Bug report: Incorrect FMA emission on non-FMA-capable targets.

[piotrtopnotch]

x86-64 icx 2025.3.1 and older versions.

With just "-O2" the following program:

#include <cmath>

double fn_test(double x, double y, double z) {
    return std::fma(x, y, z);
}​

is translated into:

fn_test(double, double, double):
        mulsd   xmm0, xmm1
        addsd   xmm0, xmm2
        ret​

which is totally wrong. Per the C++ std::fma specification, "Computes x * y + z as if to infinite precision and rounded only once to fit the result type."

The infinite precision is being lost with the sequence listed above. The compiler should emit a call to a library function that emulates FMA correctly on the targets missing that instruction, as e.g. clang does:

fn_test(double, double, double):
        jmp     fma@PLT
​

Severity: very high, as techniques like double-double immediately fail due to incorrect rounding.

 

[Viet_H_Intel]

This is due to the fact that floating-point calculations use different default settings than Clang.

By default, icx enables -fp-model=fast. Compiling with -fp-model=precise, you would see:

         jmp fma@PLT # TAILCALL

Likewise, if you compile with clang -O2 -ffp-model=fast you would see:

       mulsd   xmm0, xmm1

        addsd   xmm0, xmm2

        ret

[piotrtopnotch]

Thank you. This default is very dangerous, but otherwise your remark is valid.