nkeynes@539: /**
nkeynes@586:  * $Id$
nkeynes@539:  * 
nkeynes@926:  * Provides the implementation for the ia32 ABI variant 
nkeynes@926:  * (eg prologue, epilogue, and calling conventions). Stack frame is
nkeynes@926:  * aligned on 16-byte boundaries for the benefit of OS X (which 
nkeynes@926:  * requires it).
nkeynes@539:  *
nkeynes@539:  * Copyright (c) 2007 Nathan Keynes.
nkeynes@539:  *
nkeynes@539:  * This program is free software; you can redistribute it and/or modify
nkeynes@539:  * it under the terms of the GNU General Public License as published by
nkeynes@539:  * the Free Software Foundation; either version 2 of the License, or
nkeynes@539:  * (at your option) any later version.
nkeynes@539:  *
nkeynes@539:  * This program is distributed in the hope that it will be useful,
nkeynes@539:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
nkeynes@539:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
nkeynes@539:  * GNU General Public License for more details.
nkeynes@539:  */
nkeynes@539: 
nkeynes@926: #ifndef lxdream_ia32mac_H
nkeynes@926: #define lxdream_ia32mac_H 1
nkeynes@539: 
nkeynes@539: #define load_ptr( reg, ptr ) load_imm32( reg, (uint32_t)ptr );
nkeynes@539: 
nkeynes@539: /**
nkeynes@539:  * Note: clobbers EAX to make the indirect call - this isn't usually
nkeynes@539:  * a problem since the callee will usually clobber it anyway.
nkeynes@539:  */
nkeynes@539: static inline void call_func0( void *ptr )
nkeynes@539: {
nkeynes@926:     CALL_ptr(ptr);
nkeynes@539: }
nkeynes@539: 
nkeynes@905: #ifdef HAVE_FASTCALL
nkeynes@905: static inline void call_func1( void *ptr, int arg1 )
nkeynes@905: {
nkeynes@905:     if( arg1 != R_EAX ) {
nkeynes@905:         MOV_r32_r32( arg1, R_EAX );
nkeynes@905:     }
nkeynes@926:     CALL_ptr(ptr);
nkeynes@905: }
nkeynes@905: 
nkeynes@905: static inline void call_func2( void *ptr, int arg1, int arg2 )
nkeynes@905: {
nkeynes@905:     if( arg2 != R_EDX ) {
nkeynes@905:         MOV_r32_r32( arg2, R_EDX );
nkeynes@905:     }
nkeynes@905:     if( arg1 != R_EAX ) {
nkeynes@905:         MOV_r32_r32( arg1, R_EAX );
nkeynes@905:     }
nkeynes@926:     CALL_ptr(ptr);
nkeynes@905: }
nkeynes@905: 
nkeynes@927: static inline void call_func1_exc( void *ptr, int arg1, int pc )
nkeynes@927: {
nkeynes@927:     if( arg1 != R_EAX ) {
nkeynes@927:         MOV_r32_r32( arg1, R_EAX );
nkeynes@927:     }
nkeynes@927:     load_exc_backpatch(R_EDX);
nkeynes@927:     CALL_ptr(ptr);
nkeynes@927: }   
nkeynes@927: 
nkeynes@927: static inline void call_func2_exc( void *ptr, int arg1, int arg2, int pc )
nkeynes@927: {
nkeynes@927:     if( arg2 != R_EDX ) {
nkeynes@927:         MOV_r32_r32( arg2, R_EDX );
nkeynes@927:     }
nkeynes@927:     if( arg1 != R_EAX ) {
nkeynes@927:         MOV_r32_r32( arg1, R_EAX );
nkeynes@927:     }
nkeynes@927:     load_exc_backpatch(R_ECX);
nkeynes@927:     CALL_ptr(ptr);
nkeynes@927: }
nkeynes@927: 
nkeynes@905: /**
nkeynes@905:  * Write a double (64-bit) value into memory, with the first word in arg2a, and
nkeynes@905:  * the second in arg2b
nkeynes@905:  */
nkeynes@905: static inline void MEM_WRITE_DOUBLE( int addr, int arg2a, int arg2b )
nkeynes@905: {
nkeynes@926:     MOV_r32_esp8(addr, 0);
nkeynes@926:     MOV_r32_esp8(arg2b, 4);
nkeynes@905:     call_func2(sh4_write_long, addr, arg2a);
nkeynes@926:     MOV_esp8_r32(0, R_EAX);
nkeynes@926:     MOV_esp8_r32(4, R_EDX);
nkeynes@905:     ADD_imm8s_r32(4, R_EAX);
nkeynes@905:     call_func0(sh4_write_long);
nkeynes@905: }
nkeynes@905: 
nkeynes@905: /**
nkeynes@905:  * Read a double (64-bit) value from memory, writing the first word into arg2a
nkeynes@905:  * and the second into arg2b. The addr must not be in EAX
nkeynes@905:  */
nkeynes@905: static inline void MEM_READ_DOUBLE( int addr, int arg2a, int arg2b )
nkeynes@905: {
nkeynes@926:     MOV_r32_esp8(addr, 0);
nkeynes@905:     call_func1(sh4_read_long, addr);
nkeynes@926:     MOV_r32_esp8(R_EAX, 4);
nkeynes@926:     MOV_esp8_r32(0, R_EAX);
nkeynes@905:     ADD_imm8s_r32(4, R_EAX);
nkeynes@905:     call_func0(sh4_read_long);
nkeynes@905:     if( arg2b != R_EAX ) {
nkeynes@905:         MOV_r32_r32(R_EAX, arg2b);
nkeynes@905:     }
nkeynes@926:     MOV_esp8_r32(4, arg2a);
nkeynes@905: }
nkeynes@905: #else
nkeynes@539: static inline void call_func1( void *ptr, int arg1 )
nkeynes@539: {
nkeynes@926:     SUB_imm8s_r32( 12, R_ESP );
nkeynes@539:     PUSH_r32(arg1);
nkeynes@926:     CALL_ptr(ptr);
nkeynes@926:     ADD_imm8s_r32( 16, R_ESP );
nkeynes@539: }
nkeynes@539: 
nkeynes@539: static inline void call_func2( void *ptr, int arg1, int arg2 )
nkeynes@539: {
nkeynes@926:     SUB_imm8s_r32( 8, R_ESP );
nkeynes@539:     PUSH_r32(arg2);
nkeynes@539:     PUSH_r32(arg1);
nkeynes@926:     CALL_ptr(ptr);
nkeynes@926:     ADD_imm8s_r32( 16, R_ESP );
nkeynes@539: }
nkeynes@539: 
nkeynes@539: /**
nkeynes@539:  * Write a double (64-bit) value into memory, with the first word in arg2a, and
nkeynes@539:  * the second in arg2b
nkeynes@539:  */
nkeynes@539: static inline void MEM_WRITE_DOUBLE( int addr, int arg2a, int arg2b )
nkeynes@539: {
nkeynes@926:     SUB_imm8s_r32( 8, R_ESP );
nkeynes@539:     PUSH_r32(arg2b);
nkeynes@926:     LEA_r32disp8_r32( addr, 4, arg2b );
nkeynes@926:     PUSH_r32(arg2b);
nkeynes@926:     SUB_imm8s_r32( 8, R_ESP );
nkeynes@539:     PUSH_r32(arg2a);
nkeynes@539:     PUSH_r32(addr);
nkeynes@926:     CALL_ptr(sh4_write_long);
nkeynes@926:     ADD_imm8s_r32( 16, R_ESP );
nkeynes@926:     CALL_ptr(sh4_write_long);
nkeynes@926:     ADD_imm8s_r32( 16, R_ESP );
nkeynes@539: }
nkeynes@539: 
nkeynes@539: /**
nkeynes@539:  * Read a double (64-bit) value from memory, writing the first word into arg2a
nkeynes@539:  * and the second into arg2b. The addr must not be in EAX
nkeynes@539:  */
nkeynes@539: static inline void MEM_READ_DOUBLE( int addr, int arg2a, int arg2b )
nkeynes@539: {
nkeynes@926:     SUB_imm8s_r32( 12, R_ESP );
nkeynes@539:     PUSH_r32(addr);
nkeynes@926:     CALL_ptr(sh4_read_long);
nkeynes@926:     MOV_r32_esp8(R_EAX, 4);
nkeynes@926:     ADD_imm8s_esp8(4, 0);
nkeynes@926:     CALL_ptr(sh4_read_long);
nkeynes@926:     if( arg2b != R_EAX ) {
nkeynes@926:         MOV_r32_r32( R_EAX, arg2b );
nkeynes@926:     }
nkeynes@926:     MOV_esp8_r32( 4, arg2a );
nkeynes@926:     ADD_imm8s_r32( 16, R_ESP );
nkeynes@539: }
nkeynes@926: 
nkeynes@905: #endif
nkeynes@539: 
nkeynes@539: /**
nkeynes@539:  * Emit the 'start of block' assembly. Sets up the stack frame and save
nkeynes@539:  * SI/DI as required
nkeynes@926:  * Allocates 8 bytes for local variables, which also has the convenient
nkeynes@926:  * side-effect of aligning the stack.
nkeynes@539:  */
nkeynes@901: void enter_block( ) 
nkeynes@539: {
nkeynes@539:     PUSH_r32(R_EBP);
nkeynes@669:     load_ptr( R_EBP, ((uint8_t *)&sh4r) + 128 );
nkeynes@926:     SUB_imm8s_r32( 8, R_ESP ); 
nkeynes@926: }
nkeynes@736: 
nkeynes@926: static inline void exit_block( )
nkeynes@926: {
nkeynes@926:     ADD_imm8s_r32( 8, R_ESP );
nkeynes@926:     POP_r32(R_EBP);
nkeynes@926:     RET();
nkeynes@539: }
nkeynes@539: 
nkeynes@539: /**
nkeynes@926:  * Exit the block with sh4r.new_pc written with the target pc
nkeynes@539:  */
nkeynes@586: void exit_block_pcset( sh4addr_t pc )
nkeynes@539: {
nkeynes@539:     load_imm32( R_ECX, ((pc - sh4_x86.block_start_pc)>>1)*sh4_cpu_period ); // 5
nkeynes@539:     ADD_r32_sh4r( R_ECX, REG_OFFSET(slice_cycle) );    // 6
nkeynes@590:     load_spreg( R_EAX, R_PC );
nkeynes@590:     if( sh4_x86.tlb_on ) {
nkeynes@736:         call_func1(xlat_get_code_by_vma,R_EAX);
nkeynes@590:     } else {
nkeynes@736:         call_func1(xlat_get_code,R_EAX);
nkeynes@926:     }
nkeynes@926:     exit_block();
nkeynes@590: }
nkeynes@590: 
nkeynes@590: /**
nkeynes@590:  * Exit the block with sh4r.new_pc written with the target pc
nkeynes@590:  */
nkeynes@590: void exit_block_newpcset( sh4addr_t pc )
nkeynes@590: {
nkeynes@590:     load_imm32( R_ECX, ((pc - sh4_x86.block_start_pc)>>1)*sh4_cpu_period ); // 5
nkeynes@590:     ADD_r32_sh4r( R_ECX, REG_OFFSET(slice_cycle) );    // 6
nkeynes@590:     load_spreg( R_EAX, R_NEW_PC );
nkeynes@590:     store_spreg( R_EAX, R_PC );
nkeynes@586:     if( sh4_x86.tlb_on ) {
nkeynes@736:         call_func1(xlat_get_code_by_vma,R_EAX);
nkeynes@586:     } else {
nkeynes@736:         call_func1(xlat_get_code,R_EAX);
nkeynes@926:     }
nkeynes@926:     exit_block();
nkeynes@539: }
nkeynes@539: 
nkeynes@586: 
nkeynes@539: /**
nkeynes@539:  * Exit the block to an absolute PC
nkeynes@539:  */
nkeynes@926: void exit_block_abs( sh4addr_t pc, sh4addr_t endpc )
nkeynes@539: {
nkeynes@539:     load_imm32( R_ECX, pc );                            // 5
nkeynes@539:     store_spreg( R_ECX, REG_OFFSET(pc) );               // 3
nkeynes@586:     if( IS_IN_ICACHE(pc) ) {
nkeynes@736:         MOV_moff32_EAX( xlat_get_lut_entry(GET_ICACHE_PHYS(pc)) ); // 5
nkeynes@926:         AND_imm8s_r32( 0xFC, R_EAX ); // 3
nkeynes@586:     } else if( sh4_x86.tlb_on ) {
nkeynes@736:         call_func1(xlat_get_code_by_vma,R_ECX);
nkeynes@586:     } else {
nkeynes@736:         call_func1(xlat_get_code,R_ECX);
nkeynes@586:     }
nkeynes@586:     load_imm32( R_ECX, ((endpc - sh4_x86.block_start_pc)>>1)*sh4_cpu_period ); // 5
nkeynes@586:     ADD_r32_sh4r( R_ECX, REG_OFFSET(slice_cycle) );     // 6
nkeynes@926:     exit_block();
nkeynes@586: }
nkeynes@586: 
nkeynes@586: /**
nkeynes@586:  * Exit the block to a relative PC
nkeynes@586:  */
nkeynes@586: void exit_block_rel( sh4addr_t pc, sh4addr_t endpc )
nkeynes@586: {
nkeynes@586:     load_imm32( R_ECX, pc - sh4_x86.block_start_pc );   // 5
nkeynes@586:     ADD_sh4r_r32( R_PC, R_ECX );
nkeynes@586:     store_spreg( R_ECX, REG_OFFSET(pc) );               // 3
nkeynes@586:     if( IS_IN_ICACHE(pc) ) {
nkeynes@736:         MOV_moff32_EAX( xlat_get_lut_entry(GET_ICACHE_PHYS(pc)) ); // 5
nkeynes@926:         AND_imm8s_r32( 0xFC, R_EAX ); // 3
nkeynes@586:     } else if( sh4_x86.tlb_on ) {
nkeynes@736:         call_func1(xlat_get_code_by_vma,R_ECX);
nkeynes@586:     } else {
nkeynes@736:         call_func1(xlat_get_code,R_ECX);
nkeynes@586:     }
nkeynes@539:     load_imm32( R_ECX, ((endpc - sh4_x86.block_start_pc)>>1)*sh4_cpu_period ); // 5
nkeynes@539:     ADD_r32_sh4r( R_ECX, REG_OFFSET(slice_cycle) );     // 6
nkeynes@926:     exit_block();
nkeynes@539: }
nkeynes@539: 
nkeynes@539: /**
nkeynes@539:  * Write the block trailer (exception handling block)
nkeynes@539:  */
nkeynes@539: void sh4_translate_end_block( sh4addr_t pc ) {
nkeynes@539:     if( sh4_x86.branch_taken == FALSE ) {
nkeynes@736:         // Didn't exit unconditionally already, so write the termination here
nkeynes@736:         exit_block_rel( pc, pc );
nkeynes@539:     }
nkeynes@539:     if( sh4_x86.backpatch_posn != 0 ) {
nkeynes@736:         unsigned int i;
nkeynes@736:         // Raise exception
nkeynes@736:         uint8_t *end_ptr = xlat_output;
nkeynes@736:         MOV_r32_r32( R_EDX, R_ECX );
nkeynes@736:         ADD_r32_r32( R_EDX, R_ECX );
nkeynes@736:         ADD_r32_sh4r( R_ECX, R_PC );
nkeynes@736:         MOV_moff32_EAX( &sh4_cpu_period );
nkeynes@736:         MUL_r32( R_EDX );
nkeynes@736:         ADD_r32_sh4r( R_EAX, REG_OFFSET(slice_cycle) );
nkeynes@539: 
nkeynes@926:         POP_r32(R_EAX);
nkeynes@907:         call_func1( sh4_raise_exception, R_EAX );
nkeynes@736:         load_spreg( R_EAX, R_PC );
nkeynes@736:         if( sh4_x86.tlb_on ) {
nkeynes@736:             call_func1(xlat_get_code_by_vma,R_EAX);
nkeynes@736:         } else {
nkeynes@736:             call_func1(xlat_get_code,R_EAX);
nkeynes@736:         }
nkeynes@926:         exit_block();
nkeynes@539: 
nkeynes@736:         // Exception already raised - just cleanup
nkeynes@736:         uint8_t *preexc_ptr = xlat_output;
nkeynes@736:         MOV_r32_r32( R_EDX, R_ECX );
nkeynes@736:         ADD_r32_r32( R_EDX, R_ECX );
nkeynes@736:         ADD_r32_sh4r( R_ECX, R_SPC );
nkeynes@736:         MOV_moff32_EAX( &sh4_cpu_period );
nkeynes@736:         MUL_r32( R_EDX );
nkeynes@736:         ADD_r32_sh4r( R_EAX, REG_OFFSET(slice_cycle) );
nkeynes@736:         load_spreg( R_EAX, R_PC );
nkeynes@736:         if( sh4_x86.tlb_on ) {
nkeynes@736:             call_func1(xlat_get_code_by_vma,R_EAX);
nkeynes@736:         } else {
nkeynes@736:             call_func1(xlat_get_code,R_EAX);
nkeynes@736:         }
nkeynes@926:         exit_block();
nkeynes@586: 
nkeynes@736:         for( i=0; i< sh4_x86.backpatch_posn; i++ ) {
nkeynes@736:             uint32_t *fixup_addr = (uint32_t *)&xlat_current_block->code[sh4_x86.backpatch_list[i].fixup_offset];
nkeynes@736:             if( sh4_x86.backpatch_list[i].exc_code < 0 ) {
nkeynes@927:                 if( sh4_x86.backpatch_list[i].exc_code == -2 ) {
nkeynes@927:                     *fixup_addr = (uint32_t)xlat_output;
nkeynes@927:                 } else {
nkeynes@927:                     *fixup_addr += xlat_output - (uint8_t *)&xlat_current_block->code[sh4_x86.backpatch_list[i].fixup_offset] - 4;
nkeynes@927:                 }
nkeynes@736:                 load_imm32( R_EDX, sh4_x86.backpatch_list[i].fixup_icount );
nkeynes@736:                 int rel = preexc_ptr - xlat_output;
nkeynes@736:                 JMP_rel(rel);
nkeynes@736:             } else {
nkeynes@927:                 *fixup_addr += xlat_output - (uint8_t *)&xlat_current_block->code[sh4_x86.backpatch_list[i].fixup_offset] - 4;
nkeynes@736:                 PUSH_imm32( sh4_x86.backpatch_list[i].exc_code );
nkeynes@736:                 load_imm32( R_EDX, sh4_x86.backpatch_list[i].fixup_icount );
nkeynes@736:                 int rel = end_ptr - xlat_output;
nkeynes@736:                 JMP_rel(rel);
nkeynes@736:             }
nkeynes@736:         }
nkeynes@539:     }
nkeynes@539: }
nkeynes@539: 
nkeynes@926: 
nkeynes@899: /**
nkeynes@899:  * The unwind methods only work if we compiled with DWARF2 frame information
nkeynes@899:  * (ie -fexceptions), otherwise we have to use the direct frame scan.
nkeynes@899:  */
nkeynes@899: #ifdef HAVE_EXCEPTIONS
nkeynes@899: #include <unwind.h>
nkeynes@899: 
nkeynes@899: struct UnwindInfo {
nkeynes@926:     uintptr_t block_start;
nkeynes@926:     uintptr_t block_end;
nkeynes@926:     void *pc;
nkeynes@899: };
nkeynes@899: 
nkeynes@899: _Unwind_Reason_Code xlat_check_frame( struct _Unwind_Context *context, void *arg )
nkeynes@899: {
nkeynes@926:     struct UnwindInfo *info = arg;
nkeynes@926:     void *pc = (void *)_Unwind_GetIP(context);
nkeynes@926:     if( ((uintptr_t)pc) >= info->block_start && ((uintptr_t)pc) < info->block_end ) {
nkeynes@926:         info->pc = pc;
nkeynes@899:         return _URC_NORMAL_STOP;
nkeynes@899:     }
nkeynes@899: 
nkeynes@899:     return _URC_NO_REASON;
nkeynes@899: }
nkeynes@899: 
nkeynes@906: void *xlat_get_native_pc( void *code, uint32_t code_size )
nkeynes@899: {
nkeynes@899:     struct _Unwind_Exception exc;
nkeynes@899:     struct UnwindInfo info;
nkeynes@899: 
nkeynes@926:     info.pc = NULL;
nkeynes@926:     info.block_start = (uintptr_t)code;
nkeynes@926:     info.block_end = info.block_start + code_size;
nkeynes@899:     void *result = NULL;
nkeynes@899:     _Unwind_Backtrace( xlat_check_frame, &info );
nkeynes@926:     return info.pc;
nkeynes@899: }
nkeynes@899: #else 
nkeynes@906: void *xlat_get_native_pc( void *code, uint32_t code_size )
nkeynes@586: {
nkeynes@586:     void *result = NULL;
nkeynes@586:     asm(
nkeynes@736:         "mov %%ebp, %%eax\n\t"
nkeynes@736:         "mov $0x8, %%ecx\n\t"
nkeynes@736:         "mov %1, %%edx\n"
nkeynes@736:         "frame_loop: test %%eax, %%eax\n\t"
nkeynes@736:         "je frame_not_found\n\t"
nkeynes@736:         "cmp (%%eax), %%edx\n\t"
nkeynes@736:         "je frame_found\n\t"
nkeynes@736:         "sub $0x1, %%ecx\n\t"
nkeynes@736:         "je frame_not_found\n\t"
nkeynes@736:         "movl (%%eax), %%eax\n\t"
nkeynes@736:         "jmp frame_loop\n"
nkeynes@736:         "frame_found: movl 0x4(%%eax), %0\n"
nkeynes@736:         "frame_not_found:"
nkeynes@736:         : "=r" (result)
nkeynes@736:         : "r" (((uint8_t *)&sh4r) + 128 )
nkeynes@736:         : "eax", "ecx", "edx" );
nkeynes@586:     return result;
nkeynes@586: }
nkeynes@899: #endif
nkeynes@586: 
nkeynes@926: #endif /* !lxdream_ia32mac.h */
nkeynes@539: 
nkeynes@539: