On 5/6/25 10:50, Richard Henderson wrote:
+ *(uint64_t *)&x = 0x4ff0000000000000UL;
+ *(uint64_t *)&y = 0x4ff0000000000000UL;
0x1.0p256
+ *(uint64_t *)&z = 0x2ff0000000000000UL;
0x1.0p-256
+
+ fesetround(FE_DOWNWARD);
+ asm("fnmsub.d %[x], %[x], %[y], %[z]\n\t"
+ :[x]"+f"(x)
+ :[y]"f"(y), [z]"f"(z));
+
+ assert(*(uint64_t *)&x == 0xdfefffffffffffffUL);
-(0x1.0p512 - epsilon)
That really should have worked as-is. I'll have a quick look.
The negate_result is applied early, so we're computing rounding on the -(1p512 - epsilon),
and since that's negative, round_down produces -1p512. Whereas what the ISA wants is the
rounding on (1p512 - epsilon), rounding down to 1.fffffffffffffp511, and only afterward
negating.
I'm thinking that the current placement of the float_muladd_negate_result test is
incorrect, since we're not negating the result, we're negating the unrounded intermediate.
There are only 3 targets which use float_muladd_negate_result: loongarch, ppc, and sparc.
I will use this test case to test those other targets as well.
r~
