[Bug tree-optimization/49709] New: Bad code gen with -fipa-sra

["ian at airs dot com" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49709

           Summary: Bad code gen with -fipa-sra
           Product: gcc
           Version: 4.7.0
            Status: UNCONFIRMED
          Keywords: wrong-code
          Severity: major
          Priority: P3
         Component: tree-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: ian@airs.com


Created attachment 24741
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24741
Test case

I have attached a C test case.  When compiled for x86 with -m32 -O2, the test
case aborts.  When compiled with -m32, or with -m32 -O2 -fno-ipa-sra it runs to
completion.

I believe the bug is that ipa-sra leaves a call which refers to function
parameters using MEM_REF expressions which refer to fields of the parameters. 
These are translated into MEM expressions which load the address of the
parameter into a register and then use a MEM_REF to load an offset from that
register.  This MEM does not refer to the internal_arg_pointer, but to a
register which was set based on the internal_arg_pointer.  That means that
mem_overlaps_already_clobbered_arg_p in calls.c does not recognize that the MEM
refers to the argument area.  That in turn means that gcc thinks that it can
make a tail call, but when setting up the arguments for the call it clobbers
arguments that it has yet to read.

The fix may be to make mem_overlaps_already_clobbered_arg_p much smarter, or to
disable tail calls when applying SRA to a function and passes parameters to it.