| Message ID | 20231001192742.145518-1-pinskia@gmail.com |
|---|---|
| State | New |
| Headers | show |
| Series | [COMMITTED/13] Fix PR 110386: backprop vs ABSU_EXPR | expand |
On Sun, Oct 1, 2023 at 12:28 PM Andrew Pinski <pinskia@gmail.com> wrote: > > From: Andrew Pinski <apinski@marvell.com> > > The issue here is that when backprop tries to go > and strip sign ops, it skips over ABSU_EXPR but > ABSU_EXPR not only does an ABS, it also changes the > type to unsigned. > Since strip_sign_op_1 is only supposed to strip off > sign changing operands and not ones that change types, > removing ABSU_EXPR here is correct. We don't handle > nop conversions so this does cause any missed optimizations either. > > Committed to the GCC 13 branch after bootstrapped and > tested on x86_64-linux-gnu with no regressions. And to the GCC 12 and 11 branches too. Thanks, Andrew > > PR tree-optimization/110386 > > gcc/ChangeLog: > > * gimple-ssa-backprop.cc (strip_sign_op_1): Remove ABSU_EXPR. > > gcc/testsuite/ChangeLog: > > * gcc.c-torture/compile/pr110386-1.c: New test. > * gcc.c-torture/compile/pr110386-2.c: New test. > > (cherry picked from commit 2bbac12ea7bd8a3eef5382e1b13f6019df4ec03f) > --- > gcc/gimple-ssa-backprop.cc | 1 - > gcc/testsuite/gcc.c-torture/compile/pr110386-1.c | 9 +++++++++ > gcc/testsuite/gcc.c-torture/compile/pr110386-2.c | 11 +++++++++++ > 3 files changed, 20 insertions(+), 1 deletion(-) > create mode 100644 gcc/testsuite/gcc.c-torture/compile/pr110386-1.c > create mode 100644 gcc/testsuite/gcc.c-torture/compile/pr110386-2.c > > diff --git a/gcc/gimple-ssa-backprop.cc b/gcc/gimple-ssa-backprop.cc > index 65a65590017..dcb15ed4f61 100644 > --- a/gcc/gimple-ssa-backprop.cc > +++ b/gcc/gimple-ssa-backprop.cc > @@ -694,7 +694,6 @@ strip_sign_op_1 (tree rhs) > switch (gimple_assign_rhs_code (assign)) > { > case ABS_EXPR: > - case ABSU_EXPR: > case NEGATE_EXPR: > return gimple_assign_rhs1 (assign); > > diff --git a/gcc/testsuite/gcc.c-torture/compile/pr110386-1.c b/gcc/testsuite/gcc.c-torture/compile/pr110386-1.c > new file mode 100644 > index 00000000000..4fcc977ad16 > --- /dev/null > +++ b/gcc/testsuite/gcc.c-torture/compile/pr110386-1.c > @@ -0,0 +1,9 @@ > + > +int f(int a) > +{ > + int c = c < 0 ? c : -c; > + c = -c; > + unsigned b = c; > + unsigned t = b*a; > + return t*t; > +} > diff --git a/gcc/testsuite/gcc.c-torture/compile/pr110386-2.c b/gcc/testsuite/gcc.c-torture/compile/pr110386-2.c > new file mode 100644 > index 00000000000..c60e1b6994b > --- /dev/null > +++ b/gcc/testsuite/gcc.c-torture/compile/pr110386-2.c > @@ -0,0 +1,11 @@ > +/* { dg-do compile { target i?86-*-* x86_64-*-* } } */ > +/* { dg-options "-mavx" } */ > + > +#include <immintrin.h> > + > +__m128i do_stuff(__m128i XMM0) { > + __m128i ABS0 = _mm_abs_epi32(XMM0); > + __m128i MUL0 = _mm_mullo_epi32(ABS0, XMM0); > + __m128i MUL1 = _mm_mullo_epi32(MUL0, MUL0); > + return MUL1; > +} > -- > 2.39.3 >
diff --git a/gcc/gimple-ssa-backprop.cc b/gcc/gimple-ssa-backprop.cc index 65a65590017..dcb15ed4f61 100644 --- a/gcc/gimple-ssa-backprop.cc +++ b/gcc/gimple-ssa-backprop.cc @@ -694,7 +694,6 @@ strip_sign_op_1 (tree rhs) switch (gimple_assign_rhs_code (assign)) { case ABS_EXPR: - case ABSU_EXPR: case NEGATE_EXPR: return gimple_assign_rhs1 (assign); diff --git a/gcc/testsuite/gcc.c-torture/compile/pr110386-1.c b/gcc/testsuite/gcc.c-torture/compile/pr110386-1.c new file mode 100644 index 00000000000..4fcc977ad16 --- /dev/null +++ b/gcc/testsuite/gcc.c-torture/compile/pr110386-1.c @@ -0,0 +1,9 @@ + +int f(int a) +{ + int c = c < 0 ? c : -c; + c = -c; + unsigned b = c; + unsigned t = b*a; + return t*t; +} diff --git a/gcc/testsuite/gcc.c-torture/compile/pr110386-2.c b/gcc/testsuite/gcc.c-torture/compile/pr110386-2.c new file mode 100644 index 00000000000..c60e1b6994b --- /dev/null +++ b/gcc/testsuite/gcc.c-torture/compile/pr110386-2.c @@ -0,0 +1,11 @@ +/* { dg-do compile { target i?86-*-* x86_64-*-* } } */ +/* { dg-options "-mavx" } */ + +#include <immintrin.h> + +__m128i do_stuff(__m128i XMM0) { + __m128i ABS0 = _mm_abs_epi32(XMM0); + __m128i MUL0 = _mm_mullo_epi32(ABS0, XMM0); + __m128i MUL1 = _mm_mullo_epi32(MUL0, MUL0); + return MUL1; +}
From: Andrew Pinski <apinski@marvell.com> The issue here is that when backprop tries to go and strip sign ops, it skips over ABSU_EXPR but ABSU_EXPR not only does an ABS, it also changes the type to unsigned. Since strip_sign_op_1 is only supposed to strip off sign changing operands and not ones that change types, removing ABSU_EXPR here is correct. We don't handle nop conversions so this does cause any missed optimizations either. Committed to the GCC 13 branch after bootstrapped and tested on x86_64-linux-gnu with no regressions. PR tree-optimization/110386 gcc/ChangeLog: * gimple-ssa-backprop.cc (strip_sign_op_1): Remove ABSU_EXPR. gcc/testsuite/ChangeLog: * gcc.c-torture/compile/pr110386-1.c: New test. * gcc.c-torture/compile/pr110386-2.c: New test. (cherry picked from commit 2bbac12ea7bd8a3eef5382e1b13f6019df4ec03f) --- gcc/gimple-ssa-backprop.cc | 1 - gcc/testsuite/gcc.c-torture/compile/pr110386-1.c | 9 +++++++++ gcc/testsuite/gcc.c-torture/compile/pr110386-2.c | 11 +++++++++++ 3 files changed, 20 insertions(+), 1 deletion(-) create mode 100644 gcc/testsuite/gcc.c-torture/compile/pr110386-1.c create mode 100644 gcc/testsuite/gcc.c-torture/compile/pr110386-2.c