SVC_ME_FunTest failed on loongarch64 platform with openh264 2.5.0

[wszqkzqk]

With openh264 2.5.0 or 2.6.0's test process, SVC_ME_FunTest failed on loongarch64 platform:

test/encoder/EncUT_SVC_me.cpp:143: Failure
Expected equality of these values:
  iRes[0]
    Which is: 7850
  iRes[1]
    Which is: 8361
[  FAILED  ] SVC_ME_FunTest.SumOf8x8SingleBlock_lsx (0 ms)

[wszqkzqk]

Additional info: This error only occur while built with LTO (either bfd or mold).

/cc @jinboson @wangluls

[jinboson]

Thanks, we will check what's going on.

[wszqkzqk]

There is a warning in LTO:

./codec/processing/src/vaacalc/vaacalculation.h:148:25: warning: type of ‘VAACalcSad_lsx’ does not match original declaration [-Wlto-type-mismatch]
  148 | VAACalcSadFunc          VAACalcSad_lsx;
      |                         ^
codec/processing/src/loongarch/vaa_lsx.c:46:6: note: type mismatch in parameter 8
   46 | void VAACalcSad_lsx (const uint8_t* pCurData, const uint8_t* pRefData,
      |      ^
codec/processing/src/loongarch/vaa_lsx.c:46:6: note: ‘VAACalcSad_lsx’ was previously declared here

[xry111]

There is a warning in LTO:

./codec/processing/src/vaacalc/vaacalculation.h:148:25: warning: type of ‘VAACalcSad_lsx’ does not match original declaration [-Wlto-type-mismatch]
  148 | VAACalcSadFunc          VAACalcSad_lsx;
      |                         ^
codec/processing/src/loongarch/vaa_lsx.c:46:6: note: type mismatch in parameter 8
   46 | void VAACalcSad_lsx (const uint8_t* pCurData, const uint8_t* pRefData,
      |      ^
codec/processing/src/loongarch/vaa_lsx.c:46:6: note: ‘VAACalcSad_lsx’ was previously declared here

I submitted #3864 to fix this, but it's not enough to fix the test failure.

I've spotted a compiler bug related to __lsx_vldx which is used by SumOf8x8SingleBlock_lsx: https://gcc.gnu.org/PR119084 but I don't know if the failure is really caused by the compiler bug.

[xry111]

I've spotted a compiler bug related to __lsx_vldx which is used by SumOf8x8SingleBlock_lsx: https://gcc.gnu.org/PR119084 but I don't know if the failure is really caused by the compiler bug.

Building openh264 with a patched GCC has fixed the issue.