RFR (XXS) 8234610: MaxVectorSize set wrongly when UseAVX=3 is specified after JDK-8221092

[Viswanathan, Sandhya]

Hi Vladimir,

The bug happens like below:

* User specifies -XX:UseAVX=3 on command line.
* On Skylake platform due to the following lines:
    __ lea(rsi, Address(rbp, in_bytes(VM_Version::std_cpuid1_offset())));
    __ movl(rax, Address(rsi, 0));
    __ cmpl(rax, 0x50654);              // If it is Skylake
    __ jcc(Assembler::equal, legacy_setup);
   The zmm registers are not saved/restored.
* This results in os_supports_avx_vectors() returning false when called from VM_Version::get_processor_features():
  int max_vector_size = 0;
  if (UseSSE < 2) {
    // Vectors (in XMM) are only supported with SSE2+
    // SSE is always 2 on x64.
    max_vector_size = 0;
  } else if (UseAVX == 0 || !os_supports_avx_vectors()) {
    // 16 byte vectors (in XMM) are supported with SSE2+
    max_vector_size = 16;                              ====> This is the point where max_vector_size is set to 16
  } else else if (UseAVX == 1 || UseAVX == 2) {   
     ...
  }
* And so we get UseAVX=3 and max_vector_size = 16.

So the fix is to save/restore zmm registers when the flag is not default and user specifies UseAVX > 2.

On your question regarding why 8221092 needs the code you conditionally exclude:
This was introduced so as not to do any AVX512 execution if not required.  ZMM register save/restore uses AVX512 instruction. 

Best Regards,
Sandhya


-----Original Message-----
From: Vladimir Ivanov <vladimir.x.ivanov at oracle.com> 
Sent: Friday, November 22, 2019 4:39 AM
To: Viswanathan, Sandhya <sandhya.viswanathan at intel.com>; hotspot compiler <hotspot-compiler-dev at openjdk.java.net>
Subject: Re: RFR (XXS) 8234610: MaxVectorSize set wrongly when UseAVX=3 is specified after JDK-8221092

Hi Sandhya,

> On Skylake platform the JVM by default sets the UseAVX level to 2 and accordingly sets MaxVectorSize to 32 bytes as per JDK-8221092<https://bugs.openjdk.java.net/browse/JDK-8221092>.
> 
> When the user explicitly wants to use AVX3 on Skylake platform, they need to invoke the JVM with -XX:UseAVX=3 command line argument.
> This should automatically result in MaxVectorSize being set to 64 bytes.
> 
> However post JDK-8221092<https://bugs.openjdk.java.net/browse/JDK-8221092>, when -XX:UseAVX=3 is given on command line, the MaxVectorSize is being wrongly set to 16 bytes.

Please, elaborate how it happens and how legacy_setup affects it?

Why 8221092 needs the code you conditionally exclude.
Why the following isn't enough?

    if (FLAG_IS_DEFAULT(UseAVX)) {
      FLAG_SET_DEFAULT(UseAVX, use_avx_limit);
+    if (is_intel_family_core() && _model == CPU_MODEL_SKYLAKE && 
_stepping < 5) {
+      FLAG_SET_DEFAULT(UseAVX, 2);  //Set UseAVX=2 for Skylake
+    }
    } else if (UseAVX > use_avx_limit) {

Best regards,
Vladimir Ivanov