https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127851
Bug ID: 127851
Summary: combine can reorder two volatile reads by folding them
into one insn
Product: gcc
Version: 16.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: rtl-optimization
Assignee: vineetg at gcc dot gnu.org
Reporter: vineetg at gcc dot gnu.org
Target Milestone: ---
Consider the following test:
```
extern volatile int x, y;
int
f (void)
{
int a = x; /* sequence point after this full expression */
int b = y;
return a < b;
}
```
x must be read before y: the individual assignment constitute the sequence
points.
However on x86 the generated code is
-O2
f:
movl y(%rip), %eax # y 1st
cmpl %eax, x(%rip) # x 2nd
setl %al
movzbl %al, %eax
ret
FWIW at -O2, -ffuse-ops-with-volatile-access is enabled, but as documented it
can't change the semantics which above clearly does.
```
Allow limited optimization of operations with volatile memory access
when doing so does not change the semantics outlined in ....
```
This can cause subtle reordering accesses in Linux kernel where multiple
READ_ONCE could be reordered due to this bug.
Not seen on aarch64 because the it doesnt have `(compare (mem/v x) (mem/v y))'
form which x86 supports.
