diff mbox

Fix PR target/62631

Message ID 15644575.zb4QZkAvgm@polaris
State New
Headers show

Commit Message

Eric Botcazou Feb. 5, 2015, 3:53 p.m. UTC 
Hi,

this is the failure of gcc.dg/tree-ssa/ivopts-lt-2.c on SPARC 64-bit, but I 
think that the problem is actually generic: get_shiftadd_cost only calls 
shiftadd_cost to get the cost of a "shift-add", which is actually a multiply 
by a power-of-2 operation.  Since SPARC doesn't have a lea-like instruction, 
shiftadd_cost returns a very high cost, about 3 times the cost of an add+shift 
sequence which is what the RTL expander will generate for this operation.

So I think that get_shiftadd_cost should mimic the algorithms in expmed.c and 
use the minimum of shiftadd_cost and (add_cost + shift_cost) for the cost of 
the shift-add operation.

Tested on x86-64, SPARC and SPARC64, OK for the mainline?


2015-02-05  Eric Botcazou  <ebotcazou@adacore.com>

	PR target/62631
	* tree-ssa-loop-ivopts.c (get_shiftadd_cost): Use the mininum of costs
	of shift-add and (add + shift) operations.  Rename local variable.

Comments

Jeff Law Feb. 6, 2015, 7:56 a.m. UTC | #1 
On 02/05/15 08:53, Eric Botcazou wrote:
> Hi,
>
> this is the failure of gcc.dg/tree-ssa/ivopts-lt-2.c on SPARC 64-bit, but I
> think that the problem is actually generic: get_shiftadd_cost only calls
> shiftadd_cost to get the cost of a "shift-add", which is actually a multiply
> by a power-of-2 operation.  Since SPARC doesn't have a lea-like instruction,
> shiftadd_cost returns a very high cost, about 3 times the cost of an add+shift
> sequence which is what the RTL expander will generate for this operation.
>
> So I think that get_shiftadd_cost should mimic the algorithms in expmed.c and
> use the minimum of shiftadd_cost and (add_cost + shift_cost) for the cost of
> the shift-add operation.
>
> Tested on x86-64, SPARC and SPARC64, OK for the mainline?
>
>
> 2015-02-05  Eric Botcazou  <ebotcazou@adacore.com>
>
> 	PR target/62631
> 	* tree-ssa-loop-ivopts.c (get_shiftadd_cost): Use the mininum of costs
> 	of shift-add and (add + shift) operations.  Rename local variable.
OK

jeff
diff mbox

Patch

--- tree-ssa-loop-ivopts.c      (revision 220343)
+++ tree-ssa-loop-ivopts.c      (working copy)
@@ -3597,22 +3597,26 @@  get_shiftadd_cost (tree expr, machine_mo
   tree multop = TREE_OPERAND (mult, 0);
   int m = exact_log2 (int_cst_value (cst));
   int maxm = MIN (BITS_PER_WORD, GET_MODE_BITSIZE (mode));
-  int sa_cost;
-  bool equal_p = false;
+  int as_cost, sa_cost;
+  bool mult_in_op1;
 
   if (!(m >= 0 && m < maxm))
     return false;
 
-  if (operand_equal_p (op1, mult, 0))
-    equal_p = true;
+  mult_in_op1 = operand_equal_p (op1, mult, 0);
 
+  as_cost = add_cost (speed, mode) + shift_cost (speed, mode, m);
+
+  /* If the target has a cheap shift-and-add or shift-and-sub instruction,
+     use that in preference to a shift insn followed by an add insn.  */
   sa_cost = (TREE_CODE (expr) != MINUS_EXPR
              ? shiftadd_cost (speed, mode, m)
-             : (equal_p
+             : (mult_in_op1
                 ? shiftsub1_cost (speed, mode, m)
                 : shiftsub0_cost (speed, mode, m)));
-  res = new_cost (sa_cost, 0);
-  res = add_costs (res, equal_p ? cost0 : cost1);
+
+  res = new_cost (MIN (as_cost, sa_cost), 0);
+  res = add_costs (res, mult_in_op1 ? cost0 : cost1);
 
   STRIP_NOPS (multop);
   if (!is_gimple_val (multop))