (match_operand:SI 0 "register_operand") > (unspec:SI > diff --git a/gcc/testsuite/gcc.target/i386/sse2-pr98461-2.c > b/gcc/testsuite/gcc.target/i386/sse2-pr98461-2.c > new file mode 100644 > index 00000000000..60fc1f3e9c1 > --- /dev/null > +++ b/gcc/testsuite/gcc.target/i386/sse2-pr98461-2.c > @@ -0,0 +1,13 @@ > +/* PR target/98461 */ > +/* { dg-do compile } */ > +/* { dg-options "-O2 -msse2 -mno-sse3 -masm=att" } */ > +/* { dg-final { scan-assembler-times "\tpmovmskb\t" 1 } } */ > +/* { dg-final { scan-assembler-not "\tmovzwl" } } */ > +/* { dg-final { scan-assembler-times "\tnotl" 1 } } */ > + > +#include <immintrin.h> > + > +unsigned int movemask_not1(__m128i logical) { > + unsigned short res = (unsigned short)(_mm_movemask_epi8(logical)); > + return ~res; > +} > -- > 2.18.1 > > > -- > BR, > Hongtao