| Issue |
242593
|
| Summary |
[X86][LVI] Debug records can cause duplicate LFENCE insertion
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
yqtian-se
|
With `-mlvi-hardening`, compiling this function with `-g` emits two
consecutive `lfence` instructions where `-g0` emits one.
```c
int *g(int);
int f(int **a, int **b, volatile int *n, int c) {
int *p;
do { p = *a; a++; } while (*n);
int *q;
if (c) { q = *b; } else { q = g(c); }
return q[*p];
}
```
```sh
clang --target=x86_64-linux-gnu -O2 -mlvi-hardening -g -S input.c -o with-g.s
clang --target=x86_64-linux-gnu -O2 -mlvi-hardening -g0 -S input.c -o without-g.s
grep -n lfence with-g.s without-g.s
```
The relevant difference is:
```diff
movq (%rsi), %rax
+ lfence
lfence
```
The first difference appears in `x86-lvi-load`. `insertFences` checks the raw
previous machine instruction when deciding whether a fence is already present.
A `DBG_VALUE` between the existing `LFENCE` and the branch hides the fence, so
the pass inserts another one.
Debug records should not defeat the adjacent-fence check. This was reproduced
with clang `7517bcdb0b87`; the responsible code remains unchanged on LLVM main
`2451a1eadfa1`.
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs