kmeakin (Jul 21 2023 at 21:27): kmeakin (Jul 21 2023 at 21:27):I want to extend the cprop.isle rules to cover i128. But iconst.i128 isn't supported. Is it not possible to have i128 constants in CLIF?
fitzgen (he/him) (Jul 21 2023 at 21:40):iconst.i128 indeed doesn't exist right now. the way to create constants is to do two iconst.i64s and then iconcat, or if the high bits are just sign/zero bits, then iconst.i64 followed by an extend
fitzgen (he/him) (Jul 21 2023 at 21:41):there was talk in some issue, I forget which and don't have it on hand at the moment, about creating a const ISLE helper to automatically do this if necessary, or else just use an iconst when possible
Afonso Bordado (Jul 21 2023 at 21:41):We had an issue a while back about emitting the various constants from a single helper which would be really usefull to have, and would probably allow us to enable most of the integer optimizations again
fitzgen (he/him) (Jul 21 2023 at 21:41):yes, that one, thanks @Afonso Bordado
Chris Fallin (Jul 21 2023 at 21:44):To add the reasoning here for completeness: we don't have iconst.i128 because InstructionData is currently 16 bytes (128 bits) total, and that has to include the enum discriminant and opcode as well as the payload. We could just create a new enum variant with an i128 but the increased size would penalize compile time and memory usage by a measurable amount
Chris Fallin (Jul 21 2023 at 21:44):There was talk last time this came up about creating a side-table; that's the route I'd suggest if anyone tackles this
Chris Fallin (Jul 21 2023 at 21:44):or using the existing constant machinery for vconst
kmeakin (Jul 22 2023 at 21:22):Hmm, I could try adding a side-table for i128 constants, but I'm not sure where to start
Last updated: Nov 22 2024 at 16:03 UTC