----------------------------------------------------------------------------- -- | -- Module : Data.SBV.Examples.BitPrecise.Legato -- Copyright : (c) Levent Erkok -- License : BSD3 -- Maintainer : erkokl@gmail.com -- Stability : experimental -- -- An encoding and correctness proof of Legato's multiplier in Haskell. Bill Legato came -- up with an interesting way to multiply two 8-bit numbers on Mostek, as described here: -- <http://www.cs.utexas.edu/~moore/acl2/workshop-2004/contrib/legato/Weakest-Preconditions-Report.pdf> -- -- Here's Legato's algorithm, as coded in Mostek assembly: -- -- @ -- step1 : LDX #8 ; load X immediate with the integer 8 -- step2 : LDA #0 ; load A immediate with the integer 0 -- step3 : LOOP ROR F1 ; rotate F1 right circular through C -- step4 : BCC ZCOEF ; branch to ZCOEF if C = 0 -- step5 : CLC ; set C to 0 -- step6 : ADC F2 ; set A to A+F2+C and C to the carry -- step7 : ZCOEF ROR A ; rotate A right circular through C -- step8 : ROR LOW ; rotate LOW right circular through C -- step9 : DEX ; set X to X-1 -- step10: BNE LOOP ; branch to LOOP if Z = 0 -- @ -- -- This program came to be known as the Legato's challenge in the community, where -- the challenge was to prove that it indeed does perform multiplication. This file -- formalizes the Mostek architecture in Haskell and proves that Legato's algorithm -- is indeed correct. ----------------------------------------------------------------------------- {-# LANGUAGE DeriveGeneric #-} {-# LANGUAGE DeriveAnyClass #-} module Data.SBV.Examples.BitPrecise.Legato where import Data.Array (Array, Ix(..), (!), (//), array) import Data.SBV import Data.SBV.Tools.CodeGen import Data.SBV.Internals (Timing(PrintTiming)) import GHC.Generics (Generic) ------------------------------------------------------------------ -- * Mostek architecture ------------------------------------------------------------------ -- | The memory is addressed by 32-bit words. type Address = SWord32 -- | We model only two registers of Mostek that is used in the above algorithm, can add more. data Register = RegX | RegA deriving (Eq, Ord, Ix, Bounded, Enum) -- | The carry flag ('FlagC') and the zero flag ('FlagZ') data Flag = FlagC | FlagZ deriving (Eq, Ord, Ix, Bounded, Enum) -- | Mostek was an 8-bit machine. type Value = SWord8 -- | Convenient synonym for symbolic machine bits. type Bit = SBool -- | Register bank type Registers = Array Register Value -- | Flag bank type Flags = Array Flag Bit -- | The memory maps 32-bit words to 8-bit words. (The 'Model' data-type is -- defined later, depending on the verification model used.) type Memory = Model Word32 Word8 -- Model defined later -- | Abstraction of the machine: The CPU consists of memory, registers, and flags. -- Unlike traditional hardware, we assume the program is stored in some other memory area that -- we need not model. (No self modifying programs!) -- -- 'Mostek' is equipped with an automatically derived 'Mergeable' instance -- because each field is 'Mergeable'. data Mostek = Mostek { memory :: Memory , registers :: Registers , flags :: Flags } deriving (Generic, Mergeable) -- | Given a machine state, compute a value out of it type Extract a = Mostek -> a -- | Programs are essentially state transformers (on the machine state) type Program = Mostek -> Mostek ------------------------------------------------------------------ -- * Low-level operations ------------------------------------------------------------------ -- | Get the value of a given register getReg :: Register -> Extract Value getReg r m = registers m ! r -- | Set the value of a given register setReg :: Register -> Value -> Program setReg r v m = m {registers = registers m // [(r, v)]} -- | Get the value of a flag getFlag :: Flag -> Extract Bit getFlag f m = flags m ! f -- | Set the value of a flag setFlag :: Flag -> Bit -> Program setFlag f b m = m {flags = flags m // [(f, b)]} -- | Read memory peek :: Address -> Extract Value peek a m = readArray (memory m) a -- | Write to memory poke :: Address -> Value -> Program poke a v m = m {memory = writeArray (memory m) a v} -- | Checking overflow. In Legato's multipler the @ADC@ instruction -- needs to see if the expression x + y + c overflowed, as checked -- by this function. Note that we verify the correctness of this check -- separately below in `checkOverflowCorrect`. checkOverflow :: SWord8 -> SWord8 -> SBool -> SBool checkOverflow x y c = s .< x ||| s .< y ||| s' .< s where s = x + y s' = s + ite c 1 0 -- | Correctness theorem for our `checkOverflow` implementation. -- -- We have: -- -- >>> checkOverflowCorrect -- Q.E.D. checkOverflowCorrect :: IO ThmResult checkOverflowCorrect = checkOverflow === overflow where -- Reference spec for overflow. We do the addition -- using 16 bits and check that it's larger than 255 overflow :: SWord8 -> SWord8 -> SBool -> SBool overflow x y c = (0 # x) + (0 # y) + ite c 1 0 .> 255 ------------------------------------------------------------------ -- * Instruction set ------------------------------------------------------------------ -- | An instruction is modeled as a 'Program' transformer. We model -- mostek programs in direct continuation passing style. type Instruction = Program -> Program -- | LDX: Set register @X@ to value @v@ ldx :: Value -> Instruction ldx v k = k . setReg RegX v -- | LDA: Set register @A@ to value @v@ lda :: Value -> Instruction lda v k = k . setReg RegA v -- | CLC: Clear the carry flag clc :: Instruction clc k = k . setFlag FlagC false -- | ROR, memory version: Rotate the value at memory location @a@ -- to the right by 1 bit, using the carry flag as a transfer position. -- That is, the final bit of the memory location becomes the new carry -- and the carry moves over to the first bit. This very instruction -- is one of the reasons why Legato's multiplier is quite hard to understand -- and is typically presented as a verification challenge. rorM :: Address -> Instruction rorM a k m = k . setFlag FlagC c' . poke a v' $ m where v = peek a m c = getFlag FlagC m v' = setBitTo (v `rotateR` 1) 7 c c' = sTestBit v 0 -- | ROR, register version: Same as 'rorM', except through register @r@. rorR :: Register -> Instruction rorR r k m = k . setFlag FlagC c' . setReg r v' $ m where v = getReg r m c = getFlag FlagC m v' = setBitTo (v `rotateR` 1) 7 c c' = sTestBit v 0 -- | BCC: branch to label @l@ if the carry flag is false bcc :: Program -> Instruction bcc l k m = ite (c .== false) (l m) (k m) where c = getFlag FlagC m -- | ADC: Increment the value of register @A@ by the value of memory contents -- at address @a@, using the carry-bit as the carry-in for the addition. adc :: Address -> Instruction adc a k m = k . setFlag FlagZ (v' .== 0) . setFlag FlagC c' . setReg RegA v' $ m where v = peek a m ra = getReg RegA m c = getFlag FlagC m v' = v + ra + ite c 1 0 c' = checkOverflow v ra c -- | DEX: Decrement the value of register @X@ dex :: Instruction dex k m = k . setFlag FlagZ (x .== 0) . setReg RegX x $ m where x = getReg RegX m - 1 -- | BNE: Branch if the zero-flag is false bne :: Program -> Instruction bne l k m = ite (z .== false) (l m) (k m) where z = getFlag FlagZ m -- | The 'end' combinator "stops" our program, providing the final continuation -- that does nothing. end :: Program end = id ------------------------------------------------------------------ -- * Legato's algorithm in Haskell/SBV ------------------------------------------------------------------ -- | Parameterized by the addresses of locations of the factors (@F1@ and @F2@), -- the following program multiplies them, storing the low-byte of the result -- in the memory location @lowAddr@, and the high-byte in register @A@. The -- implementation is a direct transliteration of Legato's algorithm given -- at the top, using our notation. legato :: Address -> Address -> Address -> Program legato f1Addr f2Addr lowAddr = start where start = ldx 8 $ lda 0 $ loop loop = rorM f1Addr $ bcc zeroCoef $ clc $ adc f2Addr $ zeroCoef zeroCoef = rorR RegA $ rorM lowAddr $ dex $ bne loop $ end ------------------------------------------------------------------ -- * Verification interface ------------------------------------------------------------------ -- | Given address/value pairs for F1 and F2, and the location of where the low-byte -- of the result should go, @runLegato@ takes an arbitrary machine state @m@ and -- returns the high and low bytes of the multiplication. runLegato :: (Address, Value) -> (Address, Value) -> Address -> Mostek -> (Value, Value) runLegato (f1Addr, f1Val) (f2Addr, f2Val) loAddr m = (getReg RegA mFinal, peek loAddr mFinal) where m0 = poke f1Addr f1Val $ poke f2Addr f2Val m mFinal = legato f1Addr f2Addr loAddr m0 -- | Helper synonym for capturing relevant bits of Mostek type InitVals = ( Value -- Content of Register X , Value -- Content of Register A , Bit -- Value of FlagC , Bit -- Value of FlagZ ) -- | Create an instance of the Mostek machine, initialized by the memory and the relevant -- values of the registers and the flags initMachine :: Memory -> InitVals -> Mostek initMachine mem (rx, ra, fc, fz) = Mostek { memory = mem , registers = array (minBound, maxBound) [(RegX, rx), (RegA, ra)] , flags = array (minBound, maxBound) [(FlagC, fc), (FlagZ, fz)] } -- | The correctness theorem. For all possible memory configurations, the factors (@x@ and @y@ below), the location -- of the low-byte result and the initial-values of registers and the flags, this function will return True only if -- running Legato's algorithm does indeed compute the product of @x@ and @y@ correctly. legatoIsCorrect :: Memory -> (Address, Value) -> (Address, Value) -> Address -> InitVals -> SBool legatoIsCorrect mem (addrX, x) (addrY, y) addrLow initVals = distinct [addrX, addrY, addrLow] -- note the conditional: addresses must be distinct! ==> result .== expected where (hi, lo) = runLegato (addrX, x) (addrY, y) addrLow (initMachine mem initVals) -- NB. perform the comparison over 16 bit values to avoid overflow! -- If Value changes to be something else, modify this accordingly. result, expected :: SWord16 result = 256 * (0 # hi) + (0 # lo) expected = (0 # x) * (0 # y) ------------------------------------------------------------------ -- * Verification ------------------------------------------------------------------ -- | Choose the appropriate array model to be used for modeling the memory. (See 'Memory'.) -- The 'SFunArray' is the function based model. 'SArray' is the SMT-Lib array's based model. type Model = SFunArray -- type Model = SArray -- | The correctness theorem. -- On a 2011 MacBook, this proof takes about 1 minute 45 seconds with the 'SFunArray' memory model -- using boolector as the solver. correctnessTheorem :: IO ThmResult correctnessTheorem = proveWith boolector{timing = PrintTiming} $ do let mem = mkSFunArray (const 0) addrX <- sWord32 "addrX" x <- sWord8 "x" addrY <- sWord32 "addrY" y <- sWord8 "y" addrLow <- sWord32 "addrLow" regX <- sWord8 "regX" regA <- sWord8 "regA" flagC <- sBool "flagC" flagZ <- sBool "flagZ" return $ legatoIsCorrect mem (addrX, x) (addrY, y) addrLow (regX, regA, flagC, flagZ) ------------------------------------------------------------------ -- * C Code generation ------------------------------------------------------------------ -- | Generate a C program that implements Legato's algorithm automatically. legatoInC :: IO () legatoInC = compileToC Nothing "runLegato" $ do x <- cgInput "x" y <- cgInput "y" let (hi, lo) = runLegato (0, x) (1, y) 2 (initMachine (mkSFunArray (const 0)) (0, 0, false, false)) cgOutput "hi" hi cgOutput "lo" lo {-# ANN legato ("HLint: ignore Redundant $" :: String) #-} {-# ANN module ("HLint: ignore Reduce duplication" :: String) #-}