Documentation Archive Developer
Search

Next: , Previous: Edges, Up: Control Flow


13.3 Profile information

In many cases a compiler must make a choice whether to trade speed in one part of code for speed in another, or to trade code size for code speed. In such cases it is useful to know information about how often some given block will be executed. That is the purpose for maintaining profile within the flow graph. GCC can handle profile information obtained through profile feedback, but it can also estimate branch probabilities based on statics and heuristics.

The feedback based profile is produced by compiling the program with instrumentation, executing it on a train run and reading the numbers of executions of basic blocks and edges back to the compiler while re-compiling the program to produce the final executable. This method provides very accurate information about where a program spends most of its time on the train run. Whether it matches the average run of course depends on the choice of train data set, but several studies have shown that the behavior of a program usually changes just marginally over different data sets.

When profile feedback is not available, the compiler may be asked to attempt to predict the behavior of each branch in the program using a set of heuristics (see predict.def for details) and compute estimated frequencies of each basic block by propagating the probabilities over the graph.

Each basic_block contains two integer fields to represent profile information: frequency and count. The frequency is an estimation how often is basic block executed within a function. It is represented as an integer scaled in the range from 0 to BB_FREQ_BASE. The most frequently executed basic block in function is initially set to BB_FREQ_BASE and the rest of frequencies are scaled accordingly. During optimization, the frequency of the most frequent basic block can both decrease (for instance by loop unrolling) or grow (for instance by cross-jumping optimization), so scaling sometimes has to be performed multiple times.

The count contains hard-counted numbers of execution measured during training runs and is nonzero only when profile feedback is available. This value is represented as the host's widest integer (typically a 64 bit integer) of the special type gcov_type.

Most optimization passes can use only the frequency information of a basic block, but a few passes may want to know hard execution counts. The frequencies should always match the counts after scaling, however during updating of the profile information numerical error may accumulate into quite large errors.

Each edge also contains a branch probability field: an integer in the range from 0 to REG_BR_PROB_BASE. It represents probability of passing control from the end of the src basic block to the dest basic block, i.e. the probability that control will flow along this edge. The EDGE_FREQUENCY macro is available to compute how frequently a given edge is taken. There is a count field for each edge as well, representing same information as for a basic block.

The basic block frequencies are not represented in the instruction stream, but in the RTL representation the edge frequencies are represented for conditional jumps (via the REG_BR_PROB macro) since they are used when instructions are output to the assembly file and the flow graph is no longer maintained.

The probability that control flow arrives via a given edge to its destination basic block is called reverse probability and is not directly represented, but it may be easily computed from frequencies of basic blocks.

Updating profile information is a delicate task that can unfortunately not be easily integrated with the CFG manipulation API. Many of the functions and hooks to modify the CFG, such as redirect_edge_and_branch, do not have enough information to easily update the profile, so updating it is in the majority of cases left up to the caller. It is difficult to uncover bugs in the profile updating code, because they manifest themselves only by producing worse code, and checking profile consistency is not possible because of numeric error accumulation. Hence special attention needs to be given to this issue in each pass that modifies the CFG.

It is important to point out that REG_BR_PROB_BASE and BB_FREQ_BASE are both set low enough to be possible to compute second power of any frequency or probability in the flow graph, it is not possible to even square the count field, as modern CPUs are fast enough to execute $2^32$ operations quickly.