[Bug rtl-optimization/34999] Fallthru crossing edges in partition_hot_cold_basic_blocks are not been fixed when the section ends with call insn

["jakub at gcc dot gnu dot org" <gcc-bugzilla@gcc.gnu.org>]

------- Comment #15 from jakub at gcc dot gnu dot org  2008-02-29 08:56 -------
.LSFDE3:
        .long   .LEFDE3-.LASFDE3        # FDE Length
.LASFDE3:
        .long   .LASFDE3-.Lframe1       # FDE CIE offset
        .long   .LHOTB2 # FDE initial location
        .long   .LHOTE2-.LHOTB2 # FDE address range
        .long   .LCOLDB2        # FDE initial location
        .long   .LCOLDE2-.LCOLDB2       # FDE address range
        .uleb128 0x0    # Augmentation size
        .byte   0x4     # DW_CFA_advance_loc4
        .long   .LCFI2-.LFB3
        .byte   0xe     # DW_CFA_def_cfa_offset
        .uleb128 0x10
        .byte   0x86    # DW_CFA_offset, column 0x6
        .uleb128 0x2
        .byte   0x4     # DW_CFA_advance_loc4
        .long   .LCFI3-.LCFI2
        .byte   0xd     # DW_CFA_def_cfa_register
        .uleb128 0x6
        .align 8
.LEFDE3:

is not a valid .eh_frame FDE, the
2005-03-31  Caroline Tice  <ctice@apple.com>

        (output_call_frame_info): Add test to see if function switches text
        sections in the middle; if so, use appropriate extra hot and cold
        section labels to compute size deltas for the hot and cold sections.

change was totally broken.  There is no such thing as a two ranges FDE in
Dwarf2/3.  FDE starts with FDE length, CIE offset, FDE initial location, FDE
address range, depending on CIE optionally with augmentation size and
augmentation and then CFA instructions.  So, to describe .text/.text.unlikely
split function, you need two FDEs, one describing the .text section, the other
describing .text.unlikely section.  When the code jumps in between the
sections, you need to insert CFA instructions that reflect what really changed.
The easiest would be to use .cfi_* assembler directives that recentish gas
supports and emitting them inline in the code, rather than creating separate
.eh_frame.  But I believe we need some more help from gas - particularly more
powerful .cfi_escape and some easy way to just save current state resp. restore
it.  That way you could save the state at the jump location and restore it at
the jump target and let gas compute what CFA instructions are needed.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34999