nkeynes@539: /**
nkeynes@586:  * $Id$
nkeynes@539:  * 
nkeynes@736:  * Provides the implementation for the AMD64 ABI (eg prologue, epilogue, and
nkeynes@539:  * calling conventions)
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@736: #ifndef lxdream_ia64abi_H
nkeynes@736: #define lxdream_ia64abi_H 1
nkeynes@539: 
nkeynes@586: #include <unwind.h>
nkeynes@539: 
nkeynes@539: #define load_ptr( reg, ptr ) load_imm64( reg, (uint64_t)ptr );
nkeynes@736: 
nkeynes@991: static inline void decode_address( int addr_reg )
nkeynes@953: {
nkeynes@953:     uintptr_t base = (sh4r.xlat_sh4_mode&SR_MD) ? (uintptr_t)sh4_address_space : (uintptr_t)sh4_user_address_space;
nkeynes@991:     MOVL_r32_r32( addr_reg, REG_RCX ); 
nkeynes@991:     SHRL_imm_r32( 12, REG_RCX ); 
nkeynes@991:     MOVP_immptr_rptr( base, REG_RDI );
nkeynes@991:     MOVP_sib_rptr(3, REG_RCX, REG_RDI, 0, REG_RCX);
nkeynes@953: }
nkeynes@953: 
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:  * Size: 12 bytes
nkeynes@539:  */
nkeynes@539: #define CALL_FUNC0_SIZE 12
nkeynes@539: static inline void call_func0( void *ptr )
nkeynes@539: {
nkeynes@991:     MOVQ_imm64_r64((uint64_t)ptr, REG_RAX);
nkeynes@991:     CALL_r32(REG_RAX);
nkeynes@539: }
nkeynes@539: 
nkeynes@539: static inline void call_func1( void *ptr, int arg1 )
nkeynes@539: {
nkeynes@991:     MOVQ_r64_r64(arg1, REG_RDI);
nkeynes@539:     call_func0(ptr);
nkeynes@539: }
nkeynes@539: 
nkeynes@927: static inline void call_func1_exc( void *ptr, int arg1, int pc )
nkeynes@927: {
nkeynes@991:     MOVQ_r64_r64(arg1, REG_RDI);
nkeynes@991:     MOVP_immptr_rptr(0, REG_RSI);
nkeynes@991:     sh4_x86_add_backpatch( xlat_output, pc, -2 );
nkeynes@927:     call_func0(ptr);
nkeynes@927: }
nkeynes@927: 
nkeynes@953: static inline void call_func1_r32disp8( int preg, uint32_t disp8, int arg1 )
nkeynes@953: {
nkeynes@991:     MOVQ_r64_r64(arg1, REG_RDI);
nkeynes@991:     CALL_r32disp(preg, disp8);    
nkeynes@953: }
nkeynes@953: 
nkeynes@953: static inline void call_func1_r32disp8_exc( int preg, uint32_t disp8, int arg1, int pc )
nkeynes@953: {
nkeynes@991:     MOVQ_r64_r64(arg1, REG_RDI);
nkeynes@991:     MOVP_immptr_rptr(0, REG_RSI);
nkeynes@991:     sh4_x86_add_backpatch( xlat_output, pc, -2 );
nkeynes@991:     CALL_r32disp(preg, disp8);
nkeynes@953: }
nkeynes@953: 
nkeynes@539: static inline void call_func2( void *ptr, int arg1, int arg2 )
nkeynes@539: {
nkeynes@991:     MOVQ_r64_r64(arg1, REG_RDI);
nkeynes@991:     MOVQ_r64_r64(arg2, REG_RSI);
nkeynes@539:     call_func0(ptr);
nkeynes@539: }
nkeynes@539: 
nkeynes@953: static inline void call_func2_r32disp8( int preg, uint32_t disp8, int arg1, int arg2 )
nkeynes@539: {
nkeynes@991:     MOVQ_r64_r64(arg1, REG_RDI);
nkeynes@991:     MOVQ_r64_r64(arg2, REG_RSI);
nkeynes@991:     CALL_r32disp(preg, disp8);    
nkeynes@539: }
nkeynes@539: 
nkeynes@953: static inline void call_func2_r32disp8_exc( int preg, uint32_t disp8, int arg1, int arg2, int pc )
nkeynes@539: {
nkeynes@991:     MOVQ_r64_r64(arg1, REG_RDI);
nkeynes@991:     MOVQ_r64_r64(arg2, REG_RSI);
nkeynes@991:     MOVP_immptr_rptr(0, REG_RDX);
nkeynes@991:     sh4_x86_add_backpatch( xlat_output, pc, -2 );
nkeynes@991:     CALL_r32disp(preg, disp8);
nkeynes@539: }
nkeynes@539: 
nkeynes@539: 
nkeynes@953: 
nkeynes@539: /**
nkeynes@539:  * Emit the 'start of block' assembly. Sets up the stack frame and save
nkeynes@539:  * SI/DI as required
nkeynes@539:  */
nkeynes@901: void enter_block( ) 
nkeynes@539: {
nkeynes@991:     PUSH_r32(REG_RBP);
nkeynes@991:     load_ptr( REG_RBP, ((uint8_t *)&sh4r) + 128 );
nkeynes@926:     // Minimum aligned allocation is 16 bytes
nkeynes@991:     SUBQ_imms_r64( 16, REG_RSP );
nkeynes@926: }
nkeynes@926: 
nkeynes@926: static inline void exit_block( )
nkeynes@926: {
nkeynes@991:     ADDQ_imms_r64( 16, REG_RSP );
nkeynes@991:     POP_r32(REG_RBP);
nkeynes@926:     RET();
nkeynes@539: }
nkeynes@908: 
nkeynes@539: /**
nkeynes@539:  * Exit the block with sh4r.pc already written
nkeynes@539:  */
nkeynes@586: void exit_block_pcset( sh4addr_t pc )
nkeynes@539: {
nkeynes@991:     load_imm32( REG_ECX, ((pc - sh4_x86.block_start_pc)>>1)*sh4_cpu_period ); // 5
nkeynes@991:     ADDL_r32_rbpdisp( REG_ECX, REG_OFFSET(slice_cycle) );    // 6
nkeynes@991:     load_spreg( REG_RAX, R_PC );
nkeynes@590:     if( sh4_x86.tlb_on ) {
nkeynes@991:         call_func1(xlat_get_code_by_vma,REG_RAX);
nkeynes@590:     } else {
nkeynes@991:         call_func1(xlat_get_code,REG_RAX);
nkeynes@590:     }
nkeynes@926:     exit_block();
nkeynes@590: }
nkeynes@590: 
nkeynes@590: /**
nkeynes@590:  * Exit the block with sh4r.new_pc written with the target address
nkeynes@590:  */
nkeynes@590: void exit_block_newpcset( sh4addr_t pc )
nkeynes@590: {
nkeynes@991:     load_imm32( REG_ECX, ((pc - sh4_x86.block_start_pc)>>1)*sh4_cpu_period ); // 5
nkeynes@991:     ADDL_r32_rbpdisp( REG_ECX, REG_OFFSET(slice_cycle) );    // 6
nkeynes@991:     load_spreg( REG_RAX, R_NEW_PC );
nkeynes@991:     store_spreg( REG_RAX, R_PC );
nkeynes@586:     if( sh4_x86.tlb_on ) {
nkeynes@991:         call_func1(xlat_get_code_by_vma,REG_RAX);
nkeynes@586:     } else {
nkeynes@991:         call_func1(xlat_get_code,REG_RAX);
nkeynes@586:     }
nkeynes@926:     exit_block();
nkeynes@539: }
nkeynes@539: 
nkeynes@586: #define EXIT_BLOCK_SIZE(pc) (25 + (IS_IN_ICACHE(pc)?10:CALL_FUNC1_SIZE))
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@991:     load_imm32( REG_RCX, pc );                            // 5
nkeynes@991:     store_spreg( REG_RCX, REG_OFFSET(pc) );               // 3
nkeynes@586:     if( IS_IN_ICACHE(pc) ) {
nkeynes@991:         MOVP_moffptr_rax( xlat_get_lut_entry(pc) );
nkeynes@991:         ANDQ_imms_r64( 0xFFFFFFFC, REG_RAX ); // 4
nkeynes@586:     } else if( sh4_x86.tlb_on ) {
nkeynes@991:         call_func1(xlat_get_code_by_vma, REG_RCX);
nkeynes@586:     } else {
nkeynes@991:         call_func1(xlat_get_code,REG_RCX);
nkeynes@586:     }
nkeynes@991:     load_imm32( REG_ECX, ((endpc - sh4_x86.block_start_pc)>>1)*sh4_cpu_period ); // 5
nkeynes@991:     ADDL_r32_rbpdisp( REG_ECX, REG_OFFSET(slice_cycle) );     // 6
nkeynes@926:     exit_block();
nkeynes@539: }
nkeynes@539: 
nkeynes@539: 
nkeynes@586: #define EXIT_BLOCK_REL_SIZE(pc)  (28 + (IS_IN_ICACHE(pc)?10:CALL_FUNC1_SIZE))
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@991:     load_imm32( REG_ECX, pc - sh4_x86.block_start_pc );   // 5
nkeynes@991:     ADDL_rbpdisp_r32( R_PC, REG_ECX );
nkeynes@991:     store_spreg( REG_ECX, REG_OFFSET(pc) );               // 3
nkeynes@586:     if( IS_IN_ICACHE(pc) ) {
nkeynes@991:         MOVP_moffptr_rax( xlat_get_lut_entry(GET_ICACHE_PHYS(pc)) ); // 5
nkeynes@991:         ANDQ_imms_r64( 0xFFFFFFFC, REG_RAX ); // 4
nkeynes@586:     } else if( sh4_x86.tlb_on ) {
nkeynes@991:         call_func1(xlat_get_code_by_vma,REG_RCX);
nkeynes@586:     } else {
nkeynes@991:         call_func1(xlat_get_code,REG_RCX);
nkeynes@586:     }
nkeynes@991:     load_imm32( REG_ECX, ((endpc - sh4_x86.block_start_pc)>>1)*sh4_cpu_period ); // 5
nkeynes@991:     ADDL_r32_rbpdisp( REG_ECX, REG_OFFSET(slice_cycle) );     // 6
nkeynes@926:     exit_block();
nkeynes@586: }
nkeynes@586: 
nkeynes@539: /**
nkeynes@957:  * Exit unconditionally with a general exception
nkeynes@957:  */
nkeynes@957: void exit_block_exc( int code, sh4addr_t pc )
nkeynes@957: {
nkeynes@991:     load_imm32( REG_ECX, pc - sh4_x86.block_start_pc );   // 5
nkeynes@991:     ADDL_r32_rbpdisp( REG_ECX, R_PC );
nkeynes@991:     load_imm32( REG_ECX, ((pc - sh4_x86.block_start_pc)>>1)*sh4_cpu_period ); // 5
nkeynes@991:     ADDL_r32_rbpdisp( REG_ECX, REG_OFFSET(slice_cycle) );     // 6
nkeynes@991:     load_imm32( REG_RAX, code );
nkeynes@991:     call_func1( sh4_raise_exception, REG_RAX );
nkeynes@957:     
nkeynes@991:     load_spreg( REG_RAX, R_PC );
nkeynes@957:     if( sh4_x86.tlb_on ) {
nkeynes@991:         call_func1(xlat_get_code_by_vma,REG_RAX);
nkeynes@957:     } else {
nkeynes@991:         call_func1(xlat_get_code,REG_RAX);
nkeynes@957:     }
nkeynes@957: 
nkeynes@957:     exit_block();
nkeynes@957: }    
nkeynes@957: 
nkeynes@957: 
nkeynes@957: /**
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@991:         MOVL_r32_r32( REG_RDX, REG_RCX );
nkeynes@991:         ADDL_r32_r32( REG_RDX, REG_RCX );
nkeynes@991:         ADDL_r32_rbpdisp( REG_RCX, R_PC );
nkeynes@991:         MOVL_moffptr_eax( &sh4_cpu_period );
nkeynes@991:         MULL_r32( REG_RDX );
nkeynes@991:         ADDL_r32_rbpdisp( REG_RAX, REG_OFFSET(slice_cycle) );
nkeynes@539: 
nkeynes@736:         call_func0( sh4_raise_exception );
nkeynes@991:         load_spreg( REG_RAX, R_PC );
nkeynes@736:         if( sh4_x86.tlb_on ) {
nkeynes@991:             call_func1(xlat_get_code_by_vma,REG_RAX);
nkeynes@736:         } else {
nkeynes@991:             call_func1(xlat_get_code,REG_RAX);
nkeynes@736:         }
nkeynes@926:         exit_block();
nkeynes@926:         
nkeynes@736:         // Exception already raised - just cleanup
nkeynes@736:         uint8_t *preexc_ptr = xlat_output;
nkeynes@991:         MOVL_r32_r32( REG_EDX, REG_ECX );
nkeynes@991:         ADDL_r32_r32( REG_EDX, REG_ECX );
nkeynes@991:         ADDL_r32_rbpdisp( REG_ECX, R_SPC );
nkeynes@991:         MOVL_moffptr_eax( &sh4_cpu_period );
nkeynes@991:         MULL_r32( REG_EDX );
nkeynes@991:         ADDL_r32_rbpdisp( REG_EAX, REG_OFFSET(slice_cycle) );
nkeynes@991:         load_spreg( REG_RDI, R_PC );
nkeynes@736:         if( sh4_x86.tlb_on ) {
nkeynes@736:             call_func0(xlat_get_code_by_vma);
nkeynes@736:         } else {
nkeynes@736:             call_func0(xlat_get_code);
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:                     *((uintptr_t *)fixup_addr) = (uintptr_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@991:                 load_imm32( REG_RDX, sh4_x86.backpatch_list[i].fixup_icount );
nkeynes@736:                 int rel = preexc_ptr - xlat_output;
nkeynes@991:                 JMP_prerel(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@991:                 load_imm32( REG_RDI, sh4_x86.backpatch_list[i].exc_code );
nkeynes@991:                 load_imm32( REG_RDX, sh4_x86.backpatch_list[i].fixup_icount );
nkeynes@736:                 int rel = end_ptr - xlat_output;
nkeynes@991:                 JMP_prerel(rel);
nkeynes@736:             }
nkeynes@736:         }
nkeynes@539:     }
nkeynes@539: }
nkeynes@539: 
nkeynes@926: struct UnwindInfo {
nkeynes@926:     uintptr_t block_start;
nkeynes@926:     uintptr_t block_end;
nkeynes@926:     void *pc;
nkeynes@926: };
nkeynes@926: 
nkeynes@586: _Unwind_Reason_Code xlat_check_frame( struct _Unwind_Context *context, void *arg )
nkeynes@586: {
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@586:         return _URC_NORMAL_STOP;
nkeynes@586:     }
nkeynes@926:    
nkeynes@586:     return _URC_NO_REASON;
nkeynes@586: }
nkeynes@586: 
nkeynes@926: void *xlat_get_native_pc( void *code, uint32_t code_size )
nkeynes@586: {
nkeynes@586:     struct _Unwind_Exception exc;
nkeynes@926:     struct UnwindInfo info;
nkeynes@736: 
nkeynes@926:     info.pc = NULL;
nkeynes@926:     info.block_start = (uintptr_t)code;
nkeynes@926:     info.block_end = info.block_start + code_size;
nkeynes@586:     void *result = NULL;
nkeynes@926:     _Unwind_Backtrace( xlat_check_frame, &info );
nkeynes@926:     return info.pc;
nkeynes@586: }
nkeynes@586: 
nkeynes@736: #endif /* !lxdream_ia64abi_H */