https://gcc.gnu.org/g:f3de20480c782e6f7f13ec0da7c4c26ff71f83ed

commit r17-5977-gf3de20480c782e6f7f13ec0da7c4c26ff71f83ed
Author: Richard Earnshaw <[email protected]>
Date:   Wed Nov 4 16:23:33 2026 +0000

    arm: fix neon 16-bit lane duplication for big-endian [PR128324]
    
    There was an issue with the ordering of statements in my patch r17-2930
    when compiling for big-endian.  The call to subreg_regno_offset to pick
    the core register for the extraction needs to be made before we adjust
    the element number to map the big-endian model on the the Arm lane
    ordering.
    
            PR target/128324
    
    gcc/ChangeLog:
    
            * config/arm/neon.md
            (neon_vdup_lan<VDQW:mode>_internal): Calculate the register
            before adjusting the element.
            (neon_vdup_lane<VHFBF:mode>_internal): Likewise.
    
    gcc/testsuite/ChangeLog:
    
            * gcc.target/arm/pr128324.c: New test.

Diff:
---
 gcc/config/arm/neon.md                  |  8 ++++----
 gcc/testsuite/gcc.target/arm/pr128324.c | 22 ++++++++++++++++++++++
 2 files changed, 26 insertions(+), 4 deletions(-)

diff --git a/gcc/config/arm/neon.md b/gcc/config/arm/neon.md
index ffda37babaec..54bd359af2c9 100644
--- a/gcc/config/arm/neon.md
+++ b/gcc/config/arm/neon.md
@@ -3566,13 +3566,13 @@ if (BYTES_BIG_ENDIAN)
        (vec_duplicate:VDQW (match_dup 3)))]
   {
     unsigned HOST_WIDE_INT elt = UINTVAL (operands[2]);
-    if (BYTES_BIG_ENDIAN)
-      elt = GET_MODE_NUNITS (<V_double_vector_mode>mode) - 1 - elt;
     unsigned HOST_WIDE_INT size = GET_MODE_SIZE (<V_elem>mode);
     int base_regno = REGNO (operands[1]);
     int regno = (base_regno
                 + subreg_regno_offset (base_regno, <V_double_vector_mode>mode,
                                        elt * size, SImode));
+    if (BYTES_BIG_ENDIAN)
+      elt = GET_MODE_NUNITS (<V_double_vector_mode>mode) - 1 - elt;
     unsigned HOST_WIDE_INT offset = (elt * size) % GET_MODE_SIZE (SImode);
     if (offset != 0)
       {
@@ -3649,13 +3649,13 @@ if (BYTES_BIG_ENDIAN)
        (vec_duplicate:VHFBF (match_dup 3)))]
   {
     unsigned HOST_WIDE_INT elt = UINTVAL (operands[2]);
-    if (BYTES_BIG_ENDIAN)
-      elt = GET_MODE_NUNITS (<V_double_vector_mode>mode) - 1 - elt;
     unsigned HOST_WIDE_INT size = GET_MODE_SIZE (<V_elem>mode);
     int base_regno = REGNO (operands[1]);
     int regno = (base_regno
                 + subreg_regno_offset (base_regno, <V_double_vector_mode>mode,
                                        elt * size, SImode));
+    if (BYTES_BIG_ENDIAN)
+      elt = GET_MODE_NUNITS (<V_double_vector_mode>mode) - 1 - elt;
     unsigned HOST_WIDE_INT offset = (elt * size) % GET_MODE_SIZE (SImode);
     if (offset != 0)
       {
diff --git a/gcc/testsuite/gcc.target/arm/pr128324.c 
b/gcc/testsuite/gcc.target/arm/pr128324.c
new file mode 100644
index 000000000000..66264a3b1e8e
--- /dev/null
+++ b/gcc/testsuite/gcc.target/arm/pr128324.c
@@ -0,0 +1,22 @@
+/* { dg-do compile } */
+/* { dg-require-effective-target stdint_types_mbig_endian } */
+/* { dg-require-effective-target arm_arch_v7ve_neon_ok } */
+/* { dg-options "-O2 -mbig-endian" } */
+/* { dg-add-options arm_arch_v7ve_neon } */
+/* { dg-final { check-function-bodies "**" "" "" } } */
+
+#include <arm_neon.h>
+#include <stdint.h>
+
+int16x4_t be_lane (int16x4_t v)
+{
+  return vdup_lane_s16 (v, 3);
+}
+
+/*
+** be_lane:
+**     lsrs?   (r[0-9]*), r0, #16
+**     vdup.16 d16, \1
+**     vmov    r1, r0, d16  @ v4hi
+**     bx      lr
+*/

Reply via email to