[PATCH] D88988: [llvm-symbolizer] Add inline stack traces for Windows.

[Reid Kleckner via Phabricator via llvm-commits]

rnk added a comment.

In D88988#2322424 <https://reviews.llvm.org/D88988#2322424>, @akhuang wrote:

> Ok, as far as I can tell, all of the asan tests are failing for the same reason-- the symbolizer now outputs an extra line for __sanitizer::BufferedStackTrace::Unwind.
>
>   #0 0x7ff6f9fa7e64 in __sanitizer::BufferedStackTrace::Unwind C:\src\llvm-project\compiler-rt\lib\sanitizer_common\sanitizer_stacktrace.h:124
>   #1 0x7ff6f9fa7e64 in malloc C:\src\llvm-project\compiler-rt\lib\asan\asan_malloc_win.cpp:98

I guess I misspoke when we chatted, I think we need to avoid these extra frames. Here's the line of code that I think is executing:
https://github.com/llvm/llvm-project/blob/d784f7406911c4fb6bc559320f7f9ff134be7ff5/compiler-rt/lib/asan/asan_stack.h#L45

  stack.Unwind(StackTrace::GetCurrentPc(),                     \
               GET_CURRENT_FRAME(), nullptr, fast, max_size);  \

Here's what I think might be happening: in the MS C++ ABI, argument evaluation has to be right-to-left. On Linux, it is left to right. So, on Linux, we get assembly that looks like this:

  callq GetCurrentPC
  movq $0, %rdi # set up other args
  ...
  # inlined callsite for Unwind

But on Windows, that's rearranged like so:

  movq $0, %rcx # set up other args
  callq GetCurrentPc # set up rightmost arg last
  # inlined call site for Unwind

I guess to sort it all out, the thing to do is to get the annotated assembly produced by clang-cl, and look at the assembly stream. It will have the human-readable .cv_loc directives to help us work out where to apply the fix. We could, for example, subtract one from the result of GetCurrentPc().


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D88988/new/

https://reviews.llvm.org/D88988