Re: [PATCH v1 8/8] RISC-V: bitmanip: relax minmax to operate on GPR

[Philipp Tomsich <philipp.tomsich@vrull.eu>]

Kito,

Unless I am missing something, the problem is not the relaxation to
GPR, but rather the sign-extending pattern I had squashed into the
same patch.
If you disable "<bitmanip_optab>si3_sext", a sext.w will be have to be
emitted after the 'max' and before the return (or before the SImode
output is consumed as a DImode), pushing the REE opportunity to a
subsequent consumer (e.g. an addw).

This will generate
   foo6:
      max a0,a0,a1
      sext.w a0,a0
      ret
which (assuming that the inputs to max are properly sign-extended
SImode values living in DImode registers) will be the same as
performing the two sext.w before the max.

Having a second set of eyes on this is appreciated — let me know if
you agree and I'll revise, once I have collected feedback on the
remaining patches of the series.

Philipp.


On Thu, 11 Nov 2021 at 17:00, Kito Cheng <kito.cheng@gmail.com> wrote:
>
> Hi Philipp:
>
> We can't pretend we have SImode min/max instruction without that semantic.
> Give this testcase, x86 and rv64gc print out 8589934592 8589934591 = 0,
> but with this patch and compile with rv64gc_zba_zbb -O3, the output
> become 8589934592 8589934591 = 8589934592
>
> -------------Testcase---------------
> #include <stdio.h>
> long long __attribute__((noinline, noipa))
> foo6(long long a, long long b)
> {
>   int xa = a;
>   int xb = b;
>   return (xa > xb ? xa : xb);
> }
> int main() {
>   long long a = 0x200000000ll;
>   long long b = 0x1ffffffffl;
>   long long c = foo6(a, b);
>   printf ("%lld %lld = %lld\n", a, b, c);
>   return 0;
> }
> --------------------------------------
> v64gc_zba_zbb -O3 w/o this patch:
> foo6:
>         sext.w  a1,a1
>         sext.w  a0,a0
>         max     a0,a0,a1
>         ret
>
> --------------------------------------
> v64gc_zba_zbb -O3 w/ this patch:
> foo6:
>         max     a0,a0,a1
>         ret
>
> On Thu, Nov 11, 2021 at 10:10 PM Philipp Tomsich
> <philipp.tomsich@vrull.eu> wrote:
> >
> > While min/minu/max/maxu instructions are provided for XLEN only, these
> > can safely operate on GPRs (i.e. SImode or DImode for RV64): SImode is
> > always sign-extended, which ensures that the XLEN-wide instructions
> > can be used for signed and unsigned comparisons on SImode yielding a
> > correct ordering of value.
> >
> > This commit
> >  - relaxes the minmax pattern to express for GPR (instead of X only),
> >    providing both a si3 and di3 expansion on RV64
> >  - adds a sign-extending form for thee si3 pattern for RV64 to all REE
> >    to eliminate redundant extensions
> >  - adds test-cases for both
> >
> > gcc/ChangeLog:
> >
> >         * config/riscv/bitmanip.md: Relax minmax to GPR (i.e SImode or
> >           DImode) on RV64.
> >         * config/riscv/bitmanip.md (<bitmanip_optab>si3_sext): Add
> >           pattern for REE.
> >
> > gcc/testsuite/ChangeLog:
> >
> >         * gcc.target/riscv/zbb-min-max.c: Add testcases for SImode
> >           operands checking that no redundant sign- or zero-extensions
> >           are emitted.
> >
> > Signed-off-by: Philipp Tomsich <philipp.tomsich@vrull.eu>
> > ---
> >
> >  gcc/config/riscv/bitmanip.md                 | 14 +++++++++++---
> >  gcc/testsuite/gcc.target/riscv/zbb-min-max.c | 20 +++++++++++++++++---
> >  2 files changed, 28 insertions(+), 6 deletions(-)
> >
> > diff --git a/gcc/config/riscv/bitmanip.md b/gcc/config/riscv/bitmanip.md
> > index 000deb48b16..2a28f78f5f6 100644
> > --- a/gcc/config/riscv/bitmanip.md
> > +++ b/gcc/config/riscv/bitmanip.md
> > @@ -260,13 +260,21 @@ (define_insn "bswap<mode>2"
> >    [(set_attr "type" "bitmanip")])
> >
> >  (define_insn "<bitmanip_optab><mode>3"
> > -  [(set (match_operand:X 0 "register_operand" "=r")
> > -        (bitmanip_minmax:X (match_operand:X 1 "register_operand" "r")
> > -                          (match_operand:X 2 "register_operand" "r")))]
> > +  [(set (match_operand:GPR 0 "register_operand" "=r")
> > +        (bitmanip_minmax:GPR (match_operand:GPR 1 "register_operand" "r")
> > +                            (match_operand:GPR 2 "register_operand" "r")))]
> >    "TARGET_ZBB"
> >    "<bitmanip_insn>\t%0,%1,%2"
> >    [(set_attr "type" "bitmanip")])
> >
> > +(define_insn "<bitmanip_optab>si3_sext"
> > +  [(set (match_operand:DI 0 "register_operand" "=r")
> > +        (sign_extend:DI (bitmanip_minmax:SI (match_operand:SI 1 "register_operand" "r")
> > +                            (match_operand:SI 2 "register_operand" "r"))))]
> > +  "TARGET_64BIT && TARGET_ZBB"
> > +  "<bitmanip_insn>\t%0,%1,%2"
> > +  [(set_attr "type" "bitmanip")])
> > +
> >  ;; orc.b (or-combine) is added as an unspec for the benefit of the support
> >  ;; for optimized string functions (such as strcmp).
> >  (define_insn "orcb<mode>2"
> > diff --git a/gcc/testsuite/gcc.target/riscv/zbb-min-max.c b/gcc/testsuite/gcc.target/riscv/zbb-min-max.c
> > index f44c398ea08..7169e873551 100644
> > --- a/gcc/testsuite/gcc.target/riscv/zbb-min-max.c
> > +++ b/gcc/testsuite/gcc.target/riscv/zbb-min-max.c
> > @@ -1,5 +1,5 @@
> >  /* { dg-do compile } */
> > -/* { dg-options "-march=rv64gc_zbb -mabi=lp64 -O2" } */
> > +/* { dg-options "-march=rv64gc_zba_zbb -mabi=lp64 -O2" } */
> >
> >  long
> >  foo1 (long i, long j)
> > @@ -25,7 +25,21 @@ foo4 (unsigned long i, unsigned long j)
> >    return i > j ? i : j;
> >  }
> >
> > +unsigned int
> > +foo5(unsigned int a, unsigned int b)
> > +{
> > +  return a > b ? a : b;
> > +}
> > +
> > +int
> > +foo6(int a, int b)
> > +{
> > +  return a > b ? a : b;
> > +}
> > +
> >  /* { dg-final { scan-assembler-times "min" 3 } } */
> > -/* { dg-final { scan-assembler-times "max" 3 } } */
> > +/* { dg-final { scan-assembler-times "max" 4 } } */
> >  /* { dg-final { scan-assembler-times "minu" 1 } } */
> > -/* { dg-final { scan-assembler-times "maxu" 1 } } */
> > +/* { dg-final { scan-assembler-times "maxu" 3 } } */
> > +/* { dg-final { scan-assembler-not "zext.w" } } */
> > +/* { dg-final { scan-assembler-not "sext.w" } } */
> > --
> > 2.32.0
> >