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"
39 -- -----------------------------------------------------------------------------
43 This pass inlines assignments to temporaries that are used just
44 once. It works as follows:
46 - count uses of each temporary
47 - for each temporary that occurs just once:
48 - attempt to push it forward to the statement that uses it
49 - only push forward past assignments to other temporaries
50 (assumes that temporaries are single-assignment)
51 - if we reach the statement that uses it, inline the rhs
52 and delete the original assignment.
54 [N.B. In the Quick C-- compiler, this optimization is achieved by a
55 combination of two dataflow passes: forward substitution (peephole
56 optimization) and dead-assignment elimination. ---NR]
58 Possible generalisations: here is an example from factorial
63 if (_smi != 0) goto cmK;
72 We want to inline _smi and _smn. To inline _smn:
74 - we must be able to push forward past assignments to global regs.
75 We can do this if the rhs of the assignment we are pushing
76 forward doesn't refer to the global reg being assigned to; easy
81 - It is a trivial replacement, reg for reg, but it occurs more than
83 - We can inline trivial assignments even if the temporary occurs
84 more than once, as long as we don't eliminate the original assignment
85 (this doesn't help much on its own).
86 - We need to be able to propagate the assignment forward through jumps;
87 if we did this, we would find that it can be inlined safely in all
91 countUses :: UserOfLocalRegs a => a -> UniqFM Int
92 countUses a = foldRegsUsed (\m r -> addToUFM m r (count m r + 1)) emptyUFM a
93 where count m r = lookupWithDefaultUFM m (0::Int) r
95 cmmMiniInline :: [CmmBasicBlock] -> [CmmBasicBlock]
96 cmmMiniInline blocks = map do_inline blocks
97 where do_inline (BasicBlock id stmts)
98 = BasicBlock id (cmmMiniInlineStmts (countUses blocks) stmts)
100 cmmMiniInlineStmts :: UniqFM Int -> [CmmStmt] -> [CmmStmt]
101 cmmMiniInlineStmts uses [] = []
102 cmmMiniInlineStmts uses (stmt@(CmmAssign (CmmLocal (LocalReg u _)) expr) : stmts)
103 -- not used at all: just discard this assignment
104 | Nothing <- lookupUFM uses u
105 = cmmMiniInlineStmts uses stmts
107 -- used once: try to inline at the use site
108 | Just 1 <- lookupUFM uses u,
109 Just stmts' <- lookForInline u expr stmts
112 trace ("nativeGen: inlining " ++ showSDoc (pprStmt stmt)) $
114 cmmMiniInlineStmts uses stmts'
116 cmmMiniInlineStmts uses (stmt:stmts)
117 = stmt : cmmMiniInlineStmts uses stmts
120 -- Try to inline a temporary assignment. We can skip over assignments to
121 -- other tempoararies, because we know that expressions aren't side-effecting
122 -- and temporaries are single-assignment.
123 lookForInline u expr (stmt@(CmmAssign (CmmLocal (LocalReg u' _)) rhs) : rest)
125 = case lookupUFM (countUses rhs) u of
126 Just 1 -> Just (inlineStmt u expr stmt : rest)
127 _other -> case lookForInline u expr rest of
129 Just stmts -> Just (stmt:stmts)
131 lookForInline u expr (CmmNop : rest)
132 = lookForInline u expr rest
134 lookForInline _ _ [] = Nothing
136 lookForInline u expr (stmt:stmts)
137 = case lookupUFM (countUses stmt) u of
138 Just 1 | ok_to_inline -> Just (inlineStmt u expr stmt : stmts)
141 -- we don't inline into CmmCall if the expression refers to global
142 -- registers. This is a HACK to avoid global registers clashing with
143 -- C argument-passing registers, really the back-end ought to be able
144 -- to handle it properly, but currently neither PprC nor the NCG can
145 -- do it. See also CgForeignCall:load_args_into_temps.
146 ok_to_inline = case stmt of
147 CmmCall{} -> hasNoGlobalRegs expr
150 inlineStmt :: Unique -> CmmExpr -> CmmStmt -> CmmStmt
151 inlineStmt u a (CmmAssign r e) = CmmAssign r (inlineExpr u a e)
152 inlineStmt u a (CmmStore e1 e2) = CmmStore (inlineExpr u a e1) (inlineExpr u a e2)
153 inlineStmt u a (CmmCall target regs es srt ret)
154 = CmmCall (infn target) regs es' srt ret
155 where infn (CmmCallee fn cconv) = CmmCallee fn cconv
156 infn (CmmPrim p) = CmmPrim p
157 es' = [ (CmmHinted (inlineExpr u a e) hint) | (CmmHinted e hint) <- es ]
158 inlineStmt u a (CmmCondBranch e d) = CmmCondBranch (inlineExpr u a e) d
159 inlineStmt u a (CmmSwitch e d) = CmmSwitch (inlineExpr u a e) d
160 inlineStmt u a (CmmJump e d) = CmmJump (inlineExpr u a e) d
161 inlineStmt u a other_stmt = other_stmt
163 inlineExpr :: Unique -> CmmExpr -> CmmExpr -> CmmExpr
164 inlineExpr u a e@(CmmReg (CmmLocal (LocalReg u' _)))
167 inlineExpr u a e@(CmmRegOff (CmmLocal (LocalReg u' rep)) off)
168 | u == u' = CmmMachOp (MO_Add width) [a, CmmLit (CmmInt (fromIntegral off) width)]
171 width = typeWidth rep
172 inlineExpr u a (CmmLoad e rep) = CmmLoad (inlineExpr u a e) rep
173 inlineExpr u a (CmmMachOp op es) = CmmMachOp op (map (inlineExpr u a) es)
174 inlineExpr u a other_expr = other_expr
176 -- -----------------------------------------------------------------------------
177 -- MachOp constant folder
179 -- Now, try to constant-fold the MachOps. The arguments have already
180 -- been optimized and folded.
183 :: MachOp -- The operation from an CmmMachOp
184 -> [CmmExpr] -- The optimized arguments
187 cmmMachOpFold op arg@[CmmLit (CmmInt x rep)]
189 MO_S_Neg r -> CmmLit (CmmInt (-x) rep)
190 MO_Not r -> CmmLit (CmmInt (complement x) rep)
192 -- these are interesting: we must first narrow to the
193 -- "from" type, in order to truncate to the correct size.
194 -- The final narrow/widen to the destination type
195 -- is implicit in the CmmLit.
196 MO_SF_Conv from to -> CmmLit (CmmFloat (fromInteger x) to)
197 MO_SS_Conv from to -> CmmLit (CmmInt (narrowS from x) to)
198 MO_UU_Conv from to -> CmmLit (CmmInt (narrowU from x) to)
200 _ -> panic "cmmMachOpFold: unknown unary op"
203 -- Eliminate conversion NOPs
204 cmmMachOpFold (MO_SS_Conv rep1 rep2) [x] | rep1 == rep2 = x
205 cmmMachOpFold (MO_UU_Conv rep1 rep2) [x] | rep1 == rep2 = x
207 -- Eliminate nested conversions where possible
208 cmmMachOpFold conv_outer args@[CmmMachOp conv_inner [x]]
209 | Just (rep1,rep2,signed1) <- isIntConversion conv_inner,
210 Just (_, rep3,signed2) <- isIntConversion conv_outer
212 -- widen then narrow to the same size is a nop
213 _ | rep1 < rep2 && rep1 == rep3 -> x
214 -- Widen then narrow to different size: collapse to single conversion
215 -- but remember to use the signedness from the widening, just in case
216 -- the final conversion is a widen.
217 | rep1 < rep2 && rep2 > rep3 ->
218 cmmMachOpFold (intconv signed1 rep1 rep3) [x]
219 -- Nested widenings: collapse if the signedness is the same
220 | rep1 < rep2 && rep2 < rep3 && signed1 == signed2 ->
221 cmmMachOpFold (intconv signed1 rep1 rep3) [x]
222 -- Nested narrowings: collapse
223 | rep1 > rep2 && rep2 > rep3 ->
224 cmmMachOpFold (MO_UU_Conv rep1 rep3) [x]
226 CmmMachOp conv_outer args
228 isIntConversion (MO_UU_Conv rep1 rep2)
229 = Just (rep1,rep2,False)
230 isIntConversion (MO_SS_Conv rep1 rep2)
231 = Just (rep1,rep2,True)
232 isIntConversion _ = Nothing
234 intconv True = MO_SS_Conv
235 intconv False = MO_UU_Conv
237 -- ToDo: a narrow of a load can be collapsed into a narrow load, right?
238 -- but what if the architecture only supports word-sized loads, should
239 -- we do the transformation anyway?
241 cmmMachOpFold mop args@[CmmLit (CmmInt x xrep), CmmLit (CmmInt y _)]
243 -- for comparisons: don't forget to narrow the arguments before
244 -- comparing, since they might be out of range.
245 MO_Eq r -> CmmLit (CmmInt (if x_u == y_u then 1 else 0) wordWidth)
246 MO_Ne r -> CmmLit (CmmInt (if x_u /= y_u then 1 else 0) wordWidth)
248 MO_U_Gt r -> CmmLit (CmmInt (if x_u > y_u then 1 else 0) wordWidth)
249 MO_U_Ge r -> CmmLit (CmmInt (if x_u >= y_u then 1 else 0) wordWidth)
250 MO_U_Lt r -> CmmLit (CmmInt (if x_u < y_u then 1 else 0) wordWidth)
251 MO_U_Le r -> CmmLit (CmmInt (if x_u <= y_u then 1 else 0) wordWidth)
253 MO_S_Gt r -> CmmLit (CmmInt (if x_s > y_s then 1 else 0) wordWidth)
254 MO_S_Ge r -> CmmLit (CmmInt (if x_s >= y_s then 1 else 0) wordWidth)
255 MO_S_Lt r -> CmmLit (CmmInt (if x_s < y_s then 1 else 0) wordWidth)
256 MO_S_Le r -> CmmLit (CmmInt (if x_s <= y_s then 1 else 0) wordWidth)
258 MO_Add r -> CmmLit (CmmInt (x + y) r)
259 MO_Sub r -> CmmLit (CmmInt (x - y) r)
260 MO_Mul r -> CmmLit (CmmInt (x * y) r)
261 MO_S_Quot r | y /= 0 -> CmmLit (CmmInt (x `quot` y) r)
262 MO_S_Rem r | y /= 0 -> CmmLit (CmmInt (x `rem` y) r)
264 MO_And r -> CmmLit (CmmInt (x .&. y) r)
265 MO_Or r -> CmmLit (CmmInt (x .|. y) r)
266 MO_Xor r -> CmmLit (CmmInt (x `xor` y) r)
268 MO_Shl r -> CmmLit (CmmInt (x `shiftL` fromIntegral y) r)
269 MO_U_Shr r -> CmmLit (CmmInt (x_u `shiftR` fromIntegral y) r)
270 MO_S_Shr r -> CmmLit (CmmInt (x `shiftR` fromIntegral y) r)
272 other -> CmmMachOp mop args
281 -- When possible, shift the constants to the right-hand side, so that we
282 -- can match for strength reductions. Note that the code generator will
283 -- also assume that constants have been shifted to the right when
286 cmmMachOpFold op [x@(CmmLit _), y]
287 | not (isLit y) && isCommutableMachOp op
288 = cmmMachOpFold op [y, x]
290 -- Turn (a+b)+c into a+(b+c) where possible. Because literals are
291 -- moved to the right, it is more likely that we will find
292 -- opportunities for constant folding when the expression is
295 -- ToDo: this appears to introduce a quadratic behaviour due to the
296 -- nested cmmMachOpFold. Can we fix this?
298 -- Why do we check isLit arg1? If arg1 is a lit, it means that arg2
299 -- is also a lit (otherwise arg1 would be on the right). If we
300 -- put arg1 on the left of the rearranged expression, we'll get into a
301 -- loop: (x1+x2)+x3 => x1+(x2+x3) => (x2+x3)+x1 => x2+(x3+x1) ...
303 -- Also don't do it if arg1 is PicBaseReg, so that we don't separate the
304 -- PicBaseReg from the corresponding label (or label difference).
306 cmmMachOpFold mop1 [CmmMachOp mop2 [arg1,arg2], arg3]
307 | mop1 == mop2 && isAssociativeMachOp mop1
308 && not (isLit arg1) && not (isPicReg arg1)
309 = cmmMachOpFold mop1 [arg1, cmmMachOpFold mop2 [arg2,arg3]]
311 -- Make a RegOff if we can
312 cmmMachOpFold (MO_Add _) [CmmReg reg, CmmLit (CmmInt n rep)]
313 = CmmRegOff reg (fromIntegral (narrowS rep n))
314 cmmMachOpFold (MO_Add _) [CmmRegOff reg off, CmmLit (CmmInt n rep)]
315 = CmmRegOff reg (off + fromIntegral (narrowS rep n))
316 cmmMachOpFold (MO_Sub _) [CmmReg reg, CmmLit (CmmInt n rep)]
317 = CmmRegOff reg (- fromIntegral (narrowS rep n))
318 cmmMachOpFold (MO_Sub _) [CmmRegOff reg off, CmmLit (CmmInt n rep)]
319 = CmmRegOff reg (off - fromIntegral (narrowS rep n))
321 -- Fold label(+/-)offset into a CmmLit where possible
323 cmmMachOpFold (MO_Add _) [CmmLit (CmmLabel lbl), CmmLit (CmmInt i rep)]
324 = CmmLit (CmmLabelOff lbl (fromIntegral (narrowU rep i)))
325 cmmMachOpFold (MO_Add _) [CmmLit (CmmInt i rep), CmmLit (CmmLabel lbl)]
326 = CmmLit (CmmLabelOff lbl (fromIntegral (narrowU rep i)))
327 cmmMachOpFold (MO_Sub _) [CmmLit (CmmLabel lbl), CmmLit (CmmInt i rep)]
328 = CmmLit (CmmLabelOff lbl (fromIntegral (negate (narrowU rep i))))
331 -- Comparison of literal with widened operand: perform the comparison
332 -- at the smaller width, as long as the literal is within range.
334 -- We can't do the reverse trick, when the operand is narrowed:
335 -- narrowing throws away bits from the operand, there's no way to do
336 -- the same comparison at the larger size.
338 #if i386_TARGET_ARCH || x86_64_TARGET_ARCH
339 -- powerPC NCG has a TODO for I8/I16 comparisons, so don't try
341 cmmMachOpFold cmp [CmmMachOp conv [x], CmmLit (CmmInt i _)]
342 | -- if the operand is widened:
343 Just (rep, signed, narrow_fn) <- maybe_conversion conv,
344 -- and this is a comparison operation:
345 Just narrow_cmp <- maybe_comparison cmp rep signed,
346 -- and the literal fits in the smaller size:
348 -- then we can do the comparison at the smaller size
349 = cmmMachOpFold narrow_cmp [x, CmmLit (CmmInt i rep)]
351 maybe_conversion (MO_UU_Conv from to)
353 = Just (from, False, narrowU)
354 maybe_conversion (MO_SS_Conv from to)
356 = Just (from, True, narrowS)
358 -- don't attempt to apply this optimisation when the source
359 -- is a float; see #1916
360 maybe_conversion _ = Nothing
362 -- careful (#2080): if the original comparison was signed, but
363 -- we were doing an unsigned widen, then we must do an
364 -- unsigned comparison at the smaller size.
365 maybe_comparison (MO_U_Gt _) rep _ = Just (MO_U_Gt rep)
366 maybe_comparison (MO_U_Ge _) rep _ = Just (MO_U_Ge rep)
367 maybe_comparison (MO_U_Lt _) rep _ = Just (MO_U_Lt rep)
368 maybe_comparison (MO_U_Le _) rep _ = Just (MO_U_Le rep)
369 maybe_comparison (MO_Eq _) rep _ = Just (MO_Eq rep)
370 maybe_comparison (MO_S_Gt _) rep True = Just (MO_S_Gt rep)
371 maybe_comparison (MO_S_Ge _) rep True = Just (MO_S_Ge rep)
372 maybe_comparison (MO_S_Lt _) rep True = Just (MO_S_Lt rep)
373 maybe_comparison (MO_S_Le _) rep True = Just (MO_S_Le rep)
374 maybe_comparison (MO_S_Gt _) rep False = Just (MO_U_Gt rep)
375 maybe_comparison (MO_S_Ge _) rep False = Just (MO_U_Ge rep)
376 maybe_comparison (MO_S_Lt _) rep False = Just (MO_U_Lt rep)
377 maybe_comparison (MO_S_Le _) rep False = Just (MO_U_Le rep)
378 maybe_comparison _ _ _ = Nothing
382 -- We can often do something with constants of 0 and 1 ...
384 cmmMachOpFold mop args@[x, y@(CmmLit (CmmInt 0 _))]
395 MO_Ne r | isComparisonExpr x -> x
396 MO_Eq r | Just x' <- maybeInvertCmmExpr x -> x'
397 MO_U_Gt r | isComparisonExpr x -> x
398 MO_S_Gt r | isComparisonExpr x -> x
399 MO_U_Lt r | isComparisonExpr x -> CmmLit (CmmInt 0 wordWidth)
400 MO_S_Lt r | isComparisonExpr x -> CmmLit (CmmInt 0 wordWidth)
401 MO_U_Ge r | isComparisonExpr x -> CmmLit (CmmInt 1 wordWidth)
402 MO_S_Ge r | isComparisonExpr x -> CmmLit (CmmInt 1 wordWidth)
403 MO_U_Le r | Just x' <- maybeInvertCmmExpr x -> x'
404 MO_S_Le r | Just x' <- maybeInvertCmmExpr x -> x'
405 other -> CmmMachOp mop args
407 cmmMachOpFold mop args@[x, y@(CmmLit (CmmInt 1 rep))]
412 MO_S_Rem r -> CmmLit (CmmInt 0 rep)
413 MO_U_Rem r -> CmmLit (CmmInt 0 rep)
414 MO_Ne r | Just x' <- maybeInvertCmmExpr x -> x'
415 MO_Eq r | isComparisonExpr x -> x
416 MO_U_Lt r | Just x' <- maybeInvertCmmExpr x -> x'
417 MO_S_Lt r | Just x' <- maybeInvertCmmExpr x -> x'
418 MO_U_Gt r | isComparisonExpr x -> CmmLit (CmmInt 0 wordWidth)
419 MO_S_Gt r | isComparisonExpr x -> CmmLit (CmmInt 0 wordWidth)
420 MO_U_Le r | isComparisonExpr x -> CmmLit (CmmInt 1 wordWidth)
421 MO_S_Le r | isComparisonExpr x -> CmmLit (CmmInt 1 wordWidth)
422 MO_U_Ge r | isComparisonExpr x -> x
423 MO_S_Ge r | isComparisonExpr x -> x
424 other -> CmmMachOp mop args
426 -- Now look for multiplication/division by powers of 2 (integers).
428 cmmMachOpFold mop args@[x, y@(CmmLit (CmmInt n _))]
431 | Just p <- exactLog2 n ->
432 CmmMachOp (MO_Shl rep) [x, CmmLit (CmmInt p rep)]
434 | Just p <- exactLog2 n,
435 CmmReg _ <- x -> -- We duplicate x below, hence require
436 -- it is a reg. FIXME: remove this restriction.
437 -- shift right is not the same as quot, because it rounds
438 -- to minus infinity, whereasq uot rounds toward zero.
439 -- To fix this up, we add one less than the divisor to the
440 -- dividend if it is a negative number.
442 -- to avoid a test/jump, we use the following sequence:
443 -- x1 = x >> word_size-1 (all 1s if -ve, all 0s if +ve)
444 -- x2 = y & (divisor-1)
445 -- result = (x+x2) >>= log2(divisor)
446 -- this could be done a bit more simply using conditional moves,
447 -- but we're processor independent here.
449 -- we optimise the divide by 2 case slightly, generating
450 -- x1 = x >> word_size-1 (unsigned)
451 -- return = (x + x1) >>= log2(divisor)
453 bits = fromIntegral (widthInBits rep) - 1
454 shr = if p == 1 then MO_U_Shr rep else MO_S_Shr rep
455 x1 = CmmMachOp shr [x, CmmLit (CmmInt bits rep)]
456 x2 = if p == 1 then x1 else
457 CmmMachOp (MO_And rep) [x1, CmmLit (CmmInt (n-1) rep)]
458 x3 = CmmMachOp (MO_Add rep) [x, x2]
460 CmmMachOp (MO_S_Shr rep) [x3, CmmLit (CmmInt p rep)]
464 unchanged = CmmMachOp mop args
466 -- Anything else is just too hard.
468 cmmMachOpFold mop args = CmmMachOp mop args
470 -- -----------------------------------------------------------------------------
473 -- This algorithm for determining the $\log_2$ of exact powers of 2 comes
474 -- from GCC. It requires bit manipulation primitives, and we use GHC
475 -- extensions. Tough.
477 -- Used to be in MachInstrs --SDM.
478 -- ToDo: remove use of unboxery --SDM.
480 -- Unboxery removed in favor of FastInt; but is the function supposed to fail
481 -- on inputs >= 2147483648, or was that just an implementation artifact?
482 -- And is this speed-critical, or can we just use Integer operations
483 -- (including Data.Bits)?
486 exactLog2 :: Integer -> Maybe Integer
488 = if (x_ <= 0 || x_ >= 2147483648) then
491 case iUnbox (fromInteger x_) of { x ->
492 if (x `bitAndFastInt` negateFastInt x) /=# x then
495 Just (toInteger (iBox (pow2 x)))
498 pow2 x | x ==# _ILIT(1) = _ILIT(0)
499 | otherwise = _ILIT(1) +# pow2 (x `shiftR_FastInt` _ILIT(1))
502 -- -----------------------------------------------------------------------------
506 This is a simple pass that replaces tail-recursive functions like this:
521 the latter generates better C code, because the C compiler treats it
522 like a loop, and brings full loop optimisation to bear.
524 In my measurements this makes little or no difference to anything
525 except factorial, but what the hell.
528 cmmLoopifyForC :: RawCmmTop -> RawCmmTop
529 cmmLoopifyForC p@(CmmProc info entry_lbl []
530 (ListGraph blocks@(BasicBlock top_id _ : _)))
531 | null info = p -- only if there's an info table, ignore case alts
533 -- pprTrace "jump_lbl" (ppr jump_lbl <+> ppr entry_lbl) $
534 CmmProc info entry_lbl [] (ListGraph blocks')
535 where blocks' = [ BasicBlock id (map do_stmt stmts)
536 | BasicBlock id stmts <- blocks ]
538 do_stmt (CmmJump (CmmLit (CmmLabel lbl)) _) | lbl == jump_lbl
542 jump_lbl | tablesNextToCode = entryLblToInfoLbl entry_lbl
543 | otherwise = entry_lbl
545 cmmLoopifyForC top = top
547 -- -----------------------------------------------------------------------------
550 isLit (CmmLit _) = True
553 isComparisonExpr :: CmmExpr -> Bool
554 isComparisonExpr (CmmMachOp op _) = isComparisonMachOp op
555 isComparisonExpr _other = False
557 isPicReg (CmmReg (CmmGlobal PicBaseReg)) = True