diff --git a/libc/arch-arm/cortex-a15/bionic/__strcat_chk_common.S b/libc/arch-arm/cortex-a15/bionic/__strcat_chk_common.S
index 020ce42..a610dd5 100644
--- a/libc/arch-arm/cortex-a15/bionic/__strcat_chk_common.S
+++ b/libc/arch-arm/cortex-a15/bionic/__strcat_chk_common.S
@@ -176,7 +176,9 @@
 .L_strlen_done:
     add     r2, r3, r4
     cmp     r2, lr
-    bhi     .L_strcat_chk_failed
+    itt     hi
+    movhi   r0, lr
+    bhi     __strcat_chk_fail
 
     // Set up the registers for the memcpy code.
     mov     r1, r5
@@ -194,15 +196,4 @@
     .cfi_adjust_cfa_offset 8
     .cfi_rel_offset r4, 0
     .cfi_rel_offset r5, 4
-.L_strcat_chk_failed:
-    ldr     r0, error_message
-1:
-    add     r0, pc
-    bl      __fortify_fatal
-error_message:
-    .word   error_string-(1b+4)
 END(__strcat_chk)
-
-    .data
-error_string:
-    .string "strcat: prevented write past end of buffer"
diff --git a/libc/arch-arm/cortex-a9/bionic/__strcat_chk.S b/libc/arch-arm/cortex-a9/bionic/__strcat_chk.S
index 6245861..776c782 100644
--- a/libc/arch-arm/cortex-a9/bionic/__strcat_chk.S
+++ b/libc/arch-arm/cortex-a9/bionic/__strcat_chk.S
@@ -179,6 +179,8 @@
 .L_strlen_done:
     add     r2, r3, r4
     cmp     r2, lr
+    itt     hi
+    movhi   r0, lr
     bhi     __strcat_chk_fail
 
     // Set up the registers for the memcpy code.
@@ -194,23 +196,3 @@
 #define MEMCPY_BASE         __strcat_chk_memcpy_base
 #define MEMCPY_BASE_ALIGNED __strcat_chk_memcpy_base_aligned
 #include "memcpy_base.S"
-
-ENTRY_PRIVATE(__strcat_chk_fail)
-    .cfi_def_cfa_offset 8
-    .cfi_rel_offset r0, 0
-    .cfi_rel_offset lr, 4
-    .cfi_adjust_cfa_offset 8
-    .cfi_rel_offset r4, 0
-    .cfi_rel_offset r5, 4
-
-    ldr     r0, error_message
-1:
-    add     r0, pc
-    bl      __fortify_fatal
-error_message:
-    .word   error_string-(1b+4)
-END(__strcat_chk_fail)
-
-    .data
-error_string:
-    .string "strcat: prevented write past end of buffer"
diff --git a/libc/arch-arm/denver/bionic/__strcat_chk.S b/libc/arch-arm/denver/bionic/__strcat_chk.S
index 2fbf5d6..9f7db59 100644
--- a/libc/arch-arm/denver/bionic/__strcat_chk.S
+++ b/libc/arch-arm/denver/bionic/__strcat_chk.S
@@ -176,7 +176,9 @@
 .L_strlen_done:
     add     r2, r3, r4
     cmp     r2, lr
-    bhi     __strcat_chk_failed
+    itt     hi
+    movhi   r0, lr
+    bhi     __strcat_chk_fail
 
     // Set up the registers for the memcpy code.
     mov     r1, r5
@@ -190,23 +192,3 @@
 #define MEMCPY_BASE_ALIGNED __strcat_chk_memcpy_base_aligned
 
 #include "memcpy_base.S"
-
-ENTRY_PRIVATE(__strcat_chk_failed)
-    .cfi_def_cfa_offset 8
-    .cfi_rel_offset r0, 0
-    .cfi_rel_offset lr, 4
-    .cfi_adjust_cfa_offset 8
-    .cfi_rel_offset r4, 0
-    .cfi_rel_offset r5, 4
-
-    ldr     r0, error_message
-1:
-    add     r0, pc
-    bl      __fortify_fatal
-error_message:
-    .word   error_string-(1b+4)
-END(__strcat_chk_failed)
-
-    .data
-error_string:
-    .string "strcat: prevented write past end of buffer"
diff --git a/libc/arch-arm/krait/bionic/__strcat_chk.S b/libc/arch-arm/krait/bionic/__strcat_chk.S
index fabca24..a46ff98 100644
--- a/libc/arch-arm/krait/bionic/__strcat_chk.S
+++ b/libc/arch-arm/krait/bionic/__strcat_chk.S
@@ -176,7 +176,9 @@
 .L_strlen_done:
     add     r2, r3, r4
     cmp     r2, lr
-    bhi     .L_strcat_chk_failed
+    itt     hi
+    movhi   r0, lr
+    bhi     __strcat_chk_fail
 
     // Set up the registers for the memcpy code.
     mov     r1, r5
@@ -194,15 +196,4 @@
     .cfi_adjust_cfa_offset 8
     .cfi_rel_offset r4, 0
     .cfi_rel_offset r5, 4
-.L_strcat_chk_failed:
-    ldr     r0, error_message
-1:
-    add     r0, pc
-    bl      __fortify_fatal
-error_message:
-    .word   error_string-(1b+4)
 END(__strcat_chk)
-
-    .data
-error_string:
-    .string "strcat: prevented write past end of buffer"
diff --git a/libc/bionic/fortify.cpp b/libc/bionic/fortify.cpp
index ad7aa04..278c606 100644
--- a/libc/bionic/fortify.cpp
+++ b/libc/bionic/fortify.cpp
@@ -268,6 +268,12 @@
   return dst;
 }
 
+// strcat is performance-critical enough that we have assembler __strcat_chk implementations.
+// This function is used to give better diagnostics than we can easily do from assembler.
+extern "C" void __strcat_chk_fail(size_t dst_buf_size) {
+  __fortify_fatal("strcat: prevented write past end of %zu-byte buffer", dst_buf_size);
+}
+
 char* __strchr_chk(const char* p, int ch, size_t s_len) {
   for (;; ++p, s_len--) {
     if (__predict_false(s_len == 0)) {