Added efi64.S file for X64 and reviewed efi32.S file.

git-svn-id: https://edk2.svn.sourceforge.net/svnroot/edk2/trunk/edk2@7887 6f19259b-4bc3-4df7-8a09-765794883524

1 files changed, 786 insertions, 0 deletions

diff --git a/DuetPkg/BootSector/efi64.S b/DuetPkg/BootSector/efi64.S
new file mode 100644
index 0000000..b7b3384
--- /dev/null
+++ b/DuetPkg/BootSector/efi64.S
@@ -0,0 +1,786 @@
+#------------------------------------------------------------------------------
+#*
+#*   Copyright 2006, Intel Corporation                                                         
+#*   All rights reserved. This program and the accompanying materials                          
+#*   are licensed and made available under the terms and conditions of the BSD License         
+#*   which accompanies this distribution.  The full text of the license may be found at        
+#*   http://opensource.org/licenses/bsd-license.php                                            
+#*                                                                                             
+#*   THE PROGRAM IS DISTRIBUTED UNDER THE BSD LICENSE ON AN "AS IS" BASIS,                     
+#*   WITHOUT WARRANTIES OR REPRESENTATIONS OF ANY KIND, EITHER EXPRESS OR IMPLIED.             
+#*   
+#*    efi64.asm
+#*  
+#*   Abstract:
+#*
+#------------------------------------------------------------------------------
+
+##############################################################################
+# Now in 64-bit long mode.
+##############################################################################
+
+        .486: 
+        .stack: 
+        .code: 
+        .org 0x21000
+
+.equ                 DEFAULT_HANDLER_SIZE, INT1 - INT0
+
+.macro jmpCommonIdtEntry  
+    # jmp     commonIdtEntry - this must be hand coded to keep the assembler from
+    #                          using a 8 bit reletive jump when the entries are
+    #                          within 255 bytes of the common entry.  This must
+    #                          be done to maintain the consistency of the size
+    #                          of entry points...
+    .byte   0xe9               			 # jmp 16 bit relative
+		.long   commonIdtEntry - . - 4   # offset to jump to
+.endm
+
+
+Start:  
+
+    movl    $0x001fffe8,%esp # make final stack aligned
+
+    # set OSFXSR and OSXMMEXCPT because some code will use XMM register
+    .byte 0xf
+    .byte 0x20
+    .byte 0xe0
+#    mov rax, cr4
+    btsl $9,%eax
+    btsl $0xa,%eax
+    .byte 0xf
+    .byte 0x22
+    .byte 0xe0
+#    mov cr4, rax
+
+    call    ClearScreen
+
+    # Populate IDT with meaningful offsets for exception handlers...
+		sidt    Idtr 
+		
+
+		movl    Halt, %eax
+    movl    %eax,%ebx                   # use bx to copy 15..0 to descriptors
+    shrl    $16,%eax                    # use ax to copy 31..16 to descriptors 
+                                        # 63..32 of descriptors is 0
+    movl    $0x78,%ecx                  # 78h IDT entries to initialize with unique entry points (exceptions)
+		movl    (Idtr + 2), %esi
+    movl    (%esi),%edi
+
+LOOP_1:                                         # loop through all IDT entries exception handlers and initialize to default handler
+    movw    %bx, (%edi)                         # write bits 15..0 of offset

+    movw    $0x38, 2(%edi)                      # SYS_CODE_SEL64 from GDT

+    movw    $(0x0e00 | 0x8000), 4(%edi)     		# type = 386 interrupt gate, present

+    movw    %ax, 6(%edi)                        # write bits 31..16 of offset
+    movl    $0, 8(%edi)                         # write bits 31..16 of offset  

+    addl    $16, %edi                           # move up to next descriptor

+    addw    DEFAULT_HANDLER_SIZE, %bx           # move to next entry point

+    loopl   LOOP_1                              # loop back through again until all descriptors are initialized
+
+    ## at this point edi contains the offset of the descriptor for INT 20
+    ## and bx contains the low 16 bits of the offset of the default handler
+    ## so initialize all the rest of the descriptors with these two values...
+#    mov     ecx, 101                            ; there are 100 descriptors left (INT 20 (14h) - INT 119 (77h)
+#@@:                                             ; loop through all IDT entries exception handlers and initialize to default handler
+#    mov     word ptr [edi], bx                  ; write bits 15..0 of offset
+#    mov     word ptr [edi+2], 38h               ; SYS_CODE64_SEL from GDT
+#    mov     word ptr [edi+4], 0e00h OR 8000h    ; type = 386 interrupt gate, present
+#    mov     word ptr [edi+6], ax                ; write bits 31..16 of offset
+#    mov     dword ptr [edi+8], 0                ; write bits 63..32 of offset
+#    add     edi, 16                             ; move up to next descriptor
+#    loop    @b                                  ; loop back through again until all descriptors are initialized
+
+
+##  DUMP    location of IDT and several of the descriptors
+#    mov     ecx, 8
+#    mov     eax, [offset Idtr + 2]
+#    mov     eax, [eax]
+#    mov     edi, 0b8000h
+#    call    PrintQword
+#    mov     esi, eax
+#    mov     edi, 0b80a0h
+#    jmp     OuterLoop
+
+##    
+## just for fun, let's do a software interrupt to see if we correctly land in the exception handler...
+#    mov     eax, 011111111h
+#    mov     ebx, 022222222h
+#    mov     ecx, 033333333h
+#    mov     edx, 044444444h
+#    mov     ebp, 055555555h
+#    mov     esi, 066666666h
+#    mov     edi, 077777777h
+#    push    011111111h
+#    push    022222222h
+#    push    033333333h
+#    int     119
+
+    movl    $0x22000,%esi               # esi = 22000
+    movl    0x14(%esi),%eax             # eax = [22014]
+    addl    %eax,%esi                   # esi = 22000 + [22014] = Base of EFILDR.C
+    movl    0x3c(%esi),%ebp             # ebp = [22000 + [22014] + 3c] = NT Image Header for EFILDR.C
+    addl    %esi,%ebp
+    movl    0x30(%ebp),%edi             # edi = [[22000 + [22014] + 3c] + 2c] = ImageBase (63..32 is zero, ignore)
+    movl    0x28(%ebp),%eax             # eax = [[22000 + [22014] + 3c] + 24] = EntryPoint
+    addl    %edi,%eax                   # eax = ImageBase + EntryPoint
+		movl    %ebx, EfiLdrOffset          
+    movl    %eax, (%ebx)                # Modify far jump instruction for correct entry point
+
+		movw    6(%ebp), %bx                # bx = Number of sections
+    xorl    %eax,%eax
+		movw    0x14(%ebp), %ax             # ax = Optional Header Size
+    addl    %eax,%ebp
+    addl    $0x18,%ebp                  # ebp = Start of 1st Section
+
+SectionLoop: 
+    pushl   %esi                        # Save Base of EFILDR.C
+    pushl   %edi                        # Save ImageBase
+    addl    0x14(%ebp),%esi             # esi = Base of EFILDR.C + PointerToRawData
+    addl    0x0c(%ebp),%edi             # edi = ImageBase + VirtualAddress
+    movl    0x10(%ebp),%ecx             # ecs = SizeOfRawData
+
+    cld
+    shrl    $2,%ecx
+    rep
+    movsl
+
+    popl    %edi                        # Restore ImageBase
+    popl    %esi                        # Restore Base of EFILDR.C
+
+    addw    $0x28,%bp                   # ebp = ebp + 028h = Pointer to next section record
+    .byte 0x66
+    .byte 0xff
+    .byte 0xcb
+#    dec     bx
+    cmpw    $0,%bx
+    jne     SectionLoop
+
+		movl    (Idtr), %eax                  # get size of IDT
+		movzx		(%edx), %eax
+    .byte 0xff
+    .byte 0xc0
+#    inc     eax
+ 		addl    2(%edx), %eax								 # add to base of IDT to get location of memory map...
+    xorl    %ecx,%ecx
+    movl    %eax,%ecx                    # put argument to RCX
+
+    .byte 0x48
+    .byte 0xc7
+    .byte 0xc0
+EfiLdrOffset: 
+    .long 0x00401000                    # Offset of EFILDR
+#   mov rax, 401000h
+    .byte 0x50
+#   push rax
+
+# ret
+    .byte 0xc3
+
+#    db      "**** DEFAULT IDT ENTRY ***",0
+    .align 0x2
+Halt: 
+INT0: 
+    pushl   $0x0    # push error code place holder on the stack
+    pushl   $0x0
+    jmpCommonIdtEntry 
+#    db      0e9h                        ; jmp 16 bit reletive 
+#    dd      commonIdtEntry - $ - 4      ;  offset to jump to
+
+INT1: 
+    pushl   $0x0    # push error code place holder on the stack
+    pushl   $0x1
+    jmpCommonIdtEntry 
+
+INT2: 
+    pushl   $0x0    # push error code place holder on the stack
+    pushl   $0x2
+    jmpCommonIdtEntry 
+
+INT3: 
+    pushl   $0x0    # push error code place holder on the stack
+    pushl   $0x3
+    jmpCommonIdtEntry 
+
+INT4: 
+    pushl   $0x0    # push error code place holder on the stack
+    pushl   $0x4
+    jmpCommonIdtEntry 
+
+INT5: 
+    pushl   $0x0    # push error code place holder on the stack
+    pushl   $0x5
+    jmpCommonIdtEntry 
+
+INT6: 
+    pushl   $0x0    # push error code place holder on the stack
+    pushl   $0x6
+    jmpCommonIdtEntry 
+
+INT7: 
+    pushl   $0x0    # push error code place holder on the stack
+    pushl   $0x7
+    jmpCommonIdtEntry 
+
+INT8: 
+#   Double fault causes an error code to be pushed so no phony push necessary
+    nop
+    nop
+    pushl   $0x8
+    jmpCommonIdtEntry 
+
+INT9: 
+    pushl   $0x0    # push error code place holder on the stack
+    pushl   $0x9
+    jmpCommonIdtEntry 
+
+INT10: 
+#   Invalid TSS causes an error code to be pushed so no phony push necessary
+    nop
+    nop
+    pushl   $10
+    jmpCommonIdtEntry 
+
+INT11: 
+#   Segment Not Present causes an error code to be pushed so no phony push necessary
+    nop
+    nop
+    pushl   $11
+    jmpCommonIdtEntry 
+
+INT12: 
+#   Stack fault causes an error code to be pushed so no phony push necessary
+    nop
+    nop
+    pushl   $12
+    jmpCommonIdtEntry 
+
+INT13: 
+#   GP fault causes an error code to be pushed so no phony push necessary
+    nop
+    nop
+    pushl   $13
+    jmpCommonIdtEntry 
+
+INT14: 
+#   Page fault causes an error code to be pushed so no phony push necessary
+    nop
+    nop
+    pushl   $14
+    jmpCommonIdtEntry 
+
+INT15: 
+    pushl   $0x0    # push error code place holder on the stack
+    pushl   $15
+    jmpCommonIdtEntry 
+
+INT16: 
+    pushl   $0x0    # push error code place holder on the stack
+    pushl   $16
+    jmpCommonIdtEntry 
+
+INT17: 
+#   Alignment check causes an error code to be pushed so no phony push necessary
+    nop
+    nop
+    pushl   $17
+    jmpCommonIdtEntry 
+
+INT18: 
+    pushl   $0x0    # push error code place holder on the stack
+    pushl   $18
+    jmpCommonIdtEntry 
+
+INT19: 
+    pushl   $0x0    # push error code place holder on the stack
+    pushl   $19
+    jmpCommonIdtEntry 
+
+INTUnknown: 
+.rept  (0x78 - 20)
+    pushl   $0x0    # push error code place holder on the stack
+#    push    xxh     ; push vector number
+    .byte 0x6a
+    .byte      ( . - INTUnknown - 3 ) / 9 + 20   # vector number
+    jmpCommonIdtEntry 
+.endr
+
+commonIdtEntry: 
+    pushl   %eax
+    pushl   %ecx
+    pushl   %edx
+    pushl   %ebx
+    pushl   %esp
+    pushl   %ebp
+    pushl   %esi
+    pushl   %edi
+    .byte 0x41
+    .byte 0x50
+#    push    r8
+    .byte 0x41
+    .byte 0x51
+#    push    r9
+    .byte 0x41
+    .byte 0x52
+#    push    r10
+    .byte 0x41
+    .byte 0x53
+#    push    r11
+    .byte 0x41
+    .byte 0x54
+#    push    r12
+    .byte 0x41
+    .byte 0x55
+#    push    r13
+    .byte 0x41
+    .byte 0x56
+#    push    r14
+    .byte 0x41
+    .byte 0x57
+#    push    r15
+    .byte 0x48
+    movl    %esp,%ebp
+#    mov     rbp, rsp
+
+##
+##  At this point the stack looks like this:
+##
+##      Calling SS
+##      Calling RSP
+##      rflags
+##      Calling CS
+##      Calling RIP
+##      Error code or 0
+##      Int num or 0ffh for unknown int num
+##      rax
+##      rcx
+##      rdx
+##      rbx
+##      rsp
+##      rbp
+##      rsi
+##      rdi
+##      r8
+##      r9
+##      r10
+##      r11
+##      r12
+##      r13
+##      r14
+##      r15 <------- RSP, RBP
+##      
+
+    call    ClearScreen
+    mov     String1, %esi 
+    call    PrintString
+    .byte 0x48
+    movl    16*8(%ebp),%eax       ## move Int number into RAX 
+    .byte 0x48
+    cmpl    $18,%eax
+    ja      PrintDefaultString
+PrintExceptionString: 
+    shll    $3,%eax             ## multiply by 8 to get offset from StringTable to actual string address
+		addl    StringTable, %eax
+    movl    (%eax),%esi
+    jmp     PrintTheString
+PrintDefaultString: 
+		movl    IntUnknownString, %esi
+    # patch Int number
+    movl    %eax,%edx
+    call    A2C
+    movb    %al,1(%esi)
+    movl    %edx,%eax
+    shrl    $4,%eax
+    call    A2C
+    movb    %al,(%esi)
+PrintTheString:        
+    call    PrintString
+		movl    String2, %esi
+    call    PrintString
+    .byte 0x48
+    movl    19*8(%ebp),%eax    # CS
+    call    PrintQword
+    movb    $':', %al
+		movb    %al, (%edi)
+    addl    $2,%edi
+    .byte 0x48
+    movl    18*8(%ebp),%eax    # RIP
+    call    PrintQword
+		movl    String3, %esi
+    call    PrintString
+
+    movl    $0xb8140,%edi
+
+		movl    StringRax, %esi
+    call    PrintString
+    .byte 0x48
+    movl    15*8(%ebp),%eax
+    call    PrintQword
+
+		movl    StringRcx, %esi
+    call    PrintString
+    .byte 0x48
+    movl    14*8(%ebp),%eax
+    call    PrintQword
+
+		movl    StringRdx, %esi
+    call    PrintString
+    .byte 0x48
+    movl    13*8(%ebp),%eax
+    call    PrintQword
+
+    movl    $0xb81e0,%edi
+
+		movl    StringRbx, %esi
+    call    PrintString
+    .byte 0x48
+    movl    12*8(%ebp),%eax
+    call    PrintQword
+
+		movl    StringRsp, %esi
+    call    PrintString
+    .byte 0x48
+    movl    21*8(%ebp),%eax
+    call    PrintQword
+
+		movl    StringRbp, %esi
+    call    PrintString
+    .byte 0x48
+    movl    10*8(%ebp),%eax
+    call    PrintQword
+
+    movl    $0xb8280,%edi
+
+		movl    StringRsi, %esi
+    call    PrintString
+    .byte 0x48
+    movl    9*8(%ebp),%eax
+    call    PrintQword
+
+		movl    StringRdi, %esi
+    call    PrintString
+    .byte 0x48
+    movl    8*8(%ebp),%eax
+    call    PrintQword
+
+		movl    StringEcode, %esi
+    call    PrintString
+    .byte 0x48
+    movl    17*8(%ebp),%eax
+    call    PrintQword
+
+    movl    $0xb8320,%edi
+
+		movl    StringR8, %esi
+    call    PrintString
+    .byte 0x48
+    movl    7*8(%ebp),%eax
+    call    PrintQword
+
+		movl    StringR9, %esi
+    call    PrintString
+    .byte 0x48
+    movl    6*8(%ebp),%eax
+    call    PrintQword
+
+		movl    StringR10, %esi
+    call    PrintString
+    .byte 0x48
+    movl    5*8(%ebp),%eax
+    call    PrintQword
+
+    movl    $0xb83c0,%edi
+
+		movl    StringR11, %esi
+    call    PrintString
+    .byte 0x48
+    movl    4*8(%ebp),%eax
+    call    PrintQword
+
+		movl    StringR12, %esi
+    call    PrintString
+    .byte 0x48
+    movl    3*8(%ebp),%eax
+    call    PrintQword
+
+		movl    StringR13, %esi
+    call    PrintString
+    .byte 0x48
+    movl    2*8(%ebp),%eax
+    call    PrintQword
+
+    movl    $0xb8460,%edi
+
+		movl    StringR14, %esi
+    call    PrintString
+    .byte 0x48
+    movl    1*8(%ebp),%eax
+    call    PrintQword
+
+		movl    StringR15, %esi
+    call    PrintString
+    .byte 0x48
+    movl    0*8(%ebp),%eax
+    call    PrintQword
+
+		movl    StringSs, %esi
+    call    PrintString
+    .byte 0x48
+    movl    22*8(%ebp),%eax
+    call    PrintQword
+
+    movl    $0xb8500,%edi
+
+		movl    StringRflags, %esi
+    call    PrintString
+    .byte 0x48
+    movl    20*8(%ebp),%eax
+    call    PrintQword
+
+    movl    $0xb8640,%edi
+
+    movl    %ebp,%esi
+    addl    $23*8,%esi
+    movl    $4,%ecx
+
+
+OuterLoop: 
+    pushl   %ecx
+    movl    $4,%ecx
+    .byte 0x48
+    movl    %edi,%edx
+
+InnerLoop: 
+    .byte 0x48
+    movl    (%esi),%eax
+    call    PrintQword
+    addl    $8,%esi
+		mov			$0x00, %al 
+    movb    %al,(%edi)
+    addl    $2,%edi
+    loop    InnerLoop
+
+    popl    %ecx
+    addl    $0xa0,%edx
+    movl    %edx,%edi
+    loop    OuterLoop
+
+
+    movl    $0xb8960,%edi
+
+    .byte 0x48
+    movl    18*8(%ebp),%eax  # RIP
+    subl    $8*8,%eax
+    .byte 0x48
+    movl    %eax,%esi       # esi = rip - 8 QWORD linear (total 16 QWORD)
+
+    movl    $4,%ecx
+
+OuterLoop1: 
+    pushl   %ecx
+    movl    $4,%ecx
+    movl    %edi,%edx
+
+InnerLoop1: 
+    .byte 0x48
+    movl    (%esi),%eax
+    call    PrintQword
+    addl    $8,%esi
+		movb		$0x00, %al
+    movb    %al,(%edi)
+    addl    $2,%edi
+    loop    InnerLoop1
+
+    popl    %ecx
+    addl    $0xa0,%edx
+    movl    %edx,%edi
+    loop    OuterLoop1
+
+
+
+    #wbinvd
+LN_C1:    

+    jmp     LN_C1
+
+#
+# return
+#
+    movl    %ebp,%esp
+#    mov     rsp, rbp
+    .byte 0x41
+    .byte 0x5f
+#    pop    r15
+    .byte 0x41
+    .byte 0x5e
+#    pop    r14
+    .byte 0x41
+    .byte 0x5d
+#    pop    r13
+    .byte 0x41
+    .byte 0x5c
+#    pop    r12
+    .byte 0x41
+    .byte 0x5b
+#    pop    r11
+    .byte 0x41
+    .byte 0x5a
+#    pop    r10
+    .byte 0x41
+    .byte 0x59
+#    pop    r9
+    .byte 0x41
+    .byte 0x58
+#    pop    r8
+    popl   %edi
+    popl   %esi
+    popl   %ebp
+    popl   %eax # esp
+    popl   %ebx
+    popl   %edx
+    popl   %ecx
+    popl   %eax
+
+    .byte 0x48
+    .byte 0x83
+    .byte 0xc4
+    .byte 0x10
+#    add    esp, 16 ; error code and INT number
+
+    .byte 0x48
+    .byte 0xcf
+#    iretq
+
+PrintString: 
+    pushl   %eax
+LN_C2: 
+		movb    (%esi), %al
+    cmpb    $0,%al
+    je      LN_C3
+ 		movb    %al, (%edi)
+    .byte 0xff
+    .byte 0xc6
+#    inc     esi
+    addl    $2,%edi
+    jmp     LN_C2
+LN_C3: 
+    popl    %eax
+    ret
+
+## RAX contains qword to print
+## RDI contains memory location (screen location) to print it to
+PrintQword: 
+    pushl   %ecx
+    pushl   %ebx
+    pushl   %eax
+
+    .byte 0x48
+    .byte 0xc7
+    .byte 0xc1
+    .long 16
+#    mov     rcx, 16
+looptop: 
+    .byte 0x48
+    roll    $4,%eax
+    movb    %al,%bl
+    andb    $0xf,%bl
+    addb    $'0', %bl
+    cmpb    $'9', %bl
+    jle     @f
+    addb    $7,%bl
+@@: 
+		movb %bl, (%edi)
+    addl    $2,%edi
+    loop    looptop
+    #wbinvd
+
+    popl    %eax
+    popl    %ebx
+    popl    %ecx
+    ret
+
+ClearScreen: 
+    pushl   %eax
+    pushl   %ecx
+
+		movb 		$0x00, %al
+    movb    $0xc,%ah
+    movl    $0xb8000,%edi
+    movl    $80*24,%ecx
+LN_C4: 
+		movw	  %ax, (%edi)
+    addl    $2,%edi
+    loop    LN_C4
+    movl    $0xb8000,%edi
+
+    popl    %ecx
+    popl    %eax
+
+    ret
+
+A2C: 
+    andb    $0xf,%al
+    addb    $'0', %al
+    cmpb    $'9', %al
+    jle     @f
+    addb    $7,%al
+LN_C5: 
+    ret
+
+String1:            .asciz      "*** INT "
+
+Int0String:         .asciz      "00h Divide by 0 -"

+Int1String:         .asciz      "01h Debug exception -"

+Int2String:         .asciz      "02h NMI -"

+Int3String:         .asciz      "03h Breakpoint -"

+Int4String:         .asciz      "04h Overflow -"

+Int5String:         .asciz      "05h Bound -"

+Int6String:         .asciz      "06h Invalid opcode -"

+Int7String:         .asciz      "07h Device not available -"

+Int8String:         .asciz      "08h Double fault -"

+Int9String:         .asciz      "09h Coprocessor seg overrun (reserved) -"

+Int10String:        .asciz      "0Ah Invalid TSS -"

+Int11String:        .asciz      "0Bh Segment not present -"

+Int12String:        .asciz      "0Ch Stack fault -"

+Int13String:        .asciz      "0Dh General protection fault -"

+Int14String:        .asciz      "0Eh Page fault -"

+Int15String:        .asciz      "0Fh (Intel reserved) -"

+Int16String:        .asciz      "10h Floating point error -"

+Int17String:        .asciz      "11h Alignment check -"

+Int18String:        .asciz      "12h Machine check -"

+Int19String:        .asciz      "13h SIMD Floating-Point Exception -"

+IntUnknownString:   .asciz      "??h Unknown interrupt -"
+
+StringTable:   .long  Int0String, Int1String, Int2String, Int3String,    \

+                      Int4String, Int5String, Int6String, Int7String,    \

+                      Int8String, Int9String, Int10String, Int11String,  \

+                      Int12String, Int13String, Int14String, Int15String,\

+                      Int16String, Int17String, Int18String, Int19String
+
+String2:           .asciz  " HALT!! *** ("
+String3:           .asciz  ")"
+StringRax:         .asciz  "RAX="
+StringRcx:         .asciz  " RCX="
+StringRdx:         .asciz  " RDX="
+StringRbx:         .asciz  "RBX="
+StringRsp:         .asciz  " RSP="
+StringRbp:         .asciz  " RBP="
+StringRsi:         .asciz  "RSI="
+StringRdi:         .asciz  " RDI="
+StringEcode:       .asciz  " ECODE="
+StringR8:          .asciz  "R8 ="
+StringR9:          .asciz  " R9 ="
+StringR10:         .asciz  " R10="
+StringR11:         .asciz  "R11="
+StringR12:         .asciz  " R12="
+StringR13:         .asciz  " R13="
+StringR14:         .asciz  "R14="
+StringR15:         .asciz  " R15="
+StringSs:          .asciz  " SS ="
+StringRflags:      .asciz  "RFLAGS="
+
+Idtr:        .float  0
+            .float  0
+
+    .org 0x21ffe
+BlockSignature: 
+    .word 0xaa55
+