diff --git a/crypto/fipsmodule/modes/asm/ghash-ssse3-x86_64.pl b/crypto/fipsmodule/modes/asm/ghash-ssse3-x86_64.pl
index 830381b..1dd2519 100644
--- a/crypto/fipsmodule/modes/asm/ghash-ssse3-x86_64.pl
+++ b/crypto/fipsmodule/modes/asm/ghash-ssse3-x86_64.pl
@@ -374,16 +374,16 @@
 	.byte	5	# num_slots = 1 + 2 + 2
 	.byte	0	# no frame register
 
-	.byte	.Lgmult_seh_allocstack-.Lgmult_seh_begin
-	.byte	@{[$UWOP_ALLOC_SMALL | (((40 - 8) / 8) << 4)]}
+	.byte	.Lgmult_seh_save_xmm10-.Lgmult_seh_begin
+	.byte	@{[$UWOP_SAVE_XMM128 | (10 << 4)]}
+	.value	1
 
 	.byte	.Lgmult_seh_save_xmm6-.Lgmult_seh_begin
 	.byte	@{[$UWOP_SAVE_XMM128 | (6 << 4)]}
 	.value	0
 
-	.byte	.Lgmult_seh_save_xmm10-.Lgmult_seh_begin
-	.byte	@{[$UWOP_SAVE_XMM128 | (10 << 4)]}
-	.value	1
+	.byte	.Lgmult_seh_allocstack-.Lgmult_seh_begin
+	.byte	@{[$UWOP_ALLOC_SMALL | (((40 - 8) / 8) << 4)]}
 
 .align	8
 .Lghash_seh_info:
@@ -392,20 +392,20 @@
 	.byte	7	# num_slots = 1 + 2 + 2 + 2
 	.byte	0	# no frame register
 
-	.byte	.Lghash_seh_allocstack-.Lghash_seh_begin
-	.byte	@{[$UWOP_ALLOC_SMALL | (((56 - 8) / 8) << 4)]}
-
-	.byte	.Lghash_seh_save_xmm6-.Lghash_seh_begin
-	.byte	@{[$UWOP_SAVE_XMM128 | (6 << 4)]}
-	.value	0
+	.byte	.Lghash_seh_save_xmm11-.Lghash_seh_begin
+	.byte	@{[$UWOP_SAVE_XMM128 | (11 << 4)]}
+	.value	2
 
 	.byte	.Lghash_seh_save_xmm10-.Lghash_seh_begin
 	.byte	@{[$UWOP_SAVE_XMM128 | (10 << 4)]}
 	.value	1
 
-	.byte	.Lghash_seh_save_xmm11-.Lghash_seh_begin
-	.byte	@{[$UWOP_SAVE_XMM128 | (11 << 4)]}
-	.value	2
+	.byte	.Lghash_seh_save_xmm6-.Lghash_seh_begin
+	.byte	@{[$UWOP_SAVE_XMM128 | (6 << 4)]}
+	.value	0
+
+	.byte	.Lghash_seh_allocstack-.Lghash_seh_begin
+	.byte	@{[$UWOP_ALLOC_SMALL | (((56 - 8) / 8) << 4)]}
 ____
 }
 
diff --git a/crypto/test/asm/trampoline-x86_64.pl b/crypto/test/asm/trampoline-x86_64.pl
index 75aae45..8cb1410 100755
--- a/crypto/test/asm/trampoline-x86_64.pl
+++ b/crypto/test/asm/trampoline-x86_64.pl
@@ -447,26 +447,7 @@
 
   my $unwind_codes = "";
   my $num_slots = 0;
-  if ($stack_alloc_size <= 128) {
-    my $info = $UWOP_ALLOC_SMALL | ((($stack_alloc_size - 8) / 8) << 4);
-    $unwind_codes .= <<____;
-	.byte	.Labi_test_trampoline_seh_prolog_alloc-.Labi_test_trampoline_seh_begin
-	.byte	$info
-____
-    $num_slots++;
-  } else {
-    die "stack allocation needs three unwind slots" if ($stack_alloc_size > 512 * 1024 + 8);
-    my $info = $UWOP_ALLOC_LARGE;
-    my $value = $stack_alloc_size / 8;
-    $unwind_codes .= <<____;
-	.byte	.Labi_test_trampoline_seh_prolog_alloc-.Labi_test_trampoline_seh_begin
-	.byte	$info
-	.value	$value
-____
-    $num_slots += 2;
-  }
-
-  foreach my $reg (@caller_state) {
+  foreach my $reg (reverse @caller_state) {
     $reg = substr($reg, 1);
     die "unknown register $reg" unless exists($reg_offsets{$reg});
     if ($reg =~ /^r/) {
@@ -493,6 +474,25 @@
     }
   }
 
+  if ($stack_alloc_size <= 128) {
+    my $info = $UWOP_ALLOC_SMALL | ((($stack_alloc_size - 8) / 8) << 4);
+    $unwind_codes .= <<____;
+	.byte	.Labi_test_trampoline_seh_prolog_alloc-.Labi_test_trampoline_seh_begin
+	.byte	$info
+____
+    $num_slots++;
+  } else {
+    die "stack allocation needs three unwind slots" if ($stack_alloc_size > 512 * 1024 + 8);
+    my $info = $UWOP_ALLOC_LARGE;
+    my $value = $stack_alloc_size / 8;
+    $unwind_codes .= <<____;
+	.byte	.Labi_test_trampoline_seh_prolog_alloc-.Labi_test_trampoline_seh_begin
+	.byte	$info
+	.value	$value
+____
+    $num_slots += 2;
+  }
+
   $code .= <<____;
 .section	.pdata
 .align	4