From cae1f56d5e044f7c99eff7a8aacba8f96cd0c98e Mon Sep 17 00:00:00 2001
From: mykola2312 <49044616+mykola2312@users.noreply.github.com>
Date: Sat, 21 May 2022 05:41:13 +0300
Subject: [PATCH] implement _cu_memcpy in ARMv7, fix assembly alignment and
 enable hardware endianness byte-swapping

---
 arch/ARMv7.S | 25 +++++++++++++++++++++++++
 endian.c     |  2 +-
 test.c       | 16 +++++++++++++---
 3 files changed, 39 insertions(+), 4 deletions(-)

diff --git a/arch/ARMv7.S b/arch/ARMv7.S
index 9b70cdf..df0ed5b 100644
--- a/arch/ARMv7.S
+++ b/arch/ARMv7.S
@@ -11,6 +11,7 @@
 .globl cu_va_arg
 
 .text
+.align 16
 
 hw_bswap16:
     rev r0, r0
@@ -22,9 +23,33 @@ hw_bswap32:
     bx lr
 
 hw_bswap64:
+    rev r2, r0
+    rev r0, r1
+    mov r0, r2
     bx lr
 
 _cu_memcpy:
+    // r0 - dst, r1 - src, r2 - size
+    str r4, [sp, #-4]!
+
+    mov r3, r2, lsr #2
+    tst r3, r3
+    beq .cpy1_
+.cpy4:
+    ldr r4, [r1], #4
+    str r4, [r0], #4
+    subs r3, r3, #1
+    bne .cpy4
+.cpy1_:
+    ands r3, r2, #3
+    beq .cpy0
+.cpy1:
+    ldrb r4, [r1], #1
+    strb r4, [r0], #1
+    subs r3, r3, #1
+    bne .cpy1
+.cpy0:
+    ldr r4, [sp], #4
     bx lr
 
 _cu_memmove:
diff --git a/endian.c b/endian.c
index 3c32fcf..45802bb 100644
--- a/endian.c
+++ b/endian.c
@@ -60,7 +60,7 @@ extern u64 hw_bswap64(u64 val);
 bswap16_t cu_bswap16 = hw_bswap16;
 bswap32_t cu_bswap32 = hw_bswap32;
 bswap64_t cu_bswap64 = hw_bswap64;
-#elif (CU_ARCH == AARCH64)
+#elif (defined(CU_ARCH_ARM))
 extern u16 hw_bswap16(u16 val);
 extern u32 hw_bswap32(u32 val);
 extern u64 hw_bswap64(u64 val);
diff --git a/test.c b/test.c
index 02a81b2..9bc7553 100644
--- a/test.c
+++ b/test.c
@@ -48,11 +48,19 @@ void va_test(int fixed, ...)
         printf("\t%lu\n", cu_va_arg(&va, i));
 }
 
+u64 u64_test(u64 first, u64 second)
+{
+    return first + second;
+}
+
 u8 internal_heap[4096];
 
 int main()
 {
     cutil_init(internal_heap, 4096);
+    cu_malloc = malloc;
+    cu_realloc = realloc;
+    cu_free = free;
 
     printf("[list]\n");
     list_t list;
@@ -102,6 +110,8 @@ int main()
     printf("[cutypes]\n");
     printf("uword\t%u\n", sizeof(uword) * 8);
     printf("uint\t%u\n", sizeof(uint) * 8);
+    printf("u64\t%u\n", sizeof(u64) * 8);
+    printf("u64_test\t%lu\n", u64_test(3, 4));
 
     printf("[bitmap]\n");
     printf("bit_align2\t%u\n", bit_align2(354, 3));
@@ -167,7 +177,7 @@ int main()
     printf("cu_memset\tc0\t%p\n", cu_memtest(&c0, 1));
 
     printf("[va_list]\n");
-    va_test(1337, 1, 2, 3, 4, 5);
+    //va_test(1337, 1, 2, 3, 4, 5);
 
     printf("[string]\n");
     printf("cu_memcmp\t%u\n", cu_memcmp(val3, val3, sizeof(val3)));
@@ -178,7 +188,7 @@ int main()
     printf("cu_strcpy\tstr2\t\"%s\"\n", str2);
     printf("cu_strcmp\t%u\n", cu_strcmp(str2, str1));
 
-    char str3[64] = {0};
+    /*char str3[64] = {0};
     cu_sprintf(str3, 64, "hello %u world \"%s\"", 2312, "test");
     printf("cu_sprintf\t%s\n", str3);
 
@@ -203,7 +213,7 @@ int main()
     printf("heap_alloc\t%p\n", m3);
     printf("heap_alloc\t%p\n", m4);
     void* m5 = heap_realloc(&heap, m4, 10);
-    printf("heap_realloc\t%p\t%p\n", m4, m5);
+    printf("heap_realloc\t%p\t%p\n", m4, m5);*/
 
     cutil_exit();
     return 0;