fxGrowChunks making very small allocations

[warner]

I've been investigating performance problems in our xsnap usage of XS, where we're seeing GC happen more frequently than expected, and the allocations being made to fxAllocateChunks are surprisingly small (32-4000 bytes at a time), even though our xsCreation settings do incrementalHeapCount = 256 * 1024.

I'm suspicious of the following code in fxGrowChunks:

moddable/xs/sources/xsMemory.c

Lines 423 to 434 in a2a38f4

	void fxGrowChunks(txMachine* the, txSize theSize)
	{
	txByte* aData;
	txBlock* aBlock;
	if (!(the->collectFlag & XS_SKIPPED_COLLECT_FLAG)) {
	txSize modulo = theSize % the->minimumChunksSize;
	if (modulo)
	fxAddChunkSizes(the, theSize, the->minimumChunksSize - modulo);
	}
	theSize = fxAddChunkSizes(the, theSize, sizeof(txBlock));
	aData = fxAllocateChunks(the, theSize);

It looks like the return value from fxAddChunkSizes() is being ignored, leaving theSize at its original value (e.g. 28 when allocating a new Function). On 17-Mar-2021, in 3070f24, this was changed:

3070f24#diff-5458ac58321f0e9dcd9ea60d1b7e2b87ea74e00cb70c5da666d1eed076c14258R380

I think the original intention was to increase theSize to the->minimumChunksSize, give or take some rounding and headers. In the new version, I think that increase got lost, resulting in small allocations, and therefore more frequent GC calls.

The fix is probably just:

	theSize = fxAddChunkSizes(the, theSize, the->minimumChunksSize - modulo);

We're building against a derivative of commit 0b451c9 (aka tag OS210406), on Linux.

thanks!
-Brian

[phoddie]

Good find. That's our bug. We'll apply that to our next SDK update. Thank you for the help. Apologies for the trouble.

[phoddie]

Closing this issue, as I believe it has been resolved.