Re: new syscall stub support for ia64 libc

[David Mosberger <davidm at napali dot hpl dot hp dot com>]

>>>>> On Thu, 13 Nov 2003 21:54:04 -0800, Ulrich Drepper <drepper@redhat.com> said:

  >> Ah, perhaps ar.pfs didn't get preserved.  Earlier versions of GCC
  >> ignored clobbers to ar.pfs, but I don't remember exactly when it
  >> was fixed.

  Uli> Looks like it.

Bingo!
With gcc-3.3.2, it looks like this:

0000000000000000 <__GI___fxstat>:
   0:   00 18 31 08 80 05       [MII]       alloc r35=ar.pfs,12,4,0
   6:   f0 e0 01 12 48 c0                   mov r15=1212
   c:   81 68 00 84                         adds r14=8,r13
  10:   02 28 01 44 00 21       [MII]       mov r37=r34
  16:   40 02 84 2c 00 00                   sxt4 r36=r33;;
  1c:   00 00 04 00                         nop.i 0x0
  20:   0b 70 00 1c 18 10       [MMI]       ld8 r14=[r14];;
  26:   00 00 00 02 00 e0                   nop.m 0x0
  2c:   e0 08 00 07                         mov b7=r14;;
  30:   1d 00 00 00 01 00       [MFB]       nop.m 0x0
  36:   00 00 00 02 00 c0                   nop.f 0x0
  3c:   78 00 80 10                         br.call.sptk.many b6=b7;;
  40:   03 70 00 02 00 24       [MII]       addl r14=0,r1
  46:   00 18 01 55 00 e0                   mov.i ar.pfs=r35;;
  4c:   f0 57 18 ec                         cmp.eq p7,p6=-1,r10;;
  50:   eb 70 00 1c 18 d0       [MMI] (p07) ld8 r14=[r14];;
  56:   e1 70 34 00 40 00             (p07) add r14=r14,r13
  5c:   00 00 04 00                         nop.i 0x0;;
  60:   f1 00 20 1c 90 d1       [MIB] (p07) st4 [r14]=r8
  66:   81 f8 f3 ff 4f 80             (p07) mov r8=-1
  6c:   08 00 84 00                         br.ret.sptk.many b0;;

  Uli> The compiler used is gcc 3.2 based.

OK, it may well be that it was fixed between 3.2 and 3.3.  Let me see
if I can track down the exact patch.

Ah, it's this one:

2003-04-25  Richard Henderson  <rth@redhat.com>

        * config/ia64/ia64.c (ia64_compute_frame_size): Allow inline asm
        to clobber ar.pfs and ar.unat.
        (ia64_expand_prologue): Force alloc instruction if ar.pfs saved;
        fix test for spilling ar.pfs to the stack.

I attached the diff below.

	--david

Index: config/ia64/ia64.c
===================================================================
RCS file: /cvsroot/gcc/gcc/gcc/config/ia64/ia64.c,v
retrieving revision 1.220
retrieving revision 1.221
diff -u -r1.220 -r1.221
--- config/ia64/ia64.c	24 Apr 2003 17:23:52 -0000	1.220
+++ config/ia64/ia64.c	25 Apr 2003 21:02:25 -0000	1.221
@@ -1878,6 +1878,17 @@
 	  spill_size += 8;
 	  n_spilled += 1;
 	}
+
+      if (regs_ever_live[AR_PFS_REGNUM])
+	{
+	  SET_HARD_REG_BIT (mask, AR_PFS_REGNUM);
+	  current_frame_info.reg_save_ar_pfs = find_gr_spill (1);
+	  if (current_frame_info.reg_save_ar_pfs == 0)
+	    {
+	      extra_spill_size += 8;
+	      n_spilled += 1;
+	    }
+	}
     }
 
   /* Unwind descriptor hackery: things are most efficient if we allocate
@@ -1916,8 +1927,10 @@
     }
 
   /* If we're forced to use st8.spill, we're forced to save and restore
-     ar.unat as well.  */
-  if (spilled_gr_p || cfun->machine->n_varargs)
+     ar.unat as well.  The check for existing liveness allows inline asm
+     to touch ar.unat.  */
+  if (spilled_gr_p || cfun->machine->n_varargs
+      || regs_ever_live[AR_UNAT_REGNUM])
     {
       regs_ever_live[AR_UNAT_REGNUM] = 1;
       SET_HARD_REG_BIT (mask, AR_UNAT_REGNUM);
@@ -2378,7 +2391,8 @@
   /* We don't need an alloc instruction if we've used no outputs or locals.  */
   if (current_frame_info.n_local_regs == 0
       && current_frame_info.n_output_regs == 0
-      && current_frame_info.n_input_regs <= current_function_args_info.int_regs)
+      && current_frame_info.n_input_regs <= current_function_args_info.int_regs
+      && !TEST_HARD_REG_BIT (current_frame_info.mask, AR_PFS_REGNUM))
     {
       /* If there is no alloc, but there are input registers used, then we
 	 need a .regstk directive.  */
@@ -2540,8 +2554,8 @@
   /* The alloc insn already copied ar.pfs into a general register.  The
      only thing we have to do now is copy that register to a stack slot
      if we'd not allocated a local register for the job.  */
-  if (current_frame_info.reg_save_ar_pfs == 0
-      && ! current_function_is_leaf)
+  if (TEST_HARD_REG_BIT (current_frame_info.mask, AR_PFS_REGNUM)
+      && current_frame_info.reg_save_ar_pfs == 0)
     {
       reg = gen_rtx_REG (DImode, AR_PFS_REGNUM);
       do_spill (gen_movdi_x, ar_pfs_save_reg, cfa_off, reg);