As pointed out by Gavin in bug #70418, when a value is already in memory we can avoid loading it to two GPRs at first and then unboxing them to a FPR.
Created attachment 111703 [details] the patch
Performance result based on current code base w/ DFG 32bit turned on (on Linux ia32 Core i7 Nehalem) - Kraken: TEST COMPARISON FROM TO DETAILS ============================================================================= ** TOTAL **: 1.087x as fast 7272.1ms +/- 0.3% 6688.4ms +/- 0.1% significant ============================================================================= ai: 1.002x as fast 797.2ms +/- 0.2% 795.3ms +/- 0.1% significant astar: 1.002x as fast 797.2ms +/- 0.2% 795.3ms +/- 0.1% significant audio: 1.116x as fast 2376.4ms +/- 0.8% 2130.0ms +/- 0.2% significant beat-detection: 1.108x as fast 702.2ms +/- 2.7% 634.0ms +/- 0.5% significant dft: 1.27x as fast 557.7ms +/- 0.2% 439.4ms +/- 0.5% significant fft: 1.123x as fast 534.6ms +/- 0.4% 476.1ms +/- 0.1% significant oscillator: - 581.9ms +/- 0.3% 580.5ms +/- 0.1% imaging: 1.129x as fast 2920.6ms +/- 0.2% 2586.2ms +/- 0.1% significant gaussian-blur: 1.35x as fast 1278.9ms +/- 0.1% 950.8ms +/- 0.1% significant darkroom: 1.003x as fast 715.8ms +/- 0.2% 713.5ms +/- 0.1% significant desaturate: - 925.9ms +/- 0.5% 921.9ms +/- 0.3% json: - 194.6ms +/- 0.5% 193.9ms +/- 0.5% parse-financial: 1.006x as fast 82.1ms +/- 0.3% 81.6ms +/- 0.5% significant stringify-tinderbox: - 112.5ms +/- 1.0% 112.3ms +/- 0.8% stanford: - 983.3ms +/- 0.2% 983.0ms +/- 0.3% crypto-aes: ?? 164.7ms +/- 0.5% 165.6ms +/- 0.6% not conclusive: might be *1.005x as slow* crypto-ccm: - 141.3ms +/- 0.7% 140.5ms +/- 0.6% crypto-pbkdf2: - 572.7ms +/- 0.2% 572.4ms +/- 0.3% crypto-sha256-iterative: - 104.6ms +/- 0.4% 104.5ms +/- 0.5%
Comment on attachment 111703 [details] the patch Clearing flags on attachment: 111703 Committed r97928: <http://trac.webkit.org/changeset/97928>
All reviewed patches have been landed. Closing bug.