Enhance Intrinsic::fmuladd translation.

[haonanya]

-cl-mad-enable is off by default, however ReplaceLLVMFmulAddWithOpenCLMad is true by default https://github.com/KhronosGroup/SPIRV-LLVM-Translator/blob/main/include/LLVMSPIRVOpts.h#L244C8-L244C40, which seems not consistent. Intrinsic::fmuladd should be translated into mad when function attribute "less-precise-fpmad" = "true" which indicates -cl-mad-enable is specified. When "less-precise-fpmad" = "false" or there is no "less-precise-fpmad", translate it into a pair of fmul + fadd.
Can we remove ReplaceLLVMFmulAddWithOpenCLMad uses and let only "less-precise-fpmad" determine how to translate fmuladd?
@AlexeySachkov, @MrSidims, @svenvh , do you think is ReplaceLLVMFmulAddWithOpenCLMad still needed and what's your opinion for above comments? Thanks very much. #2163 is drafted and need further discussion.

[wenju-he]

fmuladd should be translated to fmuladd as it is. It is likely the lack of support of the translation in spirv spec so that it is translate to mad.

[AlexeySachkov]

fmuladd should be translated to fmuladd as it is. It is likely the lack of support of the translation in spirv spec so that it is translate to mad.

That is correct. We don't have a direct equivalent of fmuladd in SPIR-V and therefore we have decided to go with mad instead.

I'm not entirely sure if we only need to use mad when -cl-mad-enable is present, because for FULL profile mad is expected to have the same semantics as fmuladd. Therefore, the current approach is valid for FULL profile. At first glance, it seems to me that if someone is targeting an embedded profile where mad could be less precise, then they should switch the value of the option.

Tagging @bashbaug for awareness as well

[bashbaug]

I agree with @AlexeySachkov - for full profile devices the OpenCL mad extended instruction has the same semantics as the LLVM fmuladd.

LLVM: "The ‘llvm.fmuladd.*’ intrinsic functions represent multiply-add expressions that can be fused if the code generator determines that (a) the target instruction set has support for a fused operation, and (b) that the fused operation is more efficient than the equivalent, separate pair of mul and add instructions."

SPIR-V: "OpExtInst mad - Implemented either as a correctly rounded fma, or as a multiply followed by an add, both of which are correctly rounded."

We have an internal Khronos tracker to improve this (SPIR-V issue 409) and support fmuladd more directly, but because it is only needed for embedded profile devices we currently consider it low priority.