From bfe5018a00555e10e7b73fd52b43242b42c113c2 Mon Sep 17 00:00:00 2001
From: Robert O'Callahan <robert@ocallahan.org>
Date: Mon, 7 Oct 2024 16:45:09 +1300
Subject: [PATCH] Fix ARM register sizes for clang builds

Resolves #3844
---
 src/preload/syscallbuf.c   | 11 +++++++----
 src/test/rseq.c            |  4 ++--
 src/test/rseq_syscallbuf.c |  4 ++--
 3 files changed, 11 insertions(+), 8 deletions(-)

diff --git a/src/preload/syscallbuf.c b/src/preload/syscallbuf.c
index 918e7555950..66415bdd664 100644
--- a/src/preload/syscallbuf.c
+++ b/src/preload/syscallbuf.c
@@ -242,6 +242,7 @@ static void local_memcpy(void* dest, const void* source, int n) {
 #elif defined(__aarch64__)
   long c1;
   long c2;
+  long n_long = n;
   __asm__ __volatile__("subs %4, %2, 16\n\t"
                        "b.lt 2f\n\t"
                        "1:\n\t"
@@ -267,7 +268,7 @@ static void local_memcpy(void* dest, const void* source, int n) {
                        "ldrb %w3, [%1]\n\t"
                        "strb %w3, [%0]\n\t"
                        "3:\n\t"
-                       : "+r"(dest), "+r"(source), "+r"(n), "=&r"(c1), "=&r"(c2)
+                       : "+r"(dest), "+r"(source), "+r"(n_long), "=&r"(c1), "=&r"(c2)
                        :
                        : "cc", "memory");
 #else
@@ -289,8 +290,9 @@ static void local_memset(void* dest, uint8_t c, int n) {
                        :
                        : "cc", "memory");
 #elif defined(__aarch64__)
-  double v1;
+  double v1 __attribute__((vector_size(16)));
   long n2;
+  long n_long = n;
   __asm__ __volatile__("subs %4, %2, 32\n\t"
                        "b.lt 2f\n\t"
                        "dup %3.16b, %w0\n"
@@ -306,7 +308,7 @@ static void local_memset(void* dest, uint8_t c, int n) {
                        "subs %2, %2, #1\n\t"
                        "b.ne 3b\n"
                        "4:\n\t"
-                       : "+r"(c), "+r"(dest), "+r"(n), "=x"(v1), "=r"(n2)
+                       : "+r"(c), "+r"(dest), "+r"(n_long), "=x"(v1), "=r"(n2)
                        :
                        : "cc", "memory");
 #else
@@ -1547,6 +1549,7 @@ static void memcpy_input_parameter(void* buf, void* src, int size) {
 #elif defined(__aarch64__)
   long c1;
   long c2;
+  long size_long = size;
   unsigned char *globals_in_replay = rr_page_replay_flag_addr();
   __asm__ __volatile__("ldrb %w3, [%5]\n\t"
                        "cmp %3, #0\n\t" // eq -> record
@@ -1580,7 +1583,7 @@ static void memcpy_input_parameter(void* buf, void* src, int size) {
                        "mov %4, xzr\n\t"
                        "mov %1, xzr\n\t"
                        : "+r"(buf), "+r"(src),
-                         "+r"(size), "=&r"(c1), "=&r"(c2), "+r"(globals_in_replay)
+                         "+r"(size_long), "=&r"(c1), "=&r"(c2), "+r"(globals_in_replay)
                        :
                        : "cc", "memory");
 #else
diff --git a/src/test/rseq.c b/src/test/rseq.c
index 76616340856..86db56a6f27 100644
--- a/src/test/rseq.c
+++ b/src/test/rseq.c
@@ -19,7 +19,7 @@ static uint64_t jump_aborts;
 static volatile uint32_t dummy;
 
 static void do_section(void) {
-  int did_abort = 0;
+  long did_abort = 0;
 
   rs_ptr->rseq_cs = (uint64_t)(uintptr_t)&rs_cs;
 #if defined(__x86_64__) || defined(__i386__)
@@ -37,7 +37,7 @@ static void do_section(void) {
     "1:\n\t"
     : : "m"(dummy), "m"(did_abort));
 #elif defined(__aarch64__)
-  int dummy2;
+  long dummy2;
   __asm__ __volatile__ (
     "start_ip:\n\t"
     "mov %1, 1234\n\t"
diff --git a/src/test/rseq_syscallbuf.c b/src/test/rseq_syscallbuf.c
index 60502d520a8..9088ad5a4ec 100644
--- a/src/test/rseq_syscallbuf.c
+++ b/src/test/rseq_syscallbuf.c
@@ -19,7 +19,7 @@ static uint64_t jump_aborts;
 static volatile uint32_t dummy;
 
 static void do_section(void) {
-  int did_abort = 0;
+  long did_abort = 0;
 
   rs_ptr->rseq_cs = (uint64_t)(uintptr_t)&rs_cs;
 #if defined(__x86_64__) || defined(__i386__)
@@ -37,7 +37,7 @@ static void do_section(void) {
     "1:\n\t"
     : : "m"(dummy), "m"(did_abort));
 #elif defined(__aarch64__)
-  int dummy2;
+  long dummy2;
   __asm__ __volatile__ (
     "start_ip:\n\t"
     "mov %1, 1234\n\t"