RTC Thumb2 JIT enhancements

[Andrew Dinn]

On 17/05/12 18:53, Andrew Haley wrote:
> This is the big patch from Mar 2010:
> http://mail.openjdk.java.net/pipermail/distro-pkg-dev/2010-March/008662.html
> merged into the HS22 codebase.  I've done basic testing on it, and this
> includes a bootstrap build.  I'm sorry it's so big, but it's not really
> possible to separate things.

Ok, so this has been reviewed by eyeballing the patch and by building it
and running SpecJVM.

>From inspection the code seems to be ok (in so far as one can assess
such a thing by eye) modulo one tiny error in thumb2.cpp at line 6141
(that's post patch)

@@ -6558,7 +6137,7 @@
 	    r = JSTACK_REG(jstack);
 	    PUSH(jstack, r);
 	    if (tos_type == btos)
-	      ldrsb_imm(jinfo->codebuf, r, r_obj, field_offset, 1, 0);
+	      ldrb_imm(jinfo->codebuf, r, r_obj, field_offset, 1, 0);
 	    else if (tos_type == ctos)
 	      ldrh_imm(jinfo->codebuf, r, r_obj, field_offset, 1, 0);
 	    else if (tos_type == stos)

For some reason the byte load has been changed to unsigned but it really
ought to be signed.

Running SpecJVM unearthed an error which appears to have been there all
along and so is only being made manifest by the patch, viz that stacked
return values are not copied through from old to new (nursery) space if
we enter a garbage collect at a return point in either compiled /or/
interpreted code. They need to be written back to the frame before a
safepoint is entered. Other than this the tests manifested no errors.

The patch improve some SpecJVM scores buut not to any great degree.
Whether this indicates something about the weakness of benchmarks or the
lack of benefit provided by the patch is an open question.

Given the above results (and modulo a fix to the code indicated above) I
suggest pushing this patch into 7.2.1

regards,


Andrew Dinn
-----------