https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63900
--- Comment #9 from Christopher Head <headch at gmail dot com> ---
I had to use slightly different code because I only have an ARM cross-compiler
version 8.2.0 installed, so I used this:
void g(unsigned char x);
struct MyStruct {
char foo[8];
};
unsigned char buffer[100], buffer2[100];
void f(void) {
buffer2[5] = 'X';
struct MyStruct *p = (void *) buffer;
asm volatile("nop" : "=m" (*p));
g(buffer2[5]);
}
No matter what I changed the size of the foo array in MyStruct to (I tried 8,
15, 13, and 257), this was always the disassembly:
00000000 <f>:
0: 4a03 ldr r2, [pc, #12] ; (10 <f+0x10>)
2: 2058 movs r0, #88 ; 0x58
4: 4b03 ldr r3, [pc, #12] ; (14 <f+0x14>)
6: 7150 strb r0, [r2, #5]
8: bf00 nop
a: f7ff bffe b.w 0 <g>
a: R_ARM_THM_JUMP24 g
e: bf00 nop
Notice that r0 is loaded at offset 2, then passed into g at offset a, and is
not reloaded from the array. If I change the asm to have a full memory clobber,
an ldrb is inserted between the nop and the b.w, as expected, since then
buffer2 is also clobbered.
So, doesn’t this mean that it *is* properly clobbering only buffer, not
buffer2, regardless of the size of foo?
