On Fri, Jul 7, 2023 at 7:31 AM liuhongt <hongtao....@intel.com> wrote:
>
> > Please split the above pattern into two, one emitting UNSPEC_IEEE_MAX
> > and the other emitting UNSPEC_IEEE_MIN.
> Splitted.
>
> > The test involves blendv instruction, which is SSE4.1, so it is
> > pointless to test it without -msse4.1. Please add -msse4.1 instead of
> > -march=x86_64 and use sse4_runtime target selector, as is the case
> > with gcc.target/i386/pr90358.c.
> Changed.
>
> > Please also use -msse4.1 instead of -march here. With -mfpmath=sse,
> > the test is valid also for 32bit targets, you should use -msseregparm
> > additional options for ia32 (please see gcc.target/i386/pr43546.c
> > testcase) in the same way as -mregparm to pass SSE arguments in
> > registers.
> 32-bit target still failed to do condition elimination for DFmode due to
> below code in rtx_cost
>
> /* A size N times larger than UNITS_PER_WORD likely needs N times as
> many insns, taking N times as long. */
> factor = mode_size > UNITS_PER_WORD ? mode_size / UNITS_PER_WORD : 1;
>
> It looks like a separate issue for DFmode operation under 32-bit target.
>
> I've enable 32-bit for the testcase, but only scan for minss/maxss
> currently.
>
> Here's updated patch.
> Bootstrapped and regtested on x86_64-pc-linux-gnu{-m32,}.
> Ok for trunk?
>
> We have ix86_expand_sse_fp_minmax to detect min/max sematics, but
> it requires rtx_equal_p for cmp_op0/cmp_op1 and if_true/if_false, for
> the testcase in the PR, there's an extra move from cmp_op0 to if_true,
> and it failed ix86_expand_sse_fp_minmax.
>
> This patch adds pre_reload splitter to detect the min/max pattern.
>
> Operands order in MINSS matters for signed zero and NANs, since the
> instruction always returns second operand when any operand is NAN or
> both operands are zero.
>
> gcc/ChangeLog:
>
> PR target/110170
> * config/i386/i386.md (*ieee_max<mode>3_1): New pre_reload
> splitter to detect fp max pattern.
> (*ieee_min<mode>3_1): Ditto, but for fp min pattern.
>
> gcc/testsuite/ChangeLog:
>
> * g++.target/i386/pr110170.C: New test.
> * gcc.target/i386/pr110170.c: New test.
OK with a testcase fix below.
Uros.
> ---
> gcc/config/i386/i386.md | 43 +++++++++++++
> gcc/testsuite/g++.target/i386/pr110170.C | 78 ++++++++++++++++++++++++
> gcc/testsuite/gcc.target/i386/pr110170.c | 21 +++++++
> 3 files changed, 142 insertions(+)
> create mode 100644 gcc/testsuite/g++.target/i386/pr110170.C
> create mode 100644 gcc/testsuite/gcc.target/i386/pr110170.c
>
> diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md
> index a82cc353cfd..6f415f899ae 100644
> --- a/gcc/config/i386/i386.md
> +++ b/gcc/config/i386/i386.md
> @@ -23163,6 +23163,49 @@ (define_insn "*ieee_s<ieee_maxmin><mode>3"
> (set_attr "type" "sseadd")
> (set_attr "mode" "<MODE>")])
>
> +;; Operands order in min/max instruction matters for signed zero and NANs.
> +(define_insn_and_split "*ieee_max<mode>3_1"
> + [(set (match_operand:MODEF 0 "register_operand")
> + (unspec:MODEF
> + [(match_operand:MODEF 1 "register_operand")
> + (match_operand:MODEF 2 "register_operand")
> + (lt:MODEF
> + (match_operand:MODEF 3 "register_operand")
> + (match_operand:MODEF 4 "register_operand"))]
> + UNSPEC_BLENDV))]
> + "SSE_FLOAT_MODE_P (<MODE>mode) && TARGET_SSE_MATH
> + && (rtx_equal_p (operands[1], operands[3])
> + && rtx_equal_p (operands[2], operands[4]))
> + && ix86_pre_reload_split ()"
> + "#"
> + "&& 1"
> + [(set (match_dup 0)
> + (unspec:MODEF
> + [(match_dup 2)
> + (match_dup 1)]
> + UNSPEC_IEEE_MAX))])
> +
> +(define_insn_and_split "*ieee_min<mode>3_1"
> + [(set (match_operand:MODEF 0 "register_operand")
> + (unspec:MODEF
> + [(match_operand:MODEF 1 "register_operand")
> + (match_operand:MODEF 2 "register_operand")
> + (lt:MODEF
> + (match_operand:MODEF 3 "register_operand")
> + (match_operand:MODEF 4 "register_operand"))]
> + UNSPEC_BLENDV))]
> + "SSE_FLOAT_MODE_P (<MODE>mode) && TARGET_SSE_MATH
> + && (rtx_equal_p (operands[1], operands[4])
> + && rtx_equal_p (operands[2], operands[3]))
> + && ix86_pre_reload_split ()"
> + "#"
> + "&& 1"
> + [(set (match_dup 0)
> + (unspec:MODEF
> + [(match_dup 2)
> + (match_dup 1)]
> + UNSPEC_IEEE_MIN))])
> +
> ;; Make two stack loads independent:
> ;; fld aa fld aa
> ;; fld %st(0) -> fld bb
> diff --git a/gcc/testsuite/g++.target/i386/pr110170.C
> b/gcc/testsuite/g++.target/i386/pr110170.C
> new file mode 100644
> index 00000000000..5d6842270d0
> --- /dev/null
> +++ b/gcc/testsuite/g++.target/i386/pr110170.C
> @@ -0,0 +1,78 @@
> +/* { dg-do run } */
> +/* { dg-options " -O2 -msse4.1 -mfpmath=sse -std=gnu++20" } */
Please either change the first line to:
{ dg-do run { target sse4_runtime } }
or add
{ dg-require-effective-target sse4_runtime }
to the runtime test.
> +#include <math.h>
> +
> +void
> +__attribute__((noinline))
> +__cond_swap(double* __x, double* __y) {
> + bool __r = (*__x < *__y);
> + auto __tmp = __r ? *__x : *__y;
> + *__y = __r ? *__y : *__x;
> + *__x = __tmp;
> +}
> +
> +auto test1() {
> + double nan = -0.0;
> + double x = 0.0;
> + __cond_swap(&nan, &x);
> + return x == -0.0 && nan == 0.0;
> +}
> +
> +auto test1r() {
> + double nan = NAN;
> + double x = 1.0;
> + __cond_swap(&x, &nan);
> + return isnan(x) && signbit(x) == 0 && nan == 1.0;
> +}
> +
> +auto test2() {
> + double nan = NAN;
> + double x = -1.0;
> + __cond_swap(&nan, &x);
> + return isnan(x) && signbit(x) == 0 && nan == -1.0;
> +}
> +
> +auto test2r() {
> + double nan = NAN;
> + double x = -1.0;
> + __cond_swap(&x, &nan);
> + return isnan(x) && signbit(x) == 0 && nan == -1.0;
> +}
> +
> +auto test3() {
> + double nan = -NAN;
> + double x = 1.0;
> + __cond_swap(&nan, &x);
> + return isnan(x) && signbit(x) == 1 && nan == 1.0;
> +}
> +
> +auto test3r() {
> + double nan = -NAN;
> + double x = 1.0;
> + __cond_swap(&x, &nan);
> + return isnan(x) && signbit(x) == 1 && nan == 1.0;
> +}
> +
> +auto test4() {
> + double nan = -NAN;
> + double x = -1.0;
> + __cond_swap(&nan, &x);
> + return isnan(x) && signbit(x) == 1 && nan == -1.0;
> +}
> +
> +auto test4r() {
> + double nan = -NAN;
> + double x = -1.0;
> + __cond_swap(&x, &nan);
> + return isnan(x) && signbit(x) == 1 && nan == -1.0;
> +}
> +
> +
> +int main() {
> + if (
> + !test1() || !test1r()
> + || !test2() || !test2r()
> + || !test3() || !test4r()
> + || !test4() || !test4r()
> + ) __builtin_abort();
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr110170.c
> b/gcc/testsuite/gcc.target/i386/pr110170.c
> new file mode 100644
> index 00000000000..c72f73398a1
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr110170.c
> @@ -0,0 +1,21 @@
> +/* { dg-do compile } */
> +/* { dg-options " -O2 -msse4.1 -mfpmath=sse" } */
> +/* { dg-final { scan-assembler-times {(?n)mins[sd]} 2 { target { ! ia32 } }
> } } */
> +/* { dg-final { scan-assembler-times {(?n)maxs[sd]} 2 { target { ! ia32 } }
> } } */
> +/* Ideally cond_swap_df is also optimized to minsd/maxsd. */
> +/* { dg-final { scan-assembler-times {(?n)mins[sd]} 1 { target ia32 } } } */
> +/* { dg-final { scan-assembler-times {(?n)maxs[sd]} 1 { target ia32 } } } */
> +
> +void __cond_swap_df(double* __x, double* __y) {
> + _Bool __r = (*__x < *__y);
> + double __tmp = __r ? *__x : *__y;
> + *__y = __r ? *__y : *__x;
> + *__x = __tmp;
> +}
> +
> +void __cond_swap_sf(float* __x, float* __y) {
> + _Bool __r = (*__x < *__y);
> + float __tmp = __r ? *__x : *__y;
> + *__y = __r ? *__y : *__x;
> + *__x = __tmp;
> +}
> --
> 2.39.1.388.g2fc9e9ca3c
>
