On Thu, Jun 2, 2022 at 3:10 AM liuhongt via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> Similar for (v + B) * C + D -> C * v + BCD.
> Don't simplify it when there's overflow and overflow is UB for type v.
>
> There's new failure
>
> gcc.dg/vect/slp-11a.c scan-tree-dump-times vect "vectorizing stmts using SLP" > 0
>
> It's because the patch simplify different operations to mult + add and enables
> SLP. So I adjust the testcase to prevent simplication by making the
> multiplication result overflow.
>
> Also with -fwrapv, benchmark in the PR now 100% faster than
> origin(scalar version).
>
> Bootstrapped and regtested on x86_64-pc-linux-gnu{-m32,}
> Ok for trunk?
>
> gcc/ChangeLog:
>
> PR tree-optimization/53533
> * match.pd: Simplify (B * v + C) * D -> BD * v + CD and
> (v + B) * C + D -> C * v + BCD when B,C,D are all INTEGER_CST,
> and there's no overflow or TYPE_OVERFLOW_WRAP.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/i386/pr53533-1.c: New test.
> * gcc.target/i386/pr53533-2.c: New test.
> * gcc.target/i386/pr53533-3.c: New test.
> * gcc.target/i386/pr53533-4.c: New test.
> * gcc.dg/vect/slp-11a.c: Adjust testcase.
> ---
> gcc/match.pd | 36 ++++++++++++++++++
> gcc/testsuite/gcc.dg/vect/slp-11a.c | 10 ++---
> gcc/testsuite/gcc.target/i386/pr53533-1.c | 23 ++++++++++++
> gcc/testsuite/gcc.target/i386/pr53533-2.c | 46 +++++++++++++++++++++++
> gcc/testsuite/gcc.target/i386/pr53533-3.c | 24 ++++++++++++
> gcc/testsuite/gcc.target/i386/pr53533-4.c | 46 +++++++++++++++++++++++
> 6 files changed, 180 insertions(+), 5 deletions(-)
> create mode 100644 gcc/testsuite/gcc.target/i386/pr53533-1.c
> create mode 100644 gcc/testsuite/gcc.target/i386/pr53533-2.c
> create mode 100644 gcc/testsuite/gcc.target/i386/pr53533-3.c
> create mode 100644 gcc/testsuite/gcc.target/i386/pr53533-4.c
>
> diff --git a/gcc/match.pd b/gcc/match.pd
> index 88c6c414881..b753f7bda3c 100644
> --- a/gcc/match.pd
> +++ b/gcc/match.pd
> @@ -489,6 +489,42 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> (if (!overflow || TYPE_OVERFLOW_WRAPS (type))
> (mult @0 { wide_int_to_tree (type, mul); }))))
>
> +/* Similar to above, but there could be an extra add/sub between
> + successive multuiplications. */
> +(simplify
> + (mult:c (plus:c@4 (mult:c@5 @0 INTEGER_CST@1) INTEGER_CST@2) INTEGER_CST@3)
since canonicalization puts INTEGER_CSTs last the :c should not be necessary.
> + (if (single_use (@4)
> + && single_use (@5))
since the resulting expression is not simple using :s instead of
single_use (..) should
work as well.
> + (with {
> + wi::overflow_type overflow;
> + wi::overflow_type overflow2;
> + wide_int mul = wi::mul (wi::to_wide (@1), wi::to_wide (@3),
> + TYPE_SIGN (type), &overflow);
> + wide_int add = wi::mul (wi::to_wide (@2), wi::to_wide (@3),
> + TYPE_SIGN (type), &overflow2);
> + }
> + /* Skip folding on overflow. */
> + (if (!(overflow || overflow2) || TYPE_OVERFLOW_WRAPS (type))
> + (plus (mult @0 { wide_int_to_tree (type, mul); })
> + { wide_int_to_tree (type, add); })))))
> +
> +/* Similar to above, but a multiplication between successive additions. */
> +(simplify
> + (plus:c (mult:c@4 (plus:c@5 @0 INTEGER_CST@1) INTEGER_CST@2) INTEGER_CST@3)
Likewise for :c and :s
> + (if (single_use (@4)
> + && single_use (@5))
> + (with {
> + wi::overflow_type overflow;
> + wi::overflow_type overflow2;
> + wide_int mul = wi::mul (wi::to_wide (@1), wi::to_wide (@2),
> + TYPE_SIGN (type), &overflow);
> + wide_int add = wi::add (mul, wi::to_wide (@3),
> + TYPE_SIGN (type), &overflow2);
> + }
> + /* Skip folding on overflow. */
> + (if (!(overflow || overflow2) || TYPE_OVERFLOW_WRAPS (type))
> + (plus (mult @0 @2) { wide_int_to_tree (type, add); })))))
when we go from (a + CST1) * CST2 to a * CST2 + CST1*CST2 we have
to worry about CST1 == -a which would make (a+CST1) * INT_MAX
not overflow but a * INT_MAX + CST1 * INT_MAX might. Is the
overflow check for CST1 * INT_MAX sufficient to rule out
that a * CST2 does not overflow when (a + CST1) * CST2 does not
overflow? Consider a == 2, CST1 == -1, CST2 == INT_MAX,
here 1 * INT_MAX does not overflow, nor does -1 * INT_MAX, but
2 * INT_MAX overflows and thus the resulting expression invokes
undefined behavior.
The same issue probably arises for the first pattern outer half
which looks like (a' + CST2) * CST3 with a' = a * CST1?
The appropriate solution might be to perform the arithmetic
in an unsigned type with the implication that has on value-range
analysis.
Richard.
> +
> /* Optimize A / A to 1.0 if we don't care about
> NaNs or Infinities. */
> (simplify
> diff --git a/gcc/testsuite/gcc.dg/vect/slp-11a.c
> b/gcc/testsuite/gcc.dg/vect/slp-11a.c
> index bcd3c861ca4..e6632fa77be 100644
> --- a/gcc/testsuite/gcc.dg/vect/slp-11a.c
> +++ b/gcc/testsuite/gcc.dg/vect/slp-11a.c
> @@ -9,14 +9,14 @@ int
> main1 ()
> {
> int i;
> - unsigned int out[N*8], a0, a1, a2, a3, a4, a5, a6, a7, b1, b0, b2, b3, b4,
> b5, b6, b7;
> - unsigned int in[N*8] =
> {0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63};
> + int out[N*8], a0, a1, a2, a3, a4, a5, a6, a7, b1, b0, b2, b3, b4, b5, b6,
> b7;
> + int in[N*8] =
> {0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63};
>
> /* Different operations - not SLPable. */
> for (i = 0; i < N; i++)
> {
> a0 = in[i*8] + 5;
> - a1 = in[i*8 + 1] * 6;
> + a1 = in[i*8 + 1] * 51072;
> a2 = in[i*8 + 2] + 7;
> a3 = in[i*8 + 3] + 8;
> a4 = in[i*8 + 4] + 9;
> @@ -25,7 +25,7 @@ main1 ()
> a7 = in[i*8 + 7] + 12;
>
> b0 = a0 * 3;
> - b1 = a1 * 2;
> + b1 = a1 * 51072;
> b2 = a2 * 12;
> b3 = a3 * 5;
> b4 = a4 * 8;
> @@ -47,7 +47,7 @@ main1 ()
> for (i = 0; i < N; i++)
> {
> if (out[i*8] != (in[i*8] + 5) * 3 - 2
> - || out[i*8 + 1] != (in[i*8 + 1] * 6) * 2 - 3
> + || out[i*8 + 1] != (in[i*8 + 1] * 51072) * 51072 - 3
> || out[i*8 + 2] != (in[i*8 + 2] + 7) * 12 - 2
> || out[i*8 + 3] != (in[i*8 + 3] + 8) * 5 - 1
> || out[i*8 + 4] != (in[i*8 + 4] + 9) * 8 - 8
> diff --git a/gcc/testsuite/gcc.target/i386/pr53533-1.c
> b/gcc/testsuite/gcc.target/i386/pr53533-1.c
> new file mode 100644
> index 00000000000..095de665366
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr53533-1.c
> @@ -0,0 +1,23 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O1" } */
> +/* { dg-final { scan-assembler-times "imull\[ \t\]" "1" } } */
> +/* { dg-final { scan-assembler-times "(?:addl|subl)\[ \t\]" "1" { target { !
> ia32 } } } } */
> +
> +void
> +__attribute__((noipa))
> +foo (unsigned a[256], unsigned b[256])
> +{
> + int i;
> + for (i = 0; i < 256; ++i)
> + {
> + unsigned tmp = a[i] + 12345U;
> + tmp *= 914237U;
> + tmp += 12332U;
> + tmp *= 914237U;
> + tmp += 12332U;
> + tmp *= 914237U;
> + tmp -= 13U;
> + tmp *= 8000U;
> + b[i] = tmp;
> + }
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr53533-2.c
> b/gcc/testsuite/gcc.target/i386/pr53533-2.c
> new file mode 100644
> index 00000000000..c31b6ff4dec
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr53533-2.c
> @@ -0,0 +1,46 @@
> +/* { dg-do run } */
> +/* { dg-options "-O2" } */
> +
> +#include "pr53533-1.c"
> +
> +void
> +__attribute__((optimize("-O0")))
> +foo1 (unsigned a[256], unsigned b[256])
> +{
> + int i;
> + for (i = 0; i < 256; ++i)
> + {
> + unsigned tmp = a[i] + 12345U;
> + tmp *= 914237U;
> + tmp += 12332U;
> + tmp *= 914237U;
> + tmp += 12332U;
> + tmp *= 914237U;
> + tmp -= 13U;
> + tmp *= 8000U;
> + b[i] = tmp;
> + }
> +}
> +
> +int main()
> +{
> + unsigned int a[256];
> + unsigned int b[256];
> + unsigned int c[256];
> + for (unsigned int i = 0; i != 256; i++)
> + {
> + b[i] = 0;
> + c[i] = 1;
> + a[i] = i * i - 10 * i + 33;
> + }
> + foo (a, b);
> + foo1 (a, c);
> +
> + for (unsigned int i = 0; i != 256; i++)
> + {
> + if (b[i] != c[i])
> + __builtin_abort ();
> + }
> +
> + return 0;
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr53533-3.c
> b/gcc/testsuite/gcc.target/i386/pr53533-3.c
> new file mode 100644
> index 00000000000..3b260d134e9
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr53533-3.c
> @@ -0,0 +1,24 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O1 -fwrapv" } */
> +/* { dg-final { scan-assembler-times "imull\[ \t\]" "1" } } */
> +/* { dg-final { scan-assembler-times "(?:addl|subl)\[ \t\]" "1" { target { !
> ia32 } } } } */
> +
> +void
> +__attribute__((noipa))
> +foo (int a[256], int b[256])
> +{
> + int i;
> + for (i = 0; i < 256; ++i)
> + {
> + int tmp = a[i] + 12345;
> + tmp *= 914237;
> + tmp += 12332;
> + tmp *= 914237;
> + tmp += 12332;
> + tmp *= 914237;
> + tmp -= 13;
> + tmp *= 8000;
> + b[i] = tmp;
> + }
> +}
> +
> diff --git a/gcc/testsuite/gcc.target/i386/pr53533-4.c
> b/gcc/testsuite/gcc.target/i386/pr53533-4.c
> new file mode 100644
> index 00000000000..c29f90a44dc
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr53533-4.c
> @@ -0,0 +1,46 @@
> +/* { dg-do run } */
> +/* { dg-options "-O2 -fwrapv" } */
> +
> +#include "pr53533-3.c"
> +
> +void
> +__attribute__((optimize("-O0")))
> +foo1 (int a[256], int b[256])
> +{
> + int i;
> + for (i = 0; i < 256; ++i)
> + {
> + int tmp = a[i] + 12345;
> + tmp *= 914237;
> + tmp += 12332;
> + tmp *= 914237;
> + tmp += 12332;
> + tmp *= 914237;
> + tmp -= 13;
> + tmp *= 8000;
> + b[i] = tmp;
> + }
> +}
> +
> +int main()
> +{
> + int a[256];
> + int b[256];
> + int c[256];
> + for (int i = 0; i != 256; i++)
> + {
> + b[i] = 0;
> + c[i] = 1;
> + a[i] = i * i - 10 * i + 33;
> + }
> + foo (a, b);
> + foo1 (a, c);
> +
> + for (unsigned int i = 0; i != 256; i++)
> + {
> + if (b[i] != c[i])
> + __builtin_abort ();
> + }
> +
> + return 0;
> +}
> --
> 2.18.1
>
