gte: parametrize remaining ops

at least the ones from C source, it looks like there could be more.

4 files changed, 253 insertions, 44 deletions

diff --git a/libpcsxcore/gte.c b/libpcsxcore/gte.c
index 38e6308..dc56b7f 100644
--- a/libpcsxcore/gte.c
+++ b/libpcsxcore/gte.c
@@ -26,6 +26,11 @@
 #include "gte.h"
 #include "psxmem.h"
 
+typedef struct psxCP2Regs {
+	psxCP2Data CP2D; 	/* Cop2 data registers */
+	psxCP2Ctrl CP2C; 	/* Cop2 control registers */
+} psxCP2Regs;
+
 #define VX(n) (n < 3 ? regs->CP2D.p[n << 1].sw.l : regs->CP2D.p[9].sw.l)
 #define VY(n) (n < 3 ? regs->CP2D.p[n << 1].sw.h : regs->CP2D.p[10].sw.l)
 #define VZ(n) (n < 3 ? regs->CP2D.p[(n << 1) + 1].sw.l : regs->CP2D.p[11].sw.l)
@@ -956,3 +961,131 @@ void gteCDP(psxCP2Regs *regs) {
 	gteG2 = limC2(gteMAC2 >> 4);
 	gteB2 = limC3(gteMAC3 >> 4);
 }
+
+/* decomposed/parametrized versions for the recompiler */
+
+#ifndef FLAGLESS
+
+void gteSQR_part_noshift(psxCP2Regs *regs) {
+	gteFLAG = 0;
+
+	gteMAC1 = gteIR1 * gteIR1;
+	gteMAC2 = gteIR2 * gteIR2;
+	gteMAC3 = gteIR3 * gteIR3;
+}
+
+void gteSQR_part_shift(psxCP2Regs *regs) {
+	gteFLAG = 0;
+
+	gteMAC1 = (gteIR1 * gteIR1) >> 12;
+	gteMAC2 = (gteIR2 * gteIR2) >> 12;
+	gteMAC3 = (gteIR3 * gteIR3) >> 12;
+}
+
+void gteOP_part_noshift(psxCP2Regs *regs) {
+	gteFLAG = 0;
+
+	gteMAC1 = (gteR22 * gteIR3) - (gteR33 * gteIR2);
+	gteMAC2 = (gteR33 * gteIR1) - (gteR11 * gteIR3);
+	gteMAC3 = (gteR11 * gteIR2) - (gteR22 * gteIR1);
+}
+
+void gteOP_part_shift(psxCP2Regs *regs) {
+	gteFLAG = 0;
+
+	gteMAC1 = ((gteR22 * gteIR3) - (gteR33 * gteIR2)) >> 12;
+	gteMAC2 = ((gteR33 * gteIR1) - (gteR11 * gteIR3)) >> 12;
+	gteMAC3 = ((gteR11 * gteIR2) - (gteR22 * gteIR1)) >> 12;
+}
+
+void gteDCPL_part(psxCP2Regs *regs) {
+	s32 RIR1 = ((s32)gteR * gteIR1) >> 8;
+	s32 GIR2 = ((s32)gteG * gteIR2) >> 8;
+	s32 BIR3 = ((s32)gteB * gteIR3) >> 8;
+
+	gteFLAG = 0;
+
+	gteMAC1 = RIR1 + ((gteIR0 * limB1(A1U((s64)gteRFC - RIR1), 0)) >> 12);
+	gteMAC2 = GIR2 + ((gteIR0 * limB1(A2U((s64)gteGFC - GIR2), 0)) >> 12);
+	gteMAC3 = BIR3 + ((gteIR0 * limB1(A3U((s64)gteBFC - BIR3), 0)) >> 12);
+}
+
+void gteGPF_part_noshift(psxCP2Regs *regs) {
+	gteFLAG = 0;
+
+	gteMAC1 = gteIR0 * gteIR1;
+	gteMAC2 = gteIR0 * gteIR2;
+	gteMAC3 = gteIR0 * gteIR3;
+}
+
+void gteGPF_part_shift(psxCP2Regs *regs) {
+	gteFLAG = 0;
+
+	gteMAC1 = (gteIR0 * gteIR1) >> 12;
+	gteMAC2 = (gteIR0 * gteIR2) >> 12;
+	gteMAC3 = (gteIR0 * gteIR3) >> 12;
+}
+
+#endif // !FLAGLESS
+
+void gteGPL_part_noshift(psxCP2Regs *regs) {
+	gteFLAG = 0;
+
+	gteMAC1 = A1((s64)gteMAC1 + (gteIR0 * gteIR1));
+	gteMAC2 = A2((s64)gteMAC2 + (gteIR0 * gteIR2));
+	gteMAC3 = A3((s64)gteMAC3 + (gteIR0 * gteIR3));
+}
+
+void gteGPL_part_shift(psxCP2Regs *regs) {
+	gteFLAG = 0;
+
+	gteMAC1 = A1((s64)gteMAC1 + ((gteIR0 * gteIR1) >> 12));
+	gteMAC2 = A2((s64)gteMAC2 + ((gteIR0 * gteIR2) >> 12));
+	gteMAC3 = A3((s64)gteMAC3 + ((gteIR0 * gteIR3) >> 12));
+}
+
+void gteDPCS_part_noshift(psxCP2Regs *regs) {
+	int shift = 0;
+
+	gteFLAG = 0;
+
+	gteMAC1 = ((gteR << 16) + (gteIR0 * limB1(A1U((s64)gteRFC - (gteR << 4)) << (12 - shift), 0))) >> 12;
+	gteMAC2 = ((gteG << 16) + (gteIR0 * limB2(A2U((s64)gteGFC - (gteG << 4)) << (12 - shift), 0))) >> 12;
+	gteMAC3 = ((gteB << 16) + (gteIR0 * limB3(A3U((s64)gteBFC - (gteB << 4)) << (12 - shift), 0))) >> 12;
+}
+
+void gteDPCS_part_shift(psxCP2Regs *regs) {
+	int shift = 12;
+
+	gteFLAG = 0;
+
+	gteMAC1 = ((gteR << 16) + (gteIR0 * limB1(A1U((s64)gteRFC - (gteR << 4)) << (12 - shift), 0))) >> 12;
+	gteMAC2 = ((gteG << 16) + (gteIR0 * limB2(A2U((s64)gteGFC - (gteG << 4)) << (12 - shift), 0))) >> 12;
+	gteMAC3 = ((gteB << 16) + (gteIR0 * limB3(A3U((s64)gteBFC - (gteB << 4)) << (12 - shift), 0))) >> 12;
+}
+
+void gteINTPL_part_noshift(psxCP2Regs *regs) {
+	gteFLAG = 0;
+
+	gteMAC1 = ((gteIR1 << 12) + (gteIR0 * limB1(A1U((s64)gteRFC - gteIR1), 0)));
+	gteMAC2 = ((gteIR2 << 12) + (gteIR0 * limB2(A2U((s64)gteGFC - gteIR2), 0)));
+	gteMAC3 = ((gteIR3 << 12) + (gteIR0 * limB3(A3U((s64)gteBFC - gteIR3), 0)));
+}
+
+void gteINTPL_part_shift(psxCP2Regs *regs) {
+	gteFLAG = 0;
+
+	gteMAC1 = ((gteIR1 << 12) + (gteIR0 * limB1(A1U((s64)gteRFC - gteIR1), 0))) >> 12;
+	gteMAC2 = ((gteIR2 << 12) + (gteIR0 * limB2(A2U((s64)gteGFC - gteIR2), 0))) >> 12;
+	gteMAC3 = ((gteIR3 << 12) + (gteIR0 * limB3(A3U((s64)gteBFC - gteIR3), 0))) >> 12;
+}
+
+void gteMACtoRGB(psxCP2Regs *regs) {
+	gteRGB0 = gteRGB1;
+	gteRGB1 = gteRGB2;
+	gteCODE2 = gteCODE;
+	gteR2 = limC1(gteMAC1 >> 4);
+	gteG2 = limC2(gteMAC2 >> 4);
+	gteB2 = limC3(gteMAC3 >> 4);
+}
+
diff --git a/libpcsxcore/gte.h b/libpcsxcore/gte.h
index 96f8542..7646226 100644
--- a/libpcsxcore/gte.h
+++ b/libpcsxcore/gte.h
@@ -44,6 +44,15 @@
 #define gteGPL gteGPL_nf
 #define gteNCCT gteNCCT_nf
 
+#define gteGPL_part_noshift gteGPL_part_noshift_nf
+#define gteGPL_part_shift gteGPL_part_shift_nf
+#define gteDPCS_part_noshift gteDPCS_part_noshift_nf
+#define gteDPCS_part_shift gteDPCS_part_shift_nf
+#define gteINTPL_part_noshift gteINTPL_part_noshift_nf
+#define gteINTPL_part_shift gteINTPL_part_shift_nf
+#define gteMACtoRGB gteMACtoRGB_nf
+
+#undef __GTE_H__
 #endif
 
 #ifndef __GTE_H__
@@ -56,10 +65,7 @@ extern "C" {
 #include "psxcommon.h"
 #include "r3000a.h"
 
-typedef struct {
-	psxCP2Data CP2D; 	/* Cop2 data registers */
-	psxCP2Ctrl CP2C; 	/* Cop2 control registers */
-} psxCP2Regs;
+struct psxCP2Regs;
 
 void gteMFC2();
 void gteCFC2();
@@ -68,28 +74,44 @@ void gteCTC2();
 void gteLWC2();
 void gteSWC2();
 
-void gteRTPS(psxCP2Regs *regs);
-void gteOP(psxCP2Regs *regs);
-void gteNCLIP(psxCP2Regs *regs);
-void gteDPCS(psxCP2Regs *regs);
-void gteINTPL(psxCP2Regs *regs);
-void gteMVMVA(psxCP2Regs *regs);
-void gteNCDS(psxCP2Regs *regs);
-void gteNCDT(psxCP2Regs *regs);
-void gteCDP(psxCP2Regs *regs);
-void gteNCCS(psxCP2Regs *regs);
-void gteCC(psxCP2Regs *regs);
-void gteNCS(psxCP2Regs *regs);
-void gteNCT(psxCP2Regs *regs);
-void gteSQR(psxCP2Regs *regs);
-void gteDCPL(psxCP2Regs *regs);
-void gteDPCT(psxCP2Regs *regs);
-void gteAVSZ3(psxCP2Regs *regs);
-void gteAVSZ4(psxCP2Regs *regs);
-void gteRTPT(psxCP2Regs *regs);
-void gteGPF(psxCP2Regs *regs);
-void gteGPL(psxCP2Regs *regs);
-void gteNCCT(psxCP2Regs *regs);
+void gteRTPS(struct psxCP2Regs *regs);
+void gteOP(struct psxCP2Regs *regs);
+void gteNCLIP(struct psxCP2Regs *regs);
+void gteDPCS(struct psxCP2Regs *regs);
+void gteINTPL(struct psxCP2Regs *regs);
+void gteMVMVA(struct psxCP2Regs *regs);
+void gteNCDS(struct psxCP2Regs *regs);
+void gteNCDT(struct psxCP2Regs *regs);
+void gteCDP(struct psxCP2Regs *regs);
+void gteNCCS(struct psxCP2Regs *regs);
+void gteCC(struct psxCP2Regs *regs);
+void gteNCS(struct psxCP2Regs *regs);
+void gteNCT(struct psxCP2Regs *regs);
+void gteSQR(struct psxCP2Regs *regs);
+void gteDCPL(struct psxCP2Regs *regs);
+void gteDPCT(struct psxCP2Regs *regs);
+void gteAVSZ3(struct psxCP2Regs *regs);
+void gteAVSZ4(struct psxCP2Regs *regs);
+void gteRTPT(struct psxCP2Regs *regs);
+void gteGPF(struct psxCP2Regs *regs);
+void gteGPL(struct psxCP2Regs *regs);
+void gteNCCT(struct psxCP2Regs *regs);
+
+void gteSQR_part_noshift(struct psxCP2Regs *regs);
+void gteSQR_part_shift(struct psxCP2Regs *regs);
+void gteOP_part_noshift(struct psxCP2Regs *regs);
+void gteOP_part_shift(struct psxCP2Regs *regs);
+void gteDCPL_part(struct psxCP2Regs *regs);
+void gteGPF_part_noshift(struct psxCP2Regs *regs);
+void gteGPF_part_shift(struct psxCP2Regs *regs);
+
+void gteGPL_part_noshift(struct psxCP2Regs *regs);
+void gteGPL_part_shift(struct psxCP2Regs *regs);
+void gteDPCS_part_noshift(struct psxCP2Regs *regs);
+void gteDPCS_part_shift(struct psxCP2Regs *regs);
+void gteINTPL_part_noshift(struct psxCP2Regs *regs);
+void gteINTPL_part_shift(struct psxCP2Regs *regs);
+void gteMACtoRGB(struct psxCP2Regs *regs);
 
 #ifdef __cplusplus
 }
diff --git a/libpcsxcore/new_dynarec/assem_arm.c b/libpcsxcore/new_dynarec/assem_arm.c
index 0c16fc2..6b663e5 100644
--- a/libpcsxcore/new_dynarec/assem_arm.c
+++ b/libpcsxcore/new_dynarec/assem_arm.c
@@ -20,6 +20,10 @@
  * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */
 
 #ifdef PCSX
+#include "../gte.h"
+#define FLAGLESS
+#include "../gte.h"
+#undef FLAGLESS
 #include "../gte_arm.h"
 #include "../gte_neon.h"
 #include "pcnt.h"
@@ -4456,15 +4460,34 @@ static void c2op_epilogue(u_int op,u_int reglist)
   restore_regs_all(reglist);
 }
 
+static void c2op_call_MACtoIR(int lm,int need_flags)
+{
+  if(need_flags)
+    emit_call((int)(lm?gteMACtoIR_lm1:gteMACtoIR_lm0));
+  else
+    emit_call((int)(lm?gteMACtoIR_lm1_nf:gteMACtoIR_lm0_nf));
+}
+
+static void c2op_call_rgb_func(void *func,int lm,int need_ir,int need_flags)
+{
+  emit_call((int)func);
+  // func is C code and trashes r0
+  emit_addimm(FP,(int)&psxRegs.CP2D.r[0]-(int)&dynarec_local,0);
+  if(need_flags||need_ir)
+    c2op_call_MACtoIR(lm,need_flags);
+  emit_call((int)(need_flags?gteMACtoRGB:gteMACtoRGB_nf));
+}
+
 static void c2op_assemble(int i,struct regstat *i_regs)
 {
   signed char temp=get_reg(i_regs->regmap,-1);
   u_int c2op=source[i]&0x3f;
-  u_int hr,reglist=0;
+  u_int hr,reglist_full=0,reglist;
   int need_flags,need_ir;
   for(hr=0;hr<HOST_REGS;hr++) {
-    if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
+    if(i_regs->regmap[hr]>=0) reglist_full|=1<<hr;
   }
+  reglist=reglist_full&0x100f;
 
   if (gte_handlers[c2op]!=NULL) {
     need_flags=!(gte_unneeded[i+1]>>63); // +1 because of how liveness detection works
@@ -4475,14 +4498,14 @@ static void c2op_assemble(int i,struct regstat *i_regs)
     // let's take more risk here
     need_flags=need_flags&&gte_reads_flags;
 #endif
+    int shift = (source[i] >> 19) & 1;
+    int lm = (source[i] >> 10) & 1;
     switch(c2op) {
       case GTE_MVMVA: {
-        int shift = (source[i] >> 19) & 1;
         int v  = (source[i] >> 15) & 3;
         int cv = (source[i] >> 13) & 3;
         int mx = (source[i] >> 17) & 3;
-        int lm = (source[i] >> 10) & 1;
-        reglist&=0x10ff; // +{r4-r7}
+        reglist=reglist_full&0x10ff; // +{r4-r7}
         c2op_prologue(c2op,reglist);
         /* r4,r5 = VXYZ(v) packed; r6 = &MX11(mx); r7 = &CV1(cv) */
         if(v<3)
@@ -4515,21 +4538,52 @@ static void c2op_assemble(int i,struct regstat *i_regs)
           emit_movimm(shift,1);
           emit_call((int)(need_flags?gteMVMVA_part_arm:gteMVMVA_part_nf_arm));
         }
-        if(need_flags||need_ir) {
-          if(need_flags)
-            emit_call((int)(lm?gteMACtoIR_lm1:gteMACtoIR_lm0));
-          else
-            emit_call((int)(lm?gteMACtoIR_lm1_nf:gteMACtoIR_lm0_nf)); // lm0 borked
-        }
+        if(need_flags||need_ir)
+          c2op_call_MACtoIR(lm,need_flags);
 #endif
         break;
       }
+      case GTE_OP:
+        c2op_prologue(c2op,reglist);
+        emit_call((int)(shift?gteOP_part_shift:gteOP_part_noshift));
+        if(need_flags||need_ir) {
+          emit_addimm(FP,(int)&psxRegs.CP2D.r[0]-(int)&dynarec_local,0);
+          c2op_call_MACtoIR(lm,need_flags);
+        }
+        break;
+      case GTE_DPCS:
+        c2op_prologue(c2op,reglist);
+        c2op_call_rgb_func(shift?gteDPCS_part_shift:gteDPCS_part_noshift,lm,need_ir,need_flags);
+        break;
+      case GTE_INTPL:
+        c2op_prologue(c2op,reglist);
+        c2op_call_rgb_func(shift?gteINTPL_part_shift:gteINTPL_part_noshift,lm,need_ir,need_flags);
+        break;
+      case GTE_SQR:
+        c2op_prologue(c2op,reglist);
+        emit_call((int)(shift?gteSQR_part_shift:gteSQR_part_noshift));
+        if(need_flags||need_ir) {
+          emit_addimm(FP,(int)&psxRegs.CP2D.r[0]-(int)&dynarec_local,0);
+          c2op_call_MACtoIR(lm,need_flags);
+        }
+        break;
+      case GTE_DCPL:
+        c2op_prologue(c2op,reglist);
+        c2op_call_rgb_func(gteDCPL_part,lm,need_ir,need_flags);
+        break;
+      case GTE_GPF:
+        c2op_prologue(c2op,reglist);
+        c2op_call_rgb_func(shift?gteGPF_part_shift:gteGPF_part_noshift,lm,need_ir,need_flags);
+        break;
+      case GTE_GPL:
+        c2op_prologue(c2op,reglist);
+        c2op_call_rgb_func(shift?gteGPL_part_shift:gteGPL_part_noshift,lm,need_ir,need_flags);
+        break;
 
       default:
-        reglist&=0x100f;
         c2op_prologue(c2op,reglist);
-        emit_movimm(source[i],1); // opcode
-        emit_writeword(1,(int)&psxRegs.code);
+        //emit_movimm(source[i],1); // opcode
+        //emit_writeword(1,(int)&psxRegs.code);
         emit_call((int)(need_flags?gte_handlers[c2op]:gte_handlers_nf[c2op]));
         break;
     }
diff --git a/libpcsxcore/psxinterpreter.c b/libpcsxcore/psxinterpreter.c
index f9833a4..66730d9 100644
--- a/libpcsxcore/psxinterpreter.c
+++ b/libpcsxcore/psxinterpreter.c
@@ -46,7 +46,7 @@ void (*psxBSC[64])();
 void (*psxSPC[64])();
 void (*psxREG[32])();
 void (*psxCP0[32])();
-void (*psxCP2[64])(psxCP2Regs *regs);
+void (*psxCP2[64])(struct psxCP2Regs *regs);
 void (*psxCP2BSC[32])();
 
 static void delayRead(int reg, u32 bpc) {
@@ -833,10 +833,10 @@ void psxCOP0() {
 }
 
 void psxCOP2() {
-	psxCP2[_Funct_]((psxCP2Regs *)&psxRegs.CP2D);
+	psxCP2[_Funct_]((struct psxCP2Regs *)&psxRegs.CP2D);
 }
 
-void psxBASIC(psxCP2Regs *regs) {
+void psxBASIC(struct psxCP2Regs *regs) {
 	psxCP2BSC[_Rs_]();
 }
 
@@ -882,7 +882,7 @@ void (*psxCP0[32])() = {
 	psxNULL, psxNULL, psxNULL, psxNULL, psxNULL, psxNULL, psxNULL, psxNULL
 };
 
-void (*psxCP2[64])(psxCP2Regs *regs) = {
+void (*psxCP2[64])(struct psxCP2Regs *regs) = {
 	psxBASIC, gteRTPS , psxNULL , psxNULL, psxNULL, psxNULL , gteNCLIP, psxNULL, // 00
 	psxNULL , psxNULL , psxNULL , psxNULL, gteOP  , psxNULL , psxNULL , psxNULL, // 08
 	gteDPCS , gteINTPL, gteMVMVA, gteNCDS, gteCDP , psxNULL , gteNCDT , psxNULL, // 10