[ESIMD] Break @llvm.fmuladd back into mul/add to satisfy VC BE.

[kbobrovs]

VC BE does not support many of std llvm intrinsics for now, so need a w/a.
This patch handles fmuladd.
This fixes regression on mandelbrot.cpp E2E test.
Signed-off-by: Konstantin S Bobrovsky [email protected]

[kbobrovs]

With the latest pulldown default FE handling of FP changed and it started to generate fmuladd here.

[v-klochkov]

Please check what is wrong with ESIMD/regression/dgetrf.cpp - it failed in CI, couldn't that be caused by this patch?

[kbobrovs]

ESIMD failures in Jenkins are "unexpected pass"-es, which are expected - this patch fixes the regressions on them.
The rest of the failures are non-ESIMD, so they can't be related to this patch because it only touches llvm/lib/SYCLLowerIR/LowerESIMD.cpp.

[2021-12-03T06:26:18.803Z] Failed Tests (5):
[2021-12-03T06:26:18.803Z] SYCL :: DeviceLib/built-ins/vector_relational.cpp
[2021-12-03T06:26:18.803Z] SYCL :: DeviceLib/cmath_fp64_test.cpp
[2021-12-03T06:26:18.803Z] SYCL :: DeviceLib/cmath_test.cpp
[2021-12-03T06:26:18.803Z] SYCL :: DeviceLib/math_fp64_test.cpp
[2021-12-03T06:26:18.803Z] SYCL :: DeviceLib/math_test.cpp

[2021-12-03T06:09:03.533Z] Unexpectedly Passed Tests (3):
[2021-12-03T06:09:03.533Z] LLVM :: ESIMD/mandelbrot/mandelbrot.cpp
[2021-12-03T06:09:03.534Z] LLVM :: ESIMD/mandelbrot/mandelbrot_spec.cpp
[2021-12-03T06:09:03.534Z] LLVM :: ESIMD/regression/dgetrf.cpp

[kbobrovs]

Please check what is wrong with ESIMD/regression/dgetrf.cpp - it failed in CI, couldn't that be caused by this patch?

It is also unexpected XPASS, the reason as the same as for mandelbrot - it started to generate fmuladd, was marked as XFAIL, now this patch fixes it.

[v-klochkov]

Please check what is wrong with ESIMD/regression/dgetrf.cpp - it failed in CI, couldn't that be caused by this patch?

It is also unexpected XPASS, the reason as the same as for mandelbrot - it started to generate fmuladd, was marked as XFAIL, now this patch fixes it.

Ok, understood. BTW, then it makes sense fixing the LIT tests (remove XFAIL) and run the command/comment int this 5075 such line: "/verify with REFERENCE_TO_LIT_FIX_PR"

[kbobrovs]

After re-running

Please check what is wrong with ESIMD/regression/dgetrf.cpp - it failed in CI, couldn't that be caused by this patch?

It is also unexpected XPASS, the reason as the same as for mandelbrot - it started to generate fmuladd, was marked as XFAIL, now this patch fixes it.

Ok, understood. BTW, then it makes sense fixing the LIT tests (remove XFAIL) and run the command/comment int this 5075 such line: "/verify with REFERENCE_TO_LIT_FIX_PR"

good point, will do

[kbobrovs]

Rerunning the precommit left only the 3 "unexpected pass" failures above.

[kbobrovs]

Accompanying test fix: intel/llvm-test-suite#600

[kbobrovs]

"/verify with #5075" in intel/llvm-test-suite#600 did not work - it did not make CI pick this my patch for testing.

Nevertheless, "Unexepected pass" on the 3 tests w/o regression is 100% evidence that the patch does the right thing, so I'm merging.