[Mlir-commits] [mlir] [MLIR][LLVM] Shard generated operation definitions (NFC) (PR #221813)

[Mehdi Amini]

joker-eph wrote:

Some stats for posterity:

 The baseline monolithic Ninja edge took 136.706s, with 136.035s of ExecuteCompiler time. Its inclusive trace totals were 43.841s frontend, 91.698s backend, 50.093s optimizer, 41.403s code generation, 20.889s function instantiation, 10.441s class instantiation, 19.228s source processing, and 4.842s class parsing.

The eight shards plus the remaining dialect TU totaled 94.125s of ExecuteCompiler time. Their inclusive totals were 48.068s frontend, 45.428s backend, 27.581s optimizer, 17.787s code generation, 15.988s function instantiation, 14.330s class instantiation, 30.823s source processing, and 20.776s class parsing. The slowest new compile edge was 13.291s.

Aggregate compiler time fell 30.8%, backend time fell 50.5%, optimizer time fell 44.9%, code-generation time fell 57.0%, and function-instantiation time fell 23.5%. The longest compile edge fell 90.3%. Repeated work increased frontend time 9.6%, class-instantiation time 37.3%, source-processing time 60.3%, and class-parsing time 329.1%.

The affected-target build ran 18 steps in 14.16s wall time, using 91.22s user time and 4.12s system time at 673% CPU, with 852,772 KiB peak RSS and zero swaps. The baseline monolithic object contained 7,939,259 text bytes; the remaining dialect object plus eight shard objects contain 8,487,922 text bytes, an increase of 6.9%. The static archive grew from 34,919,416 to 36,426,416 bytes, an increase of 4.3%.

Rebuilding mlir-opt and running four focused suites took 9.29s wall time, 46.71s user time, and 16.71s system time at 682% CPU, with 1,633,484 KiB peak RSS and zero swaps. The suites ran 4 tests in 0.16s, 78 in 0.44s, 452 in 0.53s, and 68 in 0.31s; all 602 tests passed.


https://github.com/llvm/llvm-project/pull/221813