Re: [GCC][PATCH][mid-end] Optimize x * copysign (1.0, y) [Patch (1/2)]

[Andrew Pinski <pinskia@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 3289 bytes --]

On Sat, Jun 24, 2017 at 4:53 PM, Andrew Pinski <pinskia@gmail.com> wrote:
> On Mon, Jun 12, 2017 at 12:56 AM, Tamar Christina
> <Tamar.Christina@arm.com> wrote:
>> Hi All,
>>
>> this patch implements a optimization rewriting
>>
>> x * copysign (1.0, y) and
>> x * copysign (-1.0, y)
>
>
> This reminds me:
> copysign(-1.0, y) can be just optimized to:
> copysign(1.0, y)
>
> I did that in my patch here:
> https://gcc.gnu.org/ml/gcc-patches/2017-06/msg01860.html

I updated the patch to handle all constants and not just -1.0.

>
> This should allow you to reduce the number of patterns needed to match here.
> Note I still think we could do this in expand without a new
> builtin/internal function.
> I might go and code that up soonish.

Also something like attached (NOTE this is NOT a full patch and needs
the xorsign optabs part of your patch) should work for the expand side
rather than creating a new builtin.  There still needs to handling of
the vector based copysign.  But you should get the general idea.  I
would like to see more of these special expand patterns really.

NOTE you can remove the target hook part and just check if xorsign
optab is there.  I don't know if that is what we want to do if not
allow for generic expanding of this.

Thanks,
Andrew Pinski


>
> Thanks,
> Andrew
>
>>
>> to:
>>
>> x ^ (y & (1 << sign_bit_position))
>>
>> This is done by creating a special builtin during matching and generate the
>> appropriate instructions during expand. This new builtin is called XORSIGN.
>>
>> The expansion of xorsign depends on if the backend has an appropriate optab
>> available. If this is not the case then we use a modified version of the existing
>> copysign which does not take the abs value of the first argument as a fall back.
>>
>> This patch is a revival of a previous patch
>> https://gcc.gnu.org/ml/gcc-patches/2015-10/msg00069.html
>>
>> Bootstrapped on both aarch64-none-linux-gnu and x86_64 with no issues.
>> Regression done on aarch64-none-linux-gnu and no regressions.
>>
>> Ok for trunk?
>>
>> gcc/
>> 2017-06-07  Tamar Christina  <tamar.christina@arm.com>
>>
>>         * builtins.def (BUILT_IN_XORSIGN, BUILT_IN_XORSIGNF): New.
>>         (BUILT_IN_XORSIGNL, BUILT_IN_XORSIGN_FLOAT_NX): Likewise.
>>         * match.pd (mult (COPYSIGN:s real_onep @0) @1): New simplifier.
>>         (mult (COPYSIGN:s real_mus_onep @0) @1): Likewise.
>>         (copysigns @0 (negate @1)): Likewise.
>>         * builtins.c (expand_builtin_copysign): Promoted local to argument.
>>         (expand_builtin): Added CASE_FLT_FN_FLOATN_NX (BUILT_IN_XORSIGN) and
>>         CASE_FLT_FN (BUILT_IN_XORSIGN).
>>         (BUILT_IN_COPYSIGN): Updated function call.
>>         * optabs.h (expand_copysign): New bool.
>>         (expand_xorsign): New.
>>         * optabs.def (xorsign_optab): New.
>>         * optabs.c (expand_copysign): New parameter.
>>         * fortran/f95-lang.c (xorsignl, xorsign, xorsignf): New.
>>         * fortran/mathbuiltins.def (XORSIGN): New.
>>
>> gcc/testsuite/
>> 2017-06-07  Tamar Christina  <tamar.christina@arm.com>
>>
>>         * gcc.dg/tree-ssa/xorsign.c: New.
>>         * gcc.dg/xorsign_exec.c: New.
>>         * gcc.dg/vec-xorsign_exec.c: New.
>>         * gcc.dg/tree-ssa/reassoc-39.c (f2, f3): Updated constant to 2.

[-- Attachment #2: startmult-copysign.diff.txt --]
[-- Type: text/plain, Size: 2493 bytes --]

Index: gcc/expr.c
===================================================================
--- gcc/expr.c	(revision 249619)
+++ gcc/expr.c	(working copy)
@@ -8182,6 +8182,59 @@
   return NULL_RTX;
 }
 
+static bool
+is_copysign_call_with_1 (gimple *call)
+{
+  if (!is_gimple_call (call))
+    return false;
+
+  if (gimple_call_builtin_p (call, BUILT_IN_NORMAL))
+    {
+      gcall *c = as_a<gcall*> (call);
+      tree decl = gimple_call_fndecl (call);
+      switch (DECL_FUNCTION_CODE (decl))
+	{
+        CASE_FLT_FN (BUILT_IN_COPYSIGN):
+	CASE_FLT_FN_FLOATN_NX (BUILT_IN_COPYSIGN):
+	  return real_one_p (gimple_call_arg (c, 0));
+	default:
+	  return false;
+	}
+    }
+}
+
+static rtx
+maybe_expand_mult_copysign (tree treeop0, tree treeop1, rtx target)
+{
+  tree type = TREE_TYPE (treeop0);
+  rtx op0, op1;
+
+  if (!SCALAR_FLOAT_TYPE_P (type)
+      && VECTOR_FLOAT_TYPE_P (type))
+    return NULL;
+
+  if (HONOR_SNANS (type))
+    return NULL;
+
+  if (!targetm.expand_mult_copysign_xor ())
+    return NULL;
+
+  if (TREE_CODE (treeop0) == SSA_NAME)
+    {
+      gimple *call0 = SSA_NAME_DEF_STMT (treeop0);
+      if (is_copysign_call_with_1 (call0))
+	{
+	  gcall *c = as_a<gcall*> (call0);
+	  treeop0 = gimple_call_arg (c, 1);
+	  expand_operands (treeop1, treeop0, NULL_RTX, &op0, &op1, EXPAND_NORMAL);
+	  return expand_copysign (op0, op1, target, true);
+	}
+    }
+
+  return NULL;
+}
+
+
 rtx
 expand_expr_real_2 (sepops ops, rtx target, machine_mode tmode,
 		    enum expand_modifier modifier)
@@ -8791,6 +8844,10 @@
       if (modifier == EXPAND_STACK_PARM)
 	target = 0;
 
+      temp = maybe_expand_mult_copysign (treeop0, treeop1);
+      if (temp)
+	return temp;
+
       expand_operands (treeop0, treeop1, subtarget, &op0, &op1, EXPAND_NORMAL);
       return REDUCE_BIT_FIELD (expand_mult (mode, op0, op1, target, unsignedp));
 
Index: gcc/target.def
===================================================================
--- gcc/target.def	(revision 249619)
+++ gcc/target.def	(working copy)
@@ -2640,6 +2640,13 @@
  default_have_conditional_execution)
 
 DEFHOOK
+(expand_mult_copysign_xor,
+ "This target hook returns true if the target wants to use xor's\n
+to expand x * copysign (1., y) as x ^ (y & (1 << sign_bit_position)).\n"
+ bool, (void),
+ hook_bool_void_false)
+
+DEFHOOK
 (gen_ccmp_first,
  "This function prepares to emit a comparison insn for the first compare in a\n\
  sequence of conditional comparisions.  It returns an appropriate comparison\n\