[llvm] [DebugInfo][DWARF] Add heapallocsite information (PR #132073)

[via llvm-commits]

thejh wrote:

@dwblaikie You mean a size comparison of clang built with debuginfo at a specific revision, built once using unmodified clang and once using clang with this patch applied?

Here you go - I built clang twice at git commit `5475834737663c0da3c444fcf1c8ab5567c39136`, once using original clang (from git commit `5475834737663c0da3c444fcf1c8ab5567c39136`), once using modified clang with my commit applied, both times using `-DCMAKE_BUILD_TYPE=RelWithDebInfo -DLLVM_ENABLE_PROJECTS=clang`:
```
$ [...]/bloaty [...] [...]/llvmp-build-2-debug-modcompiler/bin/clang -- [...]/llvmp-build-2-debug-origcompiler/bin/clang
    FILE SIZE        VM SIZE
 --------------  --------------
  +0.0%  +256Ki  [ = ]       0    .debug_info
  +0.3% +94.2Ki  [ = ]       0    .debug_addr
  +0.2% +23.4Ki  [ = ]       0    .debug_abbrev
  +0.0% +4.85Ki  [ = ]       0    .debug_loclists
  +0.0% +2.63Ki  [ = ]       0    .debug_str_offsets
  +0.0% +1.45Ki  [ = ]       0    .debug_str
  +0.0% +1.03Ki  [ = ]       0    .debug_rnglists
  -0.0%      -2  [ = ]       0    .debug_line_str
  +0.0%  +384Ki  [ = ]       0    TOTAL
$ ls -Ll [...]/llvmp-build-2-debug-{orig,mod}compiler/bin/clang
-rwxr-x--- 1 [...] 1454958680 Apr  3 20:54 [...]/llvmp-build-2-debug-modcompiler/bin/clang
-rwxr-x--- 1 [...] 1454564968 Apr  3 19:05 [...]/llvmp-build-2-debug-origcompiler/bin/clang
```
Grepping the output of running `llvm-dwarfdump` on `llvmp-build-2-debug-modcompiler/bin/clang`, it looks like there are 17639 `DW_AT_LLVM_heapallocsite` DWARF entries.

If you think that's too much to add in by default when there is currently no tooling that can make use of this, I'd be happy to gate this on some new default-disabled flag.

Yes, I have a planned usecase - I would like to eventually use this feature for augmenting Linux kernel memory access traces, and maybe also ASAN crash reports (which already report allocator call site addresses), with type information. However, that usecase requires more LLVM/clang patches on top of this, since Linux kernel code uses something like `malloc(sizeof(<type>))` rather than C++ `new`. I have not yet talked about this with upstream Linux developers, and figured I'd maybe simultaneously talk about this idea on the Linux mailing list and in https://discourse.llvm.org/ once I have some draft patch(es) for handling some more complicated versions of patterns like `malloc(sizeof())`.

Another usecase I had in the past, for which I originally wrote this patch, was to analyze heap memory usage of a C++ codebase (like seeing which classes use how much memory, or seeing statistics for what the most common values of some class member are) - for that usecase, this patch alone mostly did the job (combined with allocator changes to record callsites). Since this patch was already independently useful for that kind of usecase, I figured I might as well land this first.
(And since https://llvm.org/docs/Contributing.html#how-to-submit-a-patch says that the preferred way of sending patches is to send one commit at a time, I figured you might also prefer getting piecemeal patches as long as they're independently useful.)

https://github.com/llvm/llvm-project/pull/132073