| Summary: | [4.8 Regression] wrong code generated by gcc 4.8.0 on i686 | ||
|---|---|---|---|
| Product: | gcc | Reporter: | Mattias Engdegård <mattiase> |
| Component: | rtl-optimization | Assignee: | Not yet assigned to anyone <unassigned> |
| Status: | RESOLVED FIXED | ||
| Severity: | normal | CC: | eraman, jakub, kha, vmakarov, vmakarov |
| Priority: | P1 | Keywords: | wrong-code |
| Version: | 4.8.0 | ||
| Target Milestone: | 4.8.1 | ||
| Host: | Target: | i?86-*-* | |
| Build: | Known to work: | 4.7.2, 4.9.0 | |
| Known to fail: | 4.8.0 | Last reconfirmed: | 2013-04-23 00:00:00 |
| Attachments: |
Test case including driver demonstrating the bug
Single-file test case. |
||
Created attachment 29918 [details]
Single-file test case.
I can reproduce the wrong-code on x86_64-linux with gcc 4.9-20130421 and 4.8-20130418, using -m32 -O2 -Wall. gcc 4.7 and 4.6 generate correct code.
Confirmed. Started with http://gcc.gnu.org/r192719 aka LRA merge, the problematic function is emac_operation. We have after the get_value call:
(insn 73 30 32 6 (set (reg:SI 76 [ D.1441 ])
(reg:SI 0 ax)) pr57046.c:42 85 {*movsi_internal}
(expr_list:REG_DEAD (reg:SI 0 ax)
(nil)))
(insn 32 73 33 6 (parallel [
(set (reg:SI 73 [ D.1443 ])
(ashift:SI (subreg:SI (reg:DI 60 [ D.1441 ]) 0)
(const_int 2 [0x2])))
(clobber (reg:CC 17 flags))
]) 502 {*ashlsi3_1}
(expr_list:REG_DEAD (reg:DI 60 [ D.1441 ])
(expr_list:REG_UNUSED (reg:CC 17 flags)
(nil))))
and IRA decides to put pseudo 76 into %ebx and pseudo 60 into %ecx. Either it is an IRA bug, or IRA takes into account that we only really need the low 32-bits of pseudo 60 at that point. In any case, reload loads SImode %ecx from the stack and uses it in the shift, while LRA loads full DImode %ecx (i.e. %ecx and %ebx) from the stack, then uses just the low bits from that (i.e. %ecx) in the shift. So the LRA generated code clobbers the value in %ebx, and get_value call is even later on DCEd because of it.
(In reply to comment #4) > We have after the get_value call: > (insn 73 30 32 6 (set (reg:SI 76 [ D.1441 ]) > (reg:SI 0 ax)) pr57046.c:42 85 {*movsi_internal} > (expr_list:REG_DEAD (reg:SI 0 ax) > (nil))) > (insn 32 73 33 6 (parallel [ > (set (reg:SI 73 [ D.1443 ]) > (ashift:SI (subreg:SI (reg:DI 60 [ D.1441 ]) 0) > (const_int 2 [0x2]))) > (clobber (reg:CC 17 flags)) > ]) 502 {*ashlsi3_1} > (expr_list:REG_DEAD (reg:DI 60 [ D.1441 ]) > (expr_list:REG_UNUSED (reg:CC 17 flags) > (nil)))) > > and IRA decides to put pseudo 76 into %ebx and pseudo 60 into %ecx. Either it > is an IRA bug, or IRA takes into account that we only really need the low > 32-bits of pseudo 60 at that point. In any case, reload loads SImode %ecx from > the stack and uses it in the shift, while LRA loads full DImode %ecx (i.e. %ecx > and %ebx) from the stack, then uses just the low bits from that (i.e. %ecx) in > the shift. So the LRA generated code clobbers the value in %ebx, and get_value > call is even later on DCEd because of it. It seems like a discrepancy in IRA which allocates in terms of subregisters and LRA splitting (including call save/restore as in this case) in terms of pseudos. I guess fixing this might take about week. Author: vmakarov Date: Wed Apr 24 20:27:33 2013 New Revision: 198263 URL: http://gcc.gnu.org/viewcvs?rev=198263&root=gcc&view=rev Log: 2013-04-24 Vladimir Makarov <vmakarov@redhat.com> PR rtl-optimizations/57046 * lra-constraints (split_reg): Set up lra_risky_transformations_p for multi-reg splits. 2013-04-24 Vladimir Makarov <vmakarov@redhat.com> PR rtl-optimizations/57046 * gcc.target/i386/pr57046.c: New test. Added: trunk/gcc/testsuite/gcc.target/i386/pr57046.c Modified: trunk/gcc/ChangeLog trunk/gcc/lra-constraints.c trunk/gcc/testsuite/ChangeLog Author: vmakarov Date: Thu May 2 17:52:45 2013 New Revision: 198556 URL: http://gcc.gnu.org/viewcvs?rev=198556&root=gcc&view=rev Log: 2013-05-02 Vladimir Makarov <vmakarov@redhat.com> Backport from mainline 2013-04-24 Vladimir Makarov <vmakarov@redhat.com> PR rtl-optimizations/57046 * lra-constraints (split_reg): Set up lra_risky_transformations_p for multi-reg splits. 2013-05-02 Vladimir Makarov <vmakarov@redhat.com> Backport from mainline 2013-04-24 Vladimir Makarov <vmakarov@redhat.com> PR rtl-optimizations/57046 * gcc.target/i386/pr57046.c: New test. Added: branches/gcc-4_8-branch/gcc/testsuite/gcc.target/i386/pr57046.c Modified: branches/gcc-4_8-branch/gcc/ChangeLog branches/gcc-4_8-branch/gcc/lra-constraints.c branches/gcc-4_8-branch/gcc/testsuite/ChangeLog |
Created attachment 29917 [details] Test case including driver demonstrating the bug Gcc 4.8.0 silently miscompiles the attached short test case (emac.c) on 32-bit x86 with -O2. It appears that the value returned from get_value is thrown away. gcc -V output: Using built-in specs. COLLECT_GCC=/usr/local/gcc/4.8.0/bin/gcc COLLECT_LTO_WRAPPER=/home/local/linux/gcc/4.8.0/bin/../libexec/gcc/i686-pc-linux-gnu/4.8.0/lto-wrapper Target: i686-pc-linux-gnu Configured with: ../gcc-4.8.0/configure --prefix=/usr/local/gcc/4.8.0 --with-mpc=/tmp/extra --with-gmp=/tmp/extra --with-mpfr=/tmp/extra --with-isl=/tmp/extra --with-cloog=/tmp/extra --with-as=/usr/local/binutils/2.23.2/bin/as --with-ld=/usr/local/binutils/2.23.2/bin/ld.gold --enable-languages=c,c++,objc,go Thread model: posix gcc version 4.8.0 (GCC)