Verify if CFI made any impact on the style change perf |
||
Issue descriptionPer Elliott suggestion, the following (currently manual) test case should be done: 1. Open https://en.wikipedia.org/wiki/Cat 2. Open DevTools console (with the Elements panel *not* open) 3. Run the following: var t = performance.now(); document.body.style.display = "none"; document.body.offsetTop; document.body.style.display = ""; document.body.offsetTop; performance.now() - t; It will display a number that is expected to be about the same for CFI and non-CFI (if no regression happened).
,
Sep 8 2016
For #1 I used the following snippet: var t = performance.now(); document.body.style.display = "none"; getComputedStyle(document.body).color document.body.style.display = ""; getComputedStyle(document.body).color; performance.now() - t; The numbers are in the spreadsheet: https://docs.google.com/spreadsheets/d/1KSJILMx7OXufaFMXgXenaLKGZnMj2jnwWvBOStjFg3g/edit?usp=sharing For #0, the numbers a consistent (mean values): non-LTO, non-CFI: 222.3 LTO, non-CFI: 199.1 LTO, CFI: 216.0 There's some CFI-related slowdown, but it's smaller than what we gain from LTO. The overhead will become even smaller, when the upcoming optimizations are implemented. While there're some methods I could find and disable CFI for, I don't think it makes sense to do. For #1 the numbers don't make any sense to me. Probably, a hint that this is a noisy benchmark (means): non-LTO, non-CFI: 73.998 LTO, non-CFI: 74.965 LTO, CFI: 70.867 Here, we see that LTO makes things slower, but then CFI speeds them up. That is unlikely to be truth, and I would rather say, these numbers are equal modulo noise. |
||
►
Sign in to add a comment |
||
Comment 1 by esprehn@chromium.org
, Sep 7 2016