Recognize unsafe unaligned getters/setters

[nbhuiyan]

This PR re-introduces changes in #19420 and addresses the review comments received there.

Contrary to the discussion in #19420, I have not found any evidence that unaligned access is not permitted in Z only. Existing CG flag SupportsAlignedAccessOnly has been used for some versions of Power as well as aarch64 and arm32. Supporting unaligned access is something that most platforms are capable of, and therefore the default behaviour should be to transform the unaligned access operation methods, unless a back-end sets the SupportsAlignedAccessOnly flag.

[nbhuiyan]

@jdmpapin Requesting review.

[nbhuiyan]

@r30shah I'd appreciate your input on whether unaligned access is permitted on Z or not. If it is not permitted, was there a reason SupportsAlignedAccessOnly is not set during Z codegen initialization?

[r30shah]

Memory references on Z are required to be aligned to operand size to ensure that all bytes within the operands are read as block concurrent as observed by other CPUs and channel programs. If the atomicity is not required - operand access can be done with unaligned address. So That is why I think code-gen on Z does not set SupportsAlignedAccessOnly.

[jdmpapin]

I just remembered some feedback I had originally sent directly to Jack. I had applied the changes to my working copy, and I was unable to see the transformation kick in. Here are a few pointers from when I was poking around looking at it

inlineUnsafeCall() is called only when the method satisfies isInlineableJNI()

openj9/runtime/compiler/optimizer/InlinerTempForJ9.cpp

Lines 2337 to 2350 in 6e1e36f

	if (isInlineableJNI(calleeSymbol->getResolvedMethod(),callNode))
	{
	if (performTransformation(comp(), "%sInlining jni %s into %s\n", OPT_DETAILS, calleeSymbol->signature(comp()->trMemory()), callerSymbol->signature(comp()->trMemory())))
	{
	if (calltarget->_myCallSite->isIndirectCall())
	return true;
	if (inlineGetClassAccessFlags(calleeSymbol, callerSymbol, callNodeTreeTop, callNode))
	guard->_kind = TR_NoGuard;
	else if (inlineUnsafeCall(calleeSymbol, callerSymbol, callNodeTreeTop, callNode))
	guard->_kind = TR_NoGuard;
	}
	return true;
	}

In isInlineableJNI(), here's where getInt(), etc., would be accepted:

openj9/runtime/compiler/optimizer/InlinerTempForJ9.cpp

Lines 2109 to 2118 in 6e1e36f

	if (method->convertToMethod()->isUnsafeWithObjectArg(comp) || method->convertToMethod()->isUnsafeCAS(comp))
	{
	// In Java9 sun/misc/Unsafe methods are simple Java wrappers to JNI
	// methods in jdk.internal, and the enum values above match both. Only
	// return true for the methods that are native.
	if (!TR::Compiler->om.canGenerateArraylets() || (callNode && callNode->isUnsafeGetPutCASCallOnNonArray()))
	return method->isNative();
	else
	return false;
	}

Two problems:

• isUnsafeWithObjectArg() doesn't know about the ...Unaligned() methods, so it won't return true
• the method isn't native

So I think that's what's involved, but in any case please confirm that you can see the transformation happening before we move on

[nbhuiyan]

@jdmpapin I have addressed your comments and have verified that the transformations are taking place with the current implementation.

[jdmpapin]

Jenkins test sanity all jdk11,jdk21

[jdmpapin]

The AArch64 Linux JDK11 sanity.openjdk failure is #13756

The Windows JDK21 sanity.openjdk failure is #18463

[jdmpapin]

Jenkins compile xlinux jdk8