Causing jvm crash when entering C2 compiling when using panama memaccess on macos

[Maurizio Cimadamore]

Thanks for the report, this seems related to long loop optimizations. 
CC'ing Roland.

Maurizio

On 18/08/2023 14:43, 刘希晨 wrote:
>
> Hi guys, I came into a jvm crash problem when using native memory 
> access when on my macbook with m1 pro, here is the minimum 
> reproducible codes:
>
> public class CrashExample {
>      private static final VarHandlebyteHandle = MethodHandles.memorySegmentViewVarHandle(ValueLayout.JAVA_BYTE);
>      private static final long len =100;
>      private static final byte CR = (byte)'\r';
>      private static final byte LF = (byte)'\n';
>
>      public static void main(String[] args) {
>          for(int round =0; round <1000; round++) {
>              try(Arena arena = Arena.ofConfined()) {
>                  MemorySegment memorySegment = arena.allocateArray(ValueLayout.JAVA_BYTE,len);
>                  for(int i =0; i <100000; i++) {
>                      int index = i %99;
>                      byteHandle.set(memorySegment, index,CR);
>                      byteHandle.set(memorySegment, index +1,LF);
>                      byte[] bytes =readUntil(memorySegment,CR,LF);
>                  }
>              }
>          }
>      }
>
>      private static byte[]readUntil(MemorySegment segment,byte... separators) {
>          for(long cur =0; cur <= segment.byteSize() - separators.length; cur++) {
>              if(matches(segment, cur, separators)) {
>                  return segment.asSlice(0, cur).toArray(ValueLayout.JAVA_BYTE);
>              }
>          }
>          return null;
>      }
>
>      public static boolean matches(MemorySegment m,long offset,byte[] bytes) {
>          for(int index =0; index < bytes.length; index++) {
>              if ((byte)byteHandle.get(m, offset + index) != bytes[index]) {
>                  return false;
>              }
>          }
>          return true;
>      }
> }
>
> the matches() method try to judge if the memorysegment has the same 
> content as bytes from offset index, the readUntil() method try to 
> traverse the memorysegment until it finds the target separators, 
> return the bytes it has traversed.
>
> in the main() method,  readUntil() method was tested multiple times to 
> reach the C2 compiler, then it will cause the JVM to crash.
>
> Here is what I have find out:
> 1. Calling readUntil() method with a single byte will not cause the 
> JVM to crash ( In our example, CR and LF were used, so there are two 
> bytes)
> 2. On Windows and Linux, it works well, this example will only crash 
> on MacOS, I don't know if x64 or arm CPU makes a difference, since I 
> only have a arm-based macbook
> 3. It will only crash when entering C2 compiler phase, when the 
> methods are not called so much times, they behave normally
> 4. crash seems to always happen in the # V [libjvm.dylib+0x139a2c] 
> AddNode::IdealIL(PhaseGVN*, bool, BasicType)+0x48 or # V 
> [libjvm.dylib+0x958270] SubLNode::Ideal(PhaseGVN*, bool)+0xfc
>
> I have attached the crash report to the mail, I don't know if it's a 
> problem about C2 compiler or Panama memaccess, cheers
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/panama-dev/attachments/20230818/a7f1566b/attachment-0001.htm>