2 -- The above warning supression flag is a temporary kludge.
3 -- While working on this module you are encouraged to remove it and fix
4 -- any warnings in the module. See
5 -- http://hackage.haskell.org/trac/ghc/wiki/Commentary/CodingStyle#Warnings
8 -----------------------------------------------------------------------------
12 -- (c) The University of Glasgow 2006
14 -----------------------------------------------------------------------------
22 #include "HsVersions.h"
38 -- -----------------------------------------------------------------------------
42 This pass inlines assignments to temporaries that are used just
43 once. It works as follows:
45 - count uses of each temporary
46 - for each temporary that occurs just once:
47 - attempt to push it forward to the statement that uses it
48 - only push forward past assignments to other temporaries
49 (assumes that temporaries are single-assignment)
50 - if we reach the statement that uses it, inline the rhs
51 and delete the original assignment.
53 [N.B. In the Quick C-- compiler, this optimization is achieved by a
54 combination of two dataflow passes: forward substitution (peephole
55 optimization) and dead-assignment elimination. ---NR]
57 Possible generalisations: here is an example from factorial
62 if (_smi != 0) goto cmK;
71 We want to inline _smi and _smn. To inline _smn:
73 - we must be able to push forward past assignments to global regs.
74 We can do this if the rhs of the assignment we are pushing
75 forward doesn't refer to the global reg being assigned to; easy
80 - It is a trivial replacement, reg for reg, but it occurs more than
82 - We can inline trivial assignments even if the temporary occurs
83 more than once, as long as we don't eliminate the original assignment
84 (this doesn't help much on its own).
85 - We need to be able to propagate the assignment forward through jumps;
86 if we did this, we would find that it can be inlined safely in all
90 countUses :: UserOfLocalRegs a => a -> UniqFM Int
91 countUses a = foldRegsUsed (\m r -> addToUFM m r (count m r + 1)) emptyUFM a
92 where count m r = lookupWithDefaultUFM m (0::Int) r
94 cmmMiniInline :: [CmmBasicBlock] -> [CmmBasicBlock]
95 cmmMiniInline blocks = map do_inline blocks
96 where do_inline (BasicBlock id stmts)
97 = BasicBlock id (cmmMiniInlineStmts (countUses blocks) stmts)
99 cmmMiniInlineStmts :: UniqFM Int -> [CmmStmt] -> [CmmStmt]
100 cmmMiniInlineStmts uses [] = []
101 cmmMiniInlineStmts uses (stmt@(CmmAssign (CmmLocal (LocalReg u _)) expr) : stmts)
102 -- not used at all: just discard this assignment
103 | Nothing <- lookupUFM uses u
104 = cmmMiniInlineStmts uses stmts
106 -- used once: try to inline at the use site
107 | Just 1 <- lookupUFM uses u,
108 Just stmts' <- lookForInline u expr stmts
111 trace ("nativeGen: inlining " ++ showSDoc (pprStmt stmt)) $
113 cmmMiniInlineStmts uses stmts'
115 cmmMiniInlineStmts uses (stmt:stmts)
116 = stmt : cmmMiniInlineStmts uses stmts
118 lookForInline u expr (stmt : rest)
119 | Just 1 <- lookupUFM (countUses stmt) u, ok_to_inline
120 = Just (inlineStmt u expr stmt : rest)
123 = case lookForInline u expr rest of
125 Just stmts -> Just (stmt:stmts)
131 -- we don't inline into CmmCall if the expression refers to global
132 -- registers. This is a HACK to avoid global registers clashing with
133 -- C argument-passing registers, really the back-end ought to be able
134 -- to handle it properly, but currently neither PprC nor the NCG can
135 -- do it. See also CgForeignCall:load_args_into_temps.
136 ok_to_inline = case stmt of
137 CmmCall{} -> hasNoGlobalRegs expr
140 -- We can skip over assignments to other tempoararies, because we
141 -- know that expressions aren't side-effecting and temporaries are
142 -- single-assignment.
143 ok_to_skip = case stmt of
146 CmmAssign (CmmLocal (LocalReg u' _)) rhs | u' /= u -> True
147 CmmAssign g@(CmmGlobal _) rhs -> not (g `regUsedIn` expr)
151 inlineStmt :: Unique -> CmmExpr -> CmmStmt -> CmmStmt
152 inlineStmt u a (CmmAssign r e) = CmmAssign r (inlineExpr u a e)
153 inlineStmt u a (CmmStore e1 e2) = CmmStore (inlineExpr u a e1) (inlineExpr u a e2)
154 inlineStmt u a (CmmCall target regs es srt ret)
155 = CmmCall (infn target) regs es' srt ret
156 where infn (CmmCallee fn cconv) = CmmCallee (inlineExpr u a fn) cconv
157 infn (CmmPrim p) = CmmPrim p
158 es' = [ (CmmHinted (inlineExpr u a e) hint) | (CmmHinted e hint) <- es ]
159 inlineStmt u a (CmmCondBranch e d) = CmmCondBranch (inlineExpr u a e) d
160 inlineStmt u a (CmmSwitch e d) = CmmSwitch (inlineExpr u a e) d
161 inlineStmt u a (CmmJump e d) = CmmJump (inlineExpr u a e) d
162 inlineStmt u a other_stmt = other_stmt
164 inlineExpr :: Unique -> CmmExpr -> CmmExpr -> CmmExpr
165 inlineExpr u a e@(CmmReg (CmmLocal (LocalReg u' _)))
168 inlineExpr u a e@(CmmRegOff (CmmLocal (LocalReg u' rep)) off)
169 | u == u' = CmmMachOp (MO_Add width) [a, CmmLit (CmmInt (fromIntegral off) width)]
172 width = typeWidth rep
173 inlineExpr u a (CmmLoad e rep) = CmmLoad (inlineExpr u a e) rep
174 inlineExpr u a (CmmMachOp op es) = CmmMachOp op (map (inlineExpr u a) es)
175 inlineExpr u a other_expr = other_expr
177 -- -----------------------------------------------------------------------------
178 -- MachOp constant folder
180 -- Now, try to constant-fold the MachOps. The arguments have already
181 -- been optimized and folded.
184 :: MachOp -- The operation from an CmmMachOp
185 -> [CmmExpr] -- The optimized arguments
188 cmmMachOpFold op arg@[CmmLit (CmmInt x rep)]
190 MO_S_Neg r -> CmmLit (CmmInt (-x) rep)
191 MO_Not r -> CmmLit (CmmInt (complement x) rep)
193 -- these are interesting: we must first narrow to the
194 -- "from" type, in order to truncate to the correct size.
195 -- The final narrow/widen to the destination type
196 -- is implicit in the CmmLit.
197 MO_SF_Conv from to -> CmmLit (CmmFloat (fromInteger x) to)
198 MO_SS_Conv from to -> CmmLit (CmmInt (narrowS from x) to)
199 MO_UU_Conv from to -> CmmLit (CmmInt (narrowU from x) to)
201 _ -> panic "cmmMachOpFold: unknown unary op"
204 -- Eliminate conversion NOPs
205 cmmMachOpFold (MO_SS_Conv rep1 rep2) [x] | rep1 == rep2 = x
206 cmmMachOpFold (MO_UU_Conv rep1 rep2) [x] | rep1 == rep2 = x
208 -- Eliminate nested conversions where possible
209 cmmMachOpFold conv_outer args@[CmmMachOp conv_inner [x]]
210 | Just (rep1,rep2,signed1) <- isIntConversion conv_inner,
211 Just (_, rep3,signed2) <- isIntConversion conv_outer
213 -- widen then narrow to the same size is a nop
214 _ | rep1 < rep2 && rep1 == rep3 -> x
215 -- Widen then narrow to different size: collapse to single conversion
216 -- but remember to use the signedness from the widening, just in case
217 -- the final conversion is a widen.
218 | rep1 < rep2 && rep2 > rep3 ->
219 cmmMachOpFold (intconv signed1 rep1 rep3) [x]
220 -- Nested widenings: collapse if the signedness is the same
221 | rep1 < rep2 && rep2 < rep3 && signed1 == signed2 ->
222 cmmMachOpFold (intconv signed1 rep1 rep3) [x]
223 -- Nested narrowings: collapse
224 | rep1 > rep2 && rep2 > rep3 ->
225 cmmMachOpFold (MO_UU_Conv rep1 rep3) [x]
227 CmmMachOp conv_outer args
229 isIntConversion (MO_UU_Conv rep1 rep2)
230 = Just (rep1,rep2,False)
231 isIntConversion (MO_SS_Conv rep1 rep2)
232 = Just (rep1,rep2,True)
233 isIntConversion _ = Nothing
235 intconv True = MO_SS_Conv
236 intconv False = MO_UU_Conv
238 -- ToDo: a narrow of a load can be collapsed into a narrow load, right?
239 -- but what if the architecture only supports word-sized loads, should
240 -- we do the transformation anyway?
242 cmmMachOpFold mop args@[CmmLit (CmmInt x xrep), CmmLit (CmmInt y _)]
244 -- for comparisons: don't forget to narrow the arguments before
245 -- comparing, since they might be out of range.
246 MO_Eq r -> CmmLit (CmmInt (if x_u == y_u then 1 else 0) wordWidth)
247 MO_Ne r -> CmmLit (CmmInt (if x_u /= y_u then 1 else 0) wordWidth)
249 MO_U_Gt r -> CmmLit (CmmInt (if x_u > y_u then 1 else 0) wordWidth)
250 MO_U_Ge r -> CmmLit (CmmInt (if x_u >= y_u then 1 else 0) wordWidth)
251 MO_U_Lt r -> CmmLit (CmmInt (if x_u < y_u then 1 else 0) wordWidth)
252 MO_U_Le r -> CmmLit (CmmInt (if x_u <= y_u then 1 else 0) wordWidth)
254 MO_S_Gt r -> CmmLit (CmmInt (if x_s > y_s then 1 else 0) wordWidth)
255 MO_S_Ge r -> CmmLit (CmmInt (if x_s >= y_s then 1 else 0) wordWidth)
256 MO_S_Lt r -> CmmLit (CmmInt (if x_s < y_s then 1 else 0) wordWidth)
257 MO_S_Le r -> CmmLit (CmmInt (if x_s <= y_s then 1 else 0) wordWidth)
259 MO_Add r -> CmmLit (CmmInt (x + y) r)
260 MO_Sub r -> CmmLit (CmmInt (x - y) r)
261 MO_Mul r -> CmmLit (CmmInt (x * y) r)
262 MO_U_Quot r | y /= 0 -> CmmLit (CmmInt (x_u `quot` y_u) r)
263 MO_U_Rem r | y /= 0 -> CmmLit (CmmInt (x_u `rem` y_u) r)
264 MO_S_Quot r | y /= 0 -> CmmLit (CmmInt (x `quot` y) r)
265 MO_S_Rem r | y /= 0 -> CmmLit (CmmInt (x `rem` y) r)
267 MO_And r -> CmmLit (CmmInt (x .&. y) r)
268 MO_Or r -> CmmLit (CmmInt (x .|. y) r)
269 MO_Xor r -> CmmLit (CmmInt (x `xor` y) r)
271 MO_Shl r -> CmmLit (CmmInt (x `shiftL` fromIntegral y) r)
272 MO_U_Shr r -> CmmLit (CmmInt (x_u `shiftR` fromIntegral y) r)
273 MO_S_Shr r -> CmmLit (CmmInt (x `shiftR` fromIntegral y) r)
275 other -> CmmMachOp mop args
284 -- When possible, shift the constants to the right-hand side, so that we
285 -- can match for strength reductions. Note that the code generator will
286 -- also assume that constants have been shifted to the right when
289 cmmMachOpFold op [x@(CmmLit _), y]
290 | not (isLit y) && isCommutableMachOp op
291 = cmmMachOpFold op [y, x]
293 -- Turn (a+b)+c into a+(b+c) where possible. Because literals are
294 -- moved to the right, it is more likely that we will find
295 -- opportunities for constant folding when the expression is
298 -- ToDo: this appears to introduce a quadratic behaviour due to the
299 -- nested cmmMachOpFold. Can we fix this?
301 -- Why do we check isLit arg1? If arg1 is a lit, it means that arg2
302 -- is also a lit (otherwise arg1 would be on the right). If we
303 -- put arg1 on the left of the rearranged expression, we'll get into a
304 -- loop: (x1+x2)+x3 => x1+(x2+x3) => (x2+x3)+x1 => x2+(x3+x1) ...
306 -- Also don't do it if arg1 is PicBaseReg, so that we don't separate the
307 -- PicBaseReg from the corresponding label (or label difference).
309 cmmMachOpFold mop1 [CmmMachOp mop2 [arg1,arg2], arg3]
310 | mop2 `associates_with` mop1
311 && not (isLit arg1) && not (isPicReg arg1)
312 = cmmMachOpFold mop2 [arg1, cmmMachOpFold mop1 [arg2,arg3]]
314 MO_Add{} `associates_with` MO_Sub{} = True
315 mop1 `associates_with` mop2 =
316 mop1 == mop2 && isAssociativeMachOp mop1
318 -- special case: (a - b) + c ==> a + (c - b)
319 cmmMachOpFold mop1@(MO_Add{}) [CmmMachOp mop2@(MO_Sub{}) [arg1,arg2], arg3]
320 | not (isLit arg1) && not (isPicReg arg1)
321 = cmmMachOpFold mop1 [arg1, cmmMachOpFold mop2 [arg3,arg2]]
323 -- Make a RegOff if we can
324 cmmMachOpFold (MO_Add _) [CmmReg reg, CmmLit (CmmInt n rep)]
325 = CmmRegOff reg (fromIntegral (narrowS rep n))
326 cmmMachOpFold (MO_Add _) [CmmRegOff reg off, CmmLit (CmmInt n rep)]
327 = CmmRegOff reg (off + fromIntegral (narrowS rep n))
328 cmmMachOpFold (MO_Sub _) [CmmReg reg, CmmLit (CmmInt n rep)]
329 = CmmRegOff reg (- fromIntegral (narrowS rep n))
330 cmmMachOpFold (MO_Sub _) [CmmRegOff reg off, CmmLit (CmmInt n rep)]
331 = CmmRegOff reg (off - fromIntegral (narrowS rep n))
333 -- Fold label(+/-)offset into a CmmLit where possible
335 cmmMachOpFold (MO_Add _) [CmmLit (CmmLabel lbl), CmmLit (CmmInt i rep)]
336 = CmmLit (CmmLabelOff lbl (fromIntegral (narrowU rep i)))
337 cmmMachOpFold (MO_Add _) [CmmLit (CmmInt i rep), CmmLit (CmmLabel lbl)]
338 = CmmLit (CmmLabelOff lbl (fromIntegral (narrowU rep i)))
339 cmmMachOpFold (MO_Sub _) [CmmLit (CmmLabel lbl), CmmLit (CmmInt i rep)]
340 = CmmLit (CmmLabelOff lbl (fromIntegral (negate (narrowU rep i))))
343 -- Comparison of literal with widened operand: perform the comparison
344 -- at the smaller width, as long as the literal is within range.
346 -- We can't do the reverse trick, when the operand is narrowed:
347 -- narrowing throws away bits from the operand, there's no way to do
348 -- the same comparison at the larger size.
350 #if i386_TARGET_ARCH || x86_64_TARGET_ARCH
351 -- powerPC NCG has a TODO for I8/I16 comparisons, so don't try
353 cmmMachOpFold cmp [CmmMachOp conv [x], CmmLit (CmmInt i _)]
354 | -- if the operand is widened:
355 Just (rep, signed, narrow_fn) <- maybe_conversion conv,
356 -- and this is a comparison operation:
357 Just narrow_cmp <- maybe_comparison cmp rep signed,
358 -- and the literal fits in the smaller size:
360 -- then we can do the comparison at the smaller size
361 = cmmMachOpFold narrow_cmp [x, CmmLit (CmmInt i rep)]
363 maybe_conversion (MO_UU_Conv from to)
365 = Just (from, False, narrowU)
366 maybe_conversion (MO_SS_Conv from to)
368 = Just (from, True, narrowS)
370 -- don't attempt to apply this optimisation when the source
371 -- is a float; see #1916
372 maybe_conversion _ = Nothing
374 -- careful (#2080): if the original comparison was signed, but
375 -- we were doing an unsigned widen, then we must do an
376 -- unsigned comparison at the smaller size.
377 maybe_comparison (MO_U_Gt _) rep _ = Just (MO_U_Gt rep)
378 maybe_comparison (MO_U_Ge _) rep _ = Just (MO_U_Ge rep)
379 maybe_comparison (MO_U_Lt _) rep _ = Just (MO_U_Lt rep)
380 maybe_comparison (MO_U_Le _) rep _ = Just (MO_U_Le rep)
381 maybe_comparison (MO_Eq _) rep _ = Just (MO_Eq rep)
382 maybe_comparison (MO_S_Gt _) rep True = Just (MO_S_Gt rep)
383 maybe_comparison (MO_S_Ge _) rep True = Just (MO_S_Ge rep)
384 maybe_comparison (MO_S_Lt _) rep True = Just (MO_S_Lt rep)
385 maybe_comparison (MO_S_Le _) rep True = Just (MO_S_Le rep)
386 maybe_comparison (MO_S_Gt _) rep False = Just (MO_U_Gt rep)
387 maybe_comparison (MO_S_Ge _) rep False = Just (MO_U_Ge rep)
388 maybe_comparison (MO_S_Lt _) rep False = Just (MO_U_Lt rep)
389 maybe_comparison (MO_S_Le _) rep False = Just (MO_U_Le rep)
390 maybe_comparison _ _ _ = Nothing
394 -- We can often do something with constants of 0 and 1 ...
396 cmmMachOpFold mop args@[x, y@(CmmLit (CmmInt 0 _))]
407 MO_Ne r | isComparisonExpr x -> x
408 MO_Eq r | Just x' <- maybeInvertCmmExpr x -> x'
409 MO_U_Gt r | isComparisonExpr x -> x
410 MO_S_Gt r | isComparisonExpr x -> x
411 MO_U_Lt r | isComparisonExpr x -> CmmLit (CmmInt 0 wordWidth)
412 MO_S_Lt r | isComparisonExpr x -> CmmLit (CmmInt 0 wordWidth)
413 MO_U_Ge r | isComparisonExpr x -> CmmLit (CmmInt 1 wordWidth)
414 MO_S_Ge r | isComparisonExpr x -> CmmLit (CmmInt 1 wordWidth)
415 MO_U_Le r | Just x' <- maybeInvertCmmExpr x -> x'
416 MO_S_Le r | Just x' <- maybeInvertCmmExpr x -> x'
417 other -> CmmMachOp mop args
419 cmmMachOpFold mop args@[x, y@(CmmLit (CmmInt 1 rep))]
424 MO_S_Rem r -> CmmLit (CmmInt 0 rep)
425 MO_U_Rem r -> CmmLit (CmmInt 0 rep)
426 MO_Ne r | Just x' <- maybeInvertCmmExpr x -> x'
427 MO_Eq r | isComparisonExpr x -> x
428 MO_U_Lt r | Just x' <- maybeInvertCmmExpr x -> x'
429 MO_S_Lt r | Just x' <- maybeInvertCmmExpr x -> x'
430 MO_U_Gt r | isComparisonExpr x -> CmmLit (CmmInt 0 wordWidth)
431 MO_S_Gt r | isComparisonExpr x -> CmmLit (CmmInt 0 wordWidth)
432 MO_U_Le r | isComparisonExpr x -> CmmLit (CmmInt 1 wordWidth)
433 MO_S_Le r | isComparisonExpr x -> CmmLit (CmmInt 1 wordWidth)
434 MO_U_Ge r | isComparisonExpr x -> x
435 MO_S_Ge r | isComparisonExpr x -> x
436 other -> CmmMachOp mop args
438 -- Now look for multiplication/division by powers of 2 (integers).
440 cmmMachOpFold mop args@[x, y@(CmmLit (CmmInt n _))]
443 | Just p <- exactLog2 n ->
444 cmmMachOpFold (MO_Shl rep) [x, CmmLit (CmmInt p rep)]
446 | Just p <- exactLog2 n ->
447 cmmMachOpFold (MO_U_Shr rep) [x, CmmLit (CmmInt p rep)]
449 | Just p <- exactLog2 n,
450 CmmReg _ <- x -> -- We duplicate x below, hence require
451 -- it is a reg. FIXME: remove this restriction.
452 -- shift right is not the same as quot, because it rounds
453 -- to minus infinity, whereasq quot rounds toward zero.
454 -- To fix this up, we add one less than the divisor to the
455 -- dividend if it is a negative number.
457 -- to avoid a test/jump, we use the following sequence:
458 -- x1 = x >> word_size-1 (all 1s if -ve, all 0s if +ve)
459 -- x2 = y & (divisor-1)
460 -- result = (x+x2) >>= log2(divisor)
461 -- this could be done a bit more simply using conditional moves,
462 -- but we're processor independent here.
464 -- we optimise the divide by 2 case slightly, generating
465 -- x1 = x >> word_size-1 (unsigned)
466 -- return = (x + x1) >>= log2(divisor)
468 bits = fromIntegral (widthInBits rep) - 1
469 shr = if p == 1 then MO_U_Shr rep else MO_S_Shr rep
470 x1 = CmmMachOp shr [x, CmmLit (CmmInt bits rep)]
471 x2 = if p == 1 then x1 else
472 CmmMachOp (MO_And rep) [x1, CmmLit (CmmInt (n-1) rep)]
473 x3 = CmmMachOp (MO_Add rep) [x, x2]
475 cmmMachOpFold (MO_S_Shr rep) [x3, CmmLit (CmmInt p rep)]
479 unchanged = CmmMachOp mop args
481 -- Anything else is just too hard.
483 cmmMachOpFold mop args = CmmMachOp mop args
485 -- -----------------------------------------------------------------------------
488 -- This algorithm for determining the $\log_2$ of exact powers of 2 comes
489 -- from GCC. It requires bit manipulation primitives, and we use GHC
490 -- extensions. Tough.
492 -- Used to be in MachInstrs --SDM.
493 -- ToDo: remove use of unboxery --SDM.
495 -- Unboxery removed in favor of FastInt; but is the function supposed to fail
496 -- on inputs >= 2147483648, or was that just an implementation artifact?
497 -- And is this speed-critical, or can we just use Integer operations
498 -- (including Data.Bits)?
501 exactLog2 :: Integer -> Maybe Integer
503 = if (x_ <= 0 || x_ >= 2147483648) then
506 case iUnbox (fromInteger x_) of { x ->
507 if (x `bitAndFastInt` negateFastInt x) /=# x then
510 Just (toInteger (iBox (pow2 x)))
513 pow2 x | x ==# _ILIT(1) = _ILIT(0)
514 | otherwise = _ILIT(1) +# pow2 (x `shiftR_FastInt` _ILIT(1))
517 -- -----------------------------------------------------------------------------
521 This is a simple pass that replaces tail-recursive functions like this:
536 the latter generates better C code, because the C compiler treats it
537 like a loop, and brings full loop optimisation to bear.
539 In my measurements this makes little or no difference to anything
540 except factorial, but what the hell.
543 cmmLoopifyForC :: RawCmmTop -> RawCmmTop
544 cmmLoopifyForC p@(CmmProc info entry_lbl
545 (ListGraph blocks@(BasicBlock top_id _ : _)))
546 | null info = p -- only if there's an info table, ignore case alts
548 -- pprTrace "jump_lbl" (ppr jump_lbl <+> ppr entry_lbl) $
549 CmmProc info entry_lbl (ListGraph blocks')
550 where blocks' = [ BasicBlock id (map do_stmt stmts)
551 | BasicBlock id stmts <- blocks ]
553 do_stmt (CmmJump (CmmLit (CmmLabel lbl)) _) | lbl == jump_lbl
557 jump_lbl | tablesNextToCode = entryLblToInfoLbl entry_lbl
558 | otherwise = entry_lbl
560 cmmLoopifyForC top = top
562 -- -----------------------------------------------------------------------------
565 isLit (CmmLit _) = True
568 isComparisonExpr :: CmmExpr -> Bool
569 isComparisonExpr (CmmMachOp op _) = isComparisonMachOp op
570 isComparisonExpr _other = False
572 isPicReg (CmmReg (CmmGlobal PicBaseReg)) = True