1.1 --- a/src/sh4/ia32abi.h	Wed Nov 21 11:40:15 2007 +0000
     1.2 +++ b/src/sh4/ia32abi.h	Tue Jan 15 20:50:23 2008 +0000
     1.3 @@ -1,5 +1,5 @@
     1.4  /**
     1.5 - * $Id: sh4x86.in,v 1.20 2007-11-08 11:54:16 nkeynes Exp $
     1.6 + * $Id$
     1.7   * 
     1.8   * Provides the implementation for the ia32 ABI (eg prologue, epilogue, and
     1.9   * calling conventions)
    1.10 @@ -80,19 +80,16 @@
    1.11  {
    1.12      PUSH_r32(addr);
    1.13      call_func0(sh4_read_long);
    1.14 -    POP_r32(addr);
    1.15 +    POP_r32(R_ECX);
    1.16      PUSH_r32(R_EAX);
    1.17 -    ADD_imm8s_r32( 4, addr );
    1.18 -    PUSH_r32(addr);
    1.19 +    ADD_imm8s_r32( 4, R_ECX );
    1.20 +    PUSH_r32(R_ECX);
    1.21      call_func0(sh4_read_long);
    1.22      ADD_imm8s_r32( 4, R_ESP );
    1.23      MOV_r32_r32( R_EAX, arg2b );
    1.24      POP_r32(arg2a);
    1.25  }
    1.26  
    1.27 -#define EXIT_BLOCK_SIZE 29
    1.28 -
    1.29 -
    1.30  /**
    1.31   * Emit the 'start of block' assembly. Sets up the stack frame and save
    1.32   * SI/DI as required
    1.33 @@ -108,7 +105,9 @@
    1.34      sh4_x86.fpuen_checked = FALSE;
    1.35      sh4_x86.branch_taken = FALSE;
    1.36      sh4_x86.backpatch_posn = 0;
    1.37 +    sh4_x86.recovery_posn = 0;
    1.38      sh4_x86.block_start_pc = pc;
    1.39 +    sh4_x86.tlb_on = IS_MMU_ENABLED();
    1.40      sh4_x86.tstate = TSTATE_NONE;
    1.41  #ifdef STACK_ALIGN
    1.42  	sh4_x86.stack_posn = 8;
    1.43 @@ -119,16 +118,23 @@
    1.44   * Exit the block with sh4r.pc already written
    1.45   * Bytes: 15
    1.46   */
    1.47 -void exit_block_pcset( pc )
    1.48 +void exit_block_pcset( sh4addr_t pc )
    1.49  {
    1.50      load_imm32( R_ECX, ((pc - sh4_x86.block_start_pc)>>1)*sh4_cpu_period ); // 5
    1.51      ADD_r32_sh4r( R_ECX, REG_OFFSET(slice_cycle) );    // 6
    1.52      load_spreg( R_EAX, REG_OFFSET(pc) );
    1.53 -    call_func1(xlat_get_code,R_EAX);
    1.54 +    if( sh4_x86.tlb_on ) {
    1.55 +	call_func1(xlat_get_code_by_vma,R_EAX);
    1.56 +    } else {
    1.57 +	call_func1(xlat_get_code,R_EAX);
    1.58 +    } 
    1.59      POP_r32(R_EBP);
    1.60      RET();
    1.61  }
    1.62  
    1.63 +#define EXIT_BLOCK_SIZE(pc)  (24 + (IS_IN_ICACHE(pc)?5:CALL_FUNC1_SIZE))
    1.64 +
    1.65 +
    1.66  /**
    1.67   * Exit the block to an absolute PC
    1.68   */
    1.69 @@ -136,7 +142,37 @@
    1.70  {
    1.71      load_imm32( R_ECX, pc );                            // 5
    1.72      store_spreg( R_ECX, REG_OFFSET(pc) );               // 3
    1.73 -    MOV_moff32_EAX( xlat_get_lut_entry(pc) ); // 5
    1.74 +    if( IS_IN_ICACHE(pc) ) {
    1.75 +	MOV_moff32_EAX( xlat_get_lut_entry(GET_ICACHE_PHYS(pc)) ); // 5
    1.76 +    } else if( sh4_x86.tlb_on ) {
    1.77 +	call_func1(xlat_get_code_by_vma,R_ECX);
    1.78 +    } else {
    1.79 +	call_func1(xlat_get_code,R_ECX);
    1.80 +    }
    1.81 +    AND_imm8s_r32( 0xFC, R_EAX ); // 3
    1.82 +    load_imm32( R_ECX, ((endpc - sh4_x86.block_start_pc)>>1)*sh4_cpu_period ); // 5
    1.83 +    ADD_r32_sh4r( R_ECX, REG_OFFSET(slice_cycle) );     // 6
    1.84 +    POP_r32(R_EBP);
    1.85 +    RET();
    1.86 +}
    1.87 +
    1.88 +#define EXIT_BLOCK_REL_SIZE(pc)  (27 + (IS_IN_ICACHE(pc)?5:CALL_FUNC1_SIZE))
    1.89 +
    1.90 +/**
    1.91 + * Exit the block to a relative PC
    1.92 + */
    1.93 +void exit_block_rel( sh4addr_t pc, sh4addr_t endpc )
    1.94 +{
    1.95 +    load_imm32( R_ECX, pc - sh4_x86.block_start_pc );   // 5
    1.96 +    ADD_sh4r_r32( R_PC, R_ECX );
    1.97 +    store_spreg( R_ECX, REG_OFFSET(pc) );               // 3
    1.98 +    if( IS_IN_ICACHE(pc) ) {
    1.99 +	MOV_moff32_EAX( xlat_get_lut_entry(GET_ICACHE_PHYS(pc)) ); // 5
   1.100 +    } else if( sh4_x86.tlb_on ) {
   1.101 +	call_func1(xlat_get_code_by_vma,R_ECX);
   1.102 +    } else {
   1.103 +	call_func1(xlat_get_code,R_ECX);
   1.104 +    }
   1.105      AND_imm8s_r32( 0xFC, R_EAX ); // 3
   1.106      load_imm32( R_ECX, ((endpc - sh4_x86.block_start_pc)>>1)*sh4_cpu_period ); // 5
   1.107      ADD_r32_sh4r( R_ECX, REG_OFFSET(slice_cycle) );     // 6
   1.108 @@ -150,48 +186,87 @@
   1.109  void sh4_translate_end_block( sh4addr_t pc ) {
   1.110      if( sh4_x86.branch_taken == FALSE ) {
   1.111  	// Didn't exit unconditionally already, so write the termination here
   1.112 -	exit_block( pc, pc );
   1.113 +	exit_block_rel( pc, pc );
   1.114      }
   1.115      if( sh4_x86.backpatch_posn != 0 ) {
   1.116 +	unsigned int i;
   1.117 +	// Raise exception
   1.118  	uint8_t *end_ptr = xlat_output;
   1.119 -	// Exception termination. Jump block for various exception codes:
   1.120 -	PUSH_imm32( EXC_DATA_ADDR_READ );
   1.121 -	JMP_rel8( 33, target1 );
   1.122 -	PUSH_imm32( EXC_DATA_ADDR_WRITE );
   1.123 -	JMP_rel8( 26, target2 );
   1.124 -	PUSH_imm32( EXC_ILLEGAL );
   1.125 -	JMP_rel8( 19, target3 );
   1.126 -	PUSH_imm32( EXC_SLOT_ILLEGAL ); 
   1.127 -	JMP_rel8( 12, target4 );
   1.128 -	PUSH_imm32( EXC_FPU_DISABLED ); 
   1.129 -	JMP_rel8( 5, target5 );
   1.130 -	PUSH_imm32( EXC_SLOT_FPU_DISABLED );
   1.131 -	// target
   1.132 -	JMP_TARGET(target1);
   1.133 -	JMP_TARGET(target2);
   1.134 -	JMP_TARGET(target3);
   1.135 -	JMP_TARGET(target4);
   1.136 -	JMP_TARGET(target5);
   1.137 -	// Raise exception
   1.138 -	load_spreg( R_ECX, REG_OFFSET(pc) );
   1.139 +	MOV_r32_r32( R_EDX, R_ECX );
   1.140  	ADD_r32_r32( R_EDX, R_ECX );
   1.141 -	ADD_r32_r32( R_EDX, R_ECX );
   1.142 -	store_spreg( R_ECX, REG_OFFSET(pc) );
   1.143 +	ADD_r32_sh4r( R_ECX, R_PC );
   1.144  	MOV_moff32_EAX( &sh4_cpu_period );
   1.145  	MUL_r32( R_EDX );
   1.146  	ADD_r32_sh4r( R_EAX, REG_OFFSET(slice_cycle) );
   1.147  
   1.148  	call_func0( sh4_raise_exception );
   1.149  	ADD_imm8s_r32( 4, R_ESP );
   1.150 -	load_spreg( R_EAX, REG_OFFSET(pc) );
   1.151 -	call_func1(xlat_get_code,R_EAX);
   1.152 +	load_spreg( R_EAX, R_PC );
   1.153 +	if( sh4_x86.tlb_on ) {
   1.154 +	    call_func1(xlat_get_code_by_vma,R_EAX);
   1.155 +	} else {
   1.156 +	    call_func1(xlat_get_code,R_EAX);
   1.157 +	}
   1.158  	POP_r32(R_EBP);
   1.159  	RET();
   1.160  
   1.161 -	sh4_x86_do_backpatch( end_ptr );
   1.162 +	// Exception already raised - just cleanup
   1.163 +	uint8_t *preexc_ptr = xlat_output;
   1.164 +	MOV_r32_r32( R_EDX, R_ECX );
   1.165 +	ADD_r32_r32( R_EDX, R_ECX );
   1.166 +	ADD_r32_sh4r( R_ECX, R_SPC );
   1.167 +	MOV_moff32_EAX( &sh4_cpu_period );
   1.168 +	MUL_r32( R_EDX );
   1.169 +	ADD_r32_sh4r( R_EAX, REG_OFFSET(slice_cycle) );
   1.170 +	load_spreg( R_EAX, R_PC );
   1.171 +	if( sh4_x86.tlb_on ) {
   1.172 +	    call_func1(xlat_get_code_by_vma,R_EAX);
   1.173 +	} else {
   1.174 +	    call_func1(xlat_get_code,R_EAX);
   1.175 +	}
   1.176 +	POP_r32(R_EBP);
   1.177 +	RET();
   1.178 +
   1.179 +	for( i=0; i< sh4_x86.backpatch_posn; i++ ) {
   1.180 +	    *sh4_x86.backpatch_list[i].fixup_addr =
   1.181 +		xlat_output - ((uint8_t *)sh4_x86.backpatch_list[i].fixup_addr) - 4;
   1.182 +	    if( sh4_x86.backpatch_list[i].exc_code == -1 ) {
   1.183 +		load_imm32( R_EDX, sh4_x86.backpatch_list[i].fixup_icount );
   1.184 +		int rel = preexc_ptr - xlat_output;
   1.185 +		JMP_rel(rel);
   1.186 +	    } else {
   1.187 +		PUSH_imm32( sh4_x86.backpatch_list[i].exc_code );
   1.188 +		load_imm32( R_EDX, sh4_x86.backpatch_list[i].fixup_icount );
   1.189 +		int rel = end_ptr - xlat_output;
   1.190 +		JMP_rel(rel);
   1.191 +	    }
   1.192 +	}
   1.193      }
   1.194  }
   1.195  
   1.196 +void *xlat_get_native_pc()
   1.197 +{
   1.198 +    void *result = NULL;
   1.199 +    asm(
   1.200 +	"mov %%ebp, %%eax\n\t"
   1.201 +	"mov $0x8, %%ecx\n\t"
   1.202 +	"mov %1, %%edx\n"
   1.203 +"frame_loop: test %%eax, %%eax\n\t"
   1.204 +	"je frame_not_found\n\t"
   1.205 +	"cmp (%%eax), %%edx\n\t"
   1.206 +	"je frame_found\n\t"
   1.207 +	"sub $0x1, %%ecx\n\t"
   1.208 +	"je frame_not_found\n\t"
   1.209 +	"movl (%%eax), %%eax\n\t"
   1.210 +	"jmp frame_loop\n"
   1.211 +"frame_found: movl 0x4(%%eax), %0\n"
   1.212 +"frame_not_found:"
   1.213 +	: "=r" (result)
   1.214 +	: "r" (&sh4r)
   1.215 +	: "eax", "ecx", "edx" );
   1.216 +    return result;
   1.217 +}
   1.218 +
   1.219  #endif
   1.220  
   1.221