diff options
Diffstat (limited to 'docs/appendices/optimizer-tuning.md')
| -rw-r--r-- | docs/appendices/optimizer-tuning.md | 65 |
1 files changed, 65 insertions, 0 deletions
diff --git a/docs/appendices/optimizer-tuning.md b/docs/appendices/optimizer-tuning.md new file mode 100644 index 00000000..11dbcdf3 --- /dev/null +++ b/docs/appendices/optimizer-tuning.md @@ -0,0 +1,65 @@ +# Optimizer and backend tuning + +These measurements are from the numbered snapshots before the 2026-10 split. +In this historical table, Ch 24 means the retired combined SCCP snapshot and +Ch 25 means the retired combined modules/constructor/import snapshot. +They are not new measurements of the extracted chapters. The implementation +continues in [Separate Compilation](../../chapter28b/README.md). + +## RegAlloc improvements: try cold loop splits first + +A loop Phi and its backedge can belong to one live range even though their +values must coexist at some instruction. This self-conflict requires splitting +the range. Chapter 23 split the Phi's definition and incoming values right away. +This chapter first leaves the hot definition/backedge alone and tries the +outside-loop copies. If coloring succeeds, it avoids moves on every iteration. +Other conflicts, including two-address operations, keep their existing handling. + +The allocator remembers which Phis have had this chance. If the same Phi +self-conflicts again, the next attempt uses the ordinary aggressive splitting. +The record uses Phi identity because live ranges are rebuilt each round. It +also covers a backedge-only conflict where the first attempt inserts no copy: +that attempt must still consume the one permitted delay. A regression checks +both the cheaper first attempt and the mandatory fallback. + +## RegAlloc improvements: live-range area and split cost + +Chapter 24 tried cold copies before splitting a loop Phi's hot edges. This +chapter also estimates how much scheduled code a spill frees from register +pressure, divided by the estimated cost of the new copies. Callee-save ranges +cover a whole function. Cloneable constants can be rebuilt nearer their uses. +Copies inside loops cost more: the estimate weights them by `8^loopDepth`. +These are inexpensive approximations, especially for values with many uses. + +A single-use function address in the same block gets priority for +rematerialization when it is separated from its use. Moving that address +calculation next to the use adds no instruction and avoids preserving a +callee-save register merely to hold the constant. Adjacent uses and multi-use +addresses retain the area/cost ranking. + +Rows are fixed test cohorts; columns are compiler chapters. Each cell gives +**Ops / RA / X**: executed ARM+RISC-V instructions, the subset emitted for +allocator copies/rematerializations, and x86's loop-weighted split-move estimate. +**Cases D / X** counts measured dynamic program/target cases and x86 compilations; +membership and inputs stay fixed across columns. Native-library work is excluded. +Chapter 20 has no encoder, hence no dynamic counts. Lower is better within each +metric; see the [measurement details](../regalloc-spills.md) for coverage. + +<!-- spill-matrix:start --> +| Test cohort | Cases D / X | Ch 20: Ops / RA / X | Ch 21: Ops / RA / X | Ch 22: Ops / RA / X | Ch 23: Ops / RA / X | Ch 24: Ops / RA / X | Ch 25: Ops / RA / X | +|---|---:|---:|---:|---:|---:|---:|---:| +| Ch 20 | 22 / 13 | — / — / 161 | 55,173 / 5,407 / 174 | 58,751 / 9,099 / 175 | 58,736 / 9,084 / 184 | 58,720 / 9,068 / 183 | 58,965 / 9,076 / 208 | +| Ch 21 | 20 / 18 | | 6,207 / 261 / 461 | 6,324 / 379 / 475 | 6,326 / 381 / 475 | 6,325 / 380 / 470 | 6,196 / 219 / 534 | +| Ch 22 | 14 / 8 | | | 198 / 60 / 7 | 194 / 56 / 7 | 194 / 56 / 7 | 252 / 92 / 11 | +| Ch 23 | 14 / 8 | | | | 630 / 160 / 36 | 630 / 160 / 36 | 864 / 358 / 36 | +| Ch 24 | 36 / 20 | | | | | 4,923 / 1,341 / 790 | 4,646 / 1,335 / 609 | +| Ch 25 | 2 / 12 | | | | | | 20 / 0 / 1,856 | +<!-- spill-matrix:end --> + + +Chapter 25 introduces several features, which when run on the older tests +generate a lot more code, which in turn generates more spills. + +Class initialization, inlining, escape analysis, and library adaptations +all change the graphs. To help with this, we adjust the inlining +heuristics to try to inline and optimize the boilerplate. |