http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54953
Bug #: 54953
Summary: [4.8 Regression] New sra-1.c FAILs on powerpc
Classification: Unclassified
Product: gcc
Version: 4.8.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: debug
AssignedTo: unassig...@gcc.gnu.org
ReportedBy: ja...@gcc.gnu.org
CC: aol...@gcc.gnu.org
Target: powerpc64-linux
When backporting the valtrack.c addition and related changes to 4.7 branch,
I've noticed it regresses some sra-1.c tests that used to pass, in particular
FAIL: gcc.dg/guality/sra-1.c -O1 line 43 a.i == 4
(ditto for all other > -O1 levels).
The problem is that before fwprop1 we have:
(insn 11 10 12 2 (set (reg:HI 130 [ a$i ])
(asm_operands:HI ("") ("=r") 0 [
(reg:HI 131)
]
[
(asm_input:HI ("0") (null):0)
]
[] sra-1.c:40)) sra-1.c:40 -1
(nil))
(insn 12 11 13 2 (set (reg:DI 127 [ a$i+-6 ])
(sign_extend:DI (reg:HI 130 [ a$i ]))) sra-1.c:40 21 {*rs6000.md:447}
(nil))
(debug_insn 13 12 14 2 (var_location:HI a$i (subreg:HI (reg:DI 127 [ a$i+-6 ])
6)) sra-1.c:40 -1
(nil))
and reg:HI 130 is really dead, and fwprop1 decides to propagate into the
debug_insn, so after fwprop1 debug_insn 13 contains (reg:HI 130) instead of the
subreg. Then cprop1 comes and performs fast DCE, which decides to put a debug
temp for reg:HI 130 using DEBUG_TEMP_BEFORE_WITH_VALUE before the asm
instruction, but obviously asm_operands inside of var_location is never going
to expand to anything useful. dce.c seems to be the only user of
DEBUG_TEMP_BEFORE_WITH_VALUE mode of insertion. If the insn it is adding it to
is not marked (!marked_insn_p (insn)), then that is certainly the right thing
to do, the insn is going to be removed and we have no other option. But if it
is
being kept, just debug temp is being added because of debug uses after the
pseudo became dead, then this is counterproductive.
