+ CmmReg (CmmGlobal mid)
+ -- Replace register leaves with appropriate StixTrees for
+ -- the given target. MagicIds which map to a reg on this
+ -- arch are left unchanged. For the rest, BaseReg is taken
+ -- to mean the address of the reg table in MainCapability,
+ -- and for all others we generate an indirection to its
+ -- location in the register table.
+ -> case get_GlobalReg_reg_or_addr mid of
+ Left realreg -> return expr
+ Right baseRegAddr
+ -> case mid of
+ BaseReg -> cmmExprConFold False baseRegAddr
+ other -> cmmExprConFold False (CmmLoad baseRegAddr
+ (globalRegRep mid))
+ -- eliminate zero offsets
+ CmmRegOff reg 0
+ -> cmmExprConFold False (CmmReg reg)
+
+ CmmRegOff (CmmGlobal mid) offset
+ -- RegOf leaves are just a shorthand form. If the reg maps
+ -- to a real reg, we keep the shorthand, otherwise, we just
+ -- expand it and defer to the above code.
+ -> case get_GlobalReg_reg_or_addr mid of
+ Left realreg -> return expr
+ Right baseRegAddr
+ -> cmmExprConFold False (CmmMachOp (MO_Add wordRep) [
+ CmmReg (CmmGlobal mid),
+ CmmLit (CmmInt (fromIntegral offset)
+ wordRep)])
+ other
+ -> return other
+
+
+-- -----------------------------------------------------------------------------
+-- 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 -> 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
+
+-- ToDo: eliminate multiple conversions. Be careful though: can't remove
+-- a narrowing, and can't remove conversions to/from floating point types.
+
+-- ToDo: eliminate nested comparisons:
+-- CmmMachOp MO_Lt [CmmMachOp MO_Eq [x,y], CmmLit (CmmInt 0 _)]
+-- turns into a simple equality test.
+
+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) ...
+--
+cmmMachOpFold mop1 [CmmMachOp mop2 [arg1,arg2], arg3]
+ | mop1 == mop2 && isAssociativeMachOp mop1 && not (isLit arg1)
+ = 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))))
+
+-- 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
+ MO_Eq r | Just x' <- maybeInvertConditionalExpr x -> x'
+ 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)
+ MO_U_Le r | Just x' <- maybeInvertConditionalExpr x -> x'
+ MO_S_Le r | Just x' <- maybeInvertConditionalExpr x -> x'
+ 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)
+ MO_Ne r | Just x' <- maybeInvertConditionalExpr x -> x'
+ MO_Eq r | isComparisonExpr x -> x
+ MO_U_Lt r | Just x' <- maybeInvertConditionalExpr x -> x'
+ MO_S_Lt r | Just x' <- maybeInvertConditionalExpr x -> x'
+ 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
+ -> case exactLog2 n of
+ Nothing -> unchanged
+ Just p -> CmmMachOp (MO_Shl rep) [x, CmmLit (CmmInt p rep)]
+ MO_S_Quot rep
+ -> case exactLog2 n of
+ Nothing -> unchanged
+ Just p -> CmmMachOp (MO_S_Shr rep) [x, CmmLit (CmmInt p rep)]
+ other
+ -> 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.