Merge pull request #2411 from IBMJimmyk/SIMDStringHashCode

Fix recognition of hashCode function on Power
eclipse-openj9 · Jul 27, 2018 · 56684b8 · 56684b8
2 parents d43cf1e + a5a0930
commit 56684b8
Show file tree

Hide file tree

Showing 4 changed files with 27 additions and 1 deletion.
diff --git a/runtime/compiler/env/ProcessorDetection.cpp b/runtime/compiler/env/ProcessorDetection.cpp
@@ -562,6 +562,16 @@ TR_J9VMBase::getPPCProcessorType()
    return portLibCall_getProcessorType();
    }
 
+bool
+TR_J9VMBase::getPPCSupportsVSXRegisters()
+   {
+#if defined(TR_TARGET_POWER)
+   return TR::Compiler->target.cpu.getPPCSupportsVSX();
+#else
+   return false;
+#endif // TR_TARGET_POWER
+   }
+
 // -----------------------------------------------------------------------------
 
 extern TR_X86CPUIDBuffer *queryX86TargetCPUID(void * javaVM);

diff --git a/runtime/compiler/env/VMJ9.cpp b/runtime/compiler/env/VMJ9.cpp
@@ -3109,6 +3109,21 @@ bool TR_J9VMBase::supressInliningRecognizedInitialCallee(TR_CallSite* callsite,
                }
             break;
          case TR::java_lang_String_hashCodeImplDecompressed:
+            /*
+             * X86 and z want to avoid inlining both java_lang_String_hashCodeImplDecompressed and java_lang_String_hashCodeImplCompressed
+             * so they can be recognized and replaced with a custom fast implementation.
+             * Power currently only has the custom fast implementation for java_lang_String_hashCodeImplDecompressed.
+             * As a result, Power only wants to prevent inlining of java_lang_String_hashCodeImplDecompressed.
+             * When Power gets a fast implementation of TR::java_lang_String_hashCodeImplCompressed, this case can be merged into the case
+             * for java_lang_String_hashCodeImplCompressed instead of using a fallthrough.
+             */
+            if (!TR::Compiler->om.canGenerateArraylets() &&
+                TR::Compiler->target.cpu.isPower() && TR::Compiler->target.cpu.id() >= TR_PPCp8 && getPPCSupportsVSXRegisters() && !comp->compileRelocatableCode())
+                  {
+                  dontInlineRecognizedMethod = true;
+                  break;
+                  }
+            // Intentional fallthrough here.
          case TR::java_lang_String_hashCodeImplCompressed:
             if (!TR::Compiler->om.canGenerateArraylets()){
                if ((TR::Compiler->target.cpu.isX86() && getX86SupportsSSE4_1()) ||

diff --git a/runtime/compiler/env/VMJ9.h b/runtime/compiler/env/VMJ9.h
@@ -273,6 +273,7 @@ class TR_J9VMBase : public TR_FrontEnd
    static TR_J9VMBase * get(J9JITConfig *, J9VMThread *, VM_TYPE vmType=DEFAULT_VM);
    static char *getJ9FormattedName(J9JITConfig *, J9PortLibrary *, char *, int32_t, char *, char *, bool suffix=false);
    static TR_Processor getPPCProcessorType();
+   virtual bool getPPCSupportsVSXRegisters();
    static void initializeX86ProcessorInfo(J9JITConfig *);
 
    static bool isBigDecimalClass(J9UTF8 * className);

diff --git a/runtime/compiler/p/codegen/J9TreeEvaluator.cpp b/runtime/compiler/p/codegen/J9TreeEvaluator.cpp
@@ -12112,7 +12112,7 @@ J9::Power::CodeGenerator::inlineDirectCall(TR::Node *node, TR::Register *&result
          return true;
 
       case TR::java_lang_String_hashCodeImplDecompressed:
-         if (TR::Compiler->target.cpu.id() >= TR_PPCp8 && TR::Compiler->target.cpu.getPPCSupportsVSX() && !cg->comp()->compileRelocatableCode())
+         if (!TR::Compiler->om.canGenerateArraylets() && TR::Compiler->target.cpu.id() >= TR_PPCp8 && TR::Compiler->target.cpu.getPPCSupportsVSX() && !cg->comp()->compileRelocatableCode())
             {
             resultReg = inlineStringHashcode(node, cg);
             return true;