Movatterモバイル変換

• llvm
• BranchProbabilityInfo

Analysis providing branch probability information.

This is a function analysis which provides information on the relative probabilities of each "edge" in the function's CFG where such an edge is defined by a pair (PredBlock and an index in the successors). The probability of an edge from one block is always relative to the probabilities of other edges from the block. The probabilites of all edges from a block sum to exactly one (100%). We use a pair (PredBlock and an index in the successors) to uniquely identify an edge, since we can have multiple edges from Src to Dst. As an example, we can have a switch which jumps to Dst with value 0 and value 10.

Process of computing branch probabilities can be logically viewed as three step process:

First, if there is a profile information associated with the branch then it is trivially translated to branch probabilities. There is one exception from this rule though. Probabilities for edges leading to "unreachable" blocks (blocks with the estimated weight not greater than UNREACHABLE_WEIGHT) are evaluated according to static estimation and override profile information. If no branch probabilities were calculated on this step then take the next one.

Second, estimate absolute execution weights for each block based on statically known information. Roots of such information are "cold", "unreachable", "noreturn" and "unwind" blocks. Those blocks get their weights set to BlockExecWeight::COLD, BlockExecWeight::UNREACHABLE, BlockExecWeight::NORETURN and BlockExecWeight::UNWIND respectively. Then the weights are propagated to the other blocks up the domination line. In addition, if all successors have estimated weights set then maximum of these weights assigned to the block itself (while this is not ideal heuristic in theory it's simple and works reasonably well in most cases) and the process repeats. Once the process of weights propagation converges branch probabilities are set for all such branches that have at least one successor with the weight set. Default execution weight (BlockExecWeight::DEFAULT) is used for any successors which doesn't have its weight set. For loop back branches we use their weights scaled by loop trip count equal to 'LBH_TAKEN_WEIGHT/LBH_NOTTAKEN_WEIGHT'.

Here is a simple example demonstrating how the described algorithm works.

     BB1    /   \   v     v BB2     BB3/   \

v v ColdBB UnreachBB

Initially, ColdBB is associated with COLD_WEIGHT and UnreachBB with UNREACHABLE_WEIGHT. COLD_WEIGHT is set to BB2 as maximum between its successors. BB1 and BB3 has no explicit estimated weights and assumed to have DEFAULT_WEIGHT. Based on assigned weights branches will have the following probabilities: P(BB1->BB2) = COLD_WEIGHT/(COLD_WEIGHT + DEFAULT_WEIGHT) = 0xffff / (0xffff + 0xfffff) = 0.0588(5.9%) P(BB1->BB3) = DEFAULT_WEIGHT_WEIGHT/(COLD_WEIGHT + DEFAULT_WEIGHT) = 0xfffff / (0xffff + 0xfffff) = 0.941(94.1%) P(BB2->ColdBB) = COLD_WEIGHT/(COLD_WEIGHT + UNREACHABLE_WEIGHT) = 1(100%) P(BB2->UnreachBB) = UNREACHABLE_WEIGHT/(COLD_WEIGHT+UNREACHABLE_WEIGHT) = 0(0%)

If no branch probabilities were calculated on this step then take the next one.

Third, apply different kinds of local heuristics for each individual branch until first match. For example probability of a pointer to be null is estimated as PH_TAKEN_WEIGHT/(PH_TAKEN_WEIGHT + PH_NONTAKEN_WEIGHT). If no local heuristic has been matched then branch is left with no explicit probability set and assumed to have default probability.

Definition at line112 of fileBranchProbabilityInfo.h.