GitLab

Marge Bot pushed to branch master at Glasgow Haskell Compiler / GHC

Commits:

eee8ec5b

by Cheng Shao at 2026-07-14T18:00:20-04:00

compiler: fix miscompiled %load_relaxed, add missing %store_relaxed

This patch fixes the %load_relaxed cmm primop compilation logic to
correctly use relaxed memory ordering, and adds the missing
%store_relaxed primop. Parsing logic of %load/%store with explicit
ordering is covered in the AtomicFetch test case. Fixes #27483.

Co-authored-by: Codex <codex@openai.com>

4 changed files:

+ changelog.d/fix-cmm-atomic-load-store
compiler/GHC/Cmm/Parser.y
testsuite/tests/cmm/should_run/AtomicFetch.hs
testsuite/tests/cmm/should_run/AtomicFetch_cmm.cmm

Changes:

changelog.d/fix-cmm-atomic-load-store

 +section: cmm
 +synopsis: Fix miscompiled %load_relaxed primop, add missing %store_relaxed
 +issues: #27483
 +mrs: !16320

compiler/GHC/Cmm/Parser.y

@@ -1210,9 +1210,10 @@ callishMachOps platform = listToUFM $
      , allWidths "pext" MO_Pext
      , allWidths "cmpxchg" MO_Cmpxchg
      , allWidths "xchg" MO_Xchg
 -    , allWidths "load_relaxed" (\w -> MO_AtomicRead w MemOrderAcquire)
 +    , allWidths "load_relaxed" (\w -> MO_AtomicRead w MemOrderRelaxed)
      , allWidths "load_acquire" (\w -> MO_AtomicRead w MemOrderAcquire)
      , allWidths "load_seqcst" (\w -> MO_AtomicRead w MemOrderSeqCst)
 +    , allWidths "store_relaxed" (\w -> MO_AtomicWrite w MemOrderRelaxed)
      , allWidths "store_release" (\w -> MO_AtomicWrite w MemOrderRelease)
      , allWidths "store_seqcst" (\w -> MO_AtomicWrite w MemOrderSeqCst)
      , allWidths "fetch_add" (\w -> MO_AtomicRMW w AMO_Add)

testsuite/tests/cmm/should_run/AtomicFetch.hs

@@ -6,6 +6,7 @@
  -- This is not a test of atomic semantics,
  -- just checking that GHC can parse %fetch_fooXX
 +-- and %load/%store with explicit ordering
  import GHC.Exts
  import GHC.Int

testsuite/tests/cmm/should_run/AtomicFetch_cmm.cmm

@@ -2,6 +2,7 @@
  // This is not a test of atomic semantics,
  // just checking that GHC can parse %fetch_fooXX
 +// and %load/%store with explicit ordering
  cmm_foo64 (P_ p)
+ {
@@ -19,6 +20,10 @@ cmm_foo64 (P_ p)
    (x) = prim %fetch_xor64(q, 33 :: I64);
    (x) = prim %fetch_nand64(q, 127 :: I64);
    (x) = prim %load_seqcst64(q);
 +  prim %store_relaxed64(q, x);
 +  (x) = prim %load_relaxed64(q);
 +  prim %store_release64(q, x);
 +  (x) = prim %load_acquire64(q);
    return (x);
+ }
@@ -38,6 +43,10 @@ cmm_foo32 (P_ p)
    (x) = prim %fetch_xor32(q, 33 :: I32);
    (x) = prim %fetch_nand32(q, 127 :: I32);
    (x) = prim %load_seqcst32(q);
 +  prim %store_relaxed32(q, x);
 +  (x) = prim %load_relaxed32(q);
 +  prim %store_release32(q, x);
 +  (x) = prim %load_acquire32(q);
    return (x);
+ }
@@ -57,6 +66,10 @@ cmm_foo16 (P_ p)
    (x) = prim %fetch_xor16(q, 33 :: I16);
    (x) = prim %fetch_nand16(q, 127 :: I16);
    (x) = prim %load_seqcst16(q);
 +  prim %store_relaxed16(q, x);
 +  (x) = prim %load_relaxed16(q);
 +  prim %store_release16(q, x);
 +  (x) = prim %load_acquire16(q);
    return (x);
+ }
@@ -76,5 +89,9 @@ cmm_foo8 (P_ p)
    (x) = prim %fetch_xor8(q, 33 :: I8);
    (x) = prim %fetch_nand8(q, 127 :: I8);
    (x) = prim %load_seqcst8(q);
 +  prim %store_relaxed8(q, x);
 +  (x) = prim %load_relaxed8(q);
 +  prim %store_release8(q, x);
 +  (x) = prim %load_acquire8(q);
    return (x);
+ }