CmmOpt.hs 19.5 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 17 18 19 20 21 22 23 24
-----------------------------------------------------------------------------
--
-- Cmm optimisation
--
-- (c) The University of Glasgow 2006
--
-----------------------------------------------------------------------------

module CmmOpt (
	cmmMiniInline,
	cmmMachOpFold,
	cmmLoopifyForC,
 ) where

#include "HsVersions.h"

import Cmm
25
import CmmExpr
Simon Marlow's avatar
Simon Marlow committed
26 27
import CmmUtils
import CLabel
28
import MachOp
29
import StaticFlags
30 31

import UniqFM
Simon Marlow's avatar
Simon Marlow committed
32
import Unique
33 34 35

import Outputable

Simon Marlow's avatar
Simon Marlow committed
36 37 38 39
import Data.Bits
import Data.Word
import Data.Int
import GHC.Exts
40 41 42 43

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

Simon Marlow's avatar
Simon Marlow committed
44 45 46 47 48 49 50 51 52 53 54 55
{-
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.

56 57 58 59
[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
60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92
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.
-}

93 94 95 96
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

97 98
cmmMiniInline :: [CmmBasicBlock] -> [CmmBasicBlock]
cmmMiniInline blocks = map do_inline blocks 
99 100
  where do_inline (BasicBlock id stmts)
          = BasicBlock id (cmmMiniInlineStmts (countUses blocks) stmts)
101 102 103

cmmMiniInlineStmts :: UniqFM Int -> [CmmStmt] -> [CmmStmt]
cmmMiniInlineStmts uses [] = []
104
cmmMiniInlineStmts uses (stmt@(CmmAssign (CmmLocal (LocalReg u _ _)) expr) : stmts)
105 106 107 108 109 110 111 112 113 114 115 116 117 118 119
  | 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


-- Try to inline a temporary assignment.  We can skip over assignments to
-- other tempoararies, because we know that expressions aren't side-effecting
-- and temporaries are single-assignment.
120
lookForInline u expr (stmt@(CmmAssign (CmmLocal (LocalReg u' _ _)) rhs) : rest)
121
  | u /= u' 
122
  = case lookupUFM (countUses rhs) u of
123 124 125 126 127 128 129 130
	Just 1 -> Just (inlineStmt u expr stmt : rest)
	_other -> case lookForInline u expr rest of
		     Nothing    -> Nothing
		     Just stmts -> Just (stmt:stmts)

lookForInline u expr (CmmNop : rest)
  = lookForInline u expr rest

131 132
lookForInline _ _ [] = Nothing

133
lookForInline u expr (stmt:stmts)
134
  = case lookupUFM (countUses stmt) u of
135
	Just 1 | ok_to_inline -> Just (inlineStmt u expr stmt : stmts)
136
	_other -> Nothing
137 138 139 140 141 142 143 144 145
  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
146 147 148 149

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)
150 151
inlineStmt u a (CmmCall target regs es srt ret)
   = CmmCall (infn target) regs es' srt ret
152
   where infn (CmmCallee fn cconv) = CmmCallee fn cconv
153 154 155 156 157 158 159 160
	 infn (CmmPrim p) = CmmPrim p
	 es' = [ (inlineExpr u a e, hint) | (e,hint) <- es ]
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
161
inlineExpr u a e@(CmmReg (CmmLocal (LocalReg u' _ _)))
162 163
  | u == u' = a
  | otherwise = e
164
inlineExpr u a e@(CmmRegOff (CmmLocal (LocalReg u' rep _)) off)
165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300
  | u == u' = CmmMachOp (MO_Add rep) [a, CmmLit (CmmInt (fromIntegral off) rep)]
  | otherwise = e
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.
      MO_S_Conv from to
	   | isFloatingRep to -> CmmLit (CmmFloat (fromInteger x) to)
	   | otherwise        -> CmmLit (CmmInt (narrowS from x) to)
      MO_U_Conv from to -> CmmLit (CmmInt (narrowU from x) to)

      _ -> panic "cmmMachOpFold: unknown unary op"


-- Eliminate conversion NOPs
cmmMachOpFold (MO_S_Conv rep1 rep2) [x] | rep1 == rep2 = x
cmmMachOpFold (MO_U_Conv rep1 rep2) [x] | rep1 == rep2 = x

-- 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 ->
	    cmmMachOpFold (MO_U_Conv rep1 rep3) [x]
	| otherwise ->
	    CmmMachOp conv_outer args
  where
	isIntConversion (MO_U_Conv rep1 rep2) 
	  | not (isFloatingRep rep1) && not (isFloatingRep rep2) 
	  = Just (rep1,rep2,False)
	isIntConversion (MO_S_Conv rep1 rep2)
	  | not (isFloatingRep rep1) && not (isFloatingRep rep2) 
	  = Just (rep1,rep2,True)
	isIntConversion _ = Nothing

	intconv True  = MO_S_Conv
	intconv False = MO_U_Conv

-- 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.
    	MO_Eq r   -> CmmLit (CmmInt (if x_u == y_u then 1 else 0) wordRep)
    	MO_Ne r   -> CmmLit (CmmInt (if x_u /= y_u then 1 else 0) wordRep)

    	MO_U_Gt r -> CmmLit (CmmInt (if x_u >  y_u then 1 else 0) wordRep)
    	MO_U_Ge r -> CmmLit (CmmInt (if x_u >= y_u then 1 else 0) wordRep)
    	MO_U_Lt r -> CmmLit (CmmInt (if x_u <  y_u then 1 else 0) wordRep)
    	MO_U_Le r -> CmmLit (CmmInt (if x_u <= y_u then 1 else 0) wordRep)

    	MO_S_Gt r -> CmmLit (CmmInt (if x_s >  y_s then 1 else 0) wordRep) 
    	MO_S_Ge r -> CmmLit (CmmInt (if x_s >= y_s then 1 else 0) wordRep)
    	MO_S_Lt r -> CmmLit (CmmInt (if x_s <  y_s then 1 else 0) wordRep)
    	MO_S_Le r -> CmmLit (CmmInt (if x_s <= y_s then 1 else 0) wordRep)

    	MO_Add r -> CmmLit (CmmInt (x + y) r)
    	MO_Sub r -> CmmLit (CmmInt (x - y) r)
    	MO_Mul r -> CmmLit (CmmInt (x * y) r)
    	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) ...
--
301 302 303
-- Also don't do it if arg1 is PicBaseReg, so that we don't separate the
-- PicBaseReg from the corresponding label (or label difference).
--
304
cmmMachOpFold mop1 [CmmMachOp mop2 [arg1,arg2], arg3]
305 306
   | mop1 == mop2 && isAssociativeMachOp mop1
     && not (isLit arg1) && not (isPicReg arg1)
307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327
   = cmmMachOpFold mop1 [arg1, cmmMachOpFold mop2 [arg2,arg3]]

-- 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))))

328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343

-- Comparison of literal with narrowed/widened operand: perform
-- the comparison at a different width, as long as the literal is
-- within range.

#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 _)]
  | Just (rep, narrow) <- maybe_conversion conv,
    Just narrow_cmp <- maybe_comparison cmp rep,
    let narrow_i = narrow rep i,
    narrow_i == i
  = cmmMachOpFold narrow_cmp [x, CmmLit (CmmInt narrow_i rep)]
 where
    maybe_conversion (MO_U_Conv from _) = Just (from, narrowU)
344 345
    maybe_conversion (MO_S_Conv from _)
        | not (isFloatingRep from) = Just (from, narrowS)
Simon Marlow's avatar
Simon Marlow committed
346 347
        -- don't attempt to apply this optimisation when the source
        -- is a float; see #1916
348 349 350 351 352 353 354 355 356 357 358 359 360 361 362
    maybe_conversion _ = Nothing
    
    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_S_Gt _) rep = Just (MO_S_Gt rep)
    maybe_comparison (MO_S_Ge _) rep = Just (MO_S_Ge rep)
    maybe_comparison (MO_S_Lt _) rep = Just (MO_S_Lt rep)
    maybe_comparison (MO_S_Le _) rep = Just (MO_S_Le rep)
    maybe_comparison (MO_Eq   _) rep = Just (MO_Eq   rep)
    maybe_comparison _ _ = Nothing

#endif

363 364 365 366 367 368 369 370 371 372 373 374 375 376
-- 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
377
	MO_Eq    r | Just x' <- maybeInvertCmmExpr x -> x'
378 379 380 381 382 383
	MO_U_Gt  r | isComparisonExpr x -> x
	MO_S_Gt  r | isComparisonExpr x -> x
	MO_U_Lt  r | isComparisonExpr x -> CmmLit (CmmInt 0 wordRep)
	MO_S_Lt  r | isComparisonExpr x -> CmmLit (CmmInt 0 wordRep)
	MO_U_Ge  r | isComparisonExpr x -> CmmLit (CmmInt 1 wordRep)
	MO_S_Ge  r | isComparisonExpr x -> CmmLit (CmmInt 1 wordRep)
384 385
	MO_U_Le  r | Just x' <- maybeInvertCmmExpr x -> x'
	MO_S_Le  r | Just x' <- maybeInvertCmmExpr x -> x'
386 387 388 389 390 391 392 393 394
    	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)
395
        MO_Ne    r | Just x' <- maybeInvertCmmExpr x -> x'
396
	MO_Eq    r | isComparisonExpr x -> x
397 398
	MO_U_Lt  r | Just x' <- maybeInvertCmmExpr x -> x'
	MO_S_Lt  r | Just x' <- maybeInvertCmmExpr x -> x'
399 400 401 402 403 404 405 406 407 408 409 410 411
	MO_U_Gt  r | isComparisonExpr x -> CmmLit (CmmInt 0 wordRep)
	MO_S_Gt  r | isComparisonExpr x -> CmmLit (CmmInt 0 wordRep)
	MO_U_Le  r | isComparisonExpr x -> CmmLit (CmmInt 1 wordRep)
	MO_S_Le  r | isComparisonExpr x -> CmmLit (CmmInt 1 wordRep)
	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
412 413
	   | Just p <- exactLog2 n ->
                 CmmMachOp (MO_Shl rep) [x, CmmLit (CmmInt p rep)]
414
    	MO_S_Quot rep
415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442
	   | 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
		-- to minus infinity, whereasq uot rounds toward zero.
		-- 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 
		    bits = fromIntegral (machRepBitWidth rep) - 1
		    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
                CmmMachOp (MO_S_Shr rep) [x3, CmmLit (CmmInt p rep)]
    	other
443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522
           -> 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.

w2i x = word2Int# x
i2w x = int2Word# x

exactLog2 :: Integer -> Maybe Integer
exactLog2 x
  = if (x <= 0 || x >= 2147483648) then
       Nothing
    else
       case fromInteger x of { I# x# ->
       if (w2i ((i2w x#) `and#` (i2w (0# -# x#))) /=# x#) then
	  Nothing
       else
	  Just (toInteger (I# (pow2 x#)))
       }
  where
    pow2 x# | x# ==# 1# = 0#
            | otherwise = 1# +# pow2 (w2i (i2w x# `shiftRL#` 1#))


-- -----------------------------------------------------------------------------
-- widening / narrowing

narrowU :: MachRep -> Integer -> Integer
narrowU I8  x = fromIntegral (fromIntegral x :: Word8)
narrowU I16 x = fromIntegral (fromIntegral x :: Word16)
narrowU I32 x = fromIntegral (fromIntegral x :: Word32)
narrowU I64 x = fromIntegral (fromIntegral x :: Word64)
narrowU _ _ = panic "narrowTo"

narrowS :: MachRep -> Integer -> Integer
narrowS I8  x = fromIntegral (fromIntegral x :: Int8)
narrowS I16 x = fromIntegral (fromIntegral x :: Int16)
narrowS I32 x = fromIntegral (fromIntegral x :: Int32)
narrowS I64 x = fromIntegral (fromIntegral x :: Int64)
narrowS _ _ = panic "narrowTo"

-- -----------------------------------------------------------------------------
-- 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.
-}

523
cmmLoopifyForC :: RawCmmTop -> RawCmmTop
524
cmmLoopifyForC p@(CmmProc info entry_lbl [] (ListGraph blocks@(BasicBlock top_id _ : _)))
525 526 527
  | null info = p  -- only if there's an info table, ignore case alts
  | otherwise =  
--  pprTrace "jump_lbl" (ppr jump_lbl <+> ppr entry_lbl) $
528
  CmmProc info entry_lbl [] (ListGraph blocks')
529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550
  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

551
isPicReg (CmmReg (CmmGlobal PicBaseReg)) = True
552
isPicReg _ = False
553 554 555

_unused :: FS.FastString -- stops a warning
_unused = undefined