GitLab

Cheng Shao pushed to branch wip/fix-llvm-split-sections at Glasgow Haskell Compiler / GHC

Commits:

41074dd7

by Cheng Shao at 2026-01-14T03:28:25+01:00

llvm: fix split sections for llvm backend

This patch fixes split sections for llvm backend:

- Pass missing `--data-sections`/`--function-sections` flags to
  llc/opt.
- Use `@llvm.compiler.used` instead of `@llvm.used` to avoid sections
  being unnecessarily retained at link-time.

Fixes #26770.

-------------------------
Metric Decrease:
    libdir
    size_hello_artifact
    size_hello_unicode
-------------------------

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

3 changed files:

compiler/GHC/CmmToLlvm.hs
compiler/GHC/CmmToLlvm/Base.hs
compiler/GHC/Driver/Pipeline/Execute.hs

Changes:

compiler/GHC/CmmToLlvm.hs

@@ -271,7 +271,7 @@ cmmUsedLlvmGens = do
    -- used if we didn't provide these hints. This will generate a
    -- definition of the form
    --
 -  --   @llvm.used = appending global [42 x i8*] [i8* bitcast <var> to i8*, ...]
 +  --   @llvm.compiler.used = appending global [42 x i8*] [i8* bitcast <var> to i8*, ...]
    --
    -- Which is the LLVM way of protecting them against getting removed.
    ivars <- getUsedVars
@@ -279,7 +279,7 @@ cmmUsedLlvmGens = do
        ty     = LMArray (length ivars) i8Ptr
        usedArray = LMStaticArray (map cast ivars) ty
        sectName  = Just $ fsLit "llvm.metadata"
 -      lmUsedVar = LMGlobalVar (fsLit "llvm.used") ty Appending sectName Nothing Constant
 +      lmUsedVar = LMGlobalVar (fsLit "llvm.compiler.used") ty Appending sectName Nothing Constant
        lmUsed    = LMGlobal lmUsedVar (Just usedArray)
    if null ivars
       then return ()

compiler/GHC/CmmToLlvm/Base.hs

@@ -286,7 +286,7 @@ data LlvmEnv = LlvmEnv
    , envUniqMeta  :: UniqFM Unique MetaId   -- ^ Global metadata nodes
    , envFunMap    :: LlvmEnvMap       -- ^ Global functions so far, with type
    , envAliases   :: UniqSet LMString -- ^ Globals that we had to alias, see [Llvm Forward References]
 -  , envUsedVars  :: [LlvmVar]        -- ^ Pointers to be added to llvm.used (see @cmmUsedLlvmGens@)
 +  , envUsedVars  :: [LlvmVar]        -- ^ Pointers to be added to llvm.compiler.used (see @cmmUsedLlvmGens@)
      -- the following get cleared for every function (see @withClearVars@)
    , envVarMap    :: LlvmEnvMap       -- ^ Local variables so far, with type

compiler/GHC/Driver/Pipeline/Execute.hs

@@ -934,6 +934,14 @@ llvmOptions llvm_config llvm_version dflags =
         [("-relocation-model=" ++ rmodel
          ,"-relocation-model=" ++ rmodel) | not (null rmodel)]
 +    -- Both llc/opt need these flags for split sections
 +    ++ [ ("--data-sections", "--data-sections")
 +       | gopt Opt_SplitSections dflags
 +       ]
 +    ++ [ ("--function-sections", "--function-sections")
 +       | gopt Opt_SplitSections dflags
 +       ]
++
      -- Additional llc flags
      ++ [("", "-mcpu=" ++ mcpu)   | not (null mcpu)
                                   , not (any (isInfixOf "-mcpu") (getOpts dflags opt_lc)) ]