Skip to content

Lint callish MachOps in Cmm lint, and add type annotations to AtomicFetch_cmm.cmm

sheaf requested to merge sheaf/ghc:llvm-atomicfetch into master

There were missing type annotations for literals in AtomicFetch_cmm.cmm, which lead the LLVM backend to confusion. It assumed the literals were 64 bits wide, which made us emit invalid LLVM, for example:

error: atomicrmw value and pointer type do not match
  %ln1R = atomicrmw add i16*  %ln1Q, i64  %ln1P seq_cst

This MR fixes the test, but also adds a new check to Cmm lint to ensure we catch such invalid Cmm at the source.

Edited by sheaf

Merge request reports