[i386] : Workaround PR 63620 by allowing only direct immediate values in FP push patterns

Hello!

Attached patch works around PR 63620. Patched gcc will allow only a
subset of CONST_DOUBLEs in FP push patterns. This subset will result
in direct immediates that can always be encoded directly in the insn
stream. These won't be reloaded through memory, and will avoid PIC
register re-materialization issues. OTOH, the patch will generate
non-optimal push sequences when general registers are available.

2014-10-31  Uros Bizjak  <ubizjak@gmail.com>

    PR target/63620
    * config/i386/i386.md (*pushtf): Allow only CONST_DOUBLES that won't
    be reloaded through memory.
    (*pushxf): Ditto.
    (*pushdf): Ditto.

The patch was bootstrapped and regression tested on x86_64-linux-gnu
{-m32}. The patch was also checked with Darwin crosscompiler on the
testcases in the PR.

Committed to mainline.

Uros

On Fri, Oct 31, 2014 at 08:45:54PM +0100, Uros Bizjak wrote:
> Attached patch works around PR 63620. Patched gcc will allow only a
> subset of CONST_DOUBLEs in FP push patterns. This subset will result
> in direct immediates that can always be encoded directly in the insn
> stream. These won't be reloaded through memory, and will avoid PIC
> register re-materialization issues. OTOH, the patch will generate
> non-optimal push sequences when general registers are available.
> 
> 2014-10-31  Uros Bizjak  <ubizjak@gmail.com>
> 
>     PR target/63620
>     * config/i386/i386.md (*pushtf): Allow only CONST_DOUBLES that won't
>     be reloaded through memory.
>     (*pushxf): Ditto.
>     (*pushdf): Ditto.
> 
> The patch was bootstrapped and regression tested on x86_64-linux-gnu
> {-m32}. The patch was also checked with Darwin crosscompiler on the
> testcases in the PR.

Can't you limit the workaround to ix86_use_pseudo_pic_reg () only
if it is about PIC register changes?

> --- config/i386/i386.md	(revision 216983)
> +++ config/i386/i386.md	(working copy)
> @@ -2733,10 +2733,14 @@
>  
>  ;; Floating point push instructions.
>  
> +;; %%% Remove CONST_DOUBLE workaround after PR63620 is fixed!
>  (define_insn "*pushtf"
>    [(set (match_operand:TF 0 "push_operand" "=<,<")
>  	(match_operand:TF 1 "general_no_elim_operand" "x,*roF"))]
> -  "TARGET_64BIT || TARGET_SSE"
> +  "(TARGET_64BIT || TARGET_SSE)
> +   && (!can_create_pseudo_p ()
> +       || GET_CODE (operands[1]) != CONST_DOUBLE
> +       || standard_sse_constant_p (operands[1]))"
>  {
>    /* This insn should be already split before reg-stack.  */
>    gcc_unreachable ();
> @@ -2758,10 +2762,13 @@
>    operands[0] = replace_equiv_address (operands[0], stack_pointer_rtx);
>  })
>  
> +;; %%% Remove CONST_DOUBLE workaround after PR63620 is fixed!
>  (define_insn "*pushxf"
>    [(set (match_operand:XF 0 "push_operand" "=<,<")
>  	(match_operand:XF 1 "general_no_elim_operand" "f,Yx*roF"))]
> -  ""
> +  "!can_create_pseudo_p ()
> +   || GET_CODE (operands[1]) != CONST_DOUBLE
> +   || standard_80387_constant_p (operands[1]) > 0"
>  {
>    /* This insn should be already split before reg-stack.  */
>    gcc_unreachable ();
> @@ -2789,10 +2796,16 @@
>    operands[0] = replace_equiv_address (operands[0], stack_pointer_rtx);
>  })
>  
> +;; %%% Remove CONST_DOUBLE workaround after PR63620 is fixed!
>  (define_insn "*pushdf"
>    [(set (match_operand:DF 0 "push_operand" "=<,<,<,<")
>  	(match_operand:DF 1 "general_no_elim_operand" "f,Yd*roF,rmF,x"))]
> -  ""
> +  "!can_create_pseudo_p ()
> +   || GET_CODE (operands[1]) != CONST_DOUBLE
> +   || (!(TARGET_SSE2 && TARGET_SSE_MATH)
> +       && standard_80387_constant_p (operands[1]) > 0)
> +   || (TARGET_SSE2 && TARGET_SSE_MATH
> +       && standard_sse_constant_p (operands[1]))"
>  {
>    /* This insn should be already split before reg-stack.  */
>    gcc_unreachable ();


	Jakub