RFR: 8309894: compiler/vectorapi/VectorLogicalOpIdentityTest.java fails on SVE system with UseSVE=0

Emanuel Peter epeter at openjdk.org
Wed Jun 21 08:58:02 UTC 2023


On Mon, 19 Jun 2023 01:49:57 GMT, Xiaohong Gong <xgong at openjdk.org> wrote:

> This test fails with several IR check failures when run on ARM SVE systems with vm option `-XX:UseSVE=0`. Here is one of the failed IR rule:
> 
> 
> @IR(counts = {IRNode.AND_V, "1"}, applyIfCPUFeatureOr = {"sve", "true", "avx512", "true"})
> public static void testAndMaskSameValue1()
> 
> The specified IR in the test depends on the platform's predicate feature. Hence the IR check can be applied only on ARM SVE or X86 AVX512 systems. But with `-XX:UseSVE=0` on ARM SVE machines, JVM will disable SVE feature for compiler. But the CPU feature is not changed. To guarantee the IR rule is run with SVE as expected, it has to add another condition like `applyIf = {"UseSVE", ">0"}`. Consider `UseSVE` is an ARM specific option, it can be used only on ARM CPUs. So the right IR rules should be:
> 
> 
> @IR(counts = {IRNode.AND_V, "1"}, applyIfCPUFeature = {"sve", "true"}, applyIf = {"UseSVE", "> 0"})
> @IR(counts = {IRNode.AND_V, "1"}, applyIfCPUFeature = {"avx512", "true"})
> public static void testAndMaskSameValue1()
> 
> 
> This patch also changes the check order of conditions for a IR rule. It's better to check `applyIfCPUFeature` before `applyIf`, in case the vm option is invalid on running hardware, which makes test throw exception and abort.
> 
> Verified on X86 systems with `UseAVX=1/2/3` by removing the test from ProblemList.txt, and SVE systems with `UseSVE=0/1`.

@XiaohongGong Thanks for looking into this. But it seems to me this is not the same approach as we are taking with x86 SSE and AVX, where the `UseAVX` and `UseSSE` flags affect both the VM features and also the `applyIfCPUFeature` from the IR framework. We check in all sorts of places for `sve` feature, so why do you now only change it for this particular test?

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

PR Comment: https://git.openjdk.org/jdk/pull/14533#issuecomment-1600454917


More information about the hotspot-compiler-dev mailing list