| Summary: | Wrong code generated (assignment incorrectly optimized out) | ||
|---|---|---|---|
| Product: | gcc | Reporter: | Marcin 'Qrczak' Kowalczyk <qrczak> |
| Component: | middle-end | Assignee: | Not yet assigned to anyone <unassigned> |
| Status: | RESOLVED DUPLICATE | ||
| Severity: | normal | CC: | cdfrey, chiabaut, fang, gcc-bugs, pinskia, qrczak, sorenj, takahisa.yokota |
| Priority: | P3 | Keywords: | alias, wrong-code |
| Version: | 4.2.0 | ||
| Target Milestone: | --- | ||
| Host: | i686-pc-linux-gnu | Target: | i686-pc-linux-gnu |
| Build: | i686-pc-linux-gnu | Known to work: | |
| Known to fail: | Last reconfirmed: | ||
| Bug Depends on: | 28003 | ||
| Bug Blocks: | |||
| Attachments: | Test case | ||
Created attachment 12075 [details]
Test case
The problem seems to be caused by illegal type punning. Rewriting the code to avoid pointer casts cures the bug, and -fno-strict-aliasing cures it too, so gcc is probably correct. Thanks to Michael Veksler for pointing this out. |
The attached module, when compiled with -O2, has wrong code generated for the last set_global_field call. The assembly shows: leal 4(%edx), %eax cmpl marked_limit, %eax movl BadArguments, %ecx movl $record_12420+4, (%edx) movl %eax, marked_top jbe .L28 movl %ecx, record_12420+4 call gc .L28: leave xorl %eax, %eax ret while the assignment of BadArguments (from %ecx) to record_12420+4 should be performed unconditionally, as can be easily deduced from the source. This is gcc-4.2.0-0.20060806r115974. The bug doesn't manifest without -O2 or when the number of set_global_field calls is reduced.