https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86520
Bug ID: 86520
Summary: AArch64: Two 8-bit accesses coalesced into a single
16-bit access
Product: gcc
Version: 7.3.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: swarren at nvidia dot com
Target Milestone: ---
When compiling for AArch64 (64-bit ARM), the following code:
void func(char *str)
{
*str = '0';
str++;
*str = 'x';
str++;
}
... I believe should always perform two separate 8-bit accesses, simply because
the compiler can't know whether str is aligned to anything more than byte
alignment.
With recent compilers without optimization, and with older compilers, the
following code is generated:
0000000000000000 <func>:
0: 52800601 mov w1, #0x30 // #48
4: 39000001 strb w1, [x0]
8: 52800f01 mov w1, #0x78 // #120
c: 39000401 strb w1, [x0,#1]
10: d65f03c0 ret
However, with any all Linaro gcc 7.x releases and -O3, those two byte writes
are combined:
0000000000000000 <func>:
0: 528f0601 mov w1, #0x7830 // #30768
4: 79000001 strh w1, [x0]
8: d65f03c0 ret
When this happens in the real code I'm compiling/execution, an alignment
exception is thrown because str (x0) is only byte aligned not 16-bit aligned.
I have not tested gcc-8.x yet, since I'm using Linaro binary releases not
building from source.
