[EntryExitInstrumenter][AArch64][RISCV][Loongarch] return address should be passed into `_mcount`

[dtcxzyw]

Reproducer:

; bin/clang -pg test.c && ./a.out && gprof -b a.out gmon.out
int fib(int n) {return n <=1 ? 1 : fib(n-1) + fib(n-2); }
int main() {fib(10); return 0;}

The call graph is missing if it is compiled with clang:

Flat profile:

Each sample counts as 0.01 seconds.
 no time accumulated

  %   cumulative   self              self     total           
 time   seconds   seconds    calls  Ts/call  Ts/call  name

With gcc, it works as expected:

Flat profile:

Each sample counts as 0.01 seconds.
 no time accumulated

  %   cumulative   self              self     total           
 time   seconds   seconds    calls  Ts/call  Ts/call  name    
  0.00      0.00     0.00        1     0.00     0.00  fib


                        Call graph


granularity: each sample hit covers 2 byte(s) no time propagated

index % time    self  children    called     name
                                 176             fib [1]
                0.00    0.00       1/1           main [6]
[1]      0.0    0.00    0.00       1+176     fib [1]
                                 176             fib [1]
-----------------------------------------------


Index by function name

   [1] fib

This issue can be reproduced on AArch64/RISCV/Loongarch platforms.

After investing this issue, I found that we should pass the return address into _mcount:
https://github.com/gcc-mirror/gcc/blob/c6b7d0343f2c0eb3102341c9326fe314a047bee6/gcc/config/riscv/riscv.h#L466-L473
https://github.com/bminor/glibc/blob/757ac24f8f2344e5f8afd2465c808d0f6adcc948/sysdeps/riscv/machine-gmon.h

This issue was originally reported by @wzssyqa.

[llvmbot]

@llvm/issue-subscribers-backend-risc-v

Author: Yingwei Zheng (dtcxzyw)

Reproducer: ``` ; bin/clang -pg test.c && ./a.out && gprof -b a.out gmon.out int fib(int n) {return n <=1 ? 1 : fib(n-1) + fib(n-2); } int main() {fib(10); return 0;} ``` The call graph is missing if it is compiled with clang: ``` Flat profile:

Each sample counts as 0.01 seconds.
no time accumulated

% cumulative self self total
time seconds seconds calls Ts/call Ts/call name

With gcc, it works as expected:

Flat profile:

Each sample counts as 0.01 seconds.
no time accumulated

% cumulative self self total
time seconds seconds calls Ts/call Ts/call name
0.00 0.00 0.00 1 0.00 0.00 fib

                    Call graph

granularity: each sample hit covers 2 byte(s) no time propagated

index % time self children called name
176 fib [1]
0.00 0.00 1/1 main [6]
[1] 0.0 0.00 0.00 1+176 fib [1]
176 fib [1]

Index by function name

[1] fib

This issue can be reproduced on AArch64/RISCV/Loongarch platforms.

After investing this issue, I found that we should pass the return address into `_mcount`:
https://github.com/gcc-mirror/gcc/blob/c6b7d0343f2c0eb3102341c9326fe314a047bee6/gcc/config/riscv/riscv.h#L466-L473
https://github.com/bminor/glibc/blob/757ac24f8f2344e5f8afd2465c808d0f6adcc948/sysdeps/riscv/machine-gmon.h

This issue was originally reported by @<!-- -->wzssyqa.

</details>

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: Yingwei Zheng (dtcxzyw)

Reproducer: ``` ; bin/clang -pg test.c && ./a.out && gprof -b a.out gmon.out int fib(int n) {return n <=1 ? 1 : fib(n-1) + fib(n-2); } int main() {fib(10); return 0;} ``` The call graph is missing if it is compiled with clang: ``` Flat profile:

Each sample counts as 0.01 seconds.
no time accumulated

% cumulative self self total
time seconds seconds calls Ts/call Ts/call name

With gcc, it works as expected:

Flat profile:

Each sample counts as 0.01 seconds.
no time accumulated

% cumulative self self total
time seconds seconds calls Ts/call Ts/call name
0.00 0.00 0.00 1 0.00 0.00 fib

                    Call graph

granularity: each sample hit covers 2 byte(s) no time propagated

index % time self children called name
176 fib [1]
0.00 0.00 1/1 main [6]
[1] 0.0 0.00 0.00 1+176 fib [1]
176 fib [1]

Index by function name

[1] fib

This issue can be reproduced on AArch64/RISCV/Loongarch platforms.

After investing this issue, I found that we should pass the return address into `_mcount`:
https://github.com/gcc-mirror/gcc/blob/c6b7d0343f2c0eb3102341c9326fe314a047bee6/gcc/config/riscv/riscv.h#L466-L473
https://github.com/bminor/glibc/blob/757ac24f8f2344e5f8afd2465c808d0f6adcc948/sysdeps/riscv/machine-gmon.h

This issue was originally reported by @<!-- -->wzssyqa.

</details>

[dtcxzyw]

@wzssyqa Please file a separate issue for MIPS target.

[SixWeining]

Are the 2 arguments (frompc and selfpc) to mcount_internal the same? Both are __builtin_return_address(0)?

[dtcxzyw]

Are the 2 arguments (frompc and selfpc) to mcount_internal the same? Both are __builtin_return_address(0)?

No. Please refer to the generic implementation: https://github.com/bminor/glibc/blob/757ac24f8f2344e5f8afd2465c808d0f6adcc948/sysdeps/generic/machine-gmon.h.
frompc should be __builtin_return_address(1). Since frompc is passed from the caller of _mcount on these platforms, we should use __builtin_return_address(0) instead.