iii-i wrote:
There is special byval handling in MSan:
```
void visitCallBase(CallBase &CB, IRBuilder<> &IRB) override {
...
for (const auto &[ArgNo, A] : llvm::enumerate(CB.args())) {
...
bool IsByVal = CB.paramHasAttr(ArgNo, Attribute::ByVal);
```
If a target does not make use of byval, I would argue that this part of MSan is
broken there.
Unfortunately this applies to SystemZ at the moment: `param_tls_limit.cpp` is
marked as XFAIL because of this (the rationale that is written there, which
says that the test is not applicable, is not correct).
A quick experiment shows that x86_64 uses byval:
```
$ uname -m
x86_64
$ clang --version
clang version 16.0.6 (Fedora 16.0.6-2.fc38)
$ cat 1.c
struct foo { char x[800]; };
void bar(struct foo) {};
$ clang -mllvm -print-after-all -c 1.c
[...]
define dso_local void @bar(ptr noundef byval(%struct.foo) align 8 %0) #0 {
[...]
```
As for the reasons why this all is the case, I think this is because of how
arguments' shadow TLS is defined: it contains shadows of all argument values,
where arguments are understood in terms of C, and not in terms of the ABI. This
definition is then used in the runtime. See, for example, the handling of long
doubles in `dfsan_custom.cpp`:
```
static int format_buffer(char *str, size_t size, const char *fmt,
dfsan_label *va_labels, dfsan_label *ret_label,
dfsan_origin *va_origins, dfsan_origin *ret_origin,
va_list ap) {
[...]
case 'G':
if (*(formatter.fmt_cur - 1) == 'L') {
retval = formatter.format(va_arg(ap, long double));
} else {
retval = formatter.format(va_arg(ap, double));
}
if (va_origins == nullptr)
dfsan_set_label(*va_labels++, formatter.str_cur(),
formatter.num_written_bytes(retval));
else
dfsan_set_label_origin(*va_labels++, *va_origins++,
formatter.str_cur(),
formatter.num_written_bytes(retval));
end_fmt = true;
break;
```
Here is wants the shadow of the long double itself, even if the target ABI
requires passing it through a pointer.
https://github.com/llvm/llvm-project/pull/66404
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
