CmmOpt.hs 23.8 KB
Newer Older
1 2 3 4 5 6 7
{-# OPTIONS -w #-}
-- The above warning supression flag is a temporary kludge.
-- While working on this module you are encouraged to remove it and fix
-- any warnings in the module. See
--     http://hackage.haskell.org/trac/ghc/wiki/Commentary/CodingStyle#Warnings
-- for details

8 9 10 11 12 13 14 15 16
-----------------------------------------------------------------------------
--
-- Cmm optimisation
--
-- (c) The University of Glasgow 2006
--
-----------------------------------------------------------------------------

module CmmOpt (
17
	cmmEliminateDeadBlocks,
18 19 20 21 22 23 24
	cmmMiniInline,
	cmmMachOpFold,
	cmmLoopifyForC,
 ) where

#include "HsVersions.h"

25
import OldCmm
Simon Marlow's avatar
Simon Marlow committed
26 27
import CmmUtils
import CLabel
28
import StaticFlags
29 30

import UniqFM
Simon Marlow's avatar
Simon Marlow committed
31
import Unique
32
import FastTypes
33
import Outputable
34
import BlockId
35

Simon Marlow's avatar
Simon Marlow committed
36 37 38
import Data.Bits
import Data.Word
import Data.Int
39
import Data.Maybe
40
import Data.List
41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

import Compiler.Hoopl hiding (Unique)

-- -----------------------------------------------------------------------------
-- Eliminates dead blocks

{-
We repeatedly expand the set of reachable blocks until we hit a
fixpoint, and then prune any blocks that were not in this set.  This is
actually a required optimization, as dead blocks can cause problems
for invariants in the linear register allocator (and possibly other
places.)
-}

-- Deep fold over statements could probably be abstracted out, but it
-- might not be worth the effort since OldCmm is moribund
cmmEliminateDeadBlocks :: [CmmBasicBlock] -> [CmmBasicBlock]
cmmEliminateDeadBlocks [] = []
cmmEliminateDeadBlocks blocks@(BasicBlock base_id _:_) =
    let -- Calculate what's reachable from what block
61 62 63
        reachableMap = foldl' f emptyUFM blocks -- lazy in values
            where f m (BasicBlock block_id stmts) = addToUFM m block_id (reachableFrom stmts)
        reachableFrom stmts = foldl stmt [] stmts
64 65 66 67 68 69 70 71
            where
                stmt m CmmNop = m
                stmt m (CmmComment _) = m
                stmt m (CmmAssign _ e) = expr m e
                stmt m (CmmStore e1 e2) = expr (expr m e1) e2
                stmt m (CmmCall c _ as _ _) = f (actuals m as) c
                    where f m (CmmCallee e _) = expr m e
                          f m (CmmPrim _) = m
72 73 74
                stmt m (CmmBranch b) = b:m
                stmt m (CmmCondBranch e b) = b:(expr m e)
                stmt m (CmmSwitch e bs) = catMaybes bs ++ expr m e
75 76
                stmt m (CmmJump e as) = expr (actuals m as) e
                stmt m (CmmReturn as) = actuals m as
77 78 79
                actuals m as = foldl' (\m h -> expr m (hintlessCmm h)) m as
                -- We have to do a deep fold into CmmExpr because
                -- there may be a BlockId in the CmmBlock literal.
80 81 82
                expr m (CmmLit l) = lit m l
                expr m (CmmLoad e _) = expr m e
                expr m (CmmReg _) = m
83
                expr m (CmmMachOp _ es) = foldl' expr m es
84 85
                expr m (CmmStackSlot _ _) = m
                expr m (CmmRegOff _ _) = m
86
                lit m (CmmBlock b) = b:m
87
                lit m _ = m
88 89 90 91 92 93 94 95
        -- go todo done
        reachable = go [base_id] (setEmpty :: BlockSet)
          where go []     m = m
                go (x:xs) m
                    | setMember x m = go xs m
                    | otherwise     = go (add ++ xs) (setInsert x m)
                        where add = fromMaybe (panic "cmmEliminateDeadBlocks: unknown block")
                                              (lookupUFM reachableMap x)
96
    in filter (\(BasicBlock block_id _) -> setMember block_id reachable) blocks
97 98 99 100

-- -----------------------------------------------------------------------------
-- The mini-inliner

Simon Marlow's avatar
Simon Marlow committed
101 102 103 104 105 106 107 108 109 110 111 112
{-
This pass inlines assignments to temporaries that are used just
once.  It works as follows:

  - count uses of each temporary
  - for each temporary that occurs just once:
	- attempt to push it forward to the statement that uses it
        - only push forward past assignments to other temporaries
	  (assumes that temporaries are single-assignment)
	- if we reach the statement that uses it, inline the rhs
	  and delete the original assignment.

113 114 115 116
[N.B. In the Quick C-- compiler, this optimization is achieved by a
 combination of two dataflow passes: forward substitution (peephole
 optimization) and dead-assignment elimination.  ---NR]

Simon Marlow's avatar
Simon Marlow committed
117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149
Possible generalisations: here is an example from factorial

Fac_zdwfac_entry:
    cmG:
        _smi = R2;
        if (_smi != 0) goto cmK;
        R1 = R3;
        jump I64[Sp];
    cmK:
        _smn = _smi * R3;
        R2 = _smi + (-1);
        R3 = _smn;
        jump Fac_zdwfac_info;

We want to inline _smi and _smn.  To inline _smn:

   - we must be able to push forward past assignments to global regs.
     We can do this if the rhs of the assignment we are pushing
     forward doesn't refer to the global reg being assigned to; easy
     to test.

To inline _smi:

   - It is a trivial replacement, reg for reg, but it occurs more than
     once.
   - We can inline trivial assignments even if the temporary occurs
     more than once, as long as we don't eliminate the original assignment
     (this doesn't help much on its own).
   - We need to be able to propagate the assignment forward through jumps;
     if we did this, we would find that it can be inlined safely in all
     its occurrences.
-}

150 151 152 153
countUses :: UserOfLocalRegs a => a -> UniqFM Int
countUses a = foldRegsUsed (\m r -> addToUFM m r (count m r + 1)) emptyUFM a
  where count m r = lookupWithDefaultUFM m (0::Int) r

154 155
cmmMiniInline :: [CmmBasicBlock] -> [CmmBasicBlock]
cmmMiniInline blocks = map do_inline blocks 
156 157
  where do_inline (BasicBlock id stmts)
          = BasicBlock id (cmmMiniInlineStmts (countUses blocks) stmts)
158 159 160

cmmMiniInlineStmts :: UniqFM Int -> [CmmStmt] -> [CmmStmt]
cmmMiniInlineStmts uses [] = []
161
cmmMiniInlineStmts uses (stmt@(CmmAssign (CmmLocal (LocalReg u _)) expr) : stmts)
162 163 164 165 166
        -- not used at all: just discard this assignment
  | Nothing <- lookupUFM uses u
  = cmmMiniInlineStmts uses stmts

        -- used once: try to inline at the use site
167 168 169 170 171 172 173 174 175 176 177
  | Just 1 <- lookupUFM uses u,
    Just stmts' <- lookForInline u expr stmts
  = 
#ifdef NCG_DEBUG
     trace ("nativeGen: inlining " ++ showSDoc (pprStmt stmt)) $
#endif
     cmmMiniInlineStmts uses stmts'

cmmMiniInlineStmts uses (stmt:stmts)
  = stmt : cmmMiniInlineStmts uses stmts

178 179 180 181
lookForInline u expr stmts = lookForInline' u expr regset stmts
    where regset = foldRegsUsed extendRegSet emptyRegSet expr

lookForInline' u expr regset (stmt : rest)
182 183 184 185
  | Just 1 <- lookupUFM (countUses stmt) u, ok_to_inline
  = Just (inlineStmt u expr stmt : rest)

  | ok_to_skip
186
  = case lookForInline' u expr regset rest of
187 188 189 190 191
           Nothing    -> Nothing
           Just stmts -> Just (stmt:stmts)

  | otherwise 
  = Nothing
192

193 194 195 196 197 198 199 200 201
  where
	-- we don't inline into CmmCall if the expression refers to global
	-- registers.  This is a HACK to avoid global registers clashing with
	-- C argument-passing registers, really the back-end ought to be able
	-- to handle it properly, but currently neither PprC nor the NCG can
	-- do it.  See also CgForeignCall:load_args_into_temps.
    ok_to_inline = case stmt of
		     CmmCall{} -> hasNoGlobalRegs expr
		     _ -> True
202

203 204 205 206 207 208 209 210
   -- Expressions aren't side-effecting.  Temporaries may or may not
   -- be single-assignment depending on the source (the old code
   -- generator creates single-assignment code, but hand-written Cmm
   -- and Cmm from the new code generator is not single-assignment.)
   -- So we do an extra check to make sure that the register being
   -- changed is not one we were relying on.  I don't know how much of a
   -- performance hit this is (we have to create a regset for every
   -- instruction.) -- EZY
211 212
    ok_to_skip = case stmt of
                 CmmNop -> True
213
                 CmmComment{} -> True
214
                 CmmAssign (CmmLocal r@(LocalReg u' _)) rhs | u' /= u && not (r `elemRegSet` regset) -> True
215 216 217 218
                 CmmAssign g@(CmmGlobal _) rhs -> not (g `regUsedIn` expr)
                 _other -> False


219 220 221
inlineStmt :: Unique -> CmmExpr -> CmmStmt -> CmmStmt
inlineStmt u a (CmmAssign r e) = CmmAssign r (inlineExpr u a e)
inlineStmt u a (CmmStore e1 e2) = CmmStore (inlineExpr u a e1) (inlineExpr u a e2)
222 223
inlineStmt u a (CmmCall target regs es srt ret)
   = CmmCall (infn target) regs es' srt ret
224
   where infn (CmmCallee fn cconv) = CmmCallee (inlineExpr u a fn) cconv
225
	 infn (CmmPrim p) = CmmPrim p
226
	 es' = [ (CmmHinted (inlineExpr u a e) hint) | (CmmHinted e hint) <- es ]
227 228 229 230 231 232
inlineStmt u a (CmmCondBranch e d) = CmmCondBranch (inlineExpr u a e) d
inlineStmt u a (CmmSwitch e d) = CmmSwitch (inlineExpr u a e) d
inlineStmt u a (CmmJump e d) = CmmJump (inlineExpr u a e) d
inlineStmt u a other_stmt = other_stmt

inlineExpr :: Unique -> CmmExpr -> CmmExpr -> CmmExpr
233
inlineExpr u a e@(CmmReg (CmmLocal (LocalReg u' _)))
234 235
  | u == u' = a
  | otherwise = e
236 237
inlineExpr u a e@(CmmRegOff (CmmLocal (LocalReg u' rep)) off)
  | u == u' = CmmMachOp (MO_Add width) [a, CmmLit (CmmInt (fromIntegral off) width)]
238
  | otherwise = e
239 240
  where
    width = typeWidth rep
241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264
inlineExpr u a (CmmLoad e rep) = CmmLoad (inlineExpr u a e) rep
inlineExpr u a (CmmMachOp op es) = CmmMachOp op (map (inlineExpr u a) es)
inlineExpr u a other_expr = other_expr

-- -----------------------------------------------------------------------------
-- MachOp constant folder

-- Now, try to constant-fold the MachOps.  The arguments have already
-- been optimized and folded.

cmmMachOpFold
    :: MachOp	    	-- The operation from an CmmMachOp
    -> [CmmExpr]   	-- The optimized arguments
    -> CmmExpr

cmmMachOpFold op arg@[CmmLit (CmmInt x rep)]
  = case op of
      MO_S_Neg r -> CmmLit (CmmInt (-x) rep)
      MO_Not r   -> CmmLit (CmmInt (complement x) rep)

	-- these are interesting: we must first narrow to the 
	-- "from" type, in order to truncate to the correct size.
	-- The final narrow/widen to the destination type
	-- is implicit in the CmmLit.
265 266 267
      MO_SF_Conv from to -> CmmLit (CmmFloat (fromInteger x) to)
      MO_SS_Conv from to -> CmmLit (CmmInt (narrowS from x) to)
      MO_UU_Conv from to -> CmmLit (CmmInt (narrowU from x) to)
268 269 270 271 272

      _ -> panic "cmmMachOpFold: unknown unary op"


-- Eliminate conversion NOPs
273 274
cmmMachOpFold (MO_SS_Conv rep1 rep2) [x] | rep1 == rep2 = x
cmmMachOpFold (MO_UU_Conv rep1 rep2) [x] | rep1 == rep2 = x
275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292

-- Eliminate nested conversions where possible
cmmMachOpFold conv_outer args@[CmmMachOp conv_inner [x]]
  | Just (rep1,rep2,signed1) <- isIntConversion conv_inner,
    Just (_,   rep3,signed2) <- isIntConversion conv_outer
  = case () of
	-- widen then narrow to the same size is a nop
      _ | rep1 < rep2 && rep1 == rep3 -> x
	-- Widen then narrow to different size: collapse to single conversion
	-- but remember to use the signedness from the widening, just in case
	-- the final conversion is a widen.
	| rep1 < rep2 && rep2 > rep3 ->
	    cmmMachOpFold (intconv signed1 rep1 rep3) [x]
	-- Nested widenings: collapse if the signedness is the same
	| rep1 < rep2 && rep2 < rep3 && signed1 == signed2 ->
	    cmmMachOpFold (intconv signed1 rep1 rep3) [x]
	-- Nested narrowings: collapse
	| rep1 > rep2 && rep2 > rep3 ->
293
	    cmmMachOpFold (MO_UU_Conv rep1 rep3) [x]
294 295 296
	| otherwise ->
	    CmmMachOp conv_outer args
  where
297
	isIntConversion (MO_UU_Conv rep1 rep2) 
298
	  = Just (rep1,rep2,False)
299
	isIntConversion (MO_SS_Conv rep1 rep2)
300 301 302
	  = Just (rep1,rep2,True)
	isIntConversion _ = Nothing

303 304
	intconv True  = MO_SS_Conv
	intconv False = MO_UU_Conv
305 306 307 308 309 310 311 312 313

-- ToDo: a narrow of a load can be collapsed into a narrow load, right?
-- but what if the architecture only supports word-sized loads, should
-- we do the transformation anyway?

cmmMachOpFold mop args@[CmmLit (CmmInt x xrep), CmmLit (CmmInt y _)]
  = case mop of
	-- for comparisons: don't forget to narrow the arguments before
	-- comparing, since they might be out of range.
314 315
    	MO_Eq r   -> CmmLit (CmmInt (if x_u == y_u then 1 else 0) wordWidth)
    	MO_Ne r   -> CmmLit (CmmInt (if x_u /= y_u then 1 else 0) wordWidth)
316

317 318 319 320
    	MO_U_Gt r -> CmmLit (CmmInt (if x_u >  y_u then 1 else 0) wordWidth)
    	MO_U_Ge r -> CmmLit (CmmInt (if x_u >= y_u then 1 else 0) wordWidth)
    	MO_U_Lt r -> CmmLit (CmmInt (if x_u <  y_u then 1 else 0) wordWidth)
    	MO_U_Le r -> CmmLit (CmmInt (if x_u <= y_u then 1 else 0) wordWidth)
321

322 323 324 325
    	MO_S_Gt r -> CmmLit (CmmInt (if x_s >  y_s then 1 else 0) wordWidth) 
    	MO_S_Ge r -> CmmLit (CmmInt (if x_s >= y_s then 1 else 0) wordWidth)
    	MO_S_Lt r -> CmmLit (CmmInt (if x_s <  y_s then 1 else 0) wordWidth)
    	MO_S_Le r -> CmmLit (CmmInt (if x_s <= y_s then 1 else 0) wordWidth)
326 327 328 329

    	MO_Add r -> CmmLit (CmmInt (x + y) r)
    	MO_Sub r -> CmmLit (CmmInt (x - y) r)
    	MO_Mul r -> CmmLit (CmmInt (x * y) r)
330 331
    	MO_U_Quot r | y /= 0 -> CmmLit (CmmInt (x_u `quot` y_u) r)
    	MO_U_Rem  r | y /= 0 -> CmmLit (CmmInt (x_u `rem`  y_u) r)
332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373
    	MO_S_Quot r | y /= 0 -> CmmLit (CmmInt (x `quot` y) r)
    	MO_S_Rem  r | y /= 0 -> CmmLit (CmmInt (x `rem` y) r)

	MO_And   r -> CmmLit (CmmInt (x .&. y) r)
	MO_Or    r -> CmmLit (CmmInt (x .|. y) r)
	MO_Xor   r -> CmmLit (CmmInt (x `xor` y) r)

        MO_Shl   r -> CmmLit (CmmInt (x `shiftL` fromIntegral y) r)
        MO_U_Shr r -> CmmLit (CmmInt (x_u `shiftR` fromIntegral y) r)
        MO_S_Shr r -> CmmLit (CmmInt (x `shiftR` fromIntegral y) r)

	other      -> CmmMachOp mop args

   where
	x_u = narrowU xrep x
	y_u = narrowU xrep y
	x_s = narrowS xrep x
	y_s = narrowS xrep y
	

-- When possible, shift the constants to the right-hand side, so that we
-- can match for strength reductions.  Note that the code generator will
-- also assume that constants have been shifted to the right when
-- possible.

cmmMachOpFold op [x@(CmmLit _), y]
   | not (isLit y) && isCommutableMachOp op 
   = cmmMachOpFold op [y, x]

-- Turn (a+b)+c into a+(b+c) where possible.  Because literals are
-- moved to the right, it is more likely that we will find
-- opportunities for constant folding when the expression is
-- right-associated.
--
-- ToDo: this appears to introduce a quadratic behaviour due to the
-- nested cmmMachOpFold.  Can we fix this?
--
-- Why do we check isLit arg1?  If arg1 is a lit, it means that arg2
-- is also a lit (otherwise arg1 would be on the right).  If we
-- put arg1 on the left of the rearranged expression, we'll get into a
-- loop:  (x1+x2)+x3 => x1+(x2+x3)  => (x2+x3)+x1 => x2+(x3+x1) ...
--
374 375 376
-- Also don't do it if arg1 is PicBaseReg, so that we don't separate the
-- PicBaseReg from the corresponding label (or label difference).
--
377
cmmMachOpFold mop1 [CmmMachOp mop2 [arg1,arg2], arg3]
378
   | mop2 `associates_with` mop1
379
     && not (isLit arg1) && not (isPicReg arg1)
380 381 382 383 384 385 386 387 388 389
   = cmmMachOpFold mop2 [arg1, cmmMachOpFold mop1 [arg2,arg3]]
   where
     MO_Add{} `associates_with` MO_Sub{} = True
     mop1 `associates_with` mop2 =
        mop1 == mop2 && isAssociativeMachOp mop1

-- special case: (a - b) + c  ==>  a + (c - b)
cmmMachOpFold mop1@(MO_Add{}) [CmmMachOp mop2@(MO_Sub{}) [arg1,arg2], arg3]
   | not (isLit arg1) && not (isPicReg arg1)
   = cmmMachOpFold mop1 [arg1, cmmMachOpFold mop2 [arg3,arg2]]
390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409

-- Make a RegOff if we can
cmmMachOpFold (MO_Add _) [CmmReg reg, CmmLit (CmmInt n rep)]
  = CmmRegOff reg (fromIntegral (narrowS rep n))
cmmMachOpFold (MO_Add _) [CmmRegOff reg off, CmmLit (CmmInt n rep)]
  = CmmRegOff reg (off + fromIntegral (narrowS rep n))
cmmMachOpFold (MO_Sub _) [CmmReg reg, CmmLit (CmmInt n rep)]
  = CmmRegOff reg (- fromIntegral (narrowS rep n))
cmmMachOpFold (MO_Sub _) [CmmRegOff reg off, CmmLit (CmmInt n rep)]
  = CmmRegOff reg (off - fromIntegral (narrowS rep n))

-- Fold label(+/-)offset into a CmmLit where possible

cmmMachOpFold (MO_Add _) [CmmLit (CmmLabel lbl), CmmLit (CmmInt i rep)]
  = CmmLit (CmmLabelOff lbl (fromIntegral (narrowU rep i)))
cmmMachOpFold (MO_Add _) [CmmLit (CmmInt i rep), CmmLit (CmmLabel lbl)]
  = CmmLit (CmmLabelOff lbl (fromIntegral (narrowU rep i)))
cmmMachOpFold (MO_Sub _) [CmmLit (CmmLabel lbl), CmmLit (CmmInt i rep)]
  = CmmLit (CmmLabelOff lbl (fromIntegral (negate (narrowU rep i))))

410

411 412 413 414 415 416
-- Comparison of literal with widened operand: perform the comparison
-- at the smaller width, as long as the literal is within range.

-- We can't do the reverse trick, when the operand is narrowed:
-- narrowing throws away bits from the operand, there's no way to do
-- the same comparison at the larger size.
417 418 419 420 421

#if i386_TARGET_ARCH || x86_64_TARGET_ARCH
-- powerPC NCG has a TODO for I8/I16 comparisons, so don't try

cmmMachOpFold cmp [CmmMachOp conv [x], CmmLit (CmmInt i _)]
422 423 424 425 426 427 428 429
  |     -- if the operand is widened:
    Just (rep, signed, narrow_fn) <- maybe_conversion conv,
        -- and this is a comparison operation:
    Just narrow_cmp <- maybe_comparison cmp rep signed,
        -- and the literal fits in the smaller size:
    i == narrow_fn rep i
        -- then we can do the comparison at the smaller size
  = cmmMachOpFold narrow_cmp [x, CmmLit (CmmInt i rep)]
430
 where
431
    maybe_conversion (MO_UU_Conv from to)
432 433
        | to > from
        = Just (from, False, narrowU)
434 435
    maybe_conversion (MO_SS_Conv from to)
        | to > from
436
        = Just (from, True, narrowS)
437

Simon Marlow's avatar
Simon Marlow committed
438 439
        -- don't attempt to apply this optimisation when the source
        -- is a float; see #1916
440 441
    maybe_conversion _ = Nothing
    
442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458
        -- careful (#2080): if the original comparison was signed, but
        -- we were doing an unsigned widen, then we must do an
        -- unsigned comparison at the smaller size.
    maybe_comparison (MO_U_Gt _) rep _     = Just (MO_U_Gt rep)
    maybe_comparison (MO_U_Ge _) rep _     = Just (MO_U_Ge rep)
    maybe_comparison (MO_U_Lt _) rep _     = Just (MO_U_Lt rep)
    maybe_comparison (MO_U_Le _) rep _     = Just (MO_U_Le rep)
    maybe_comparison (MO_Eq   _) rep _     = Just (MO_Eq   rep)
    maybe_comparison (MO_S_Gt _) rep True  = Just (MO_S_Gt rep)
    maybe_comparison (MO_S_Ge _) rep True  = Just (MO_S_Ge rep)
    maybe_comparison (MO_S_Lt _) rep True  = Just (MO_S_Lt rep)
    maybe_comparison (MO_S_Le _) rep True  = Just (MO_S_Le rep)
    maybe_comparison (MO_S_Gt _) rep False = Just (MO_U_Gt rep)
    maybe_comparison (MO_S_Ge _) rep False = Just (MO_U_Ge rep)
    maybe_comparison (MO_S_Lt _) rep False = Just (MO_U_Lt rep)
    maybe_comparison (MO_S_Le _) rep False = Just (MO_U_Le rep)
    maybe_comparison _ _ _ = Nothing
459 460 461

#endif

462 463 464 465 466 467 468 469 470 471 472 473 474 475
-- We can often do something with constants of 0 and 1 ...

cmmMachOpFold mop args@[x, y@(CmmLit (CmmInt 0 _))]
  = case mop of
    	MO_Add   r -> x
    	MO_Sub   r -> x
    	MO_Mul   r -> y
    	MO_And   r -> y
    	MO_Or    r -> x
    	MO_Xor   r -> x
    	MO_Shl   r -> x
    	MO_S_Shr r -> x
    	MO_U_Shr r -> x
        MO_Ne    r | isComparisonExpr x -> x
476
	MO_Eq    r | Just x' <- maybeInvertCmmExpr x -> x'
477 478
	MO_U_Gt  r | isComparisonExpr x -> x
	MO_S_Gt  r | isComparisonExpr x -> x
479 480 481 482
	MO_U_Lt  r | isComparisonExpr x -> CmmLit (CmmInt 0 wordWidth)
	MO_S_Lt  r | isComparisonExpr x -> CmmLit (CmmInt 0 wordWidth)
	MO_U_Ge  r | isComparisonExpr x -> CmmLit (CmmInt 1 wordWidth)
	MO_S_Ge  r | isComparisonExpr x -> CmmLit (CmmInt 1 wordWidth)
483 484
	MO_U_Le  r | Just x' <- maybeInvertCmmExpr x -> x'
	MO_S_Le  r | Just x' <- maybeInvertCmmExpr x -> x'
485 486 487 488 489 490 491 492 493
    	other    -> CmmMachOp mop args

cmmMachOpFold mop args@[x, y@(CmmLit (CmmInt 1 rep))]
  = case mop of
    	MO_Mul    r -> x
    	MO_S_Quot r -> x
    	MO_U_Quot r -> x
    	MO_S_Rem  r -> CmmLit (CmmInt 0 rep)
    	MO_U_Rem  r -> CmmLit (CmmInt 0 rep)
494
        MO_Ne    r | Just x' <- maybeInvertCmmExpr x -> x'
495
	MO_Eq    r | isComparisonExpr x -> x
496 497
	MO_U_Lt  r | Just x' <- maybeInvertCmmExpr x -> x'
	MO_S_Lt  r | Just x' <- maybeInvertCmmExpr x -> x'
498 499 500 501
	MO_U_Gt  r | isComparisonExpr x -> CmmLit (CmmInt 0 wordWidth)
	MO_S_Gt  r | isComparisonExpr x -> CmmLit (CmmInt 0 wordWidth)
	MO_U_Le  r | isComparisonExpr x -> CmmLit (CmmInt 1 wordWidth)
	MO_S_Le  r | isComparisonExpr x -> CmmLit (CmmInt 1 wordWidth)
502 503 504 505 506 507 508 509 510
	MO_U_Ge  r | isComparisonExpr x -> x
	MO_S_Ge  r | isComparisonExpr x -> x
    	other       -> CmmMachOp mop args

-- Now look for multiplication/division by powers of 2 (integers).

cmmMachOpFold mop args@[x, y@(CmmLit (CmmInt n _))]
  = case mop of
    	MO_Mul rep
511
	   | Just p <- exactLog2 n ->
512
                 cmmMachOpFold (MO_Shl rep) [x, CmmLit (CmmInt p rep)]
513 514
    	MO_U_Quot rep
	   | Just p <- exactLog2 n ->
515
                 cmmMachOpFold (MO_U_Shr rep) [x, CmmLit (CmmInt p rep)]
516
    	MO_S_Quot rep
517 518 519 520
	   | Just p <- exactLog2 n, 
	     CmmReg _ <- x ->	-- We duplicate x below, hence require
				-- it is a reg.  FIXME: remove this restriction.
		-- shift right is not the same as quot, because it rounds
Simon Marlow's avatar
Simon Marlow committed
521
		-- to minus infinity, whereasq quot rounds toward zero.
522 523 524 525 526 527 528 529 530 531 532 533 534 535
		-- To fix this up, we add one less than the divisor to the
		-- dividend if it is a negative number.
		--
		-- to avoid a test/jump, we use the following sequence:
		-- 	x1 = x >> word_size-1  (all 1s if -ve, all 0s if +ve)
		--      x2 = y & (divisor-1)
		--      result = (x+x2) >>= log2(divisor)
		-- this could be done a bit more simply using conditional moves,
		-- but we're processor independent here.
		--
		-- we optimise the divide by 2 case slightly, generating
		--      x1 = x >> word_size-1  (unsigned)
		--      return = (x + x1) >>= log2(divisor)
		let 
536
		    bits = fromIntegral (widthInBits rep) - 1
537 538 539 540 541 542
		    shr = if p == 1 then MO_U_Shr rep else MO_S_Shr rep
		    x1 = CmmMachOp shr [x, CmmLit (CmmInt bits rep)]
		    x2 = if p == 1 then x1 else
			 CmmMachOp (MO_And rep) [x1, CmmLit (CmmInt (n-1) rep)]
		    x3 = CmmMachOp (MO_Add rep) [x, x2]
		in
543
                cmmMachOpFold (MO_S_Shr rep) [x3, CmmLit (CmmInt p rep)]
544
    	other
545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562
           -> unchanged
    where
       unchanged = CmmMachOp mop args

-- Anything else is just too hard.

cmmMachOpFold mop args = CmmMachOp mop args

-- -----------------------------------------------------------------------------
-- exactLog2

-- This algorithm for determining the $\log_2$ of exact powers of 2 comes
-- from GCC.  It requires bit manipulation primitives, and we use GHC
-- extensions.  Tough.
-- 
-- Used to be in MachInstrs --SDM.
-- ToDo: remove use of unboxery --SDM.

563 564 565 566 567
-- Unboxery removed in favor of FastInt; but is the function supposed to fail
-- on inputs >= 2147483648, or was that just an implementation artifact?
-- And is this speed-critical, or can we just use Integer operations
-- (including Data.Bits)?
--  --Isaac Dupree
568 569

exactLog2 :: Integer -> Maybe Integer
570 571
exactLog2 x_
  = if (x_ <= 0 || x_ >= 2147483648) then
572 573
       Nothing
    else
574 575
       case iUnbox (fromInteger x_) of { x ->
       if (x `bitAndFastInt` negateFastInt x) /=# x then
576 577
	  Nothing
       else
578
	  Just (toInteger (iBox (pow2 x)))
579 580
       }
  where
581 582
    pow2 x | x ==# _ILIT(1) = _ILIT(0)
           | otherwise = _ILIT(1) +# pow2 (x `shiftR_FastInt` _ILIT(1))
583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610


-- -----------------------------------------------------------------------------
-- Loopify for C

{-
 This is a simple pass that replaces tail-recursive functions like this:

   fac() {
     ...
     jump fac();
   }

 with this:

  fac() {
   L:
     ...
     goto L;
  }

  the latter generates better C code, because the C compiler treats it
  like a loop, and brings full loop optimisation to bear.

  In my measurements this makes little or no difference to anything
  except factorial, but what the hell.
-}

611
cmmLoopifyForC :: RawCmmTop -> RawCmmTop
612
cmmLoopifyForC p@(CmmProc info entry_lbl
613
                 (ListGraph blocks@(BasicBlock top_id _ : _)))
614 615 616
  | null info = p  -- only if there's an info table, ignore case alts
  | otherwise =  
--  pprTrace "jump_lbl" (ppr jump_lbl <+> ppr entry_lbl) $
617
  CmmProc info entry_lbl (ListGraph blocks')
618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639
  where blocks' = [ BasicBlock id (map do_stmt stmts)
		  | BasicBlock id stmts <- blocks ]

        do_stmt (CmmJump (CmmLit (CmmLabel lbl)) _) | lbl == jump_lbl
		= CmmBranch top_id
	do_stmt stmt = stmt

	jump_lbl | tablesNextToCode = entryLblToInfoLbl entry_lbl
		 | otherwise        = entry_lbl

cmmLoopifyForC top = top

-- -----------------------------------------------------------------------------
-- Utils

isLit (CmmLit _) = True
isLit _          = False

isComparisonExpr :: CmmExpr -> Bool
isComparisonExpr (CmmMachOp op _) = isComparisonMachOp op
isComparisonExpr _other 	    = False

640
isPicReg (CmmReg (CmmGlobal PicBaseReg)) = True
641
isPicReg _ = False
642