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

["Uros Bizjak" <ubizjak@gmail.com>]

Hello!

Looking a bit further into the problem, I have noticed that actual
argument address becomes different from what CALL_INSN_FUNCTION_USAGE
claims in store_one_arg() function. The problem is in the call to
emit_push_insn() in line 4280. emit_push_insn() calls memory_address()
that fubars the address.

An (less intrusive?) variant of my previous patch is to change call to
memory_address in emit_push_insn() to memory_address_noforce(). I
can't see the benefit of

load reg <- (SP + off)
store (reg) <- regX

since IMO the address of the pushed argument won't be CSE'd with any
other address in any way.

BTW: I have tried to update arg->stack just after the call to
emit_push_insn in store_one_arg with SET_DEST (PATTERN (get_last_insn
())) to update CALL_INSN_FUNCTION_USAGE, but IMO, this is too hacky.

Uros.

Index: expr.c
===================================================================
--- expr.c      (revision 129152)
+++ expr.c      (working copy)
@@ -3910,13 +3910,12 @@ emit_push_insn (rtx x, enum machine_mode
 #endif
        {
          if (GET_CODE (args_so_far) == CONST_INT)
-           addr
-             = memory_address (mode,
-                               plus_constant (args_addr,
-                                              INTVAL (args_so_far)));
+           addr = plus_constant (args_addr, INTVAL (args_so_far));
          else
-           addr = memory_address (mode, gen_rtx_PLUS (Pmode, args_addr,
-                                                      args_so_far));
+           addr = gen_rtx_PLUS (Pmode, args_addr, args_so_far);
+
+         addr = memory_address_noforce (mode, addr);
+
          dest = gen_rtx_MEM (mode, addr);

          /* We do *not* set_mem_attributes here, because incoming arguments