On Sun, Jun 25, 2023 at 2:16 PM Jan Beulich <[email protected]> wrote:
>
> On 25.06.2023 07:06, Hongtao Liu wrote:
> > On Wed, Jun 21, 2023 at 2:28 PM Jan Beulich via Gcc-patches
> > <[email protected]> wrote:
> >>
> >> With respective two-operand bitwise operations now expressable by a
> >> single VPTERNLOG, add splitters to also deal with ior and xor
> >> counterparts of the original and-only case. Note that the splitters need
> >> to be separate, as the placement of "not" differs in the final insns
> >> (*iornot<mode>3, *xnor<mode>3) which are intended to pick up one half of
> >> the result.
> >>
> >> gcc/
> >>
> >> * config/i386/sse.md: New splitters to simplify
> >> not;vec_duplicate;{ior,xor} as vec_duplicate;{iornot,xnor}.
> >>
> >> gcc/testsuite/
> >>
> >> * gcc.target/i386/pr100711-4.c: New test.
> >> * gcc.target/i386/pr100711-5.c: New test.
> >>
> >> --- a/gcc/config/i386/sse.md
> >> +++ b/gcc/config/i386/sse.md
> >> @@ -17366,6 +17366,36 @@
> >> (match_dup 2)))]
> >> "operands[3] = gen_reg_rtx (<MODE>mode);")
> >>
> >> +(define_split
> >> + [(set (match_operand:VI 0 "register_operand")
> >> + (ior:VI
> >> + (vec_duplicate:VI
> >> + (not:<ssescalarmode>
> >> + (match_operand:<ssescalarmode> 1 "nonimmediate_operand")))
> >> + (match_operand:VI 2 "vector_operand")))]
> >> + "<MODE_SIZE> == 64 || TARGET_AVX512VL
> >> + || (TARGET_AVX512F && !TARGET_PREFER_AVX256)"
> >> + [(set (match_dup 3)
> >> + (vec_duplicate:VI (match_dup 1)))
> >> + (set (match_dup 0)
> >> + (ior:VI (not:VI (match_dup 3)) (match_dup 2)))]
> >> + "operands[3] = gen_reg_rtx (<MODE>mode);")
> >> +
> >> +(define_split
> >> + [(set (match_operand:VI 0 "register_operand")
> >> + (xor:VI
> >> + (vec_duplicate:VI
> >> + (not:<ssescalarmode>
> >> + (match_operand:<ssescalarmode> 1 "nonimmediate_operand")))
> >> + (match_operand:VI 2 "vector_operand")))]
> >> + "<MODE_SIZE> == 64 || TARGET_AVX512VL
> >> + || (TARGET_AVX512F && !TARGET_PREFER_AVX256)"
> >> + [(set (match_dup 3)
> >> + (vec_duplicate:VI (match_dup 1)))
> >> + (set (match_dup 0)
> >> + (not:VI (xor:VI (match_dup 3) (match_dup 2))))]
> >> + "operands[3] = gen_reg_rtx (<MODE>mode);")
> >> +
> > Can we merge this splitter(xor:not) into ior:not one with a code
> > iterator for xor,ior, They look the same except for the xor/ior.
>
> They're only almost the same: Note (ior (not )) vs (not (xor )) as
> the result of the splitting. The difference is necessary to fit
> with what patch 1 introduces (which in turn is the way it is to
> fit with what generic code transforms things to up front). (I had
> it the way you suggest initially, until I figured why one of the
> two would end up never being used.)
>
3597 /* Convert (XOR (NOT x) (NOT y)) to (XOR x y).
3598 Also convert (XOR (NOT x) y) to (NOT (XOR x y)), similarly for
3599 (NOT y). */
3600 {
3601 int num_negated = 0;
3602
3603 if (GET_CODE (op0) == NOT)
3604 num_negated++, op0 = XEXP (op0, 0);
3605 if (GET_CODE (op1) == NOT)
3606 num_negated++, op1 = XEXP (op1, 0);
It looks simplify_rtx plays the trick.
And it's documented.
8602@cindex @code{xor}, canonicalization of
8603@item
8604The only possible RTL expressions involving both bitwise exclusive-or
8605and bitwise negation are @code{(xor:@var{m} @var{x} @var{y})}
8606and @code{(not:@var{m} (xor:@var{m} @var{x} @var{y}))}.
Then the original patch LGTM.
> Jan
--
BR,
Hongtao
