diff options
Diffstat (limited to 'chapter20/README.md')
| -rw-r--r-- | chapter20/README.md | 31 |
1 files changed, 28 insertions, 3 deletions
diff --git a/chapter20/README.md b/chapter20/README.md index e463831a..271653c6 100644 --- a/chapter20/README.md +++ b/chapter20/README.md @@ -152,6 +152,11 @@ represent mutable bitsets with the allocator routinely masking off bits when accumulating a set of constraints. To help keep these uses apart, the `RegMask` class includes mutable and immutable variants. +#### Memory + +Memory edges exist to enforce ordering, but do not get any registers. +The allocator basically ignores them. + ### Callee/Caller Save Registers @@ -505,21 +510,41 @@ checks, not additional program benchmarks. | Program-test cohort | Allocator chapter | Compilations | Split moves | Loop-weighted moves | |---|---:|---:|---:|---:| -| Chapter 20 | 20 | 39 | 238 | 357 | +| Chapter 20 | 20 | 39 | 234 | 360 | `_spills` counts surviving split moves, including register-to-register copies; it does not count only stack stores. `_spillScaled` weights each move by `8^loopDepth`, a rough cost estimate rather than measured execution frequency. -The weighted subtotals are 154 for x86-64, 101 for RISC-V, and 102 for ARM. +The weighted subtotals are 161 for x86-64, 100 for RISC-V, and 99 for ARM. ### Comments on the measurements Before the correctness fixes, this cohort had 235 moves / 354 weighted moves. -The final corrected baseline costs three more of each. The final narrow x86 +The corrected baseline before lazy memory cost three more of each. The final narrow x86 store-mask backport accounts for two of those moves: an integer byte/short store cannot consume an XMM register. This is a correctness cost, not a quality gain. The fixed-neighbor color-bias correction changes no counts in this cohort. +With lazy memory and selected-read ordering, the same 39 compilations have +234 moves / 360 weighted moves, versus 238 / 357 before the port. Register +legality passes on every target; allocator heuristics are unchanged. The changed +programs are: + +| Program | Target | Moves before / after | Weighted before / after | +|---|---|---:|---:| +| Array access (`testAlloc2`) | x86-64 | 1 / 2 | 1 / 2 | +| Array prefix sum (`testArray1`) | x86-64 | 5 / 6 | 5 / 13 | +| String hash | x86-64 | 7 / 5 | 21 / 19 | +| String hash | RISC-V | 4 / 3 | 4 / 3 | +| String hash | ARM | 5 / 3 | 5 / 3 | +| Merge sort | ARM | 37 / 36 | 37 / 36 | + +The prefix sum's extra loop move follows from scheduling selected reads before +clobbering writes. A diagnostic build omitting that ordering returns to 5 / 5; +it is not a valid alternative schedule. These are changed frontend and scheduling +graphs, not a comparison of allocator heuristics. The new memory-allocation +regressions check legality but do not add entries to this measured cohort. + Later chapters will add one quality technique at a time and end with this kind of table: Chapter 21 will measure both the Chapter 20 and Chapter 21 program cohorts using its allocator, and Chapter 25 will have rows for cohorts 20-25. |