[PeterK]

Originally Posted by Toni Wilen

I did the same to 80-bit when using multi-instruction code (trigonometric/logaritmic) because "wrong" mid instruction rounding would cause side-effects. ...

It is not going to work because regs.fpcr content would become JIT compile time constant. It needs to be pointer to regs.fpcr.

Ok, so it would be better to remove the if clauses completely and also use a pointer to regs.fpcr when setting the mode to "round to nearest":

			case 0x1c: /* FACOS */
#if USE_X86_FPUCW
			{ 
			mov_l_rm (S1, uae_p32(&regs.fpcr));
			and_l_ri (S1, 0xC0);  /* use round to nearest */
			fldcw_m_indexed (S1, uae_p32(x86_fpucw));
			facos_rr (dreg, sreg);
			mov_l_rm (S1, uae_p32(&regs.fpcr));
			and_l_ri (S1, 0xf0); /* restore control word */
			fldcw_m_indexed (S1, uae_p32(x86_fpucw));
			break;
			}
#endif
			facos_rr (dreg, sreg);
			break;

... and you could also remove the already disabled if(0)/else clauses and the dead code from FINTRZ where I wrote the comment that the conditions are checked at compilation time.