[PATCH] aarch64: Fix SVE tanf mixed-lane results [BZ #34922]

Dylan Fleming Dylan.Fleming@arm.com
Wed Nov 4 15:06:40 GMT 2026


The special-case path reciprocates every lane before calling scalar tanf
for large inputs. This also changes ordinary lanes which do not need the
reciprocal, so inputs [0.0f, 0x1p15f, ...] return infinity in the first
lane instead of zero.

Fixed by passing the alternate-lane predicate to special_case and selecting
the reciprocal only where needed before handling the special lanes.
---
 SHARED-FILES                   | 2 +-
 sysdeps/aarch64/fpu/sve/tanf.c | 6 ++++--
 2 files changed, 5 insertions(+), 3 deletions(-)

diff --git a/SHARED-FILES b/SHARED-FILES
index 8f89400177..2c6e4546dd 100644
--- a/SHARED-FILES
+++ b/SHARED-FILES
@@ -573,7 +573,7 @@ arm-optimized-routines:
   sysdeps/aarch64/fpu/sve/sinpif.c
   # math/aarch64/sve/tan.c, revision 1aad38d
   sysdeps/aarch64/fpu/sve/tan.c
-  # math/aarch64/sve/tanf.c, revision 1aad38d
+  # math/aarch64/sve/tanf.c, revision 5e51e52
   sysdeps/aarch64/fpu/sve/tanf.c
   # math/aarch64/sve/tanh.c, revision 1aad38d
   sysdeps/aarch64/fpu/sve/tanh.c
diff --git a/sysdeps/aarch64/fpu/sve/tanf.c b/sysdeps/aarch64/fpu/sve/tanf.c
index 60d93fd7e2..4a627d31e0 100644
--- a/sysdeps/aarch64/fpu/sve/tanf.c
+++ b/sysdeps/aarch64/fpu/sve/tanf.c
@@ -47,8 +47,10 @@ static const struct data
 };
 
 static svfloat32_t NOINLINE
-special_case (svfloat32_t x, svfloat32_t y, svbool_t cmp)
+special_case (svfloat32_t x, svfloat32_t y, svbool_t cmp, svbool_t pred_alt)
 {
+  svfloat32_t inv_y = svdivr_x (svptrue_b32 (), y, 1.0f);
+  y = svsel (pred_alt, inv_y, y);
   return sv_call_f32 (tanf, x, y, cmp);
 }
 
@@ -107,7 +109,7 @@ svfloat32_t SV_NAME_F1 (tan) (svfloat32_t x, const svbool_t pg)
   /* Determine whether input is too large to perform fast regression.  */
   svbool_t cmp = svacge (pg, x, d->range_val);
   if (__glibc_unlikely (svptest_any (pg, cmp)))
-    return special_case (x, svdivr_x (pg, y, 1.0f), cmp);
+    return special_case (x, y, cmp, pred_alt);
 
   svfloat32_t inv_y = svdivr_x (pg, y, 1.0f);
   return svsel (pred_alt, inv_y, y);
-- 
2.43.0



More information about the Libc-alpha mailing list