* Re: Implement a build-in function with floating point exceptions.
[not found] <SG2PR06MB2697E3105BBA8B4DF5558647C8129@SG2PR06MB2697.apcprd06.prod.outlook.com>
@ 2022-11-29 14:19 ` Xi Ruoyao
0 siblings, 0 replies; only message in thread
From: Xi Ruoyao @ 2022-11-29 14:19 UTC (permalink / raw)
To: 陈 小龙, gcc-patches; +Cc: i, xuchenghua, chenglulu
On Tue, 2022-11-29 at 09:52 +0000, 陈 小龙 wrote:
> +(define_insn "<float_pattern><mode>2"
> + [(set (match_operand:ANYF 0 "register_operand" "=f")
> + (unspec:ANYF[(match_operand:ANYF 1 "register_operand" "f")]
> + FLOAT))]
> + "TARGET_HARD_FLOAT&&(flag_fp_int_builtin_inexact ||
> !flag_trapping_math)"
> +
> "ftint<float_submenmonic>.<wl>.<fmt>\t%0,%1\n\tffint.<fmt>.<wl>\t%0,%0
It's wrong. For example, consider ceil(0x1.0p+100). 0x1.0p+100 cannot
be represented by any 64-bit integer, so ftintrp.l.d will raise an
overflow exception and provide "some value" as the output. Then ffint
can't produce the correct result.
Nacked-by: Xi Ruoyao <xry111@xry111.site>
I'd suggest to add frint.{w,l}.{s,d}.{rm,rp,rz} etc. (BTW, also the
variants without raising inexact exceptions) into a future LoongArch ISA
version for this.
--
Xi Ruoyao <xry111@xry111.site>
School of Aerospace Science and Technology, Xidian University
^ permalink raw reply [flat|nested] only message in thread