[RTL/fwprop] Allow propagations from inner loop to outer loop.
Commit Message
> that's flow_loop_nested_p (loop *outer, loop *inner) which
> is implemented in O(1). Note behavior for outer == inner
> might be different (didn't check your implementation too hard)
>
Thanks, it seems flow_loop_nested_p assume outer and inner not to be
NULL. So I add some conditions to check NULL which is considered as an outer
loop of any other loop.
gcc/ChangeLog:
PR rtl/103750
* fwprop.c (forward_propagate_into): Allow propagations from
inner loop to outer loop.
gcc/testsuite/ChangeLog:
* g++.target/i386/pr103750-fwprop-1.C: New test.
---
gcc/fwprop.c | 7 +++--
.../g++.target/i386/pr103750-fwprop-1.C | 26 +++++++++++++++++++
2 files changed, 31 insertions(+), 2 deletions(-)
create mode 100644 gcc/testsuite/g++.target/i386/pr103750-fwprop-1.C
Comments
On January 6, 2022 7:51:48 AM GMT+01:00, liuhongt <hongtao.liu@intel.com> wrote:
>> that's flow_loop_nested_p (loop *outer, loop *inner) which
>> is implemented in O(1). Note behavior for outer == inner
>> might be different (didn't check your implementation too hard)
>>
>Thanks, it seems flow_loop_nested_p assume outer and inner not to be
>NULL. So I add some conditions to check NULL which is considered as an outer
> loop of any other loop.
Huh, loop_father should never be NULL. Maybe when fwprop is run after RTL loop opts you instead want to add a check for current_loops or alternelatively initialize loops in fwprop.
>
>
>gcc/ChangeLog:
>
> PR rtl/103750
> * fwprop.c (forward_propagate_into): Allow propagations from
> inner loop to outer loop.
>
>gcc/testsuite/ChangeLog:
>
> * g++.target/i386/pr103750-fwprop-1.C: New test.
>---
> gcc/fwprop.c | 7 +++--
> .../g++.target/i386/pr103750-fwprop-1.C | 26 +++++++++++++++++++
> 2 files changed, 31 insertions(+), 2 deletions(-)
> create mode 100644 gcc/testsuite/g++.target/i386/pr103750-fwprop-1.C
>
>diff --git a/gcc/fwprop.c b/gcc/fwprop.c
>index 2eab4fd4614..af2e9d1c189 100644
>--- a/gcc/fwprop.c
>+++ b/gcc/fwprop.c
>@@ -866,10 +866,13 @@ forward_propagate_into (use_info *use, bool reg_prop_only = false)
> rtx src = SET_SRC (def_set);
>
> /* Allow propagations into a loop only for reg-to-reg copies, since
>- replacing one register by another shouldn't increase the cost. */
>+ replacing one register by another shouldn't increase the cost.
>+ Propagations from inner loop to outer loop should be also ok. */
> struct loop *def_loop = def_insn->bb ()->cfg_bb ()->loop_father;
> struct loop *use_loop = use->bb ()->cfg_bb ()->loop_father;
>- if ((reg_prop_only || def_loop != use_loop)
>+ if ((reg_prop_only
>+ || (use_loop && def_loop != use_loop
>+ &&(!def_loop || !flow_loop_nested_p (use_loop, def_loop))))
> && (!reg_single_def_p (dest) || !reg_single_def_p (src)))
> return false;
>
>diff --git a/gcc/testsuite/g++.target/i386/pr103750-fwprop-1.C b/gcc/testsuite/g++.target/i386/pr103750-fwprop-1.C
>new file mode 100644
>index 00000000000..26987d307aa
>--- /dev/null
>+++ b/gcc/testsuite/g++.target/i386/pr103750-fwprop-1.C
>@@ -0,0 +1,26 @@
>+/* PR target/103750. */
>+/* { dg-do compile } */
>+/* { dg-options "-O2 -std=c++1y -march=cannonlake -fdump-rtl-fwprop1" } */
>+/* { dg-final { scan-rtl-dump-not "subreg:HI\[ \\\(\]*reg:SI\[^\n]*\n\[^\n]*UNSPEC_TZCNT" "fwprop1" } } */
>+
>+#include<immintrin.h>
>+const char16_t *qustrchr(char16_t *n, char16_t *e, char16_t c) noexcept
>+{
>+ __m256i mch256 = _mm256_set1_epi16(c);
>+ for ( ; n < e; n += 32) {
>+ __m256i data1 = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(n));
>+ __m256i data2 = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(n) + 1);
>+ __mmask16 mask1 = _mm256_cmpeq_epu16_mask(data1, mch256);
>+ __mmask16 mask2 = _mm256_cmpeq_epu16_mask(data2, mch256);
>+ if (_kortestz_mask16_u8(mask1, mask2))
>+ continue;
>+
>+ unsigned idx = _tzcnt_u32(mask1);
>+ if (mask1 == 0) {
>+ idx = __tzcnt_u16(mask2);
>+ n += 16;
>+ }
>+ return n + idx;
>+ }
>+ return e;
>+}
@@ -866,10 +866,13 @@ forward_propagate_into (use_info *use, bool reg_prop_only = false)
rtx src = SET_SRC (def_set);
/* Allow propagations into a loop only for reg-to-reg copies, since
- replacing one register by another shouldn't increase the cost. */
+ replacing one register by another shouldn't increase the cost.
+ Propagations from inner loop to outer loop should be also ok. */
struct loop *def_loop = def_insn->bb ()->cfg_bb ()->loop_father;
struct loop *use_loop = use->bb ()->cfg_bb ()->loop_father;
- if ((reg_prop_only || def_loop != use_loop)
+ if ((reg_prop_only
+ || (use_loop && def_loop != use_loop
+ &&(!def_loop || !flow_loop_nested_p (use_loop, def_loop))))
&& (!reg_single_def_p (dest) || !reg_single_def_p (src)))
return false;
new file mode 100644
@@ -0,0 +1,26 @@
+/* PR target/103750. */
+/* { dg-do compile } */
+/* { dg-options "-O2 -std=c++1y -march=cannonlake -fdump-rtl-fwprop1" } */
+/* { dg-final { scan-rtl-dump-not "subreg:HI\[ \\\(\]*reg:SI\[^\n]*\n\[^\n]*UNSPEC_TZCNT" "fwprop1" } } */
+
+#include<immintrin.h>
+const char16_t *qustrchr(char16_t *n, char16_t *e, char16_t c) noexcept
+{
+ __m256i mch256 = _mm256_set1_epi16(c);
+ for ( ; n < e; n += 32) {
+ __m256i data1 = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(n));
+ __m256i data2 = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(n) + 1);
+ __mmask16 mask1 = _mm256_cmpeq_epu16_mask(data1, mch256);
+ __mmask16 mask2 = _mm256_cmpeq_epu16_mask(data2, mch256);
+ if (_kortestz_mask16_u8(mask1, mask2))
+ continue;
+
+ unsigned idx = _tzcnt_u32(mask1);
+ if (mask1 == 0) {
+ idx = __tzcnt_u16(mask2);
+ n += 16;
+ }
+ return n + idx;
+ }
+ return e;
+}