Open-code LIR instead of calling a builtin.

Comment on attachment 462970 [details] [diff] [review] patch Cute. What's its perf now? >+ LIns *unitstr_ins = lir->ins2(LIR_addp, >+ INS_CONSTPTR(JSString::unitStringTable), >+ lir->ins2ImmI(LIR_lshp, >+ lir->insUI2P(ch_ins), >+ (sizeof(JSString) == 16) ? 4 : 5)); This is just asking to break. r+ with JS_STATIC_ASSERT(sizeof(JSString) == 16 || sizeof(JSString) == 32);

Cachegrind results: --------------------------------------------------------------- | millions of instructions executed | | total | on-trace (may overestimate) | --------------------------------------------------------------- | 111.202 111.481 (0.997x) | 20.797 20.809 (0.999x) | 3d-cube | 56.990 56.949 (1.001x) | 22.313 22.313 (------) | 3d-morph | 147.820 146.795 (1.007x) | 21.982 21.971 (------) | 3d-raytrace | 77.548 77.418 (1.002x) | 17.401 17.400 (------) | access-binary-trees | 183.728 183.440 (1.002x) | 103.707 103.704 (------) | access-fannkuch | 41.565 41.405 (1.004x) | 17.553 17.551 (------) | access-nbody | 60.014 59.983 (1.001x) | 27.013 27.012 (------) | access-nsieve | 15.688 15.681 (------) | 2.862 2.862 (------) | bitops-3bit-bits-in-byte | 43.303 43.285 (------) | 30.280 30.280 (------) | bitops-bits-in-byte | 22.773 22.771 (------) | 10.219 10.219 (------) | bitops-bitwise-and | 71.624 71.611 (------) | 39.453 39.453 (------) | bitops-nsieve-bits | 42.397 42.309 (1.002x) | 26.655 26.654 (------) | controlflow-recursive | 47.115 47.002 (1.002x) | 5.541 5.660 (0.979x) | crypto-md5 | 31.336 31.286 (1.002x) | 7.030 7.109 (0.989x) | crypto-sha1 | 127.956 127.903 (------) | 10.461 10.461 (------) | date-format-tofte | 99.144 98.884 (1.003x) | 11.118 11.115 (------) | date-format-xparb | 52.617 52.594 (------) | 27.614 27.614 (------) | math-cordic | 38.415 38.409 (------) | 6.252 6.252 (------) | math-partial-sums | 28.590 28.475 (1.004x) | 10.286 10.285 (------) | math-spectral-norm | 100.938 100.915 (------) | 75.045 75.045 (------) | regexp-dna | 44.076 40.705 (1.083x) | 8.098 10.397 (0.779x) | string-base64 | 123.213 122.919 (1.002x) | 24.544 24.543 (------) | string-fasta | 272.611 269.835 (1.010x) | 6.840 6.671 (1.025x) | string-tagcloud | 251.422 251.318 (------) | 5.637 5.635 (------) | string-unpack-code | 64.494 63.696 (1.013x) | 7.401 7.521 (0.984x) | string-validate-input ------- | 2156.590 2147.080 (1.004x) | 546.114 548.548 (0.996x) | all string-base64 is a nice result. Note that only the following tests had their codegen affected: crypto-md5.js crypto-sha1.js date-format-xparb.js string-base64.js string-tagcloud.js string-validate-input.js Variation on others is due to the property cache.