From: Matthias Kruk <m@m10k.eu>
Date: Sun, 26 Jul 2020 08:09:52 +0000 (+0900)
Subject: libc, kernel/klibc: Fix bug in memcpy() where quantities less than 4 byte wouldn... 
X-Git-Url: https://git.corax.cc/?a=commitdiff_plain;h=0a5eb7fd51eb70745b738b6b7db545b873a5f04c;p=corax

libc, kernel/klibc: Fix bug in memcpy() where quantities less than 4 byte wouldn't be copied correctly (replacing the implementation with an assembly one)
---

diff --git a/kernel/klibc/Makefile b/kernel/klibc/Makefile
index 40a2c94..ff21f07 100644
--- a/kernel/klibc/Makefile
+++ b/kernel/klibc/Makefile
@@ -1,6 +1,6 @@
 OUTPUT = klibc.a
 OBJECTS = string.o spinlock.o mutex.o posixcall.o coraxcall.o heap.o stdio.o signal.o \
-          kq.o sem.o
+          kq.o sem.o string_ia32.o
 PHONY = clean
 INCLUDES = -I../include -I../../include -I../..
 CFLAGS = -m32 -Wall -nostdlib -nodefaultlibs -nostartfiles -ffreestanding $(INCLUDES)
diff --git a/kernel/klibc/string.c b/kernel/klibc/string.c
index 657d746..aa56af3 100644
--- a/kernel/klibc/string.c
+++ b/kernel/klibc/string.c
@@ -10,33 +10,6 @@ size_t strlen(const char *s)
 	return(ret_val);
 }
 
-void* memcpy(void *dst, const void *src, size_t n)
-{
-	u32_t *d;
-	const u32_t *s;
-
-	d = (u32_t*)dst;
-	s = (const u32_t*)src;
-
-	while(n > 4) {
-		*d++ = *s++;
-		n -= 4;
-	}
-
-	if(n >= 2) {
-		*((u16_t*)d) = *((u16_t*)s);
-		d = (u32_t*)(((void*)d) + 2);
-		s = (u32_t*)(((void*)s) + 2);
-		n -= 2;
-	}
-
-	if(n >= 1) {
-		*((u8_t*)d) = *((u8_t*)s);
-	}
-
-	return(dst);
-}
-
 void* memset(void *dst, int val, u32_t n)
 {
 	char *ptr;
diff --git a/kernel/klibc/string_ia32.S b/kernel/klibc/string_ia32.S
new file mode 100644
index 0000000..b2fb623
--- /dev/null
+++ b/kernel/klibc/string_ia32.S
@@ -0,0 +1,44 @@
+#define __ASSEMBLY_SOURCE
+
+#include <config.h>
+
+.section .text
+
+.global memcpy
+
+memcpy:
+	pushl	%ebx
+
+	movl	8(%esp), %eax
+	movl	12(%esp), %ebx
+	movl	16(%esp), %ecx
+
+0:	testl	$0xfffffffc, %ecx
+	jz	1f
+
+	movl	(%ebx), %edx
+	movl	%edx, (%eax)
+	addl	$4, %eax
+	addl	$4, %ebx
+	subl	$4, %ecx
+	jmp	0b
+
+1:	testl	$0x2, %ecx
+	jz	2f
+
+	movw	(%ebx), %dx
+	movw	%dx, (%eax)
+	addl	$2, %eax
+	addl	$2, %ebx
+	subl	$2, %ecx
+
+2:	testl	$0x1, %ecx
+	jz	3f
+
+	movb	(%ebx), %dl
+	movb	%dl, (%eax)
+
+3:	movl	8(%esp), %eax
+
+	popl	%ebx
+	ret
diff --git a/libc/Makefile b/libc/Makefile
index 2d4fc3d..2383060 100644
--- a/libc/Makefile
+++ b/libc/Makefile
@@ -1,7 +1,7 @@
 OUTPUT = libc.a
-OBJECTS = syscall.o string.o signal.o start.o stdio.o
+OBJECTS = syscall.o string.o signal.o start.o stdio.o string_ia32.o
 PHONY = clean
-INCLUDES = -I../include
+INCLUDES = -I../include -I..
 CFLAGS = -m32 -Wall -nostdlib -nodefaultlibs -nostartfiles -ffreestanding $(INCLUDES)
 ASFLAGS = $(CFLAGS)
 LIBGCC = /usr/lib/gcc-cross/i686-linux-gnu/8/libgcc.a
diff --git a/libc/string.c b/libc/string.c
index 5822776..a8e548a 100644
--- a/libc/string.c
+++ b/libc/string.c
@@ -22,30 +22,3 @@ void *memset(void *s, int c, size_t n)
 
 	return(s);
 }
-
-void* memcpy(void *dst, const void *src, size_t n)
-{
-        u32_t *d;
-        const u32_t *s;
-
-        d = (u32_t*)dst;
-        s = (const u32_t*)src;
-
-        while(n > 4) {
-                *d++ = *s++;
-                n -= 4;
-        }
-
-        if(n >= 2) {
-                *((u16_t*)d) = *((u16_t*)s);
-                d = (u32_t*)(((void*)d) + 2);
-                s = (u32_t*)(((void*)s) + 2);
-                n -= 2;
-        }
-
-        if(n >= 1) {
-                *((u8_t*)d) = *((u8_t*)s);
-        }
-
-        return(dst);
-}
diff --git a/libc/string_ia32.S b/libc/string_ia32.S
new file mode 100644
index 0000000..b2fb623
--- /dev/null
+++ b/libc/string_ia32.S
@@ -0,0 +1,44 @@
+#define __ASSEMBLY_SOURCE
+
+#include <config.h>
+
+.section .text
+
+.global memcpy
+
+memcpy:
+	pushl	%ebx
+
+	movl	8(%esp), %eax
+	movl	12(%esp), %ebx
+	movl	16(%esp), %ecx
+
+0:	testl	$0xfffffffc, %ecx
+	jz	1f
+
+	movl	(%ebx), %edx
+	movl	%edx, (%eax)
+	addl	$4, %eax
+	addl	$4, %ebx
+	subl	$4, %ecx
+	jmp	0b
+
+1:	testl	$0x2, %ecx
+	jz	2f
+
+	movw	(%ebx), %dx
+	movw	%dx, (%eax)
+	addl	$2, %eax
+	addl	$2, %ebx
+	subl	$2, %ecx
+
+2:	testl	$0x1, %ecx
+	jz	3f
+
+	movb	(%ebx), %dl
+	movb	%dl, (%eax)
+
+3:	movl	8(%esp), %eax
+
+	popl	%ebx
+	ret