1.1 --- a/src/sh4/sh4x86.in	Fri Sep 28 07:27:20 2007 +0000
     1.2 +++ b/src/sh4/sh4x86.in	Sat Sep 29 05:33:02 2007 +0000
     1.3 @@ -1,5 +1,5 @@
     1.4  /**
     1.5 - * $Id: sh4x86.in,v 1.16 2007-09-28 07:27:20 nkeynes Exp $
     1.6 + * $Id: sh4x86.in,v 1.17 2007-09-29 05:33:02 nkeynes Exp $
     1.7   * 
     1.8   * SH4 => x86 translation. This version does no real optimization, it just
     1.9   * outputs straight-line x86 code - it mainly exists to provide a baseline
    1.10 @@ -42,6 +42,7 @@
    1.11      gboolean in_delay_slot;
    1.12      gboolean priv_checked; /* true if we've already checked the cpu mode. */
    1.13      gboolean fpuen_checked; /* true if we've already checked fpu enabled. */
    1.14 +    gboolean branch_taken; /* true if we branched unconditionally */
    1.15      uint32_t block_start_pc;
    1.16  
    1.17      /* Allocated memory for the (block-wide) back-patch list */
    1.18 @@ -371,6 +372,7 @@
    1.19      sh4_x86.in_delay_slot = FALSE;
    1.20      sh4_x86.priv_checked = FALSE;
    1.21      sh4_x86.fpuen_checked = FALSE;
    1.22 +    sh4_x86.branch_taken = FALSE;
    1.23      sh4_x86.backpatch_posn = 0;
    1.24      sh4_x86.block_start_pc = pc;
    1.25  }
    1.26 @@ -410,6 +412,10 @@
    1.27   * Write the block trailer (exception handling block)
    1.28   */
    1.29  void sh4_translate_end_block( sh4addr_t pc ) {
    1.30 +    if( sh4_x86.branch_taken == FALSE ) {
    1.31 +	// Didn't exit unconditionally already, so write the termination here
    1.32 +	exit_block( pc, pc );
    1.33 +    }
    1.34      if( sh4_x86.backpatch_posn != 0 ) {
    1.35  	uint8_t *end_ptr = xlat_output;
    1.36  	// Exception termination. Jump block for various exception codes:
    1.37 @@ -1282,7 +1288,6 @@
    1.38  	JNE_rel8( 30, nottaken );
    1.39  	exit_block( disp + pc + 4, pc+2 );
    1.40  	JMP_TARGET(nottaken);
    1.41 -	exit_block( pc + 2, pc + 2 );
    1.42  	return 2;
    1.43      }
    1.44  :}
    1.45 @@ -1298,7 +1303,6 @@
    1.46  	// not taken
    1.47  	*patch = (xlat_output - ((uint8_t *)patch)) - 4;
    1.48  	sh4_x86_translate_instruction(pc+2);
    1.49 -	exit_block( pc + 4, pc+4 );
    1.50  	return 4;
    1.51      }
    1.52  :}
    1.53 @@ -1309,6 +1313,7 @@
    1.54  	sh4_x86.in_delay_slot = TRUE;
    1.55  	sh4_x86_translate_instruction( pc + 2 );
    1.56  	exit_block( disp + pc + 4, pc+4 );
    1.57 +	sh4_x86.branch_taken = TRUE;
    1.58  	return 4;
    1.59      }
    1.60  :}
    1.61 @@ -1322,6 +1327,7 @@
    1.62  	sh4_x86.in_delay_slot = TRUE;
    1.63  	sh4_x86_translate_instruction( pc + 2 );
    1.64  	exit_block_pcset(pc+2);
    1.65 +	sh4_x86.branch_taken = TRUE;
    1.66  	return 4;
    1.67      }
    1.68  :}
    1.69 @@ -1334,6 +1340,7 @@
    1.70  	sh4_x86.in_delay_slot = TRUE;
    1.71  	sh4_x86_translate_instruction( pc + 2 );
    1.72  	exit_block( disp + pc + 4, pc+4 );
    1.73 +	sh4_x86.branch_taken = TRUE;
    1.74  	return 4;
    1.75      }
    1.76  :}
    1.77 @@ -1348,6 +1355,7 @@
    1.78  	sh4_x86.in_delay_slot = TRUE;
    1.79  	sh4_x86_translate_instruction( pc + 2 );
    1.80  	exit_block_pcset(pc+2);
    1.81 +	sh4_x86.branch_taken = TRUE;
    1.82  	return 4;
    1.83      }
    1.84  :}
    1.85 @@ -1359,7 +1367,6 @@
    1.86  	JE_rel8( 30, nottaken );
    1.87  	exit_block( disp + pc + 4, pc+2 );
    1.88  	JMP_TARGET(nottaken);
    1.89 -	exit_block( pc + 2, pc+2 );
    1.90  	return 2;
    1.91      }
    1.92  :}
    1.93 @@ -1375,7 +1382,6 @@
    1.94  	// not taken
    1.95  	*patch = (xlat_output - ((uint8_t *)patch)) - 4;
    1.96  	sh4_x86_translate_instruction(pc+2);
    1.97 -	exit_block( pc + 4, pc+4 );
    1.98  	return 4;
    1.99      }
   1.100  :}
   1.101 @@ -1388,6 +1394,7 @@
   1.102  	sh4_x86.in_delay_slot = TRUE;
   1.103  	sh4_x86_translate_instruction(pc+2);
   1.104  	exit_block_pcset(pc+2);
   1.105 +	sh4_x86.branch_taken = TRUE;
   1.106  	return 4;
   1.107      }
   1.108  :}
   1.109 @@ -1402,6 +1409,7 @@
   1.110  	sh4_x86.in_delay_slot = TRUE;
   1.111  	sh4_x86_translate_instruction(pc+2);
   1.112  	exit_block_pcset(pc+2);
   1.113 +	sh4_x86.branch_taken = TRUE;
   1.114  	return 4;
   1.115      }
   1.116  :}
   1.117 @@ -1419,6 +1427,7 @@
   1.118  	sh4_x86.fpuen_checked = FALSE;
   1.119  	sh4_x86_translate_instruction(pc+2);
   1.120  	exit_block_pcset(pc+2);
   1.121 +	sh4_x86.branch_taken = TRUE;
   1.122  	return 4;
   1.123      }
   1.124  :}
   1.125 @@ -1431,6 +1440,7 @@
   1.126  	sh4_x86.in_delay_slot = TRUE;
   1.127  	sh4_x86_translate_instruction(pc+2);
   1.128  	exit_block_pcset(pc+2);
   1.129 +	sh4_x86.branch_taken = TRUE;
   1.130  	return 4;
   1.131      }
   1.132  :}
   1.133 @@ -1442,6 +1452,7 @@
   1.134  	call_func0( sh4_raise_trap );
   1.135  	ADD_imm8s_r32( 4, R_ESP );
   1.136  	exit_block_pcset(pc);
   1.137 +	sh4_x86.branch_taken = TRUE;
   1.138  	return 2;
   1.139      }
   1.140  :}
   1.141 @@ -2283,7 +2294,6 @@
   1.142      call_func0( sh4_sleep );
   1.143      sh4_x86.in_delay_slot = FALSE;
   1.144      INC_r32(R_ESI);
   1.145 -    exit_block(pc+2, pc+2);
   1.146      return 2;
   1.147  :}
   1.148  STC SR, Rn {: