(original) (raw)

Is -0.0{f,d} not replaced with an immediate because the immediate then increases instruction size? Just curious ...

Thanks

On Thu, Mar 12, 2015 at 8:43 AM, Zoltán Majó <zoltan.majo@oracle.com> wrote:

Hi,


please review the following small patch.

Bug: https://bugs.openjdk.java.net/browse/JDK-8074869


Problem: On Linux, the C2 code generator can replace the value -0.0f with +0.0f (and also the value -0.0d with +0.0d). The reason is that in some \*.ad files both the value -0.0f and +0.0f is treated as being +0.0f and can therefore be replaced with an immediate +0.0f embedded into an instruction.

For example, in the sparc.ad file, the 'fpclass' function is used to decide if a float node's content is +0.0:

predicate((n->getf() == 0) && (fpclass(n->getf()) == FP\_PZERO));

On Solaris, 'fpclass' returns FP\_PZERO if the parameter is +0.0f and FP\_NZERO if the parameter is -0.0f. As a result, +0.0f and -0.0f are distinguished by the compiler.

On Linux, however, 'fpclass' is not available and therefore 'fpclassify' is used. 'fpclassify' does not distinguish between ±0.0f, it returns FP\_ZERO for both +0.0f and -0.0f.


Solution: Instead of 'fpclass', use cast float->int and double->long to check if value is +0.0f and +0.0d, respectively. This logic is already use on some architectures, for example on x86\_32 and on x86\_64.

As 'fpclass' is not used any more, remove its declarations from globalDefintions\_\*. The include of ieeefp.h must be kept as we rely on some other functionality from this header on solaris.


Webrev: http://cr.openjdk.java.net/\~zmajo/8074869/webrev.00/


Testing:
\- JPRT testing on all supported platforms (does \*not\* include aarch64 and ppc64)

\- manual testing on aarch64:

All DaCapo benchmarks with the small input size. I used the default JVM flags and tested the VM w/ and w/o the fix. All benchmarks pass except eclipse. For eclipse, the same Java-level failure appears both w/ and w/o the fix, so I think the problem with eclipse is not due to these changes.

I also tested with the "-Xcomp -XX:-TieredCompilation -server" flags. Eclipse fails in this case as well. Additionally, tradebeans and tradesoap fail with a Java-level failure. As the failure happens also with both builds (w/ and w/o the fix), I don't think the problem is caused by these changes either.

\- no testing on ppc64: I don't have access to a ppc64 machine. Could somebody with access to a ppc64 machine please build and test the VM with this patch (and then maybe confirm that it works)?


Thank you and best regards,


Zoltan