[gcc r8-10700] i386: Fix __builtin_rint with FE_DOWNWARD rounding direction [PR96793]

[Uros Bizjak <uros@gcc.gnu.org>]

https://gcc.gnu.org/g:e64a8d36074ca15b2a3fb453c9abbf9a884e9668

commit r8-10700-ge64a8d36074ca15b2a3fb453c9abbf9a884e9668
Author: Uros Bizjak <ubizjak@gmail.com>
Date:   Tue Dec 29 00:17:24 2020 +0100

    i386: Fix __builtin_rint with FE_DOWNWARD rounding direction [PR96793]
    
    x86_expand_rint expander uses x86_sse_copysign_to_positive, which
    is unable to change the sign from - to +.  When FE_DOWNWARD rounding
    direction is in effect, the expanded sequence that involves subtraction
    can trigger x - x = -0.0 special rule.  x86_sse_copysign_to_positive
    fails to change the sign of the intermediate value, assumed to always
    be positive, back to positive.
    
    The patch adds one extra fabs that strips the sign from the intermediate
    value when flag_rounding_math is in effect.
    
    2020-12-28  Uroš Bizjak  <ubizjak@gmail.com>
    
    gcc/
            PR target/96793
            * config/i386/i386.c (ix86_expand_rint):
            Remove the sign of the intermediate value for flag_rounding_math.
    
    gcc/testsuite/
            PR target/96793
            * gcc.target/i386/pr96793-2.c: New test.

Diff:
---
 gcc/config/i386/i386.c                    | 16 +++++++++-------
 gcc/testsuite/gcc.target/i386/pr96793-2.c | 28 ++++++++++++++++++++++++++++
 2 files changed, 37 insertions(+), 7 deletions(-)

diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
index 5ac453b9fed..d97d5c0f0fe 100644
--- a/gcc/config/i386/i386.c
+++ b/gcc/config/i386/i386.c
@@ -45613,7 +45613,7 @@ ix86_expand_rint (rtx operand0, rtx operand1)
         return copysign (xa, operand1);
    */
   machine_mode mode = GET_MODE (operand0);
-  rtx res, xa, TWO52, two52, mask;
+  rtx res, xa, TWO52, mask;
   rtx_code_label *label;
 
   res = gen_reg_rtx (mode);
@@ -45626,16 +45626,18 @@ ix86_expand_rint (rtx operand0, rtx operand1)
   TWO52 = ix86_gen_TWO52 (mode);
   label = ix86_expand_sse_compare_and_jump (UNLE, TWO52, xa, false);
 
-  two52 = TWO52;
   if (flag_rounding_math)
     {
-      two52 = gen_reg_rtx (mode);
-      ix86_sse_copysign_to_positive (two52, TWO52, res, mask);
+      ix86_sse_copysign_to_positive (TWO52, TWO52, res, mask);
       xa = res;
     }
 
-  xa = expand_simple_binop (mode, PLUS, xa, two52, NULL_RTX, 0, OPTAB_DIRECT);
-  xa = expand_simple_binop (mode, MINUS, xa, two52, xa, 0, OPTAB_DIRECT);
+  xa = expand_simple_binop (mode, PLUS, xa, TWO52, NULL_RTX, 0, OPTAB_DIRECT);
+  xa = expand_simple_binop (mode, MINUS, xa, TWO52, xa, 0, OPTAB_DIRECT);
+
+  /* Remove the sign with FE_DOWNWARD, where x - x = -0.0.  */
+  if (HONOR_SIGNED_ZEROS (mode) && flag_rounding_math)
+    xa = ix86_expand_sse_fabs (xa, NULL);
 
   ix86_sse_copysign_to_positive (res, xa, res, mask);
 
@@ -45948,7 +45950,7 @@ ix86_expand_truncdf_32 (rtx operand0, rtx operand1)
   tmp = expand_simple_binop (mode, MINUS,
 			     xa2, tmp, NULL_RTX, 0, OPTAB_DIRECT);
   /* Remove the sign with FE_DOWNWARD, where x - x = -0.0.  */
-  if (flag_rounding_math)
+  if (HONOR_SIGNED_ZEROS (mode) && flag_rounding_math)
     tmp = ix86_expand_sse_fabs (tmp, NULL);
 
   /* res = copysign (xa2, operand1) */
diff --git a/gcc/testsuite/gcc.target/i386/pr96793-2.c b/gcc/testsuite/gcc.target/i386/pr96793-2.c
new file mode 100644
index 00000000000..291a2a1d38d
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/pr96793-2.c
@@ -0,0 +1,28 @@
+/* PR target/96793 */
+/* { dg-do run { target *-*-linux* *-*-gnu* } } */
+/* { dg-require-effective-target sse2_runtime } */
+/* { dg-options "-O2 -frounding-math -msse2 -mno-sse4 -mfpmath=sse" } */
+
+#include <fenv.h>
+
+double
+__attribute__((noinline))
+test (double value)
+{
+  return __builtin_rint (value);
+}
+
+int
+main ()
+{
+  double result;
+
+  fesetround (FE_DOWNWARD);
+
+  result = test (0.25);
+
+  if (__builtin_signbit (result) != 0)
+    __builtin_abort ();
+
+  return 0;
+}