Don't optimize MultiplyNoFlags away

[fiigii]

Contribute to https://github.com/dotnet/coreclr/issues/21899

[fiigii]

@CarolEidt @AndyAyersMS PTAL

The code below was compiled incorrectly that MultiplyNoFlags was optimized away.

        static unsafe ulong test(ulong a, ulong b)
        {
            ulong r;
	    Bmi2.X64.MultiplyNoFlags(a, b, &r);
            return r;
        }

My current solution is to assign the intrinsic call to a local variable. Although it would have a redundant ASG IR for "normal usages" (i.e., res = Bmi2.X64.MultiplyNoFlags(a, b, &r)), the final codegen does not have redundant movs.

[mikedn]

Hmm, this intrinsic should be treated as a memory store. Is it? I guess this is another peculiar thing about it - the other memory stores don't return a value AFAIR.

[fiigii]

Right, the importer usually forces to append the expr into the IR tree if it returns VOID. But this intrinsic has a real return value and memory store semantic simultaneously.

[tannergooding]

Do we know why this isn't this covered by it both being a memory store (https://github.com/dotnet/coreclr/blob/master/src/jit/gentree.cpp#L17920) and the intrinsics having been updated to be special handled in fgComputeLifeLIR and fgCheckRemoveStmt (see https://github.com/dotnet/coreclr/issues/16088 for more details)?

[fiigii]

@tannergooding this happens in the importer and doesn’t reach that point.

[mikedn]

Check how GT_CMPXCHG is handled in the importer, maybe it helps. It has the same characteristics - writes to memory and returns a value.

[CarolEidt]

I believe that it needs to be marked as having a side-effect. In the GT_CMPXCHG case, this is the GTF_ASG flag. The method GenTree::OperRequiresAsgFlag() is where the conditions for this are checked. I suspect that any memory store intrinsic should also be handled there.

[fiigii]

I believe that it needs to be marked as having a side-effect. In the GT_CMPXCHG case, this is the GTF_ASG flag.

Ah, thank you so much, will try.

[fiigii]

Added GTF_GLOB_REF and GT_ASG flags to solve the issue. Checked JitDump that does not have redundant IRs or codegen.

@CarolEidt @AndyAyersMS @mikedn PTAL

[CarolEidt]

I think that you should also change GenTree::OperRequiresAsgFlag() to return true for these intrinsics. @erozenfeld would know better than I, but I believe that, even though these intrinsics are setting it in the constructor, we may need this to also return true for checking of flags.

[erozenfeld]

Yes, GenTree::OperRequiresAsgFlag() should also be changed. It's used both for flags validation and for recomputing the flags.

[fiigii]

@CarolEidt @erozenfeld OperRequiresAsgFlag has already had the code to check these intrinsics.

coreclr/src/jit/gentree.cpp

Lines 5063 to 5081 in 459b58a

	bool GenTree::OperRequiresAsgFlag()
	{
	if (OperIs(GT_ASG) || OperIs(GT_XADD, GT_XCHG, GT_LOCKADD, GT_CMPXCHG, GT_MEMORYBARRIER))
	{
	return true;
	}
	#ifdef FEATURE_HW_INTRINSICS
	if (gtOper == GT_HWIntrinsic)
	{
	GenTreeHWIntrinsic* hwIntrinsicNode = this->AsHWIntrinsic();
	if (hwIntrinsicNode->OperIsMemoryStore())
	{
	// A MemoryStore operation is an assignment
	return true;
	}
	}
	#endif // FEATURE_HW_INTRINSICS
	return false;
	}

[CarolEidt]

OperRequiresAsgFlag has already had the code to check these intrinsics.

Wow, right there as I was looking at it, and I missed it! Thanks.

[CarolEidt]

test Ubuntu arm Cross Checked Innerloop Build and Test

[CarolEidt]

LGTM - thanks!