GHC.CmmToAsm.CFG

Documentation

type CFG = EdgeInfoMap EdgeInfo

A control flow graph where edges have been annotated with a weight. Implemented as IntMap (IntMap <edgeData>) We must uphold the invariant that for each edge A -> B we must have: A entry B in the outer map. A entry B in the map we get when looking up A. Maintaining this invariant is useful as any failed lookup now indicates an actual error in code which might go unnoticed for a while otherwise.

data CfgEdge

Constructors

CfgEdge

Fields

• edgeFrom :: !BlockId
• edgeTo :: !BlockId
• edgeInfo :: !EdgeInfo

Instances

Eq CfgEdge

Careful! Since we assume there is at most one edge from A to B the Eq instance does not consider weight.

Defined in GHC.CmmToAsm.CFG

Methods

(==) :: CfgEdge -> CfgEdge -> Bool

(/=) :: CfgEdge -> CfgEdge -> Bool

Ord CfgEdge

Edges are sorted ascending pointwise by weight, source and destination

Defined in GHC.CmmToAsm.CFG

Methods

compare :: CfgEdge -> CfgEdge -> Ordering

(<) :: CfgEdge -> CfgEdge -> Bool

(<=) :: CfgEdge -> CfgEdge -> Bool

(>) :: CfgEdge -> CfgEdge -> Bool

(>=) :: CfgEdge -> CfgEdge -> Bool

max :: CfgEdge -> CfgEdge -> CfgEdge

min :: CfgEdge -> CfgEdge -> CfgEdge

Outputable CfgEdge

Defined in GHC.CmmToAsm.CFG

Methods

ppr :: CfgEdge -> SDoc

data EdgeInfo

Information about edges

Constructors

EdgeInfo

Fields

• transitionSource :: !TransitionSource
• edgeWeight :: !EdgeWeight

Instances

Eq EdgeInfo

Defined in GHC.CmmToAsm.CFG

Methods

(==) :: EdgeInfo -> EdgeInfo -> Bool

(/=) :: EdgeInfo -> EdgeInfo -> Bool

Outputable EdgeInfo

Defined in GHC.CmmToAsm.CFG

Methods

ppr :: EdgeInfo -> SDoc

newtype EdgeWeight

Constructors

EdgeWeight

Fields

• weightToDouble :: Double

Instances

Enum EdgeWeight

Defined in GHC.CmmToAsm.CFG

Methods

succ :: EdgeWeight -> EdgeWeight

pred :: EdgeWeight -> EdgeWeight

toEnum :: Int -> EdgeWeight

fromEnum :: EdgeWeight -> Int

enumFrom :: EdgeWeight -> [EdgeWeight]

enumFromThen :: EdgeWeight -> EdgeWeight -> [EdgeWeight]

enumFromTo :: EdgeWeight -> EdgeWeight -> [EdgeWeight]

enumFromThenTo :: EdgeWeight -> EdgeWeight -> EdgeWeight -> [EdgeWeight]

Eq EdgeWeight

Defined in GHC.CmmToAsm.CFG

Methods

(==) :: EdgeWeight -> EdgeWeight -> Bool

(/=) :: EdgeWeight -> EdgeWeight -> Bool

Fractional EdgeWeight

Defined in GHC.CmmToAsm.CFG

Methods

(/) :: EdgeWeight -> EdgeWeight -> EdgeWeight

recip :: EdgeWeight -> EdgeWeight

fromRational :: Rational -> EdgeWeight

Num EdgeWeight

Defined in GHC.CmmToAsm.CFG

Methods

(+) :: EdgeWeight -> EdgeWeight -> EdgeWeight

(-) :: EdgeWeight -> EdgeWeight -> EdgeWeight

(*) :: EdgeWeight -> EdgeWeight -> EdgeWeight

negate :: EdgeWeight -> EdgeWeight

abs :: EdgeWeight -> EdgeWeight

signum :: EdgeWeight -> EdgeWeight

fromInteger :: Integer -> EdgeWeight

Ord EdgeWeight

Defined in GHC.CmmToAsm.CFG

Methods

compare :: EdgeWeight -> EdgeWeight -> Ordering

(<) :: EdgeWeight -> EdgeWeight -> Bool

(<=) :: EdgeWeight -> EdgeWeight -> Bool

(>) :: EdgeWeight -> EdgeWeight -> Bool

(>=) :: EdgeWeight -> EdgeWeight -> Bool

max :: EdgeWeight -> EdgeWeight -> EdgeWeight

min :: EdgeWeight -> EdgeWeight -> EdgeWeight

Real EdgeWeight

Defined in GHC.CmmToAsm.CFG

Methods

toRational :: EdgeWeight -> Rational

Outputable EdgeWeight

Defined in GHC.CmmToAsm.CFG

Methods

ppr :: EdgeWeight -> SDoc

data TransitionSource

Can we trace back a edge to a specific Cmm Node or has it been introduced during assembly codegen. We use this to maintain some information which would otherwise be lost during the Cmm <-> asm transition. See also Note [Inverting Conditional Branches]

Constructors

CmmSource

Fields

• trans_cmmNode :: CmmNode O C
• trans_info :: BranchInfo

AsmCodeGen

Instances

Eq TransitionSource

Defined in GHC.CmmToAsm.CFG

Methods

(==) :: TransitionSource -> TransitionSource -> Bool

(/=) :: TransitionSource -> TransitionSource -> Bool

addWeightEdge :: BlockId -> BlockId -> EdgeWeight -> CFG -> CFG

Adds a edge with the given weight to the cfg If there already existed an edge it is overwritten. `addWeightEdge from to weight cfg`

addEdge :: BlockId -> BlockId -> EdgeInfo -> CFG -> CFG

Adds a new edge, overwrites existing edges if present

delEdge :: BlockId -> BlockId -> CFG -> CFG

delNode :: BlockId -> CFG -> CFG

addNodesBetween :: Weights -> CFG -> [(BlockId, BlockId, BlockId)] -> CFG

Insert a block in the control flow between two other blocks. We pass a list of tuples (A,B,C) where * A -> C: Old edge * A -> B -> C : New Arc, where B is the new block. It's possible that a block has two jumps to the same block in the assembly code. However we still only store a single edge for these cases. We assign the old edge info to the edge A -> B and assign B -> C the weight of an unconditional jump.

shortcutWeightMap :: LabelMap (Maybe BlockId) -> CFG -> CFG

reverseEdges :: CFG -> CFG

filterEdges :: (BlockId -> BlockId -> EdgeInfo -> Bool) -> CFG -> CFG

Filter the CFG with a custom function f. Paramaeters are `f from to edgeInfo`

addImmediateSuccessor :: Weights -> BlockId -> BlockId -> CFG -> CFG

Sometimes we insert a block which should unconditionally be executed after a given block. This function updates the CFG for these cases. So we get A -> B => A -> A' -> B -- -> C => -> C

mkWeightInfo :: EdgeWeight -> EdgeInfo

Convenience function, generate edge info based on weight not originating from cmm.

adjustEdgeWeight :: CFG -> (EdgeWeight -> EdgeWeight) -> BlockId -> BlockId -> CFG

Adjust the weight between the blocks using the given function. If there is no such edge returns the original map.

setEdgeWeight :: CFG -> EdgeWeight -> BlockId -> BlockId -> CFG

Set the weight between the blocks to the given weight. If there is no such edge returns the original map.

infoEdgeList :: CFG -> [CfgEdge]

Returns a unordered list of all edges with info

edgeList :: CFG -> [Edge]

Returns a unordered list of all edges without weights

getSuccessorEdges :: HasDebugCallStack => CFG -> BlockId -> [(BlockId, EdgeInfo)]

Get successors of a given node with edge weights.

getSuccessors :: HasDebugCallStack => CFG -> BlockId -> [BlockId]

Get successors of a given node without edge weights.

getSuccEdgesSorted :: CFG -> BlockId -> [(BlockId, EdgeInfo)]

Destinations from bid ordered by weight (descending)

getEdgeInfo :: BlockId -> BlockId -> CFG -> Maybe EdgeInfo

getCfgNodes :: CFG -> [BlockId]

hasNode :: CFG -> BlockId -> Bool

Is this block part of this graph?

loopMembers :: HasDebugCallStack => CFG -> LabelMap Bool

Determine loop membership of blocks based on SCC analysis This is faster but only gives yes/no answers.

loopLevels :: CFG -> BlockId -> LabelMap Int

loopInfo :: HasDebugCallStack => CFG -> BlockId -> LoopInfo

Determine loop membership of blocks based on Dominator analysis. This is slower but gives loop levels instead of just loop membership. However it only detects natural loops. Irreducible control flow is not recognized even if it loops. But that is rare enough that we don't have to care about that special case.

getCfg :: Platform -> Weights -> CmmGraph -> CFG

getCfgProc :: Platform -> Weights -> RawCmmDecl -> CFG

Generate weights for a Cmm proc based on some simple heuristics.

pprEdgeWeights :: CFG -> SDoc

sanityCheckCfg :: CFG -> LabelSet -> SDoc -> Bool

Check if the nodes in the cfg and the set of blocks are the same. In a case of a missmatch we panic and show the difference.

optimizeCFG :: Bool -> Weights -> RawCmmDecl -> CFG -> CFG

mkGlobalWeights :: HasDebugCallStack => BlockId -> CFG -> (LabelMap Double, LabelMap (LabelMap Double))

We take in a CFG which has on its edges weights which are relative only to other edges originating from the same node.

We return a CFG for which each edge represents a GLOBAL weight. This means edge weights are comparable across the whole graph.

For irreducible control flow results might be imprecise, otherwise they are reliable.

The algorithm is based on the Paper "Static Branch Prediction and Program Profile Analysis" by Y Wu, JR Larus The only big change is that we go over the nodes in the body of loops in reverse post order. Which is required for diamond control flow to work probably.

We also apply a few prediction heuristics (based on the same paper)

The returned result represents frequences. For blocks it's the expected number of executions and for edges is the number of traversals.