[tree-complex.c] PR tree-optimization/70291: Inline floating-point complex multiplication more aggressively

Message ID 5AE75555.6060303@foss.arm.com
State New
Headers show
Series
  • [tree-complex.c] PR tree-optimization/70291: Inline floating-point complex multiplication more aggressively
Related show

Commit Message

Kyrill Tkachov April 30, 2018, 5:41 p.m.
Hi all,

We can improve the performance of complex floating-point multiplications by inlining the expansion a bit more aggressively.
We can inline complex x = a * b as:
x = (ar*br - ai*bi) + i(ar*bi + br*ai);
if (isunordered (__real__ x, __imag__ x))
   x = __muldc3 (a, b); //Or __mulsc3 for single-precision

That way the common case where no NaNs are produced we can avoid the libgcc call and fall back to the
NaN handling stuff in libgcc if either components of the expansion are NaN.

The implementation is done in expand_complex_multiplication in tree-complex.c and the above expansion
will be done when optimising for -O1 and greater and when not optimising for size.
At -O0 and -Os the single call to libgcc will be emitted.

For the code:
__complex double
foo (__complex double a, __complex double b)
{
   return a * b;
}

We will now emit at -O2 for aarch64:
foo:
         fmul    d16, d1, d3
         fmul    d6, d1, d2
         fnmsub  d5, d0, d2, d16
         fmadd   d4, d0, d3, d6
         fcmp    d5, d4
         bvs     .L8
         fmov    d1, d4
         fmov    d0, d5
         ret
.L8:
         stp     x29, x30, [sp, -16]!
         mov     x29, sp
         bl      __muldc3
         ldp     x29, x30, [sp], 16
         ret

Instead of just a branch to __muldc3.

Bootstrapped and tested on aarch64-none-linux-gnu, arm-none-linux-gnueabihf, x86_64-unknown-linux-gnu.

Ok for trunk? (GCC 9)

Thanks,
Kyrill

2018-04-30  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>

     PR tree-optimization/70291
     * tree-complex.c (insert_complex_mult_libcall): New function.
     (expand_complex_multiplication_limited_range): Likewise.
     (expand_complex_multiplication): Expand floating-point complex
     multiplication using the above.

2018-04-30  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>

     PR tree-optimization/70291
     * gcc.dg/complex-6.c: New test.
     * gcc.dg/complex-7.c: Likewise.

Comments

Richard Biener May 2, 2018, 10:15 a.m. | #1
On Mon, Apr 30, 2018 at 7:41 PM, Kyrill  Tkachov
<kyrylo.tkachov@foss.arm.com> wrote:
> Hi all,

>

> We can improve the performance of complex floating-point multiplications by

> inlining the expansion a bit more aggressively.

> We can inline complex x = a * b as:

> x = (ar*br - ai*bi) + i(ar*bi + br*ai);

> if (isunordered (__real__ x, __imag__ x))

>   x = __muldc3 (a, b); //Or __mulsc3 for single-precision

>

> That way the common case where no NaNs are produced we can avoid the libgcc

> call and fall back to the

> NaN handling stuff in libgcc if either components of the expansion are NaN.

>

> The implementation is done in expand_complex_multiplication in

> tree-complex.c and the above expansion

> will be done when optimising for -O1 and greater and when not optimising for

> size.

> At -O0 and -Os the single call to libgcc will be emitted.

>

> For the code:

> __complex double

> foo (__complex double a, __complex double b)

> {

>   return a * b;

> }

>

> We will now emit at -O2 for aarch64:

> foo:

>         fmul    d16, d1, d3

>         fmul    d6, d1, d2

>         fnmsub  d5, d0, d2, d16

>         fmadd   d4, d0, d3, d6

>         fcmp    d5, d4

>         bvs     .L8

>         fmov    d1, d4

>         fmov    d0, d5

>         ret

> .L8:

>         stp     x29, x30, [sp, -16]!

>         mov     x29, sp

>         bl      __muldc3

>         ldp     x29, x30, [sp], 16

>         ret

>

> Instead of just a branch to __muldc3.

>

> Bootstrapped and tested on aarch64-none-linux-gnu, arm-none-linux-gnueabihf,

> x86_64-unknown-linux-gnu.

>

> Ok for trunk? (GCC 9)


+         /* If optimizing for size or not at all just do a libcall.  */
+         if (optimize == 0 || optimize_function_for_size_p (cfun))
+           {
+             expand_complex_libcall (gsi, ar, ai, br, bi, MULT_EXPR);
+             return;
+           }

use optimize_bb_for_size_p instead please (get BB from the mult stmt).

 /* Expand a complex multiplication or division to a libcall to the c99
+   compliant routines.  Unlike expand_complex_libcall create and insert
+   the call, assign it to an output variable and return that rather than
+   modifying existing statements in place.  */
+
+static tree
+insert_complex_mult_libcall (gimple_stmt_iterator *gsi, tree type, tree ar,
+                             tree ai, tree br, tree bi)
+{

can you please try merging both functions instead?

Also it shows a possible issue if with -fnon-call-exceptions the original
multiplication has EH edges.  I think you want to side-step that
by doing the libcall-only way in that case as well (stmt_can_throw_internal).

+         tree isunordered_decl = builtin_decl_explicit (BUILT_IN_ISUNORDERED);
+         tree isunordered_res = create_tmp_var (integer_type_node);
+         gimple *tmpr_unord_check
+           = gimple_build_call (isunordered_decl, 2, tmpr, tmpi);
+         gimple_call_set_lhs (tmpr_unord_check, isunordered_res);
+
+         gsi_insert_before (gsi, tmpr_unord_check, GSI_SAME_STMT);
+         gimple *check
+           = gimple_build_cond (NE_EXPR, isunordered_res, integer_zero_node,
+                                NULL_TREE, NULL_TREE);

why use BUILT_IN_ISUNORDERED but not a GIMPLE_COND with
UNORDERED_EXPR?  Note again that might trap/throw with -fsignalling-nans
so better avoid this transform for flag_signalling_nans as well...

+         /* We have a conditional block with some assignments in cond_bb.
+            Wire up the PHIs to wrap up.  */
+         if (gimple_in_ssa_p (cfun))
+           {

we are always in SSA form(?)  (probably tree-complex.c can use some TLC here).

+       /* If we are not worrying about NaNs expand to
+         (ar*br - ai*bi) + i(ar*bi + br*ai) directly.  */
+       expand_complex_multiplication_limited_range (gsi, inner_type, ar, ai,
+                                                     br, bi, &rr, &ri);

I think the function is badly worded - this isn't about limited
ranges, no?  Which
also means that we can dispatch to this simple variant not only for
flag_complex_method != 2 but for !HONOR_NANS && !HONOR_INFINITIES?
Maybe that should be done as followup.

Richard.


> Thanks,

> Kyrill

>

> 2018-04-30  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>

>

>     PR tree-optimization/70291

>     * tree-complex.c (insert_complex_mult_libcall): New function.

>     (expand_complex_multiplication_limited_range): Likewise.

>     (expand_complex_multiplication): Expand floating-point complex

>     multiplication using the above.

>

> 2018-04-30  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>

>

>     PR tree-optimization/70291

>     * gcc.dg/complex-6.c: New test.

>     * gcc.dg/complex-7.c: Likewise.
Kyrill Tkachov May 2, 2018, 3:56 p.m. | #2
Hi Richard,

On 02/05/18 11:15, Richard Biener wrote:
 > On Mon, Apr 30, 2018 at 7:41 PM, Kyrill  Tkachov

 > <kyrylo.tkachov@foss.arm.com> wrote:

 >> Hi all,

 >>

 >> We can improve the performance of complex floating-point multiplications by

 >> inlining the expansion a bit more aggressively.

 >> We can inline complex x = a * b as:

 >> x = (ar*br - ai*bi) + i(ar*bi + br*ai);

 >> if (isunordered (__real__ x, __imag__ x))

 >>   x = __muldc3 (a, b); //Or __mulsc3 for single-precision

 >>

 >> That way the common case where no NaNs are produced we can avoid the libgcc

 >> call and fall back to the

 >> NaN handling stuff in libgcc if either components of the expansion are NaN.

 >>

 >> The implementation is done in expand_complex_multiplication in

 >> tree-complex.c and the above expansion

 >> will be done when optimising for -O1 and greater and when not optimising for

 >> size.

 >> At -O0 and -Os the single call to libgcc will be emitted.

 >>

 >> For the code:

 >> __complex double

 >> foo (__complex double a, __complex double b)

 >> {

 >>   return a * b;

 >> }

 >>

 >> We will now emit at -O2 for aarch64:

 >> foo:

 >>         fmul    d16, d1, d3

 >>         fmul    d6, d1, d2

 >>         fnmsub  d5, d0, d2, d16

 >>         fmadd   d4, d0, d3, d6

 >>         fcmp    d5, d4

 >>         bvs     .L8

 >>         fmov    d1, d4

 >>         fmov    d0, d5

 >>         ret

 >> .L8:

 >>         stp     x29, x30, [sp, -16]!

 >>         mov     x29, sp

 >>         bl      __muldc3

 >>         ldp     x29, x30, [sp], 16

 >>         ret

 >>

 >> Instead of just a branch to __muldc3.

 >>

 >> Bootstrapped and tested on aarch64-none-linux-gnu, arm-none-linux-gnueabihf,

 >> x86_64-unknown-linux-gnu.

 >>

 >> Ok for trunk? (GCC 9)

 >

 > +         /* If optimizing for size or not at all just do a libcall.  */

 > +         if (optimize == 0 || optimize_function_for_size_p (cfun))

 > +           {

 > +             expand_complex_libcall (gsi, ar, ai, br, bi, MULT_EXPR);

 > +             return;

 > +           }

 >

 > use optimize_bb_for_size_p instead please (get BB from the mult stmt).

 >


Done.

 >  /* Expand a complex multiplication or division to a libcall to the c99

 > +   compliant routines.  Unlike expand_complex_libcall create and insert

 > +   the call, assign it to an output variable and return that rather than

 > +   modifying existing statements in place.  */

 > +

 > +static tree

 > +insert_complex_mult_libcall (gimple_stmt_iterator *gsi, tree type, tree ar,

 > +                             tree ai, tree br, tree bi)

 > +{

 >

 > can you please try merging both functions instead?

 >


I tried that initially, it looked a bit awkward because we sometimes want to modify
an existing assignment in-place, other times we want to insert it as a new statement.
This latest version keeps only one function and adds an inplace_p parameter.

 > Also it shows a possible issue if with -fnon-call-exceptions the original

 > multiplication has EH edges.  I think you want to side-step that

 > by doing the libcall-only way in that case as well (stmt_can_throw_internal).


Ok, done.

 >

 > +         tree isunordered_decl = builtin_decl_explicit (BUILT_IN_ISUNORDERED);

 > +         tree isunordered_res = create_tmp_var (integer_type_node);

 > +         gimple *tmpr_unord_check

 > +           = gimple_build_call (isunordered_decl, 2, tmpr, tmpi);

 > +         gimple_call_set_lhs (tmpr_unord_check, isunordered_res);

 > +

 > +         gsi_insert_before (gsi, tmpr_unord_check, GSI_SAME_STMT);

 > +         gimple *check

 > +           = gimple_build_cond (NE_EXPR, isunordered_res, integer_zero_node,

 > +                                NULL_TREE, NULL_TREE);

 >

 > why use BUILT_IN_ISUNORDERED but not a GIMPLE_COND with

 > UNORDERED_EXPR?  Note again that might trap/throw with -fsignalling-nans

 > so better avoid this transform for flag_signalling_nans as well...



No reason, I was just not familiar enough with the tree-level expressions.
Using UNORDERED_EXPR does look much better.

 >

 > +         /* We have a conditional block with some assignments in cond_bb.

 > +            Wire up the PHIs to wrap up.  */

 > +         if (gimple_in_ssa_p (cfun))

 > +           {

 >

 > we are always in SSA form(?)  (probably tree-complex.c can use some TLC here).


I think we're always in SSA form. I suspected this was an artifact from the past.
I've removed these conditionals in this latest version and testing doesn't show any problems.


 >

 > +       /* If we are not worrying about NaNs expand to

 > +         (ar*br - ai*bi) + i(ar*bi + br*ai) directly.  */

 > +       expand_complex_multiplication_limited_range (gsi, inner_type, ar, ai,

 > +                                                     br, bi, &rr, &ri);

 >

 > I think the function is badly worded - this isn't about limited

 > ranges, no?  Which

 > also means that we can dispatch to this simple variant not only for

 > flag_complex_method != 2 but for !HONOR_NANS && !HONOR_INFINITIES?

 > Maybe that should be done as followup.

 >


Hmm, I've struggled a bit to come up with a good names.
I've renamed it to expand_complex_multiplication_components.
Is that more descriptive?

I'm happy to enable this expansion path for !HONOR_NANS && !HONOR_INFINITIES as a separate patch.

Bootstrapped and tested on aarch64-none-linux-gnu, arm-none-linux-gnueabihf, x86_64-unknown-linux-gnu.

Thanks,
Kyrill

2018-05-02  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>

     PR tree-optimization/70291
     * tree-complex.c (expand_complex_libcall): Add type, inplace_p
     arguments.  Change return type to tree.  Emit libcall as a new
     statement rather than replacing existing one when inplace_p is true.
     (expand_complex_multiplication_components): New function.
     (expand_complex_multiplication): Expand floating-point complex
     multiplication using the above.
     (expand_complex_division): Rename inner_type parameter to type.
     Update expand_complex_libcall call-site.
     (expand_complex_operations_1): Update expand_complex_multiplication
     and expand_complex_division call-sites.

2018-05-02  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>

     PR tree-optimization/70291
     * gcc.dg/complex-6.c: New test.
     * gcc.dg/complex-7.c: Likewise.
diff --git a/gcc/testsuite/gcc.dg/complex-6.c b/gcc/testsuite/gcc.dg/complex-6.c
new file mode 100644
index 0000000000000000000000000000000000000000..e70322bf6f378d1d6947de4c10f671bc0a7ded49
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/complex-6.c
@@ -0,0 +1,13 @@
+/* PR tree-optimization/70291.  */
+
+/* { dg-do compile } */
+/* { dg-options "-O2 -fdump-tree-cplxlower" } */
+
+__complex float
+foo (__complex float a, __complex float b)
+{
+  return a * b;
+}
+
+/* { dg-final { scan-tree-dump-times "unord" 1 "cplxlower1" } } */
+/* { dg-final { scan-tree-dump-times "__mulsc3" 1 "cplxlower1" } } */
diff --git a/gcc/testsuite/gcc.dg/complex-7.c b/gcc/testsuite/gcc.dg/complex-7.c
new file mode 100644
index 0000000000000000000000000000000000000000..78f1a290e34af0c092a00639d979bc32b332b1da
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/complex-7.c
@@ -0,0 +1,13 @@
+/* PR tree-optimization/70291.  */
+
+/* { dg-do compile } */
+/* { dg-options "-O2 -fdump-tree-cplxlower" } */
+
+__complex double
+foo (__complex double a, __complex double b)
+{
+  return a * b;
+}
+
+/* { dg-final { scan-tree-dump-times "unord" 1 "cplxlower1" } } */
+/* { dg-final { scan-tree-dump-times "__muldc3" 1 "cplxlower1" } } */
diff --git a/gcc/tree-complex.c b/gcc/tree-complex.c
index 622b8696399b9e9d8bddcc6340d2f8d8ca852637..d194493ff859783678036ca3527ea7942eecc869 100644
--- a/gcc/tree-complex.c
+++ b/gcc/tree-complex.c
@@ -978,22 +978,22 @@ expand_complex_addition (gimple_stmt_iterator *gsi, tree inner_type,
 }
 
 /* Expand a complex multiplication or division to a libcall to the c99
-   compliant routines.  */
+   compliant routines.  TYPE is the complex type of the operation.
+   If INPLACE_P replace the statement at GSI with
+   the libcall and return NULL_TREE.  Else insert the call, assign its
+   result to an output variable and return that variable.  If INPLACE_P
+   is true then the statement being replaced should be an assignment
+   statement.  */
 
-static void
-expand_complex_libcall (gimple_stmt_iterator *gsi, tree ar, tree ai,
-			tree br, tree bi, enum tree_code code)
+static tree
+expand_complex_libcall (gimple_stmt_iterator *gsi, tree type, tree ar, tree ai,
+			tree br, tree bi, enum tree_code code, bool inplace_p)
 {
   machine_mode mode;
   enum built_in_function bcode;
-  tree fn, type, lhs;
-  gimple *old_stmt;
+  tree fn, lhs;
   gcall *stmt;
 
-  old_stmt = gsi_stmt (*gsi);
-  lhs = gimple_assign_lhs (old_stmt);
-  type = TREE_TYPE (lhs);
-
   mode = TYPE_MODE (type);
   gcc_assert (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT);
 
@@ -1008,21 +1008,65 @@ expand_complex_libcall (gimple_stmt_iterator *gsi, tree ar, tree ai,
   fn = builtin_decl_explicit (bcode);
 
   stmt = gimple_build_call (fn, 4, ar, ai, br, bi);
-  gimple_call_set_lhs (stmt, lhs);
-  update_stmt (stmt);
-  gsi_replace (gsi, stmt, false);
 
-  if (maybe_clean_or_replace_eh_stmt (old_stmt, stmt))
-    gimple_purge_dead_eh_edges (gsi_bb (*gsi));
 
-  if (gimple_in_ssa_p (cfun))
+  if (inplace_p)
     {
+      gimple *old_stmt = gsi_stmt (*gsi);
+      lhs = gimple_assign_lhs (old_stmt);
+      gimple_call_set_lhs (stmt, lhs);
+      update_stmt (stmt);
+      gsi_replace (gsi, stmt, false);
+
+      if (maybe_clean_or_replace_eh_stmt (old_stmt, stmt))
+	gimple_purge_dead_eh_edges (gsi_bb (*gsi));
+
       type = TREE_TYPE (type);
       update_complex_components (gsi, stmt,
-				 build1 (REALPART_EXPR, type, lhs),
-				 build1 (IMAGPART_EXPR, type, lhs));
+				  build1 (REALPART_EXPR, type, lhs),
+				  build1 (IMAGPART_EXPR, type, lhs));
       SSA_NAME_DEF_STMT (lhs) = stmt;
+      return NULL_TREE;
     }
+
+  lhs = create_tmp_var (type);
+  gimple_call_set_lhs (stmt, lhs);
+
+  lhs = make_ssa_name (lhs, stmt);
+  gimple_call_set_lhs (stmt, lhs);
+
+  update_stmt (stmt);
+  gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
+  return lhs;
+}
+
+/* Perform a complex multiplication on two complex constants A, B represented
+   by AR, AI, BR, BI of type TYPE.
+   The operation we want is: a * b = (ar*br - ai*bi) + i(ar*bi + br*ai).
+   Insert the GIMPLE statements into GSI.  Store the real and imaginary
+   components of the result into RR and RI.  */
+
+static void
+expand_complex_multiplication_components (gimple_stmt_iterator *gsi,
+					     tree type, tree ar, tree ai,
+					     tree br, tree bi,
+					     tree *rr, tree *ri)
+{
+  tree t1, t2, t3, t4;
+
+  t1 = gimplify_build2 (gsi, MULT_EXPR, type, ar, br);
+  t2 = gimplify_build2 (gsi, MULT_EXPR, type, ai, bi);
+  t3 = gimplify_build2 (gsi, MULT_EXPR, type, ar, bi);
+
+  /* Avoid expanding redundant multiplication for the common
+     case of squaring a complex number.  */
+  if (ar == br && ai == bi)
+    t4 = t3;
+  else
+    t4 = gimplify_build2 (gsi, MULT_EXPR, type, ai, br);
+
+  *rr = gimplify_build2 (gsi, MINUS_EXPR, type, t1, t2);
+  *ri = gimplify_build2 (gsi, PLUS_EXPR, type, t3, t4);
 }
 
 /* Expand complex multiplication to scalars:
@@ -1030,11 +1074,12 @@ expand_complex_libcall (gimple_stmt_iterator *gsi, tree ar, tree ai,
 */
 
 static void
-expand_complex_multiplication (gimple_stmt_iterator *gsi, tree inner_type,
+expand_complex_multiplication (gimple_stmt_iterator *gsi, tree type,
 			       tree ar, tree ai, tree br, tree bi,
 			       complex_lattice_t al, complex_lattice_t bl)
 {
   tree rr, ri;
+  tree inner_type = TREE_TYPE (type);
 
   if (al < bl)
     {
@@ -1080,27 +1125,79 @@ expand_complex_multiplication (gimple_stmt_iterator *gsi, tree inner_type,
     case PAIR (VARYING, VARYING):
       if (flag_complex_method == 2 && SCALAR_FLOAT_TYPE_P (inner_type))
 	{
-	  expand_complex_libcall (gsi, ar, ai, br, bi, MULT_EXPR);
-	  return;
-	}
-      else
-	{
-	  tree t1, t2, t3, t4;
-
-	  t1 = gimplify_build2 (gsi, MULT_EXPR, inner_type, ar, br);
-	  t2 = gimplify_build2 (gsi, MULT_EXPR, inner_type, ai, bi);
-	  t3 = gimplify_build2 (gsi, MULT_EXPR, inner_type, ar, bi);
-
-	  /* Avoid expanding redundant multiplication for the common
-	     case of squaring a complex number.  */
-	  if (ar == br && ai == bi)
-	    t4 = t3;
-	  else
-	    t4 = gimplify_build2 (gsi, MULT_EXPR, inner_type, ai, br);
+	  /* If optimizing for size or not at all just do a libcall.
+	     Same if there are exception-handling edges or signaling NaNs.  */
+	  if (optimize == 0 || optimize_bb_for_size_p (gsi_bb (*gsi))
+	     || stmt_can_throw_internal (gsi_stmt (*gsi))
+	     || flag_signaling_nans)
+	    {
+	      expand_complex_libcall (gsi, type, ar, ai, br, bi,
+				      MULT_EXPR, true);
+	      return;
+	    }
 
-	  rr = gimplify_build2 (gsi, MINUS_EXPR, inner_type, t1, t2);
-	  ri = gimplify_build2 (gsi, PLUS_EXPR, inner_type, t3, t4);
+	  /* Else, expand x = a * b into
+	     x = (ar*br - ai*bi) + i(ar*bi + br*ai);
+	     if (isunordered (__real__ x, __imag__ x))
+		x = __muldc3 (a, b);  */
+
+	  tree tmpr, tmpi;
+	  expand_complex_multiplication_components (gsi, inner_type, ar, ai,
+						     br, bi, &tmpr, &tmpi);
+
+	  gimple *check
+	    = gimple_build_cond (UNORDERED_EXPR, tmpr, tmpi,
+				 NULL_TREE, NULL_TREE);
+
+	  basic_block orig_bb = gsi_bb (*gsi);
+	  /* We want to keep track of the original complex multiplication
+	     statement as we're going to modify it later in
+	     update_complex_assignment.  Make sure that insert_cond_bb leaves
+	     that statement in the join block.  */
+	  gsi_prev (gsi);
+	  basic_block cond_bb
+	    = insert_cond_bb (gsi_bb (*gsi), gsi_stmt (*gsi), check,
+			      profile_probability::very_unlikely ());
+
+
+	  gimple_stmt_iterator cond_bb_gsi = gsi_last_bb (cond_bb);
+	  gsi_insert_after (&cond_bb_gsi, gimple_build_nop (), GSI_NEW_STMT);
+
+	  tree libcall_res
+	    = expand_complex_libcall (&cond_bb_gsi, type, ar, ai, br,
+				       bi, MULT_EXPR, false);
+	  tree cond_real = gimplify_build1 (&cond_bb_gsi, REALPART_EXPR,
+					    inner_type, libcall_res);
+	  tree cond_imag = gimplify_build1 (&cond_bb_gsi, IMAGPART_EXPR,
+					    inner_type, libcall_res);
+
+	  basic_block join_bb = single_succ_edge (cond_bb)->dest;
+	  *gsi = gsi_start_nondebug_after_labels_bb (join_bb);
+
+	  /* We have a conditional block with some assignments in cond_bb.
+	     Wire up the PHIs to wrap up.  */
+	  rr = create_tmp_var (inner_type);
+	  rr = make_ssa_name (rr);
+	  ri = create_tmp_var (inner_type);
+	  ri = make_ssa_name (ri);
+	  edge cond_to_join = single_succ_edge (cond_bb);
+	  edge orig_to_join = find_edge (orig_bb, join_bb);
+
+	  gphi *real_phi = create_phi_node (rr, gsi_bb (*gsi));
+	  add_phi_arg (real_phi, cond_real, cond_to_join,
+			UNKNOWN_LOCATION);
+	  add_phi_arg (real_phi, tmpr, orig_to_join, UNKNOWN_LOCATION);
+
+	  gphi *imag_phi = create_phi_node (ri, gsi_bb (*gsi));
+	  add_phi_arg (imag_phi, cond_imag, cond_to_join,
+			UNKNOWN_LOCATION);
+	  add_phi_arg (imag_phi, tmpi, orig_to_join, UNKNOWN_LOCATION);
 	}
+      else
+	/* If we are not worrying about NaNs expand to
+	  (ar*br - ai*bi) + i(ar*bi + br*ai) directly.  */
+	expand_complex_multiplication_components (gsi, inner_type, ar, ai,
+						      br, bi, &rr, &ri);
       break;
 
     default:
@@ -1308,13 +1405,14 @@ expand_complex_div_wide (gimple_stmt_iterator *gsi, tree inner_type,
 /* Expand complex division to scalars.  */
 
 static void
-expand_complex_division (gimple_stmt_iterator *gsi, tree inner_type,
+expand_complex_division (gimple_stmt_iterator *gsi, tree type,
 			 tree ar, tree ai, tree br, tree bi,
 			 enum tree_code code,
 			 complex_lattice_t al, complex_lattice_t bl)
 {
   tree rr, ri;
 
+  tree inner_type = TREE_TYPE (type);
   switch (PAIR (al, bl))
     {
     case PAIR (ONLY_REAL, ONLY_REAL):
@@ -1362,7 +1460,7 @@ expand_complex_division (gimple_stmt_iterator *gsi, tree inner_type,
 	case 2:
 	  if (SCALAR_FLOAT_TYPE_P (inner_type))
 	    {
-	      expand_complex_libcall (gsi, ar, ai, br, bi, code);
+	      expand_complex_libcall (gsi, type, ar, ai, br, bi, code, true);
 	      break;
 	    }
 	  /* FALLTHRU */
@@ -1630,7 +1728,7 @@ expand_complex_operations_1 (gimple_stmt_iterator *gsi)
       break;
 
     case MULT_EXPR:
-      expand_complex_multiplication (gsi, inner_type, ar, ai, br, bi, al, bl);
+      expand_complex_multiplication (gsi, type, ar, ai, br, bi, al, bl);
       break;
 
     case TRUNC_DIV_EXPR:
@@ -1638,7 +1736,7 @@ expand_complex_operations_1 (gimple_stmt_iterator *gsi)
     case FLOOR_DIV_EXPR:
     case ROUND_DIV_EXPR:
     case RDIV_EXPR:
-      expand_complex_division (gsi, inner_type, ar, ai, br, bi, code, al, bl);
+      expand_complex_division (gsi, type, ar, ai, br, bi, code, al, bl);
       break;
 
     case NEGATE_EXPR:

Patch

diff --git a/gcc/testsuite/gcc.dg/complex-6.c b/gcc/testsuite/gcc.dg/complex-6.c
new file mode 100644
index 0000000000000000000000000000000000000000..123b2a8206f098e7140792375830ff5f01f30cf6
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/complex-6.c
@@ -0,0 +1,13 @@ 
+/* PR tree-optimization/70291.  */
+
+/* { dg-do compile } */
+/* { dg-options "-O2 -fdump-tree-cplxlower" } */
+
+__complex float
+foo (__complex float a, __complex float b)
+{
+  return a * b;
+}
+
+/* { dg-final { scan-tree-dump-times "__builtin_isunordered" 1 "cplxlower1" } } */
+/* { dg-final { scan-tree-dump-times "__mulsc3" 1 "cplxlower1" } } */
diff --git a/gcc/testsuite/gcc.dg/complex-7.c b/gcc/testsuite/gcc.dg/complex-7.c
new file mode 100644
index 0000000000000000000000000000000000000000..7d5ba3aefb3e007824b778d716ef7f21a48c58f8
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/complex-7.c
@@ -0,0 +1,13 @@ 
+/* PR tree-optimization/70291.  */
+
+/* { dg-do compile } */
+/* { dg-options "-O2 -fdump-tree-cplxlower" } */
+
+__complex double
+foo (__complex double a, __complex double b)
+{
+  return a * b;
+}
+
+/* { dg-final { scan-tree-dump-times "__builtin_isunordered" 1 "cplxlower1" } } */
+/* { dg-final { scan-tree-dump-times "__muldc3" 1 "cplxlower1" } } */
diff --git a/gcc/tree-complex.c b/gcc/tree-complex.c
index 622b8696399b9e9d8bddcc6340d2f8d8ca852637..319c302526483ca80ecafe7e55289c1850ad6a11 100644
--- a/gcc/tree-complex.c
+++ b/gcc/tree-complex.c
@@ -978,6 +978,43 @@  expand_complex_addition (gimple_stmt_iterator *gsi, tree inner_type,
 }
 
 /* Expand a complex multiplication or division to a libcall to the c99
+   compliant routines.  Unlike expand_complex_libcall create and insert
+   the call, assign it to an output variable and return that rather than
+   modifying existing statements in place.  */
+
+static tree
+insert_complex_mult_libcall (gimple_stmt_iterator *gsi, tree type, tree ar,
+			      tree ai, tree br, tree bi)
+{
+  machine_mode mode;
+  built_in_function bcode;
+  tree fn, lhs;
+  gcall *stmt;
+
+
+  mode = TYPE_MODE (type);
+  gcc_assert (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT);
+
+  bcode = ((built_in_function)
+	     (BUILT_IN_COMPLEX_MUL_MIN + mode - MIN_MODE_COMPLEX_FLOAT));
+
+  fn = builtin_decl_explicit (bcode);
+
+  stmt = gimple_build_call (fn, 4, ar, ai, br, bi);
+  lhs = create_tmp_var (type);
+  gimple_call_set_lhs (stmt, lhs);
+  if (gimple_in_ssa_p (cfun))
+    {
+      lhs = make_ssa_name (lhs, stmt);
+      gimple_call_set_lhs (stmt, lhs);
+    }
+  update_stmt (stmt);
+  gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
+
+  return lhs;
+}
+
+/* Expand a complex multiplication or division to a libcall to the c99
    compliant routines.  */
 
 static void
@@ -1025,6 +1062,35 @@  expand_complex_libcall (gimple_stmt_iterator *gsi, tree ar, tree ai,
     }
 }
 
+/* Perform a complex multiplication assuming limited range on two
+   complex constants A, B represented by AR, AI, BR, BI of type TYPE.
+   The operation we want is: a * b = (ar*br - ai*bi) + i(ar*bi + br*ai).
+   Insert the GIMPLE statements into GSI.  Store the real and imaginary
+   components of the result into RR and RI.  */
+
+static void
+expand_complex_multiplication_limited_range (gimple_stmt_iterator *gsi,
+					     tree type, tree ar, tree ai,
+					     tree br, tree bi,
+					     tree *rr, tree *ri)
+{
+  tree t1, t2, t3, t4;
+
+  t1 = gimplify_build2 (gsi, MULT_EXPR, type, ar, br);
+  t2 = gimplify_build2 (gsi, MULT_EXPR, type, ai, bi);
+  t3 = gimplify_build2 (gsi, MULT_EXPR, type, ar, bi);
+
+  /* Avoid expanding redundant multiplication for the common
+     case of squaring a complex number.  */
+  if (ar == br && ai == bi)
+    t4 = t3;
+  else
+    t4 = gimplify_build2 (gsi, MULT_EXPR, type, ai, br);
+
+  *rr = gimplify_build2 (gsi, MINUS_EXPR, type, t1, t2);
+  *ri = gimplify_build2 (gsi, PLUS_EXPR, type, t3, t4);
+}
+
 /* Expand complex multiplication to scalars:
 	a * b = (ar*br - ai*bi) + i(ar*bi + br*ai)
 */
@@ -1080,27 +1146,98 @@  expand_complex_multiplication (gimple_stmt_iterator *gsi, tree inner_type,
     case PAIR (VARYING, VARYING):
       if (flag_complex_method == 2 && SCALAR_FLOAT_TYPE_P (inner_type))
 	{
-	  expand_complex_libcall (gsi, ar, ai, br, bi, MULT_EXPR);
-	  return;
-	}
-      else
-	{
-	  tree t1, t2, t3, t4;
-
-	  t1 = gimplify_build2 (gsi, MULT_EXPR, inner_type, ar, br);
-	  t2 = gimplify_build2 (gsi, MULT_EXPR, inner_type, ai, bi);
-	  t3 = gimplify_build2 (gsi, MULT_EXPR, inner_type, ar, bi);
+	  /* If optimizing for size or not at all just do a libcall.  */
+	  if (optimize == 0 || optimize_function_for_size_p (cfun))
+	    {
+	      expand_complex_libcall (gsi, ar, ai, br, bi, MULT_EXPR);
+	      return;
+	    }
 
-	  /* Avoid expanding redundant multiplication for the common
-	     case of squaring a complex number.  */
-	  if (ar == br && ai == bi)
-	    t4 = t3;
+	  /* Else, expand x = a * b into
+	     x = (ar*br - ai*bi) + i(ar*bi + br*ai);
+	     if (isunordered (__real__ x, __imag__ x))
+		x = __muldc3 (a, b);  */
+
+	  /* Get the top-level complex type.  We'll need it soon.  */
+	  tree top_type = TREE_TYPE (gimple_assign_lhs (gsi_stmt (*gsi)));
+
+	  tree tmpr, tmpi;
+	  expand_complex_multiplication_limited_range (gsi, inner_type, ar,
+							ai, br, bi, &tmpr,
+							&tmpi);
+
+	  tree isunordered_decl = builtin_decl_explicit (BUILT_IN_ISUNORDERED);
+	  tree isunordered_res = create_tmp_var (integer_type_node);
+	  gimple *tmpr_unord_check
+	    = gimple_build_call (isunordered_decl, 2, tmpr, tmpi);
+	  gimple_call_set_lhs (tmpr_unord_check, isunordered_res);
+
+	  gsi_insert_before (gsi, tmpr_unord_check, GSI_SAME_STMT);
+	  gimple *check
+	    = gimple_build_cond (NE_EXPR, isunordered_res, integer_zero_node,
+				 NULL_TREE, NULL_TREE);
+
+	  basic_block orig_bb = gsi_bb (*gsi);
+	  /* We want to keep track of the original complex multiplication
+	     statement as we're going to modify it later in
+	     update_complex_assignment.  Make sure that insert_cond_bb leaves
+	     that statement in the join block.  */
+	  gsi_prev (gsi);
+	  basic_block cond_bb
+	    = insert_cond_bb (gsi_bb (*gsi), gsi_stmt (*gsi), check,
+			      profile_probability::very_unlikely ());
+
+
+	  gimple_stmt_iterator cond_bb_gsi = gsi_last_bb (cond_bb);
+	  gsi_insert_after (&cond_bb_gsi, gimple_build_nop (), GSI_NEW_STMT);
+
+	  tree libcall_res
+	    = insert_complex_mult_libcall (&cond_bb_gsi, top_type, ar, ai, br,
+					    bi);
+	  tree cond_real = gimplify_build1 (&cond_bb_gsi, REALPART_EXPR,
+					    inner_type, libcall_res);
+	  tree cond_imag = gimplify_build1 (&cond_bb_gsi, IMAGPART_EXPR,
+					    inner_type, libcall_res);
+
+	  basic_block join_bb = single_succ_edge (cond_bb)->dest;
+	  *gsi = gsi_start_nondebug_after_labels_bb (join_bb);
+
+	  /* We have a conditional block with some assignments in cond_bb.
+	     Wire up the PHIs to wrap up.  */
+	  if (gimple_in_ssa_p (cfun))
+	    {
+	      rr = create_tmp_var (inner_type);
+	      rr = make_ssa_name (rr);
+	      ri = create_tmp_var (inner_type);
+	      ri = make_ssa_name (ri);
+	      edge cond_to_join = single_succ_edge (cond_bb);
+	      edge orig_to_join = find_edge (orig_bb, join_bb);
+
+	      gphi *real_phi = create_phi_node (rr, gsi_bb (*gsi));
+	      add_phi_arg (real_phi, cond_real, cond_to_join,
+			   UNKNOWN_LOCATION);
+	      add_phi_arg (real_phi, tmpr, orig_to_join, UNKNOWN_LOCATION);
+
+	      gphi *imag_phi = create_phi_node (ri, gsi_bb (*gsi));
+	      add_phi_arg (imag_phi, cond_imag, cond_to_join,
+			   UNKNOWN_LOCATION);
+	      add_phi_arg (imag_phi, tmpi, orig_to_join, UNKNOWN_LOCATION);
+	    }
 	  else
-	    t4 = gimplify_build2 (gsi, MULT_EXPR, inner_type, ai, br);
-
-	  rr = gimplify_build2 (gsi, MINUS_EXPR, inner_type, t1, t2);
-	  ri = gimplify_build2 (gsi, PLUS_EXPR, inner_type, t3, t4);
+	    {
+	      gimple *stmt = gimple_build_assign (tmpr, cond_real);
+	      gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
+	      stmt = gimple_build_assign (tmpi, cond_imag);
+	      gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
+	      rr = tmpr;
+	      ri = tmpi;
+	    }
 	}
+      else
+	/* If we are not worrying about NaNs expand to
+	  (ar*br - ai*bi) + i(ar*bi + br*ai) directly.  */
+	expand_complex_multiplication_limited_range (gsi, inner_type, ar, ai,
+						      br, bi, &rr, &ri);
       break;
 
     default: