gcc/ia64 codegenerator questions

[robert.muth@intel.com]

My group is working on a tool to optimize ia64 binaries.
It already works with images produced by intel compilers.
and I am trying to make it work for gcc as well.
I am using gcc 3.1 so my observations might already be outdated.


* break instructions  after non returning calls

  I believe that gcc tries to put a break instruction
  after a call to functions like abort.
  This is very useful to us as it keeps us from having
  "impossible" control edges in our cfg.
  I think this also necessary for "unwind" to work correctly.
  the problem is that with -O3 the break seems to get optimized
  away. This can be reproduced with the following simple program:

  #include  <stdlib.h>
  void do_abort ()
  {
   abort ();
  }

* delayed saving of ar.lc

  In the code snippet below the prolog ends rather late
  and even includes a call. Is this intentional or
  is this a bug?


        .endp efloor#
        .align 16
        .proc eldexp#
eldexp:
        .prologue 12, 36
        .mii
        .save ar.pfs, r37
        alloc r37 = ar.pfs, 3, 4, 5, 0
        .fframe 32
        adds r12 = -32, r12
        ;;
        adds r35 = 18, r12
        .mmi
        mov r39 = r32
        adds r40 = 16, r12
        mov r32 = r1
        .mib
        nop.m 0
        .save rp, r36
        mov r36 = b0
        br.call.sptk.many b0 = emovi#
        ;;
        .mii
        ld2 r8 = [r35]
        mov r1 = r32
        adds r39 = 16, r12
        .mmb
        mov r40 = r0
        mov r41 = r0
        nop.b 0
        .mii
        addl r43 = 64, r0
        .save ar.lc, r38
        mov r38 = ar.lc
        .body
        ;;
        add r3 = r33, r8
        ;;
        .mib
        mov r42 = r3
        nop.i 0


* tail calls

  It seems to me that tail calls are a little bit fragile
  with respect to "unwinding".
  In the code snippet below the alloc instruction
  just before the actual tail call will hide the registers
  r40 and r39. If I set a breakpoint at the last nop.i
  in the debugger and try to unwind from there I might
  have a problem.


  .L333:
        .mii
        addl r24 = @gprel(subflg#), gp
        addl r25 = 1, r0
        mov r32 = r37
        .mii
        nop.m 0
        mov ar.pfs = r39
        ;;
        mov ar.lc = r40
        .mii
        st4 [r24] = r25
        mov b0 = r38
        ;;
        nop.i 0
        .mib
        alloc r2 = ar.pfs, 0, 0, 3, 0
        nop.i 0
        br.sptk.many eadd1#
        ;;

* slotting of calls

  this is more question than a problem:

  Does gcc always place call instructions in the last slot
  of a bundle -- call shadows are a nasty problem to deal
  with on our side.

Robert