PUBLIC COMMENTS

Integrated fix into Tiger Update 1 (1.5.0_01) Build 03
                                     

CONVERTED DATA

BugTraq+ Release Management Values

COMMIT TO FIX:
1.5.0_01
mustang

FIXED IN:
mustang

INTEGRATED IN:
1.5.0_01
mustang


                                     

SUGGESTED FIX

The simplest fix will pin any CheckCastPP node with a rawptr input in the same block as its input.

###@###.### 2004-07-22

See this PRT webrev for the fix details:

file:/net/prt-archiver.sfbay/data/archived_workspaces/main/c2_baseline/2004/20040818125259.rasbold.c2_baseline/workspace/webrevs/webrev-2004.08.18/index.html

###@###.### 2004-08-19
                                     

EVALUATION

I don't have a Japanese version of XP, but my PC running Windows XP SP1 did not crash running the test (with JDK1.5-b57).  Are you running XP SP2 by chance?  
###@###.### 2004-07-12

I got it to reproduce on a Solaris x86 machine (foundation). 
###@###.### 2004-07-12

The C2 compiled oop maps for method50 are incomplete; they are missing
an entry for an oop spilled to the stack.

This is a result of C2 inappropriately moving a CheckCastPP node that
converts from a rawptr to an oopptr into a loop.  That code motion makes the
rawptr, for which C2 does not create oop map entries, dynamically
span a safepoint. When the register allocator gives the rawptr and the oopptr two distinct locations, the oop maps are incorrect.

The problem is not restricted to x86, nor is the problem fixed in
1.5.0.  A remark in the comments section suspected a fix in b42. This
is not the case; the change in b42 simply makes the bug latent.

###@###.### 2004-07-22