Draft: use slots smaller than word as tag for smaller unboxed sums

Saves some space, especially when storing a bunch of unpacked enums.

To clarify: This is beneficial for unboxed sums being used to {-# UNPACK #-} things into heap objects. This MR doesn't change the calling convention when passing/returning unboxed tuples/sums with smaller-than-word fields.

marked this merge request as draft

added 1 commit

• 91314118 - add some missing cases

Compare with previous version

added 1 commit

• 894c765f - remove some unused things

Compare with previous version

probably needs some tests:

• to check that 8x {-# UNPACK #-} !Bool is stored in one 64 bit word
• around the limit of 255 constructors
• alignment of bigger slots following smaller (sub-word) slots

Multiple unpacked bools are packed correctly with the current version.

added full-ci label

I suspect 32 bit platforms need slightly different test outputs, hence full-ci label.

added 1 commit

• 3a1a1c90 - adjust T25166 and fixes for 32 bit

Compare with previous version

See #25167

The new slot constructors also allow packing of more things than tags, so perhaps this also tackles #25168

currently there is no reordering of fields being done to optimize packing. perhaps we should change that.

If we're dealing with a single sum we can reorder fields by hand for optimal packing in the source code.

But I'd like the compiler to optimize data layout with multiple or nested {-# UNPACK #-}, where we don't have (or want) as much control over ordering.

Perhaps if we start to do packing we may want to distinguish between unboxed sums that are passed around directly in registers and unboxed sums that are stored inside heap object ({-# UNPACK #-} or otherwise)

The current implementation cannot overlap two Word8Slot with a Word16Slot, so I think a bit more thought is needed.

I think it's better to have some separate layout for packing into heap objects vs calling convention. Maybe keep the SlotTy things the same and return a size plus a bunch of byte offsets for the heap layout

changed the description