[0xWOLAND]

libint2's generated recurrence kernels load the same coefficients from memory. Instead we can load them into a scalar once and then reuse it across the recurrence ops. This lets us reduce memory accesses substantially. It appears to be faster cc-pVQZ (1.24x), def2-QZVP (~1.24x), and ANO-RCC (~1.41x) Gaussian basis sets.

llvm estimates

CPU model	upstream master	hoisting branch	speedup
Intel Skylake-AVX512	114,506 cycles	98,870 cycles	1.16×
AMD Zen 4	125,784	70,454	1.79×
Apple M1	125,676	70,282	1.79×
Neoverse V2	85,051	61,593	1.38×

Happy to benchmark on other machines :)

Edit: Here's the script I used to benchmarklibint using llvm-mca

[evaleev]

Thanks for this. I validated it locally and measured wall-clock timings.

Validation. I built check for both the merge-base (69f92196) and this PR. The config was 1-body, ERI, ERI3 and ERI2 to am 4, with 1st derivatives to am 3, all scalar.

• All tests pass on both: compiler coretest, eritest, unit tests (20M assertions), hf and hf++.
• 2655 of the 3393 generated files change.
• Compiling every generated .cc with -Wunused-variable -Wunused-but-set-variable gives 0 warnings.
• Integral checksums are bit-identical to master.

Timings. Apple M-series, Apple clang -O3. The workload is all unique shell sets for one water molecule. Baseline and PR ran alternately: 3 rounds × min of 5 runs, except ANO-RCC (2 rounds × 1 run). The spread between rounds was under 0.5%.

Case	master	PR	speedup
ERI cc-pVQZ	0.374 s	0.304 s	1.23×
ERI def2-QZVP	0.386 s	0.313 s	1.23×
ERI ANO-RCC	149.0 s	124.0 s	1.20×
ERI 1st derivative, cc-pVTZ	0.202 s	0.171 s	1.18×
ERI3 cc-pVQZ / cc-pVTZ-JKFIT	0.0099 s	0.0092 s	1.08×

The cc-pVQZ and def2-QZVP numbers match yours. ANO-RCC comes out at 1.20× here rather than 1.41×. The llvm-mca estimate for M1 (1.79×) doesn't show up in wall-clock time.

Follow-up commit (200b83e). Previously generate_code() found the hoisted loads by re-deriving each hv_ name and comparing strings against the symbols. That breaks silently if symbols are ever assigned along another path. Now assign_symbols() records each hoisted local together with its inteval field, and generate_code() declares exactly that list. Vertices that refer to another vertex are not hoisted. The generated library is byte-identical to the one from your commit.

---

This comment was agent-generated and human-reviewed.