Re: [PATCH, rtl-optimization]: Fix PR rtl-optimization/33638

[Eric Botcazou <ebotcazou at libertysurf dot fr>]

> It does.  The problem is that CALL_INSN_FUNCTION_USAGE has a mem with
> virtual-outgoing-args, but the actual store is done via a reg.  The
> address was copied into the reg due to -fforce-addr.

No, I don't think it does.  Apply

Index: cselib.c
===================================================================
--- cselib.c    (revision 129158)
+++ cselib.c    (working copy)
@@ -1687,7 +1687,8 @@ cselib_process_insn (rtx insn)
   if (CALL_P (insn))
     {
       for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
-       if (call_used_regs[i]
+       if ((call_used_regs[i]
+            && !(ACCUMULATE_OUTGOING_ARGS && i == STACK_POINTER_REGNUM))
            || (REG_VALUES (i) && REG_VALUES (i)->elt
                && HARD_REGNO_CALL_PART_CLOBBERED (i,
                      GET_MODE (REG_VALUES (i)->elt->val_rtx))))

and comment out the special code in dse.c for const functions that invalidates
all the stores into the frame.  Now compile the testcase without -fforce-addr 
and you'll get in cse2:

(insn 1039 1038 1040 149 comunpack.f:61 (set (mem:SF (plus:SI (reg/f:SI 7 sp)
                (const_int 4 [0x4])) [0 S4 A32])
        (reg:SF 374)) 66 {*movsf_1} (expr_list:REG_DEAD (reg:SF 374)
        (nil)))

[...]

(call_insn/u 1042 1041 1043 149 comunpack.f:61 (set (reg:SF 8 st)
        (call (mem:QI (symbol_ref:SI ("powf") [flags 0x41] <function_decl 
0x55704200 __builtin_powf>) [0 S1 A8])
            (const_int 8 [0x8]))) 611 {*call_value_0} (nil)
    (expr_list:REG_DEP_TRUE (use (mem/i:SF (reg/f:SI 7 sp) [0 S4 A32]))
        (expr_list:REG_DEP_TRUE (use (mem/i:SF (plus:SI (reg/f:SI 7 sp)
                        (const_int 4 [0x4])) [0 S4 A32]))
            (nil))))

[...]

(insn 1048 1047 1049 149 comunpack.f:62 (set (mem:SF (plus:SI (reg/f:SI 7 sp)
                (const_int 4 [0x4])) [0 S4 A32])
        (reg:SF 379)) 66 {*movsf_1} (expr_list:REG_DEAD (reg:SF 379)
        (nil)))


And then in dse1:

**scanning insn=1048
  mem: (plus:SI (reg/f:SI 7 sp)
    (const_int 4 [0x4]))

   after cselib_expand address: (plus:SI (reg/f:SI 7 sp)
    (const_int 4 [0x4]))

   after canon_rtx address: (plus:SI (reg/f:SI 7 sp)
    (const_int 4 [0x4]))
  varying cselib base=2 offset = 4
 processing cselib store [4..8)
    trying store in insn=1041 gid=-1[0..4)
    trying store in insn=1039 gid=-1[4..8)
Locally deleting insn 1039 
deferring deletion of insn with uid = 1039.
mems_found = 1, cannot_delete = false
[...]
deleting insn with uid = 1039.

and yet there is an explicit use for the store in CALL_INSN_FUNCTION_USAGE.

> dse.c is not smart enough to see that the address in the
> CALL_INSN_FUNCTION_USAGE is the same as the address in the register.

I think that dse.c doesn't even look into CALL_INSN_FUNCTION_USAGE.

-- 
Eric Botcazou