Source of Ada bootstrap failure on x86

[kenner at vlsi1 dot ultra dot nyu dot edu (Richard Kenner)]

The miscompilation is in exp_ch4.adb:Expand_N_Not_In

Look at what happens to the return value of sinfo__etype.

The reason has to do with tail calls and /u.

Here's an excerpt from exp_ch4.adb.21.regmove:

(call_insn:HI 16 15 17 0 (set (reg:SI 0 ax)
        (call (mem:QI (symbol_ref:SI ("sinfo__etype") [flags 0x41] <function_decl 0x4050115c sinfo__etype>) [0 S1 A8])
            (const_int 4 [0x4]))) 499 {*call_value_0} (nil)
    (nil)
    (nil))

(insn:HI 17 16 19 0 (set (reg/v:SI 63 [ typ ])
        (reg:SI 0 ax)) 36 {*movsi_1} (insn_list 16 (nil))
    (expr_list:REG_DEAD (reg:SI 0 ax)
        (nil)))

....
(insn:HI 28 27 29 0 (set (mem/i:SI (plus:SI (reg/f:SI 7 sp)
                (const_int 4 [0x4])) [0 S4 A32])
        (reg:SI 59 [ T.1409 ])) 36 {*movsi_1} (insn_list 25 (nil))
    (expr_list:REG_DEAD (reg:SI 59 [ T.1409 ])
        (nil)))

...
(insn:HI 42 40 43 0 (set (mem/u/i:SI (plus:SI (reg/f:SI 7 sp)
                (const_int 4 [0x4])) [0 S4 A32])
        (reg/v:SI 63 [ typ ])) 36 {*movsi_1} (insn_list 17 (nil))
    (expr_list:REG_DEAD (reg/v:SI 63 [ typ ])
        (nil)))

(insn:HI 43 42 44 0 (set (mem/u/i:SI (reg/f:SI 7 sp) [0 S4 A32])
        (reg/v/u:SI 64 [ n ])) 36 {*movsi_1} (nil)
    (expr_list:REG_DEAD (reg/v/u:SI 64 [ n ])
        (nil)))

(call_insn:HI 44 43 45 0 (call (mem:QI (symbol_ref:SI ("sem_res__analyze_and_resolve__2") [flags 0x41] <function_decl 0x40505e0c sem_res__analyze_and_resolve__2>) [0 S1 A8])
        (const_int 8 [0x8])) 363 {*call_0} (nil)
    (nil)
    (nil))

Note that insn 27 sets 4(%esp) without /u while insn 42 sets it with /u.
These memrefs don't conflict and what ends up happening is that
insn 17 gets a REG_EQUIV note for 4(%exp).

This fixed by the following:

*** expr.c	20 Jul 2004 09:57:00 -0000	1.686
--- expr.c	21 Jul 2004 12:09:33 -0000
*************** emit_push_insn (rtx x, enum machine_mode
*** 3362,3367 ****
  		 outgoing arguments and we cannot allow reordering of reads
  		 from function arguments with stores to outgoing arguments
! 		 of sibling calls.  */
  	      set_mem_alias_set (dest, 0);
  	    }
  
--- 3362,3371 ----
  		 outgoing arguments and we cannot allow reordering of reads
  		 from function arguments with stores to outgoing arguments
! 		 of sibling calls.  Likewise, we can't allow this to be
! 		 marked as unchanging.
! 
! 		 ??? So what is the call above doing?  */
  	      set_mem_alias_set (dest, 0);
+ 	      RTX_UNCHANGING_P (dest) = 0;
  	    }
  
But I'm thinking it's best simply to remove the call to set_mem_attributes
here (with appropriate comment, of course).  What do you think?

I'm going to go back to x86_64, put in the kludge to get rid of the
variable-size issue, and see how ACATS stands: it was 154 failures yesterday.