http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54963
Kazumoto Kojima <kkojima at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |olegendo at gcc dot gnu.org
Component|middle-end |target
Summary|Wrong code generated for |[4.8 Regression] Wrong code
|libgfortran/generated/eoshi |generated for
|ft3_8.c on SH |libgfortran/generated/eoshi
| |ft3_8.c on SH
--- Comment #1 from Kazumoto Kojima <kkojima at gcc dot gnu.org> 2012-10-20
06:33:06 UTC ---
It looks that the problematic code run through a wrong flow
and the above was misleading. It seems that the change
r184829 | olegendo | 2012-03-03 06:21:13 +0900 (Sat, 03 Mar 2012) | 8 lines
PR target/49486
* config/sh/sh.md (negdi2): Add TARGET_SH1 condition.
(absdi2): New expander.
(*absdi2, *negabsdi2, negdi_cond): New insns and splits.
* gcc.target/sh/pr49468-si.c: Skip unsupported test for SH64.
* gcc.target/sh/pr49468-di.c: New.
causes this. After reload, there is an insn for abs(sh):
(insn 432 431 1393 39 (set (reg:DI 2 r2)
(abs:DI (reg:DI 1 r1))) eoshift.c:167 189 {*absdi2}
(nil))
which is splitted into
(insn 1607 431 1609 39 (set (reg:SI 147 t)
(ge:SI (reg:SI 2 r2 [+4 ])
(const_int 0 [0]))) eoshift.c:167 13 {cmpgesi_t}
(nil))
(insn 1609 1607 1610 39 (set (reg:SI 2 r2)
(reg:SI 1 r1)) eoshift.c:167 234 {movsi_ie}
(nil))
(insn 1610 1609 1611 39 (set (reg:SI 3 r3 [+4 ])
(reg:SI 2 r2 [+4 ])) eoshift.c:167 234 {movsi_ie}
(nil))
(jump_insn 1611 1610 1626 39 (set (pc)
(if_then_else (ne (reg:SI 147 t)
(const_int 0 [0]))
...
The insn 1609 changes r2 before its original value is used.
