[vectorIntrinsics] RFR: 8261394: [vector] Crash with "assert(Matcher::vector_size_supported(elem_bt, length)) failed: length in range"

[Jie Fu]

On Thu, 18 Feb 2021 03:00:00 GMT, Ningsheng Jian <njian at openjdk.org> wrote:

>> The crash is introduced by [1] and happens when the compiler is making an unspported vector type:
>>   assert(Matcher::vector_size_supported(elem_bt, length)) failed: length in range
>> Before vectorization for each vector intrinsic, the hotspot will check whether current arch supports the vector operation from several aspects like the backend implementation, the vector type and length. Once one of them is not matched, the compiler will stop the vectorization and go back to the default java implementation.
>> 
>> The changes in [1] missed the check for "Op_CallLeafVector". I think the double 64-bits vector is not supported to be vectorized. However, due to the missing check, the compiler continues the progress and then the crash happens.
>> 
>> This patch fixes it by making sure the check contains all opcodes.
>> 
>> [1] https://bugs.openjdk.java.net/browse/JDK-8261267
>
> Looks good to me, but perhaps @sviswa7 may take a look as well?

For ACOSDouble64VectorTests, I think @sviswa7 would like to generate code like this:

098     B6: #   out( B14 B7 ) <- in( B5 )  Freq: 0.999977
098     # checkcastPP of RBX
098     movl    R11, [RBX + #12 (8-bit)]        # compressed ptr ! Field: jdk/internal/vm/vector/VectorSupport$VectorPayload.payload (constant)
09c     load_vector XMM0,[R12 + R11 << 3 + #16] (compressed oop addressing)
0a3     call_leaf,vector  vector_acos_double64
        No JVM State Info
        #

However, the fix will generate code like this.

095     B6: #   out( B18 B7 ) <- in( B5 )  Freq: 0.999977
095     movq    R9, RAX # spill
098     # checkcastPP of R9
098     movq    RDI, jdk/incubator/vector/DoubleVector$$Lambda$27+0x0000000801052e28:exact *    # ptr
0a2     movl    RSI, #108       # int
0a7     movq    RDX, java/lang/Class:exact *    # ptr
0b1     movq    RCX, java/lang/Class:exact *    # ptr
0bb     movl    R8, #1  # int
        nop     # 3 bytes pad for loops and calls
0c4     call,static  jdk.internal.vm.vector.VectorSupport::unaryOp
        # jdk.incubator.vector.DoubleVector::lanewiseTemplate @ bci:71 (line 554) L[0]=_ L[1]=_ L[2]=_
        # jdk.incubator.vector.Double64Vector::lanewise @ bci:2 (line 273) L[0]=_ L[1]=_
        # jdk.incubator.vector.Double64Vector::lanewise @ bci:2 (line 41) L[0]=_ L[1]=_
        # Test::ACOSDouble64VectorTests @ bci:15 (line 10) L[0]=_
        # OopMap {rbp=Oop off=204/0xcc}

Maybe, it's OK to call vector_acos_double64 for Double64Vector on x86.
Thanks.

-------------

PR: https://git.openjdk.java.net/panama-vector/pull/38