[Bug target/52352] New: [x32] - Wrong code to access addresses 0x80000000 to 0xFFFFFFFF using registers

["steffen-schmidt at siemens dot com" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52352

             Bug #: 52352
           Summary: [x32] - Wrong code to access addresses 0x80000000 to
                    0xFFFFFFFF using registers
    Classification: Unclassified
           Product: gcc
           Version: 4.6.3
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: steffen-schmidt@siemens.com


Created attachment 26727
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=26727
C code resulting in wrong instructions when compiled with -mx32 and -O2 or
higher

The example C code defines a structure in memory at address 0xFEE00000
Compilation of this example with -mx32 and -O2 or higher (seems to be related
to -fgcse) results in faulty assembler instructions.

For accessing several members of the structure loads the signed address of one
member of the structure (-18874360 = 0xFEE00008) into register %rax. This
actually results in a 64bit negative address of 0xFFFFFFFFFEE00008, which is
incorrect when later using complete %rax register for memory access.

// -mx32 -O3
movq    $-18874360, %rax
movl    (%rax), %edx
xorb    %al, %al
movl    %edx, 4(%eax)

When optimizing with -O1 register %eax is used instead of %rax which results in
correct behaviour:

// -mx32 -O1
movl    $-18874368, %eax
movl    8(%eax), %edx
movl    %edx, 4(%eax)


The x64 compiler produces correct code with -O1 and -O3 loading the address
into %eax not %rax:

// -m64 -O1
movl    $4276092928, %eax
movl    8(%rax), %edx
movl    %edx, 4(%rax)

// -m64 -O3
movl    $4276092936, %eax
movl    (%rax), %edx
xorb    %al, %al
movl    %edx, 4(%rax)