https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96573
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Ever confirmed|0 |1
Last reconfirmed| |2020-08-11
Status|UNCONFIRMED |NEW
Target Milestone|--- |10.3
CC| |jakub at gcc dot gnu.org
Component|c |tree-optimization
Summary|[Regression] Regression in |[10/11 Regression]
|optimization on x86-64 with |Regression in optimization
|-O3 from GCC 9 to 10 |on x86-64 with -O3
--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Stopped being recognized as bswap with r274922.
That changed the IL before store-merging from
_22 = MEM[(const uint8_t *)&pointer];
MEM[(uint8_t *)&newPointer + 7B] = _22;
_33 = MEM[(const uint8_t *)&pointer + 1B];
MEM[(uint8_t *)&newPointer + 6B] = _33;
_44 = MEM[(const uint8_t *)&pointer + 2B];
MEM[(uint8_t *)&newPointer + 5B] = _44;
_55 = MEM[(const uint8_t *)&pointer + 3B];
MEM[(uint8_t *)&newPointer + 4B] = _55;
_66 = MEM[(const uint8_t *)&pointer + 4B];
MEM[(uint8_t *)&newPointer + 3B] = _66;
_77 = MEM[(const uint8_t *)&pointer + 5B];
MEM[(uint8_t *)&newPointer + 2B] = _77;
_88 = MEM[(const uint8_t *)&pointer + 6B];
MEM[(uint8_t *)&newPointer + 1B] = _88;
_5 = MEM[(const uint8_t *)&pointer + 7B];
MEM[(uint8_t *)&newPointer] = _5;
_9 = newPointer;
newPointer ={v} {CLOBBER};
return _9;
to:
_22 = BIT_FIELD_REF <pointer_36(D), 8, 0>;
MEM[(uint8_t *)&newPointer + 7B] = _22;
_33 = BIT_FIELD_REF <pointer_36(D), 8, 8>;
MEM[(uint8_t *)&newPointer + 6B] = _33;
_44 = BIT_FIELD_REF <pointer_36(D), 8, 16>;
MEM[(uint8_t *)&newPointer + 5B] = _44;
_55 = BIT_FIELD_REF <pointer_36(D), 8, 24>;
MEM[(uint8_t *)&newPointer + 4B] = _55;
_66 = BIT_FIELD_REF <pointer_36(D), 8, 32>;
MEM[(uint8_t *)&newPointer + 3B] = _66;
_77 = BIT_FIELD_REF <pointer_36(D), 8, 40>;
MEM[(uint8_t *)&newPointer + 2B] = _77;
_88 = BIT_FIELD_REF <pointer_36(D), 8, 48>;
MEM[(uint8_t *)&newPointer + 1B] = _88;
_5 = BIT_FIELD_REF <pointer_36(D), 8, 56>;
MEM[(uint8_t *)&newPointer] = _5;
_9 = newPointer;
newPointer ={v} {CLOBBER};
return _9;
store merging has code to handle this, but punts on non-integral types
(POINTER_TYPE_P in this case).
Just changing:
--- gcc/gimple-ssa-store-merging.c.jj 2020-07-28 15:39:09.909757589 +0200
+++ gcc/gimple-ssa-store-merging.c 2020-08-11 20:19:19.293299266 +0200
@@ -333,7 +333,8 @@ init_symbolic_number (struct symbolic_nu
{
int size;
- if (! INTEGRAL_TYPE_P (TREE_TYPE (src)))
+ if (! INTEGRAL_TYPE_P (TREE_TYPE (src))
+ && ! POINTER_TYPE_P (TREE_TYPE (src)))
return false;
n->base_addr = n->offset = n->alias_set = n->vuse = NULL_TREE;
seems to work, though I must say I don't find much rationale for byte-swapping
pointers...