comiss vs. ucomiss
Thomas.Rodriguez at Sun.COM
Tue Oct 14 12:27:07 PDT 2008
I wouldn't expect it to be any different execution-wise. It's just odd.
On Oct 14, 2008, at 11:08 AM, Jiva, Azeem wrote:
> At least on AMD hardware latency costs are identical for comiss/comisd
> and ucomiss/ucomisd and they both use the same decode paths.
> Azeem Jiva
> AMD Java Labs
> T 512.602.0907
>> -----Original Message-----
>> From: hotspot-compiler-dev-bounces at openjdk.java.net [mailto:hotspot-
>> compiler-dev-bounces at openjdk.java.net] On Behalf Of Tom Rodriguez
>> Sent: Tuesday, October 14, 2008 1:04 PM
>> To: hotspot compiler
>> Subject: comiss vs. ucomiss
>> I was just making some comparisons between x86_32.ad and x86_64.ad
>> noticed that x86_32.ad uses comiss/comisd for floating point compares
>> with SSE but x86_64.ad uses ucomiss/ucomisd. The only difference
>> between them is that the ucom versions don't signal for quiet NaNs.
>> We have fp exception suppressed so functionally it doesn't make any
>> difference. It does seem like they should be doing the same thing
>> though. Anyone know why they emit different things?
More information about the hotspot-compiler-dev