| Summary: | Less inlining leads to CSiBE regression | ||
|---|---|---|---|
| Product: | gcc | Reporter: | Jan Hubicka <hubicka> |
| Component: | ipa | Assignee: | Not yet assigned to anyone <unassigned> |
| Status: | NEW --- | ||
| Severity: | normal | CC: | marxin |
| Priority: | P3 | Keywords: | missed-optimization |
| Version: | 4.6.0 | ||
| Target Milestone: | --- | ||
| Host: | Target: | ||
| Build: | Known to work: | ||
| Known to fail: | Last reconfirmed: | 2010-11-19 10:52:53 | |
| Attachments: | testcase flex-2.5.31/regex.c | ||
I thought partial inlining would maybe fix this? Otherwise it's really a case that needs IP analysis. > I thought partial inlining would maybe fix this? Otherwise it's really a
> case that needs IP analysis.
Not with -Os, we really know that it will optimize away.
Honza
Many things has changed since GCC 4.6 with respect to the inliner, has this been fixed? |
Created attachment 22451 [details] testcase flex-2.5.31/regex.c The loss here is not inlining regmatch_len. The catch is that the test if (m == ((void *)0) || m->rm_so < 0) is tested before all uses of regmatch_len and thus optimized out. So it simplifies into m->rm_so < 0 test and arithmetic that ends up being cheaper than call. int regmatch_len (regmatch_t * m) { if (m == ((void *)0) || m->rm_so < 0) { return 0; } return m->rm_eo - m->rm_so; } It is used as: if (m == ((void *)0) || m->rm_so < 0) return 0; if (regmatch_len (m) < 20) s = regmatch_cpy (m, buf, src); else s = regmatch_dup (m, src); Tricky. Inliner sees it as: Analyzing function body size: regmatch_len freq: 1000 size: 2 time: 2 if (m_2(D) == 0B) freq: 898 size: 1 time: 1 D.7268_3 = m_2(D)->rm_so; 50% will be eliminated by inlining freq: 898 size: 2 time: 2 if (D.7268_3 < 0) freq: 726 size: 1 time: 1 D.7270_4 = m_2(D)->rm_eo; 50% will be eliminated by inlining freq: 726 size: 1 time: 1 D.7268_5 = m_2(D)->rm_so; 50% will be eliminated by inlining freq: 726 size: 1 time: 1 D.7269_6 = D.7270_4 - D.7268_5; freq: 1000 size: 1 time: 2 return D.7269_1; will eliminated by inlining Overall function body time: 9-3 size: 11-5 With function call overhead time: 9-15 size: 11-8 I can imagine we can try to get summary based on value ranges, instead of known constants, do early VRP and work out first test well. Even optimizing the first conditoinal away won't make it inlined, it will be still considered to have size 9, so code will be expected to grow by 1 byte. Optimizing second conditoinal is even trickier. The code can be optimized away by IP-value range propagation that would be interesting optimization to have...