diff options
Diffstat (limited to 'compiler/cmm/CmmOpt.hs')
| -rw-r--r-- | compiler/cmm/CmmOpt.hs | 80 |
1 files changed, 42 insertions, 38 deletions
diff --git a/compiler/cmm/CmmOpt.hs b/compiler/cmm/CmmOpt.hs index 3cb28217f2..e837d29783 100644 --- a/compiler/cmm/CmmOpt.hs +++ b/compiler/cmm/CmmOpt.hs @@ -1,10 +1,6 @@ -{-# LANGUAGE CPP #-} - -- The default iteration limit is a bit too low for the definitions -- in this module. -#if __GLASGOW_HASKELL__ >= 800 {-# OPTIONS_GHC -fmax-pmcheck-iterations=10000000 #-} -#endif ----------------------------------------------------------------------------- -- @@ -21,7 +17,7 @@ module CmmOpt ( cmmMachOpFoldM ) where -#include "HsVersions.h" +import GhcPrelude import CmmUtils import Cmm @@ -357,35 +353,51 @@ cmmMachOpFoldM dflags mop [x, (CmmLit (CmmInt n _))] MO_U_Quot rep | Just p <- exactLog2 n -> Just (cmmMachOpFold dflags (MO_U_Shr rep) [x, CmmLit (CmmInt p rep)]) + MO_U_Rem rep + | Just _ <- exactLog2 n -> + Just (cmmMachOpFold dflags (MO_And rep) [x, CmmLit (CmmInt (n - 1) rep)]) MO_S_Quot rep | Just p <- exactLog2 n, - CmmReg _ <- x -> -- We duplicate x below, hence require + CmmReg _ <- x -> -- We duplicate x in signedQuotRemHelper, hence require + -- it is a reg. FIXME: remove this restriction. + Just (cmmMachOpFold dflags (MO_S_Shr rep) + [signedQuotRemHelper rep p, CmmLit (CmmInt p rep)]) + MO_S_Rem rep + | Just p <- exactLog2 n, + CmmReg _ <- x -> -- We duplicate x in signedQuotRemHelper, 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 quot 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 (widthInBits 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 - Just (cmmMachOpFold dflags (MO_S_Shr rep) [x3, CmmLit (CmmInt p rep)]) + -- We replace (x `rem` 2^p) by (x - (x `quot` 2^p) * 2^p). + -- Moreover, we fuse MO_S_Shr (last operation of MO_S_Quot) + -- and MO_S_Shl (multiplication by 2^p) into a single MO_And operation. + Just (cmmMachOpFold dflags (MO_Sub rep) + [x, cmmMachOpFold dflags (MO_And rep) + [signedQuotRemHelper rep p, CmmLit (CmmInt (- n) rep)]]) _ -> Nothing + where + -- In contrast with unsigned integers, for signed ones + -- shift right is not the same as quot, because it rounds + -- to minus infinity, whereas quot 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 + -- 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 + signedQuotRemHelper :: Width -> Integer -> CmmExpr + signedQuotRemHelper rep p = CmmMachOp (MO_Add rep) [x, x2] + where + bits = fromIntegral (widthInBits 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)] -- ToDo (#7116): optimise floating-point multiplication, e.g. x*2.0 -> x+x -- Unfortunately this needs a unique supply because x might not be a @@ -410,14 +422,6 @@ That's what the constant-folding operations on comparison operators do above. -- ----------------------------------------------------------------------------- -- Utils -isLit :: CmmExpr -> Bool -isLit (CmmLit _) = True -isLit _ = False - -isComparisonExpr :: CmmExpr -> Bool -isComparisonExpr (CmmMachOp op _) = isComparisonMachOp op -isComparisonExpr _ = False - isPicReg :: CmmExpr -> Bool isPicReg (CmmReg (CmmGlobal PicBaseReg)) = True isPicReg _ = False |