Support native fma on riscv64

[giordano]

This is an attempt to support native fma on riscv64, but I'm quite confused because with this PR I get at the same time

julia> Core.Intrinsics.have_fma(Float64)
false

julia> code_llvm(fma, NTuple{3,Float64}; debuginfo=:none)

; Function Signature: fma(Float64, Float64, Float64)
define double @julia_fma_1141(double %"x::Float64", double %"y::Float64", double %"z::Float64") #0 {
common.ret:
    #dbg_value(double %"x::Float64", !2, !DIExpression(), !16)
    #dbg_value(double %"y::Float64", !14, !DIExpression(), !16)
    #dbg_value(double %"z::Float64", !15, !DIExpression(), !16)
  %0 = call double @llvm.fma.f64(double %"x::Float64", double %"y::Float64", double %"z::Float64")
  ret double %0
}

Core.Intrinsics.have_fma claims native FMA instructions aren't available, but fma now uses @llvm.fma.f64 instead of going through the emulated path, as it does on master:

julia> code_llvm(fma, NTuple{3,Float64}; debuginfo=:none)

; Function Signature: fma(Float64, Float64, Float64)
define double @julia_fma_1023(double %"x::Float64", double %"y::Float64", double %"z::Float64") #0 {
common.ret:
    #dbg_value(double %"x::Float64", !3, !DIExpression(), !17)
    #dbg_value(double %"y::Float64", !15, !DIExpression(), !17)
    #dbg_value(double %"z::Float64", !16, !DIExpression(), !17)
  %0 = call double @j_fma_emulated_1026(double %"x::Float64", double %"y::Float64", double %"z::Float64")
  ret double %0
}

@gbaraldi any idea of what's going on?

[gbaraldi]

@giordano the runtime version of this needs to also check it correctly. Also how are we passing the runtime features?

[giordano]

the runtime version of this needs to also check it correctly.

You mean I need to complement a function like

julia/src/processor_x86.cpp

Lines 1094 to 1102 in 6cf2b14

	JL_DLLEXPORT jl_value_t *jl_cpu_has_fma(int bits)
	{
	TargetData<feature_sz> target = jit_targets.front();
	FeatureList<feature_sz> features = target.en.features;
	if ((bits == 32 || bits == 64) && (test_nbit(features, Feature::fma) || test_nbit(features, Feature::fma4)))
	return jl_true;
	else
	return jl_false;
	}

for riscv64?

Also how are we passing the runtime features?

Not sure what you mean.

[oscardssmith]

You mean I need to complement a function like

yes. The part you've done so far is the version for where have_fma is figured out at compile time, but the other implementation is for when it is done as a runtime check.

[giordano]

I think that overlaps quite a bit with #57049?