Disable accumulate-outgoing-args for Generic and Buldozers

On Thu, Jan 23, 2014 at 4:39 PM, Jan Hubicka <hubicka@ucw.cz> wrote:
> Hi,
> this is improved patch I am testing.  The basic idea is to remove push
> expanders for cases where we do not have push instruction anyway.
> emit_move_insns then resorts to unconditonally call move expander
> with push operand.  I expended ix86_expand_vector_move to handle
> it gracefully and for that I borrowed emit_move_resolve_push
> function from expr.c since it seemed pointless to preserve
> duplicated logic in ix86_expand_push.
>
> I can easily imagine that scheduling around function call sequences
> matters, so I also updated push/pop expanders to preserve memory attributes.
>
> Eventually I found the attributes to be blank because of logic in expr.c
> that clears alias info when sibcall is enabled.  We can now do better
> by only disabling it in functions that actually do sibcalls.
>
> Bootstrap/regtest running on x86_64-linux, OK for the non-i386 parts
> if it passes?
>
> Honza
>
>         * expr.c (emit_move_resolve_push): Export; be bit more selective
>         on when to clear alias set.
>         * expr.h (emit_move_resolve_push): Declare.
>         * function.h (struct function): Add tail_call_marked.
>         * tree-tailcall.c (optimize_tail_call): Set tail_call_marked.
>         * config/i386/i386-protos.h (ix86_expand_push): Remove.
>         * config/i386/i386.md (TImode move expander): De not call
>         ix86_expand_push.
>         (FP push expanders): Preserve memory attributes.
>         * config/i386/sse.md (push<mode>1): Remove.
>         * config/i386/i386.c (ix86_expand_vector_move): Handle push
>         operation.
>         (ix86_expand_push): Remove.
>         * config/i386/mmx.md (push<mode>1): Remove.
...

> --- config/i386/i386.md (revision 206946)
> +++ config/i386/i386.md (working copy)
> @@ -1818,8 +1818,6 @@
>  {
>    if (TARGET_64BIT)
>      ix86_expand_move (TImode, operands);
> -  else if (push_operand (operands[0], TImode))
> -    ix86_expand_push (TImode, operands[1]);
>    else
>      ix86_expand_vector_move (TImode, operands);
>    DONE;
> @@ -2665,7 +2663,11 @@
>         (match_operand:TF 1 "sse_reg_operand"))]
>    "TARGET_SSE && reload_completed"
>    [(set (reg:P SP_REG) (plus:P (reg:P SP_REG) (const_int -16)))
> -   (set (mem:TF (reg:P SP_REG)) (match_dup 1))])
> +   (set (match_dup 0) (match_dup 1))]
> +{
> +  /* Preserve memory attributes. */
> +  operands[0] = replace_equiv_address (operands[0], stack_pointer_rtx);
> +})
>
>  (define_insn "*pushxf"
>    [(set (match_operand:XF 0 "push_operand" "=<,<")
> @@ -2691,8 +2693,12 @@
>         (match_operand:XF 1 "fp_register_operand"))]
>    "reload_completed"
>    [(set (reg:P SP_REG) (plus:P (reg:P SP_REG) (match_dup 2)))
> -   (set (mem:XF (reg:P SP_REG)) (match_dup 1))]
> -  "operands[2] = GEN_INT (-GET_MODE_SIZE (XFmode));")
> +   (set (match_dup 0) (match_dup 1))]
> +{
> +  operands[2] = GEN_INT (-GET_MODE_SIZE (XFmode));
> +  /* Preserve memory attributes. */
> +  operands[0] = replace_equiv_address (operands[0], stack_pointer_rtx);
> +})
>
>  (define_insn "*pushdf"
>    [(set (match_operand:DF 0 "push_operand" "=<,<,<,<")
> @@ -2713,7 +2719,11 @@
>         (match_operand:DF 1 "any_fp_register_operand"))]
>    "reload_completed"
>    [(set (reg:P SP_REG) (plus:P (reg:P SP_REG) (const_int -8)))
> -   (set (mem:DF (reg:P SP_REG)) (match_dup 1))])
> +   (set (match_dup 0) (match_dup 1))]
> +{
> +  /* Preserve memory attributes. */
> +  operands[0] = replace_equiv_address (operands[0], stack_pointer_rtx);
> +})
>
>  (define_insn "*pushsf_rex64"
>    [(set (match_operand:SF 0 "push_operand" "=X,X,X")
> @@ -2747,8 +2757,12 @@
>         (match_operand:SF 1 "any_fp_register_operand"))]
>    "reload_completed"
>    [(set (reg:P SP_REG) (plus:P (reg:P SP_REG) (match_dup 2)))
> -   (set (mem:SF (reg:P SP_REG)) (match_dup 1))]
> -  "operands[2] = GEN_INT (-<P:MODE_SIZE>);")
> +   (set (match_dup 0) (match_dup 1))]
> +{
> +  operands[2] = GEN_INT (-<P:MODE_SIZE>);
> +  /* Preserve memory attributes. */
> +  operands[0] = replace_equiv_address (operands[0], stack_pointer_rtx);
> +})
>

This piece code is wrong for x32:

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59929

I am testing:

> This piece code is wrong for x32:
> 
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59929

Though an independent bug ;)
> 
> I am testing:
> 
> diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md
> index ddc3be6..92e8fd0 100644
> --- a/gcc/config/i386/i386.md
> +++ b/gcc/config/i386/i386.md
> @@ -2765,7 +2765,20 @@
>    "reload_completed"
>    [(set (reg:P SP_REG) (plus:P (reg:P SP_REG) (match_dup 2)))
>     (set (mem:SF (reg:P SP_REG)) (match_dup 1))]
> -  "operands[2] = GEN_INT (-<P:MODE_SIZE>);")
> +{
> +  rtx op = XEXP (operands[0], 0);
> +  if (GET_CODE (op) == PRE_DEC)
> +    {
> +      gcc_assert (!TARGET_64BIT);
> +      op = GEN_INT (-4);
> +    }
> +  else
> +    {
> +      op = XEXP (XEXP (op, 1), 1);
> +      gcc_assert (CONST_INT_P (op));
> +    }
> +  operands[2] = op;
> +})

Not pretty, but I can not think of much more compact way, so OK
if it passes.
> 
>  (define_split
>    [(set (match_operand:SF 0 "push_operand")
> 
> 
> -- 
> H.J.