tree 40caedcccf29fdc726b99902eafc21e4b314ff4c
parent 1c9d18307bdf75bee7f3ec7f7d2e421a734831f1
author David Benjamin <davidben@google.com> 1674086139 -0500
committer Boringssl LUCI CQ <boringssl-scoped@luci-project-accounts.iam.gserviceaccount.com> 1674505312 +0000

Maintain a frame pointer in aesni-gcm-x86_64.pl and add SEH unwind codes

Some profiling systems cannot unwind with CFI and benefit from having a
frame pointer. Since this code doesn't have enough register pressure to
actually need to use rbp as a general register, this change tweaks
things so that a frame pointer is preserved.

As this would invalidate the SEH handler, just replace it with proper
unwind codes, which are more profiler-friendly and supportable by our
unwind tests. Some notes on this:

- We don't currently support the automatic calling convention conversion
  with unwind codes, but this file already puts all arguments in
  registers, so I just renamed the arguments and put the last two
  arguments in RDI and RSI. Those I stashed into the parameter stack
  area because it's free storage.

- It is tedious to write the same directives in both CFI and SEH. We
  really could do with an abstraction. Although since most of our
  functions need a Windows variation anyway.

- I restored the original file's use of PUSH to save the registers.
  This matches what Clang likes to output anyway, and push is probably
  smaller than the corresponding move with offset. (And it reduces how
  much thinking about offsets I need to do.)

- Although it's an extra instruction, I restored the original file's
  separate fixed stack allocation and alloca for the sake of clarity.

- The epilog is constrained by Windows being extremely picky about
  epilogs. (Windows doesn't annotate epilogs and instead simulates
  forward.) I think other options are possible, but using LEA with an
  offset to realign the stack for the POPs both matches the examples in
  Windows and what Clang seems to like to output. The original file used
  MOV with offset, but it seems to be related to the funny SEH handler.

- The offsets in SEH directives may be surprising to someone used to CFI
  directives or a SysV RBP frame pointer. All three use slightly
  different baselines:

  CFI's canonical frame address (CFA) is RSP just before a CALL (so
  before the saved RIP in stack order). It is 16-byte aligned by ABI.

  A SysV RBP frame pointer is 16 bytes after that, after a saved RIP and
  saved RBP. It is also 16-byte aligned.

  Windows' baseline is the top of the fixed stack allocation, so
  potentially some bytes after that (all pushreg and allocstack
  directives). This too is required to be 16-byte aligned.

  Windows, however, doesn't require the frame register actually contain
  the fixed stack allocation. You can specify an offset from the value
  in the register to the actual top. But all the offsets in savereg,
  etc., directives use this baseline.

Performance difference is within measurement noise.

This does not create a stack frame for internal functions so
frame-pointer unwinding may miss a function or two, but the broad
attribution will be correct.

Change originally by Clemens Fruhwirth. Then reworked from Adam
Langley's https://boringssl-review.googlesource.com/c/boringssl/+/55945
by me to work on Windows and fix up some issues with the RBP setup.

Bug: b/33072965, 259
Change-Id: I52302635a8ad3d9272404feac125e2a4a4a5d14c
Reviewed-on: https://boringssl-review.googlesource.com/c/boringssl/+/56128
Reviewed-by: Adam Langley <agl@google.com>
Commit-Queue: David Benjamin <davidben@google.com>