Bootloader: Implement better memset and memcpy for 32 bit addresses

bootloader/bios/CMakeLists.txt

@@ -15,5 +15,6 @@ set(BOOTLOADER_SOURCES
 )
 
 add_executable(bootloader ${BOOTLOADER_SOURCES})
+include_directories(${CMAKE_CURRENT_SOURCE_DIR})
 target_link_options(bootloader PRIVATE LINKER:-T,${CMAKE_CURRENT_SOURCE_DIR}/linker.ld)
 target_link_options(bootloader PRIVATE -nostdlib)

bootloader/bios/boot.S

@@ -1,3 +1,5 @@
+.include "common.S"
+
 .code16
 
 #########################################
@@ -103,12 +105,12 @@ stage2_main:
 	movl %edx, %cr0
 
 	# jump to protected mode
-	ljmpl $0x18, $protected_mode
+	ljmpl $GDT_CODE32, $protected_mode
 
 .code32
 protected_mode:
 	# setup protected mode segments
-	movw $0x10, %dx
+	movw $GDT_DATA32, %dx
 	movw %dx, %ds
 	movw %dx, %es
 	movw %dx, %fs
@@ -127,15 +129,15 @@ enter_unreal_mode:
 	movl %cr0, %eax
 	orb $1, %al
 	movl %eax, %cr0
-	ljmpl $0x8, $.enter_unreal_mode_pmode
+	ljmpl $GDT_CODE16, $.enter_unreal_mode_pmode
 
  .enter_unreal_mode_pmode:
-	movw $0x10, %bx
+	movw $GDT_DATA32, %bx
 	movw %bx, %ds
 
 	andb $0xFE, %al
 	movl %eax, %cr0
-	ljmpl $0x0, $.enter_unreal_mode_unreal
+	ljmpl $0x00, $.enter_unreal_mode_unreal
 
  .enter_unreal_mode_unreal:
 	popw %ds

bootloader/bios/common.S

@@ -0,0 +1,3 @@
+.set GDT_CODE16, 0x08
+.set GDT_DATA32, 0x10
+.set GDT_CODE32, 0x18

bootloader/bios/elf.S

@@ -113,24 +113,6 @@ elf_validate_file_header:
 	movw $elf_validate_file_header_not_executable_msg, %si
 	jmp print_and_halt
 
-
-# sets memory to zero
-#   edi: start address
-#   ecx: byte count
-# on return
-#   edi: start address + byte count
-#   ecx: 0
-elf_memset_zero:
-	test %ecx, %ecx
-	jz .elf_memset_zero_done
- .elf_memset_zero_loop:
-	movb $0, (%edi)
-	incl %edi
-	decl %ecx
-	jnz .elf_memset_zero_loop
- .elf_memset_zero_done:
-	ret
-
 # reads memory specified by 32 bit elf_program_header to memory
 elf_read_program_header32_to_memory:
 	pushal
@@ -144,7 +126,7 @@ elf_read_program_header32_to_memory:
 	addl %ebx, %edi
 	movl (elf_program_header + p32_memsz),  %ecx
 	subl %ebx, %ecx
-	call elf_memset_zero
+	xorb %al, %al; call memset32
 
 	# read file specified in program header to memory
 	movl (elf_program_header + p32_offset), %eax
@@ -171,7 +153,7 @@ elf_read_program_header64_to_memory:
 	addl %ebx, %edi
 	movl (elf_program_header + p64_memsz),  %ecx
 	subl %ebx, %ecx
-	call elf_memset_zero
+	xorb %al, %al; call memset32
 
 	# read file specified in program header to memory
 	movl (elf_program_header + p64_offset), %eax

bootloader/bios/ext2.S

@@ -454,15 +454,7 @@ ext2_inode_read_bytes:
 	movl $ext2_block_buffer, %esi
 	addl %edx, %esi
 
-	# very dumb memcpy with 32 bit addresses
-	xorl %ebx, %ebx
- .ext2_inode_read_bytes_memcpy_partial:
-	movb (%esi, %ebx), %al
-	movb %al, (%edi, %ebx)
-	incl %ebx
-	decl %ecx
-	jnz .ext2_inode_read_bytes_memcpy_partial
-	addl %ebx, %edi
+	call memcpy32
 
 	# check if all sectors are read
 	cmpl $0, 4(%esp)
@@ -487,16 +479,8 @@ ext2_inode_read_bytes:
 	addl %ecx, 0(%esp)
 	subl %ecx, 4(%esp)
 
-	# very dumb memcpy with 32 bit addresses
 	movl $ext2_block_buffer, %esi
-	movl $0, %ebx
- .ext2_inode_read_bytes_memcpy:
-	movb (%esi, %ebx), %al
-	movb %al, (%edi, %ebx)
-	incl %ebx
-	decl %ecx
-	jnz .ext2_inode_read_bytes_memcpy
-	addl %ebx, %edi
+	call memcpy32
 
 	# read next block if more sectors remaining
 	cmpl $0, 4(%esp)

bootloader/bios/utils.S

@@ -1,3 +1,5 @@
+.include "common.S"
+
 .set SCREEN_WIDTH,	80
 .set SCREEN_HEIGHT,	25
 
@@ -273,6 +275,127 @@ isprint:
 	movb $0, %al
 	ret
 
+
+# memset with 32 bit registers
+#   edi:    destination address
+#   ecx:    bytes count
+#   al:     value to set
+# return:
+#   edi:    destination address + bytes count
+#   ecx:    0
+#   other:  preserved
+.global memset32
+memset32:
+	testl %ecx, %ecx
+	jz .memset32_done
+
+	pushf; cli
+	pushw %es
+	pushl %eax
+	pushl %ebx
+	pushl %edx
+
+	movl %cr0, %ebx
+	orb $1, %bl
+	movl %ebx, %cr0
+
+	ljmpl $GDT_CODE32, $.memset32_pmode32
+
+ .code32
+ .memset32_pmode32:
+	movw $GDT_DATA32, %dx
+	movw %dx, %es
+
+	movl %ecx, %edx
+
+	andl $3, %ecx
+	rep stosb %es:(%edi)
+
+	movl %edx, %ecx
+	shrl $2, %ecx
+
+	movb %al, %ah
+	movw %ax, %dx
+	shll $16, %eax
+	movw %dx, %ax
+	rep stosl %es:(%edi)
+
+	ljmpl $GDT_CODE16, $.memset32_pmode16
+
+ .code16
+ .memset32_pmode16:
+	andb $0xFE, %bl
+	movl %ebx, %cr0
+	ljmpl $0x00, $.memset32_rmode16
+
+ .memset32_rmode16:
+	popl %edx
+	popl %ebx
+	popl %eax
+	popw %es
+	popf
+
+ .memset32_done:
+	ret
+
+# memcpy with 32 bit registers
+#   esi:    source address
+#   edi:    destination address
+#   ecx:    bytes count
+# return:
+#   esi:    source address + bytes count
+#   edi:    destination address + bytes count
+#   ecx:    0
+#   other:  preserved
+.global memcpy32
+memcpy32:
+	testl %ecx, %ecx
+	jz .memcpy32_done
+
+	pushf; cli
+	pushw %ds
+	pushw %es
+	pushl %ebx
+	pushl %edx
+
+	movl %cr0, %ebx
+	orb $1, %bl
+	movl %ebx, %cr0
+
+	ljmpl $GDT_CODE32, $.memcpy32_pmode32
+
+ .code32
+ .memcpy32_pmode32:
+	movw $GDT_DATA32, %dx
+	movw %dx, %ds
+	movw %dx, %es
+
+	movl %ecx, %edx
+	andl $3, %ecx
+	rep movsb %ds:(%esi), %es:(%edi)
+
+	movl %edx, %ecx
+	shrl $2, %ecx
+	rep movsl %ds:(%esi), %es:(%edi)
+
+	ljmpl $GDT_CODE16, $.memcpy32_pmode16
+
+ .code16
+ .memcpy32_pmode16:
+	andb $0xFE, %bl
+	movl %ebx, %cr0
+	ljmpl $0x00, $.memcpy32_rmode16
+
+ .memcpy32_rmode16:
+	popl %edx
+	popl %ebx
+	popw %es
+	popw %ds
+	popf
+
+ .memcpy32_done:
+	ret
+
 .section .bss
 
 # enough for base 2 printing