wrong variable name for vla about kmsan HOT 9 CLOSED

Also filed https://bugs.llvm.org/show_bug.cgi?id=37207 as it's msan bug.

from kmsan.

Right now [K]MSan uses AllocaInst.getName() to obtain the local's name.
It's actually possible to query the LLVM metadata to get the exact name and declaration location for local vars.

Another idea we've just discussed is to associate the local var's DILocation with the corresponding __msan_poison_alloca() call.

from kmsan.

Querying metadata is a good idea, and it's not that hard to implement. It will only work with full debug info though, not -gmlt.

In your last sentence, are you saying that you want to describe __msan_poison_alloca call site as data in debug info?

from kmsan.

In your last sentence, are you saying that you want to describe __msan_poison_alloca call site as data in debug info?
No, the idea was to make the call __msan_poison_alloca have the same file:line info as the local variable.

from kmsan.

Sure. Either that or __msan_set_alloca_origin4.

from kmsan.

Another question is how do we represent the information about inlined variables.
E.g. in the following case:

static int bar(int index) {
  int buf[10];
  for (int i = 0; i < 10; i++) {
    buf[i] = (i > 1) ? buf[i-1] + buf[i-2] : 1;
  }
  return (index >= 0 && index < 10) ? buf[index] : 0;
}

static
int foo(int index) {
  return bar(index);
}

We could describe the inlined buf as one of the following options:

1. buf@foo (current version)
2. buf@bar (show only the original function)
3. buf@bar@foo (show all the inlining stack)

Version 3 is the most informative, but it'll create a lot of unnecessary global strings. Version 2 has only the deepest function in the inline stack, which may sometimes be obscure.

from kmsan.

Uninit come from local variable buf created at:
    bar bar.c:11
    foo foo.c:22
    baz baz.c:33

from kmsan.

Still reproducible with Clang 15.0.0.

from kmsan.

On the other hand, VLAs are gone from the kernel, so keeping the bug around in LLVM should be enough: https://bugs.llvm.org/show_bug.cgi?id=37207

from kmsan.