Message ID | 20220721112338.7CEC113A1B@imap2.suse-dmz.suse.de |
---|---|
State | New |
Headers | show |
Series | tree-optimization/106379 - add missing ~(a ^ b) folding for _Bool | expand |
On Thu, Jul 21, 2022 at 4:24 AM Richard Biener via Gcc-patches <gcc-patches@gcc.gnu.org> wrote: > > The following makes sure to fold ~(a ^ b) to a == b for truth > values (but not vectors, we'd have to check for vector support of > equality). That turns the PR106379 testcase into a ranger one. > > Note that while we arrive at ~(a ^ b) in a convoluted way from > original !a == !b one can eventually write the expression this > way directly as well. > > Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed. > > PR tree-optimization/106379 > * match.pd (~(a ^ b) -> a == b): New pattern. > > * gcc.dg/pr106379-1.c: New testcase. > --- > gcc/match.pd | 6 ++++++ > gcc/testsuite/gcc.dg/pr106379-1.c | 9 +++++++++ > 2 files changed, 15 insertions(+) > create mode 100644 gcc/testsuite/gcc.dg/pr106379-1.c > > diff --git a/gcc/match.pd b/gcc/match.pd > index 8bbc0dbd5cd..88a1a5aa9cc 100644 > --- a/gcc/match.pd > +++ b/gcc/match.pd > @@ -1938,6 +1938,12 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT) > (if (tree_nop_conversion_p (type, TREE_TYPE (@0))) > (bit_not (bit_xor (view_convert @0) @1)))) > > +/* ~(a ^ b) is a == b for truth valued a and b. */ > +(simplify > + (bit_not (bit_xor:s truth_valued_p@0 truth_valued_p@1)) > + (if (!VECTOR_TYPE_P (type)) > + (convert (eq @0 @1)))) For integers, isn't it wrong to convert ~(boolean exp) to boolean exp? > /* (x & ~m) | (y & m) -> ((x ^ y) & m) ^ x */ > (simplify > (bit_ior:c (bit_and:cs @0 (bit_not @2)) (bit_and:cs @1 @2)) > diff --git a/gcc/testsuite/gcc.dg/pr106379-1.c b/gcc/testsuite/gcc.dg/pr106379-1.c > new file mode 100644 > index 00000000000..7f2575e02dc > --- /dev/null > +++ b/gcc/testsuite/gcc.dg/pr106379-1.c > @@ -0,0 +1,9 @@ > +/* { dg-do compile } */ > +/* { dg-options "-O -fdump-tree-forwprop1" } */ > + > +_Bool foo (_Bool a, _Bool b) > +{ > + return !a == !b; > +} > + > +/* { dg-final { scan-tree-dump "\[ab\]_\[0-9\]+\\(D\\) == \[ba\]_\[0-9\]+\\(D\\)" "forwprop1" } } */ > -- > 2.35.3 -- H.J.
> Am 22.07.2022 um 22:17 schrieb H.J. Lu via Gcc-patches <gcc-patches@gcc.gnu.org>: > > On Thu, Jul 21, 2022 at 4:24 AM Richard Biener via Gcc-patches > <gcc-patches@gcc.gnu.org> wrote: >> >> The following makes sure to fold ~(a ^ b) to a == b for truth >> values (but not vectors, we'd have to check for vector support of >> equality). That turns the PR106379 testcase into a ranger one. >> >> Note that while we arrive at ~(a ^ b) in a convoluted way from >> original !a == !b one can eventually write the expression this >> way directly as well. >> >> Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed. >> >> PR tree-optimization/106379 >> * match.pd (~(a ^ b) -> a == b): New pattern. >> >> * gcc.dg/pr106379-1.c: New testcase. >> --- >> gcc/match.pd | 6 ++++++ >> gcc/testsuite/gcc.dg/pr106379-1.c | 9 +++++++++ >> 2 files changed, 15 insertions(+) >> create mode 100644 gcc/testsuite/gcc.dg/pr106379-1.c >> >> diff --git a/gcc/match.pd b/gcc/match.pd >> index 8bbc0dbd5cd..88a1a5aa9cc 100644 >> --- a/gcc/match.pd >> +++ b/gcc/match.pd >> @@ -1938,6 +1938,12 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT) >> (if (tree_nop_conversion_p (type, TREE_TYPE (@0))) >> (bit_not (bit_xor (view_convert @0) @1)))) >> >> +/* ~(a ^ b) is a == b for truth valued a and b. */ >> +(simplify >> + (bit_not (bit_xor:s truth_valued_p@0 truth_valued_p@1)) >> + (if (!VECTOR_TYPE_P (type)) >> + (convert (eq @0 @1)))) > > For integers, isn't it wrong to convert ~(boolean exp) to boolean exp? That’s what the (convert. …) should compensate for? Richard > >> /* (x & ~m) | (y & m) -> ((x ^ y) & m) ^ x */ >> (simplify >> (bit_ior:c (bit_and:cs @0 (bit_not @2)) (bit_and:cs @1 @2)) >> diff --git a/gcc/testsuite/gcc.dg/pr106379-1.c b/gcc/testsuite/gcc.dg/pr106379-1.c >> new file mode 100644 >> index 00000000000..7f2575e02dc >> --- /dev/null >> +++ b/gcc/testsuite/gcc.dg/pr106379-1.c >> @@ -0,0 +1,9 @@ >> +/* { dg-do compile } */ >> +/* { dg-options "-O -fdump-tree-forwprop1" } */ >> + >> +_Bool foo (_Bool a, _Bool b) >> +{ >> + return !a == !b; >> +} >> + >> +/* { dg-final { scan-tree-dump "\[ab\]_\[0-9\]+\\(D\\) == \[ba\]_\[0-9\]+\\(D\\)" "forwprop1" } } */ >> -- >> 2.35.3 > > > > -- > H.J.
On Fri, Jul 22, 2022 at 11:10 PM Richard Biener via Gcc-patches <gcc-patches@gcc.gnu.org> wrote: > > > > > Am 22.07.2022 um 22:17 schrieb H.J. Lu via Gcc-patches <gcc-patches@gcc.gnu.org>: > > > > On Thu, Jul 21, 2022 at 4:24 AM Richard Biener via Gcc-patches > > <gcc-patches@gcc.gnu.org> wrote: > >> > >> The following makes sure to fold ~(a ^ b) to a == b for truth > >> values (but not vectors, we'd have to check for vector support of > >> equality). That turns the PR106379 testcase into a ranger one. > >> > >> Note that while we arrive at ~(a ^ b) in a convoluted way from > >> original !a == !b one can eventually write the expression this > >> way directly as well. > >> > >> Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed. > >> > >> PR tree-optimization/106379 > >> * match.pd (~(a ^ b) -> a == b): New pattern. > >> > >> * gcc.dg/pr106379-1.c: New testcase. > >> --- > >> gcc/match.pd | 6 ++++++ > >> gcc/testsuite/gcc.dg/pr106379-1.c | 9 +++++++++ > >> 2 files changed, 15 insertions(+) > >> create mode 100644 gcc/testsuite/gcc.dg/pr106379-1.c > >> > >> diff --git a/gcc/match.pd b/gcc/match.pd > >> index 8bbc0dbd5cd..88a1a5aa9cc 100644 > >> --- a/gcc/match.pd > >> +++ b/gcc/match.pd > >> @@ -1938,6 +1938,12 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT) > >> (if (tree_nop_conversion_p (type, TREE_TYPE (@0))) > >> (bit_not (bit_xor (view_convert @0) @1)))) > >> > >> +/* ~(a ^ b) is a == b for truth valued a and b. */ > >> +(simplify > >> + (bit_not (bit_xor:s truth_valued_p@0 truth_valued_p@1)) > >> + (if (!VECTOR_TYPE_P (type)) > >> + (convert (eq @0 @1)))) > > > > For integers, isn't it wrong to convert ~(boolean exp) to boolean exp? > > That’s what the (convert. …) should compensate for? Is ~(boolean exp) == ~((int) (boolean exp)) or (int) (~(boolean exp))? > Richard > > > > >> /* (x & ~m) | (y & m) -> ((x ^ y) & m) ^ x */ > >> (simplify > >> (bit_ior:c (bit_and:cs @0 (bit_not @2)) (bit_and:cs @1 @2)) > >> diff --git a/gcc/testsuite/gcc.dg/pr106379-1.c b/gcc/testsuite/gcc.dg/pr106379-1.c > >> new file mode 100644 > >> index 00000000000..7f2575e02dc > >> --- /dev/null > >> +++ b/gcc/testsuite/gcc.dg/pr106379-1.c > >> @@ -0,0 +1,9 @@ > >> +/* { dg-do compile } */ > >> +/* { dg-options "-O -fdump-tree-forwprop1" } */ > >> + > >> +_Bool foo (_Bool a, _Bool b) > >> +{ > >> + return !a == !b; > >> +} > >> + > >> +/* { dg-final { scan-tree-dump "\[ab\]_\[0-9\]+\\(D\\) == \[ba\]_\[0-9\]+\\(D\\)" "forwprop1" } } */ > >> -- > >> 2.35.3 > > > > > > > > -- > > H.J.
On Mon, 25 Jul 2022, H.J. Lu wrote: > On Fri, Jul 22, 2022 at 11:10 PM Richard Biener via Gcc-patches > <gcc-patches@gcc.gnu.org> wrote: > > > > > > > > > Am 22.07.2022 um 22:17 schrieb H.J. Lu via Gcc-patches <gcc-patches@gcc.gnu.org>: > > > > > > On Thu, Jul 21, 2022 at 4:24 AM Richard Biener via Gcc-patches > > > <gcc-patches@gcc.gnu.org> wrote: > > >> > > >> The following makes sure to fold ~(a ^ b) to a == b for truth > > >> values (but not vectors, we'd have to check for vector support of > > >> equality). That turns the PR106379 testcase into a ranger one. > > >> > > >> Note that while we arrive at ~(a ^ b) in a convoluted way from > > >> original !a == !b one can eventually write the expression this > > >> way directly as well. > > >> > > >> Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed. > > >> > > >> PR tree-optimization/106379 > > >> * match.pd (~(a ^ b) -> a == b): New pattern. > > >> > > >> * gcc.dg/pr106379-1.c: New testcase. > > >> --- > > >> gcc/match.pd | 6 ++++++ > > >> gcc/testsuite/gcc.dg/pr106379-1.c | 9 +++++++++ > > >> 2 files changed, 15 insertions(+) > > >> create mode 100644 gcc/testsuite/gcc.dg/pr106379-1.c > > >> > > >> diff --git a/gcc/match.pd b/gcc/match.pd > > >> index 8bbc0dbd5cd..88a1a5aa9cc 100644 > > >> --- a/gcc/match.pd > > >> +++ b/gcc/match.pd > > >> @@ -1938,6 +1938,12 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT) > > >> (if (tree_nop_conversion_p (type, TREE_TYPE (@0))) > > >> (bit_not (bit_xor (view_convert @0) @1)))) > > >> > > >> +/* ~(a ^ b) is a == b for truth valued a and b. */ > > >> +(simplify > > >> + (bit_not (bit_xor:s truth_valued_p@0 truth_valued_p@1)) > > >> + (if (!VECTOR_TYPE_P (type)) > > >> + (convert (eq @0 @1)))) > > > > > > For integers, isn't it wrong to convert ~(boolean exp) to boolean exp? > > > > That’s what the (convert. …) should compensate for? > > Is ~(boolean exp) == ~((int) (boolean exp)) or (int) (~(boolean exp))? Depends on what 'boolean exp' is, and as PR106414 shows I was wrong because in C a == b has type 'int' and not type bool but we consider it "boolean". There of course ~(a == b) is not equal to a != b. Richard.
diff --git a/gcc/match.pd b/gcc/match.pd index 8bbc0dbd5cd..88a1a5aa9cc 100644 --- a/gcc/match.pd +++ b/gcc/match.pd @@ -1938,6 +1938,12 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT) (if (tree_nop_conversion_p (type, TREE_TYPE (@0))) (bit_not (bit_xor (view_convert @0) @1)))) +/* ~(a ^ b) is a == b for truth valued a and b. */ +(simplify + (bit_not (bit_xor:s truth_valued_p@0 truth_valued_p@1)) + (if (!VECTOR_TYPE_P (type)) + (convert (eq @0 @1)))) + /* (x & ~m) | (y & m) -> ((x ^ y) & m) ^ x */ (simplify (bit_ior:c (bit_and:cs @0 (bit_not @2)) (bit_and:cs @1 @2)) diff --git a/gcc/testsuite/gcc.dg/pr106379-1.c b/gcc/testsuite/gcc.dg/pr106379-1.c new file mode 100644 index 00000000000..7f2575e02dc --- /dev/null +++ b/gcc/testsuite/gcc.dg/pr106379-1.c @@ -0,0 +1,9 @@ +/* { dg-do compile } */ +/* { dg-options "-O -fdump-tree-forwprop1" } */ + +_Bool foo (_Bool a, _Bool b) +{ + return !a == !b; +} + +/* { dg-final { scan-tree-dump "\[ab\]_\[0-9\]+\\(D\\) == \[ba\]_\[0-9\]+\\(D\\)" "forwprop1" } } */