http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55346
Bug #: 55346
Summary: var-tracking generates incomplete/inaccurate debug
info
Classification: Unclassified
Product: gcc
Version: 4.8.0
Status: UNCONFIRMED
Severity: major
Priority: P3
Component: rtl-optimization
AssignedTo: unassig...@gcc.gnu.org
ReportedBy: ea...@eagercon.com
Created attachment 28703
--> http://gcc.gnu.org/bugzilla/attachment.cgi?id=28703
Test program
Var-tracking is does not record the lifetime of variables allocated to
registers correctly, resulting in gdb displaying the variable as "optimized
out" when it is live. I've confirmed this problem on x86 and arm.
The test case is a loop which parses command line options, indexed by integer
variable argno. When compiled with "-g -O2", argno is allocated to rbx, which
is never reused. The DWARF location list which is generated from
NOTE_INSN_VAR_LOCATION insns, shows disjoint allocations for argno, omitting PC
ranges where the variable is live in rbx.
The DWARF location list has argno as literal 1 beyond the instruction where it
is allocated to rbx. Argno is defined as rbx starts several instructions after
it is allocated and ends premature.
Var-tracking inserts a NOTE_INSN_VAR_LOCATION for argno as const int 1 at the
start of the BB #2, but one is not inserted when argno is allocated to rbx
later in BB #2. One is generated later, in BB #5.
Var-tracking inserts a (nil) NOTE_INSN_VAR_LOCATION for argno after BB #3
(branch to epilogue), although argno is live and is used in BB #4, immediately
following BB #3. Either this should not be generated or another
NOTE_INSN_VAR_LOCATION for argno should be generated at the start of BB #4.
