8230015: [instruction selector] generic vector operands support.

[Vladimir Ivanov]

Hi Jatin,

FTR I'm looking at:
   http://cr.openjdk.java.net/~jbhateja/genericVectorOperands/webrev.03/

Good work, Jatin. The reduction in number of instructions looks 
impressive. Also, it's nice to see excessive vec? <-> legVec? moves 
being eliminated.

High-level comments on the patch:

   - Please, separate the patch in 2 changes: generic vector type 
support and AD instruction unification. It would significantly simplify 
review.

   - Try to avoid x86-specific changes (#ifdefs) in shared files.


I'm still looking through the implementation, but here are the bugs in 
AD unification spotted so far:

====================================================================
src/hotspot/cpu/x86/x86_64.ad:

+operand vecG() %{
+  constraint(ALLOC_IN_RC(vectorg_reg));

+operand legVecG() %{
+  constraint(ALLOC_IN_RC(vectorg_reg));

legVecG definition shouldn't be equivalent to vecG: on AVX512-capable 
host legVecG should never include zmm16-zmm31 while vecG can.


====================================================================
src/hotspot/cpu/x86/x86.ad:

+instruct ReplF_zero_avx(vecG dst, immF0 zero) %{
+  predicate(UseAVX < 3 &&

UseAVX > 0 check is missing (same with ReplD_zero_avx).


====================================================================
+instruct vaddB_mem(vecG dst, vecG src, memory mem) %{
+  predicate(UseAVX && (UseAVX <= 2 || VM_Version::supports_avx512bw()) &&
+            n->as_Vector()->length() >= 4 && n->as_Vector()->length() 
<= 64);

It doesn't match the following case anymore:
   UseAVX == 3 &&
   VM_Version::supports_avx512bw() == false &&
   n->as_Vector()->length() < 64 // smaller than 512bit

There are other AD instructions  (with VMVersion::supports_avx512bw() 
check) which are affected the same way.

Best regards,
Vladimir Ivanov

On 22/08/2019 09:49, Bhateja, Jatin wrote:
> Hi All,
> 
> Please find below a patch for generic vector operands[1] support during 
> instruction selection.
> 
> Motivation behind the patch is to reduce the number of vector selection 
> patterns whose operands meagerly differ in vector lengths.
> 
> This will not only result in lesser code being generated by ADLC which 
> effectively translates to size reduction in libjvm.so but also
> 
> help in better maintenance of AD files.
> 
> Using generic operands we were able to collapse multiple vector patterns 
> over mainline
> 
>               Initial number of vector instruction patterns (vec[XYZSD] 
> + legVec[ZXYSD]   : *510*
> 
>               Reduced vector instruction patterns  (vecG + 
> legVecG)                                  : *222*
> 
> **
> 
> With this we could see around 1MB size reduction in libjvm.so.
> 
> In order to have minimal impact over downstream compiler passes, a 
> post-selection pass has been introduced (currently enabled only for X86 
> target)
> 
> which replaces these generic operands with their corresponding concreter 
> vector length variants.
> 
> JBS      : https://bugs.openjdk.java.net/browse/JDK-8230015
> 
> Patch  : 
> http://cr.openjdk.java.net/~jbhateja/genericVectorOperands/webrev.00/
> 
> Kindly review and share your feedback.
> 
> Best Regards,
> 
> Jatin Bhateja
> 
> [1] 
> http://cr.openjdk.java.net/~jbhateja/genericVectorOperands/generic_operands_support_v1.0.pdf
>