https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102125
--- Comment #5 from Richard Earnshaw <rearnsha at gcc dot gnu.org> ---
Testcase was not quite complete. Extending it to:
typedef unsigned long long uint64_t;
typedef unsigned long uint32_t;
typedef unsigned char uint8_t;
uint64_t bar64(const uint8_t *rData1)
{
uint64_t buffer;
__builtin_memcpy(&buffer, rData1, sizeof(buffer));
return buffer;
}
uint32_t bar32(const uint8_t *rData1)
{
uint32_t buffer;
__builtin_memcpy(&buffer, rData1, sizeof(buffer));
return buffer;
}
and then looking at the optimized tree output we see:
;; Function bar64 (bar64, funcdef_no=0, decl_uid=4196, cgraph_uid=1,
symbol_order=0)
uint64_t bar64 (const uint8_t * rData1)
{
uint64_t buffer;
uint64_t _4;
<bb 2> [local count: 1073741824]:
__builtin_memcpy (&buffer, rData1_2(D), 8);
_4 = buffer;
buffer ={v} {CLOBBER};
return _4;
}
;; Function bar32 (bar32, funcdef_no=1, decl_uid=4200, cgraph_uid=2,
symbol_order=1)
uint32_t bar32 (const uint8_t * rData1)
{
unsigned int _3;
<bb 2> [local count: 1073741824]:
_3 = MEM <unsigned int> [(char * {ref-all})rData1_2(D)];
return _3;
}
So in the 32-bit case we've eliminated the memcpy at the tree level, but failed
to do that for 64-bit objects.
We probably need to add 64-bit support to the movmisalign<mode> pattern.
