1.1 --- a/src/sh4/sh4x86.in	Wed Dec 14 21:51:55 2011 +1000
     1.2 +++ b/src/sh4/sh4x86.in	Wed Dec 14 23:23:32 2011 +1000
     1.3 @@ -304,18 +304,20 @@
     1.4  #define JNA_label(label) JCC_cc_rel8(X86_COND_NA,-1); MARK_JMP8(label)
     1.5  #define JNE_label(label) JCC_cc_rel8(X86_COND_NE,-1); MARK_JMP8(label)
     1.6  #define JNO_label(label) JCC_cc_rel8(X86_COND_NO,-1); MARK_JMP8(label)
     1.7 +#define JP_label(label)  JCC_cc_rel8(X86_COND_P,-1); MARK_JMP8(label)
     1.8  #define JS_label(label)  JCC_cc_rel8(X86_COND_S,-1); MARK_JMP8(label)
     1.9  #define JMP_label(label) JMP_rel8(-1); MARK_JMP8(label)
    1.10  #define JNE_exc(exc)     JCC_cc_rel32(X86_COND_NE,0); sh4_x86_add_backpatch(xlat_output, pc, exc)
    1.11  
    1.12 +#define LOAD_t() if( sh4_x86.tstate == TSTATE_NONE ) { \
    1.13 +	CMPL_imms_rbpdisp( 1, R_T ); sh4_x86.tstate = TSTATE_E; }     
    1.14 +
    1.15  /** Branch if T is set (either in the current cflags, or in sh4r.t) */
    1.16 -#define JT_label(label) if( sh4_x86.tstate == TSTATE_NONE ) { \
    1.17 -	CMPL_imms_rbpdisp( 1, R_T ); sh4_x86.tstate = TSTATE_E; } \
    1.18 +#define JT_label(label) LOAD_t() \
    1.19      JCC_cc_rel8(sh4_x86.tstate,-1); MARK_JMP8(label)
    1.20  
    1.21  /** Branch if T is clear (either in the current cflags or in sh4r.t) */
    1.22 -#define JF_label(label) if( sh4_x86.tstate == TSTATE_NONE ) { \
    1.23 -	CMPL_imms_rbpdisp( 1, R_T ); sh4_x86.tstate = TSTATE_E; } \
    1.24 +#define JF_label(label) LOAD_t() \
    1.25      JCC_cc_rel8(sh4_x86.tstate^1, -1); MARK_JMP8(label)
    1.26  
    1.27  
    1.28 @@ -1957,10 +1959,7 @@
    1.29  	    sh4_x86.branch_taken = TRUE;
    1.30  	    return 2;
    1.31  	} else {
    1.32 -	    if( sh4_x86.tstate == TSTATE_NONE ) {
    1.33 -		CMPL_imms_rbpdisp( 1, R_T );
    1.34 -		sh4_x86.tstate = TSTATE_E;
    1.35 -	    }
    1.36 +	    LOAD_t();
    1.37  	    sh4vma_t target = disp + pc + 4;
    1.38  	    JCC_cc_rel32(sh4_x86.tstate,0);
    1.39  	    uint32_t *patch = ((uint32_t *)xlat_output)-1;
    1.40 @@ -2095,10 +2094,7 @@
    1.41  	    sh4_x86.branch_taken = TRUE;
    1.42  	    return 2;
    1.43  	} else {
    1.44 -	    if( sh4_x86.tstate == TSTATE_NONE ) {
    1.45 -		CMPL_imms_rbpdisp( 1, R_T );
    1.46 -		sh4_x86.tstate = TSTATE_E;
    1.47 -	    }
    1.48 +		LOAD_t();
    1.49  	    JCC_cc_rel32(sh4_x86.tstate^1,0);
    1.50  	    uint32_t *patch = ((uint32_t *)xlat_output)-1;
    1.51  
    1.52 @@ -2431,14 +2427,15 @@
    1.53      } else {
    1.54          push_fr( FRm );
    1.55      }
    1.56 +    MOVP_immptr_rptr( &min_int, REG_ECX );
    1.57 +    FILD_r32disp( REG_ECX, 0 );
    1.58 +    FCOMIP_st(1);              
    1.59 +    JAE_label( sat );     
    1.60 +    JP_label( sat2 );       
    1.61      MOVP_immptr_rptr( &max_int, REG_ECX );
    1.62      FILD_r32disp( REG_ECX, 0 );
    1.63      FCOMIP_st(1);
    1.64 -    JNA_label( sat );
    1.65 -    MOVP_immptr_rptr( &min_int, REG_ECX );
    1.66 -    FILD_r32disp( REG_ECX, 0 );
    1.67 -    FCOMIP_st(1);              
    1.68 -    JAE_label( sat2 );            
    1.69 +    JNA_label( sat3 );
    1.70      MOVP_immptr_rptr( &save_fcw, REG_EAX );
    1.71      FNSTCW_r32disp( REG_EAX, 0 );
    1.72      MOVP_immptr_rptr( &trunc_fcw, REG_EDX );
    1.73 @@ -2449,6 +2446,7 @@
    1.74  
    1.75      JMP_TARGET(sat);
    1.76      JMP_TARGET(sat2);
    1.77 +    JMP_TARGET(sat3);
    1.78      MOVL_r32disp_r32( REG_ECX, 0, REG_ECX ); // 2
    1.79      MOVL_r32_rbpdisp( REG_ECX, R_FPUL );
    1.80      FPOP_st();
    1.81 @@ -2626,10 +2624,14 @@
    1.82          push_fr(FRm);
    1.83          push_fr(FRn);
    1.84      }
    1.85 +    XORL_r32_r32(REG_EAX, REG_EAX);
    1.86 +    XORL_r32_r32(REG_EDX, REG_EDX);
    1.87      FCOMIP_st(1);
    1.88 -    SETE_t();
    1.89 +    SETCCB_cc_r8(X86_COND_NP, REG_DL);
    1.90 +    CMOVCCL_cc_r32_r32(X86_COND_E, REG_EDX, REG_EAX);
    1.91 +    MOVL_r32_rbpdisp(REG_EAX, R_T);
    1.92      FPOP_st();
    1.93 -    sh4_x86.tstate = TSTATE_E;
    1.94 +    sh4_x86.tstate = TSTATE_NONE;
    1.95  :}
    1.96  FCMP/GT FRm, FRn {:  
    1.97      COUNT_INST(I_FCMPGT);