import CmmUtils
import MachOp ( MachRep(..), wordRep, MachOp(..), MachHint(..),
mo_wordOr, mo_wordAnd, mo_wordNe, mo_wordEq,
- mo_wordULt, machRepByteWidth )
+ mo_wordULt, mo_wordUGt, machRepByteWidth )
import ForeignCall ( CCallConv(..) )
import Literal ( Literal(..) )
import CLabel ( CLabel, mkStringLitLabel )
import Digraph ( SCC(..), stronglyConnComp )
import ListSetOps ( assocDefault )
import Util ( filterOut, sortLe )
-import Char ( ord )
+import CmdLineOpts ( DynFlags(..), HscTarget(..) )
import FastString ( LitString, FastString, unpackFS )
import Outputable
+import Char ( ord )
import DATA_BITS
-
-#include "../includes/ghcconfig.h"
- -- For WORDS_BIGENDIAN
+import Maybe ( isNothing )
-------------------------------------------------------------------------
--
cmmAndWord e1 e2 = CmmMachOp mo_wordAnd [e1, e2]
cmmNeWord e1 e2 = CmmMachOp mo_wordNe [e1, e2]
cmmEqWord e1 e2 = CmmMachOp mo_wordEq [e1, e2]
-cmmULtWord e1 e2 = CmmMachOp mo_wordULt [e1, e2]
+cmmULtWord e1 e2 = CmmMachOp mo_wordULt [e1, e2]
+cmmUGtWord e1 e2 = CmmMachOp mo_wordUGt [e1, e2]
cmmNegate :: CmmExpr -> CmmExpr
cmmNegate (CmmLit (CmmInt n rep)) = CmmLit (CmmInt (-n) rep)
--
-------------------------------------------------------------------------
-tagToClosure :: TyCon -> CmmExpr -> CmmExpr
-tagToClosure tycon tag
+tagToClosure :: DynFlags -> TyCon -> CmmExpr -> CmmExpr
+tagToClosure dflags tycon tag
= CmmLoad (cmmOffsetExprW closure_tbl tag) wordRep
- where closure_tbl = CmmLit (CmmLabel (mkClosureTblLabel (tyConName tycon)))
+ where closure_tbl = CmmLit (CmmLabel lbl)
+ lbl = mkClosureTableLabel dflags (tyConName tycon)
-------------------------------------------------------------------------
--
emitRODataLits :: CLabel -> [CmmLit] -> Code
-- Emit a read-only data block
emitRODataLits lbl lits
- = emitData ReadOnlyData (CmmDataLabel lbl : map CmmStaticLit lits)
+ = emitData section (CmmDataLabel lbl : map CmmStaticLit lits)
+ where section | any needsRelocation lits = RelocatableReadOnlyData
+ | otherwise = ReadOnlyData
+ needsRelocation (CmmLabel _) = True
+ needsRelocation (CmmLabelOff _ _) = True
+ needsRelocation _ = False
mkStringCLit :: String -> FCode CmmLit
-- Make a global definition for the string,
Nothing -> return Nothing
Just stmts -> do id <- forkCgStmts stmts; return (Just id)
+ ; dflags <- getDynFlags
+ ; let via_C | HscC <- hscTarget dflags = True
+ | otherwise = False
+
; stmts <- mk_switch tag_expr (sortLe le branches)
- mb_deflt_id lo_tag hi_tag
+ mb_deflt_id lo_tag hi_tag via_C
; emitCgStmts stmts
}
where
mk_switch :: CmmExpr -> [(ConTagZ, CgStmts)]
- -> Maybe BlockId -> ConTagZ -> ConTagZ
+ -> Maybe BlockId -> ConTagZ -> ConTagZ -> Bool
-> FCode CgStmts
-- SINGLETON TAG RANGE: no case analysis to do
-mk_switch tag_expr [(tag,stmts)] _ lo_tag hi_tag
+mk_switch tag_expr [(tag,stmts)] _ lo_tag hi_tag via_C
| lo_tag == hi_tag
= ASSERT( tag == lo_tag )
return stmts
-- SINGLETON BRANCH, NO DEFUALT: no case analysis to do
-mk_switch tag_expr [(tag,stmts)] Nothing lo_tag hi_tag
+mk_switch tag_expr [(tag,stmts)] Nothing lo_tag hi_tag via_C
= return stmts
-- The simplifier might have eliminated a case
-- so we may have e.g. case xs of
-- can't happen, so no need to test
-- SINGLETON BRANCH: one equality check to do
-mk_switch tag_expr [(tag,stmts)] (Just deflt) lo_tag hi_tag
+mk_switch tag_expr [(tag,stmts)] (Just deflt) lo_tag hi_tag via_C
= return (CmmCondBranch cond deflt `consCgStmt` stmts)
where
cond = cmmNeWord tag_expr (CmmLit (mkIntCLit tag))
-- the branches is the tag 0, because comparing '== 0' is likely to be
-- more efficient than other kinds of comparison.
--- DENSE TAG RANGE: use a switch statment
-mk_switch tag_expr branches mb_deflt lo_tag hi_tag
- | use_switch -- Use a switch
+-- DENSE TAG RANGE: use a switch statment.
+--
+-- We also use a switch uncoditionally when compiling via C, because
+-- this will get emitted as a C switch statement and the C compiler
+-- should do a good job of optimising it. Also, older GCC versions
+-- (2.95 in particular) have problems compiling the complicated
+-- if-trees generated by this code, so compiling to a switch every
+-- time works around that problem.
+--
+mk_switch tag_expr branches mb_deflt lo_tag hi_tag via_C
+ | use_switch || via_C -- Use a switch
= do { branch_ids <- mapM forkCgStmts (map snd branches)
; let
tagged_blk_ids = zip (map fst branches) (map Just branch_ids)
find_branch :: ConTagZ -> Maybe BlockId
find_branch i = assocDefault mb_deflt tagged_blk_ids i
- arms = [ find_branch (i+lo_tag) | i <- [0..n_tags-1]]
+ -- NB. we have eliminated impossible branches at
+ -- either end of the range (see below), so the first
+ -- tag of a real branch is real_lo_tag (not lo_tag).
+ arms = [ find_branch i | i <- [real_lo_tag..real_hi_tag]]
- switch_stmt = CmmSwitch (cmmOffset tag_expr (- lo_tag)) arms
+ switch_stmt = CmmSwitch (cmmOffset tag_expr (- real_lo_tag)) arms
- ; return (oneCgStmt switch_stmt)
+ ; ASSERT(not (all isNothing arms))
+ return (oneCgStmt switch_stmt)
}
+ -- if we can knock off a bunch of default cases with one if, then do so
+ | Just deflt <- mb_deflt, (lowest_branch - lo_tag) >= n_branches
+ = do { (assign_tag, tag_expr') <- assignTemp' tag_expr
+ ; let cond = cmmULtWord tag_expr' (CmmLit (mkIntCLit lowest_branch))
+ branch = CmmCondBranch cond deflt
+ ; stmts <- mk_switch tag_expr' branches mb_deflt
+ lowest_branch hi_tag via_C
+ ; return (assign_tag `consCgStmt` (branch `consCgStmt` stmts))
+ }
+
+ | Just deflt <- mb_deflt, (hi_tag - highest_branch) >= n_branches
+ = do { (assign_tag, tag_expr') <- assignTemp' tag_expr
+ ; let cond = cmmUGtWord tag_expr' (CmmLit (mkIntCLit highest_branch))
+ branch = CmmCondBranch cond deflt
+ ; stmts <- mk_switch tag_expr' branches mb_deflt
+ lo_tag highest_branch via_C
+ ; return (assign_tag `consCgStmt` (branch `consCgStmt` stmts))
+ }
+
| otherwise -- Use an if-tree
= do { (assign_tag, tag_expr') <- assignTemp' tag_expr
-- To avoid duplication
- ; lo_stmts <- mk_switch tag_expr' lo_branches mb_deflt lo_tag (mid_tag-1)
- ; hi_stmts <- mk_switch tag_expr' hi_branches mb_deflt mid_tag hi_tag
+ ; lo_stmts <- mk_switch tag_expr' lo_branches mb_deflt
+ lo_tag (mid_tag-1) via_C
+ ; hi_stmts <- mk_switch tag_expr' hi_branches mb_deflt
+ mid_tag hi_tag via_C
; lo_id <- forkCgStmts lo_stmts
; let cond = cmmULtWord tag_expr' (CmmLit (mkIntCLit mid_tag))
branch_stmt = CmmCondBranch cond lo_id
}
where
use_switch = ASSERT( n_branches > 1 && n_tags > 1 )
+ {- pprTrace "mk_switch" (ppr tag_expr <+> text "n_tags: "
+ <+> int n_tags <+> text "dense: "
+ <+> int n_branches) $ -}
n_tags > 2 && (small || dense)
-- a 2-branch switch always turns into an if.
small = n_tags <= 4
dense = n_branches > (n_tags `div` 2)
exhaustive = n_tags == n_branches
- n_tags = hi_tag - lo_tag + 1
n_branches = length branches
+ -- ignore default slots at each end of the range if there's
+ -- no default branch defined.
+ lowest_branch = fst (head branches)
+ highest_branch = fst (last branches)
+
+ real_lo_tag
+ | isNothing mb_deflt = lowest_branch
+ | otherwise = lo_tag
+
+ real_hi_tag
+ | isNothing mb_deflt = highest_branch
+ | otherwise = hi_tag
+
+ n_tags = real_hi_tag - real_lo_tag + 1
+
-- INVARIANT: Provided hi_tag > lo_tag (which is true)
-- lo_tag <= mid_tag < hi_tag
-- lo_branches have tags < mid_tag
-- Used for general literals, whose size might not be a word,
-- where there is always a default case, and where we don't know
-- the range of values for certain. For simplicity we always generate a tree.
+--
+-- ToDo: for integers we could do better here, perhaps by generalising
+-- mk_switch and using that. --SDM 15/09/2004
emitLitSwitch scrut [] deflt
= emitCgStmts deflt
emitLitSwitch scrut branches deflt_blk