clang-13 performs inlining after mcount instrumentation

[tmiasko]

In clang-13 inlining might follow the mcount instrumentation used by -pg.
The mcount relies on frame pointers to recover a call graph, so inlining after
inserting calls to mcount produces misleading end results. In the example
below, g which is called only once is recorded as being called 1025 times:

int f(int x) {
  return (x & (x << 1u)) == 0;
}

#include <stdio.h>
int f(int);
int __attribute__ ((noinline)) g() {
  int c = 0;
  for (int i = 0; i < (1 << 10); ++i)
    c += f(i);
  return c;
}

int main() {
  printf("%d\n", g());
}

$ clang-13 -g -pg -flto -O2 -c a.c
$ clang-13 -g -pg -flto -O2 -c b.c
$ clang-13 -g -pg -flto -O2 a.o b.o
$ ./a.out > /dev/null && gprof -b | grep -A2 calls
 time   seconds   seconds    calls  Ts/call  Ts/call  name    
  0.00      0.00     0.00     1025     0.00     0.00  g

$ clang-12 -g -pg -flto -O2 -c a.c
$ clang-12 -g -pg -flto -O2 -c b.c
$ clang-12 -g -pg -flto -O2 a.o b.o
$ ./a.out > /dev/null && gprof -b | grep -A2 calls
 time   seconds   seconds    calls  Ts/call  Ts/call  name    
   0.00      0.00     0.00        1     0.00     0.00  g

This seems to be result of changes from https://reviews.llvm.org/D97608,
which moved PostInlineEntryExitInstrumentationPass out of TargetPassConfig::addIRPasses.

cc @aeubanks

[aeubanks]

This is because we invoke OptimizerLastEPCallbacks multiple times during LTO, once in the first compile and then in the final optimization stage. In BackendUtil.cpp we shouldn't add the PreInlineEE pass if we're in (Thin)LTO backend compile, and we shouldn't add PostInlineEE if we're in a (Thin)LTO frontend compile.

(It's vacation time so probably won't get around to fixing this soon)

[nikic]

This is because we invoke OptimizerLastEPCallbacks multiple times during LTO, once in the first compile and then in the final optimization stage. In BackendUtil.cpp we shouldn't add the PreInlineEE pass if we're in (Thin)LTO backend compile, and we shouldn't add PostInlineEE if we're in a (Thin)LTO frontend compile.

At least for the post-inline case, wouldn't it be better undo that part of the change, and move this back into the backend pipeline? Having a pass that needs to be inserted into the post-link LTO pipeline by the frontend seems problematic, as the frontend may not control the LTO pipeline if linker plugin LTO rather than compiler LTO is used.

[alehander92]

is this fixed now? #92171 is merged and it documents that it tries to fix this issue

[pasko]

is this fixed now? #92171 is merged and it documents that it tries to fix this issue

I did not test it, but chances are it is fixed.