x86 Intrinsics for fma in Math Library

Andrew Haley aph at redhat.com
Fri Jul 22 09:08:05 UTC 2016

On 21/07/16 18:18, Vladimir Kozlov wrote:

> Good suggestion. Need a test to confirm "correct" results matching
> result of assembler instruction.

I've been looking at JDK tests for the correct fma rounding.  I've run
these to check that AArch64 fma works correctly.  It turns out that if
you run tests with inlining enabled, C2 collapses most of the code to
nothing because of constant propagation.  If you run the tests with
inlining enabled, you test the arithmetic at runtime.  I've also added
a test which makes sure that rounding is only done once on the
low-order bits.

This would form the basis for a HotSpot test, I think.


More information about the hotspot-compiler-dev mailing list