RFR: JDK-8325037: x86: enable and fix hotspot/jtreg/compiler/vectorization/TestRoundVectFloat.java

[Jatin Bhateja]

On Wed, 31 Jan 2024 12:54:18 GMT, Tobias Hartmann <thartmann at openjdk.org> wrote:

>> test/hotspot/jtreg/compiler/vectorization/TestRoundVectFloat.java line 52:
>> 
>>> 50: 
>>> 51:   @Test
>>> 52:   @IR(applyIf = {"UseAVX", " > 1"}, counts = {IRNode.ROUND_VF , " > 0 "})
>> 
>> Isn't that node only available if `UseAVX >= 2`?
>> https://github.com/openjdk/jdk/blob/f0bae7939a61a79f3e07de97451c433e91742069/src/hotspot/cpu/x86/x86.ad#L1501-L1504
>
> Sorry, misread the code. Looks good.

FTR, backend implementation of rounding first adds a broadcasted 0.5 value to incoming floating point vector lanes, register only flavor of broadcast operation is only supported by AVX2 targets. We can broadcast 0.5 into constant table emitted prior to method body to lift this limitation. Another concern is around different max vector size for floating point and integral vectors on AVX1 targets, 256 bit float vector vs 128 bit integral vector. Thus, a sharper check in match_rule_supported_vector (size_in_bits  > 128 && UseAVX < 2) should be added to support this for AVX1 targets as we are emitting an integral compare in special case handling for NaN and -/+0.0 / Inf.  

But, this can be taken up separately, if need be, current fix for PR looks ok to me.

-------------

PR Review Comment: https://git.openjdk.org/jdk/pull/17649#discussion_r1474216444