llvm-project

7a129526 - [BFI] Solve irreducible SCCs instead of splitting their headers (#215170)

[BFI] Solve irreducible SCCs instead of splitting their headers (#215170) c5a3139ebd0d (2014) approximates irreducible control flow by modelling an SCC as a loop with multiple headers, and http://reviews.llvm.org/D10348 re-distributes the loop mass across those headers in proportion to the backedge mass each one received -- one step of a power iteration, from an assumed split. Package the SCC with a single representative and solve it instead. `solveIrreducibleMass` iterates the SCC's internal chain towards its dominant eigenvector and reads the member masses, the exits and the circulating mass off that, so the entries' relative frequencies come out of the solve. Power iteration rather than a relaxation of `f = e + f*P`: the mass `e` entering the SCC is unknown here, so there is no fixed point to relax towards, only a direction. NumHeaders, the per-header BackedgeMass, getHeaderIndex, the isHeader binary search and adjustLoopHeaderMass go away with the split. Relative error against an exact rational solve has improved. ``` higher is better | relative error (lower is better) exact within 1% within 10% | p50 p90 max before 93 111 118 | 0.99 5.27 521.6 after 93 112 129 | 0.77 2.40 46.2 ``` Among the lit tests, selfloops and unequalrows converge to their exact ratios, and nonentry's worst block is off by 1.9x rather than 31.5x. Bound the iteration rather than run it to convergence: a periodic SCC oscillates and never terminates, and over that corpus the error distribution is the same at 16 iterations as at 1000, moving non-monotonically in between. The work is bounded by the size of the SCC rather than by the function. This does not subsume -use-iterative-bfi-inference (0a0800c4d10c), which relaxes the same equations over the whole function, where the entry mass is known. On yyparse_1 in profile-correlation-irreducible-loops.ll the solve alone puts b2 at 2.1x the entry against an exact 586.19x, and inference reaches 586.19x either way. Over the corpus above the two stack: enabling inference takes within 10% from 118 to 137 functions before this patch and from 129 to 144 after. CFGMST weights edges by BFI, so the instrumentation counter indices permute. Re-record Inputs/irreducible{,_entry}.proftext under the new numbering; the block counts the tests check are unchanged. LLM-assisted