FPU is not enabled in Newlib for ARC-V

[kolerov]

FPU must be enabled before using floating point instructions. Newlib's crt0.S and arcv-crt0.S (our own custom crt0) does not do this. Here is an example of how it's done in Picolibc:

#ifdef __riscv_flen
	__asm__("csrr	t0, mstatus\n"
                "li	t1, 8192\n"     	// 1 << 13 = 8192
                "or	t0, t1, t0\n"
                "csrw	mstatus, t0\n"
                "csrwi	fcsr, 0");
#endif

Here is a piece of code that could be used as a quick workaround. It's placed in .init section and executed before main.

#ifdef __riscv_flen
__attribute__((naked)) __attribute__((used)) __attribute__((aligned(4))) __attribute__((constructor))
void enable_fpu(void)
{
        __asm__("csrr   t0, mstatus\n"
                "li     t1, 8192\n"             // 1 << 13 = 8192
                "or     t0, t1, t0\n"
                "csrw   mstatus, t0\n"
                "csrwi  fcsr, 0\n"
                "ret");
}
#endif

It would be better to add support of enabling FPU right in arcv-crt0.S. Maybe it's worth placing the same code to original crt0.S and push to upstream.

[kolerov]

The fix is pushed in arc64 and arc-2024.06 branches of Newlib repository. Now FPU is enabled for configurations with f or d extensions.