OSDN Git Service

Handling new opcodes in PixelProgram and VertexProgram
[android-x86/external-swiftshader.git] / src / Shader / VertexProgram.cpp
1 // SwiftShader Software Renderer
2 //
3 // Copyright(c) 2005-2013 TransGaming Inc.
4 //
5 // All rights reserved. No part of this software may be copied, distributed, transmitted,
6 // transcribed, stored in a retrieval system, translated into any human or computer
7 // language by any means, or disclosed to third parties without the explicit written
8 // agreement of TransGaming Inc. Without such an agreement, no rights or licenses, express
9 // or implied, including but not limited to any patent rights, are granted to you.
10 //
11
12 #include "VertexProgram.hpp"
13
14 #include "Renderer.hpp"
15 #include "VertexShader.hpp"
16 #include "Vertex.hpp"
17 #include "Half.hpp"
18 #include "SamplerCore.hpp"
19 #include "Debug.hpp"
20
21 namespace sw
22 {
23         VertexProgram::VertexProgram(const VertexProcessor::State &state, const VertexShader *shader) : VertexRoutine(state, shader)
24         {
25                 ifDepth = 0;
26                 loopRepDepth = 0;
27                 breakDepth = 0;
28                 currentLabel = -1;
29                 whileTest = false;
30
31                 for(int i = 0; i < 2048; i++)
32                 {
33                         labelBlock[i] = 0;
34                 }
35         }
36
37         VertexProgram::~VertexProgram()
38         {
39                 for(int i = 0; i < VERTEX_TEXTURE_IMAGE_UNITS; i++)
40                 {
41                         delete sampler[i];
42                 }
43         }
44
45         void VertexProgram::pipeline(Registers &r)
46         {
47                 for(int i = 0; i < VERTEX_TEXTURE_IMAGE_UNITS; i++)
48                 {
49                         sampler[i] = new SamplerCore(r.constants, state.samplerState[i]);
50                 }
51
52                 if(!state.preTransformed)
53                 {
54                         program(r);
55                 }
56                 else
57                 {
58                         passThrough(r);
59                 }
60         }
61
62         void VertexProgram::program(Registers &r)
63         {
64         //      shader->print("VertexShader-%0.8X.txt", state.shaderID);
65
66                 unsigned short version = shader->getVersion();
67
68                 r.enableIndex = 0;
69                 r.stackIndex = 0;
70
71                 if(shader->containsLeaveInstruction())
72                 {
73                         r.enableLeave = Int4(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
74                 }
75
76                 // Create all call site return blocks up front
77                 for(size_t i = 0; i < shader->getLength(); i++)
78                 {
79                         const Shader::Instruction *instruction = shader->getInstruction(i);
80                         Shader::Opcode opcode = instruction->opcode;
81
82                         if(opcode == Shader::OPCODE_CALL || opcode == Shader::OPCODE_CALLNZ)
83                         {
84                                 const Dst &dst = instruction->dst;
85
86                                 ASSERT(callRetBlock[dst.label].size() == dst.callSite);
87                                 callRetBlock[dst.label].push_back(Nucleus::createBasicBlock());
88                         }
89                 }
90         
91                 for(size_t i = 0; i < shader->getLength(); i++)
92                 {
93                         const Shader::Instruction *instruction = shader->getInstruction(i);
94                         Shader::Opcode opcode = instruction->opcode;
95
96                         if(opcode == Shader::OPCODE_DCL || opcode == Shader::OPCODE_DEF || opcode == Shader::OPCODE_DEFI || opcode == Shader::OPCODE_DEFB)
97                         {
98                                 continue;
99                         }
100
101                         Dst dst = instruction->dst;
102                         Src src0 = instruction->src[0];
103                         Src src1 = instruction->src[1];
104                         Src src2 = instruction->src[2];
105
106                         bool predicate = instruction->predicate;
107                         Control control = instruction->control;
108                         bool integer = dst.type == Shader::PARAMETER_ADDR;
109                         bool pp = dst.partialPrecision;
110
111                         Vector4f d;
112                         Vector4f s0;
113                         Vector4f s1;
114                         Vector4f s2;
115
116                         if(src0.type != Shader::PARAMETER_VOID) s0 = fetchRegisterF(r, src0);
117                         if(src1.type != Shader::PARAMETER_VOID) s1 = fetchRegisterF(r, src1);
118                         if(src2.type != Shader::PARAMETER_VOID) s2 = fetchRegisterF(r, src2);
119
120                         switch(opcode)
121                         {
122                         case Shader::OPCODE_VS_1_0:                                                                             break;
123                         case Shader::OPCODE_VS_1_1:                                                                             break;
124                         case Shader::OPCODE_VS_2_0:                                                                             break;
125                         case Shader::OPCODE_VS_2_x:                                                                             break;
126                         case Shader::OPCODE_VS_2_sw:                                                                    break;
127                         case Shader::OPCODE_VS_3_0:                                                                             break;
128                         case Shader::OPCODE_VS_3_sw:                                                                    break;
129                         case Shader::OPCODE_DCL:                                                                                break;
130                         case Shader::OPCODE_DEF:                                                                                break;
131                         case Shader::OPCODE_DEFI:                                                                               break;
132                         case Shader::OPCODE_DEFB:                                                                               break;
133                         case Shader::OPCODE_NOP:                                                                                break;
134                         case Shader::OPCODE_ABS:                abs(d, s0);                                             break;
135                         case Shader::OPCODE_ADD:                add(d, s0, s1);                                 break;
136                         case Shader::OPCODE_IADD:       iadd(d, s0, s1);                break;
137                         case Shader::OPCODE_CRS:                crs(d, s0, s1);                                 break;
138                         case Shader::OPCODE_FORWARD1:   forward1(d, s0, s1, s2);                break;
139                         case Shader::OPCODE_FORWARD2:   forward2(d, s0, s1, s2);                break;
140                         case Shader::OPCODE_FORWARD3:   forward3(d, s0, s1, s2);                break;
141                         case Shader::OPCODE_FORWARD4:   forward4(d, s0, s1, s2);                break;
142                         case Shader::OPCODE_REFLECT1:   reflect1(d, s0, s1);                    break;
143                         case Shader::OPCODE_REFLECT2:   reflect2(d, s0, s1);                    break;
144                         case Shader::OPCODE_REFLECT3:   reflect3(d, s0, s1);                    break;
145                         case Shader::OPCODE_REFLECT4:   reflect4(d, s0, s1);                    break;
146                         case Shader::OPCODE_REFRACT1:   refract1(d, s0, s1, s2.x);              break;
147                         case Shader::OPCODE_REFRACT2:   refract2(d, s0, s1, s2.x);              break;
148                         case Shader::OPCODE_REFRACT3:   refract3(d, s0, s1, s2.x);              break;
149                         case Shader::OPCODE_REFRACT4:   refract4(d, s0, s1, s2.x);              break;
150                         case Shader::OPCODE_DP1:                dp1(d, s0, s1);                                 break;
151                         case Shader::OPCODE_DP2:                dp2(d, s0, s1);                                 break;
152                         case Shader::OPCODE_DP3:                dp3(d, s0, s1);                                 break;
153                         case Shader::OPCODE_DP4:                dp4(d, s0, s1);                                 break;
154                         case Shader::OPCODE_ATT:                att(d, s0, s1);                                 break;
155                         case Shader::OPCODE_EXP2X:              exp2x(d, s0, pp);                               break;
156                         case Shader::OPCODE_EXP2:               exp2(d, s0, pp);                                break;
157                         case Shader::OPCODE_EXPP:               expp(d, s0, version);                   break;
158                         case Shader::OPCODE_EXP:                exp(d, s0, pp);                                 break;
159                         case Shader::OPCODE_FRC:                frc(d, s0);                                             break;
160                         case Shader::OPCODE_TRUNC:      trunc(d, s0);                   break;
161                         case Shader::OPCODE_FLOOR:      floor(d, s0);                   break;
162                         case Shader::OPCODE_ROUND:      round(d, s0);                   break;
163                         case Shader::OPCODE_ROUNDEVEN:  roundEven(d, s0);               break;
164                         case Shader::OPCODE_CEIL:       ceil(d, s0);                    break;
165                         case Shader::OPCODE_LIT:                lit(d, s0);                                             break;
166                         case Shader::OPCODE_LOG2X:              log2x(d, s0, pp);                               break;
167                         case Shader::OPCODE_LOG2:               log2(d, s0, pp);                                break;
168                         case Shader::OPCODE_LOGP:               logp(d, s0, version);                   break;
169                         case Shader::OPCODE_LOG:                log(d, s0, pp);                                 break;
170                         case Shader::OPCODE_LRP:                lrp(d, s0, s1, s2);                             break;
171                         case Shader::OPCODE_STEP:               step(d, s0, s1);                                break;
172                         case Shader::OPCODE_SMOOTH:             smooth(d, s0, s1, s2);                  break;
173                         case Shader::OPCODE_FLOATBITSTOINT:
174                         case Shader::OPCODE_FLOATBITSTOUINT:
175                         case Shader::OPCODE_INTBITSTOFLOAT:
176                         case Shader::OPCODE_UINTBITSTOFLOAT: d = s0;                    break;
177                         case Shader::OPCODE_M3X2:               M3X2(r, d, s0, src1);                   break;
178                         case Shader::OPCODE_M3X3:               M3X3(r, d, s0, src1);                   break;
179                         case Shader::OPCODE_M3X4:               M3X4(r, d, s0, src1);                   break;
180                         case Shader::OPCODE_M4X3:               M4X3(r, d, s0, src1);                   break;
181                         case Shader::OPCODE_M4X4:               M4X4(r, d, s0, src1);                   break;
182                         case Shader::OPCODE_MAD:                mad(d, s0, s1, s2);                             break;
183                         case Shader::OPCODE_IMAD:               imad(d, s0, s1, s2);                    break;
184                         case Shader::OPCODE_MAX:                max(d, s0, s1);                                 break;
185                         case Shader::OPCODE_IMAX:       imax(d, s0, s1);                break;
186                         case Shader::OPCODE_UMAX:       umax(d, s0, s1);                break;
187                         case Shader::OPCODE_MIN:                min(d, s0, s1);                                 break;
188                         case Shader::OPCODE_IMIN:       imin(d, s0, s1);                break;
189                         case Shader::OPCODE_UMIN:       umin(d, s0, s1);                break;
190                         case Shader::OPCODE_MOV:                mov(d, s0, integer);                    break;
191                         case Shader::OPCODE_MOVA:               mov(d, s0);                                             break;
192                         case Shader::OPCODE_NEG:        neg(d, s0);                     break;
193                         case Shader::OPCODE_INEG:       ineg(d, s0);                    break;
194                         case Shader::OPCODE_F2B:                f2b(d, s0);                                             break;
195                         case Shader::OPCODE_B2F:                b2f(d, s0);                                             break;
196                         case Shader::OPCODE_F2I:        f2i(d, s0);                     break;
197                         case Shader::OPCODE_I2F:        i2f(d, s0);                     break;
198                         case Shader::OPCODE_F2U:        f2u(d, s0);                     break;
199                         case Shader::OPCODE_U2F:        u2f(d, s0);                     break;
200                         case Shader::OPCODE_I2B:        i2b(d, s0);                     break;
201                         case Shader::OPCODE_B2I:        b2i(d, s0);                     break;
202                         case Shader::OPCODE_U2B:        u2b(d, s0);                     break;
203                         case Shader::OPCODE_B2U:        b2u(d, s0);                     break;
204                         case Shader::OPCODE_MUL:                mul(d, s0, s1);                                 break;
205                         case Shader::OPCODE_IMUL:       imul(d, s0, s1);                break;
206                         case Shader::OPCODE_NRM2:               nrm2(d, s0, pp);                                break;
207                         case Shader::OPCODE_NRM3:               nrm3(d, s0, pp);                                break;
208                         case Shader::OPCODE_NRM4:               nrm4(d, s0, pp);                                break;
209                         case Shader::OPCODE_POWX:               powx(d, s0, s1, pp);                    break;
210                         case Shader::OPCODE_POW:                pow(d, s0, s1, pp);                             break;
211                         case Shader::OPCODE_RCPX:               rcpx(d, s0, pp);                                break;
212                         case Shader::OPCODE_DIV:                div(d, s0, s1);                                 break;
213                         case Shader::OPCODE_IDIV:       idiv(d, s0, s1);                break;
214                         case Shader::OPCODE_UDIV:       udiv(d, s0, s1);                break;
215                         case Shader::OPCODE_MOD:                mod(d, s0, s1);                                 break;
216                         case Shader::OPCODE_IMOD:       imod(d, s0, s1);                break;
217                         case Shader::OPCODE_UMOD:       umod(d, s0, s1);                break;
218                         case Shader::OPCODE_SHL:        shl(d, s0, s1);                 break;
219                         case Shader::OPCODE_ISHR:       ishr(d, s0, s1);                 break;
220                         case Shader::OPCODE_USHR:       ushr(d, s0, s1);                 break;
221                         case Shader::OPCODE_RSQX:               rsqx(d, s0, pp);                                break;
222                         case Shader::OPCODE_SQRT:               sqrt(d, s0, pp);                                break;
223                         case Shader::OPCODE_RSQ:                rsq(d, s0, pp);                                 break;
224                         case Shader::OPCODE_LEN2:               len2(d.x, s0, pp);                              break;
225                         case Shader::OPCODE_LEN3:               len3(d.x, s0, pp);                              break;
226                         case Shader::OPCODE_LEN4:               len4(d.x, s0, pp);                              break;
227                         case Shader::OPCODE_DIST1:              dist1(d.x, s0, s1, pp);                 break;
228                         case Shader::OPCODE_DIST2:              dist2(d.x, s0, s1, pp);                 break;
229                         case Shader::OPCODE_DIST3:              dist3(d.x, s0, s1, pp);                 break;
230                         case Shader::OPCODE_DIST4:              dist4(d.x, s0, s1, pp);                 break;
231                         case Shader::OPCODE_SGE:                step(d, s1, s0);                                break;
232                         case Shader::OPCODE_SGN:                sgn(d, s0);                                             break;
233                         case Shader::OPCODE_SINCOS:             sincos(d, s0, pp);                              break;
234                         case Shader::OPCODE_COS:                cos(d, s0, pp);                                 break;
235                         case Shader::OPCODE_SIN:                sin(d, s0, pp);                                 break;
236                         case Shader::OPCODE_TAN:                tan(d, s0);                                             break;
237                         case Shader::OPCODE_ACOS:               acos(d, s0);                                    break;
238                         case Shader::OPCODE_ASIN:               asin(d, s0);                                    break;
239                         case Shader::OPCODE_ATAN:               atan(d, s0);                                    break;
240                         case Shader::OPCODE_ATAN2:              atan2(d, s0, s1);                               break;
241                         case Shader::OPCODE_COSH:               cosh(d, s0, pp);                                break;
242                         case Shader::OPCODE_SINH:               sinh(d, s0, pp);                                break;
243                         case Shader::OPCODE_TANH:               tanh(d, s0, pp);                                break;
244                         case Shader::OPCODE_ACOSH:              acosh(d, s0, pp);                               break;
245                         case Shader::OPCODE_ASINH:              asinh(d, s0, pp);                               break;
246                         case Shader::OPCODE_ATANH:              atanh(d, s0, pp);                               break;
247                         case Shader::OPCODE_SLT:                slt(d, s0, s1);                                 break;
248                         case Shader::OPCODE_SUB:                sub(d, s0, s1);                                 break;
249                         case Shader::OPCODE_ISUB:       isub(d, s0, s1);                break;
250                         case Shader::OPCODE_BREAK:              BREAK(r);                                               break;
251                         case Shader::OPCODE_BREAKC:             BREAKC(r, s0, s1, control);             break;
252                         case Shader::OPCODE_BREAKP:             BREAKP(r, src0);                                break;
253                         case Shader::OPCODE_CONTINUE:   CONTINUE(r);                                    break;
254                         case Shader::OPCODE_TEST:               TEST();                                                 break;
255                         case Shader::OPCODE_CALL:               CALL(r, dst.label, dst.callSite);         break;
256                         case Shader::OPCODE_CALLNZ:             CALLNZ(r, dst.label, dst.callSite, src0); break;
257                         case Shader::OPCODE_ELSE:               ELSE(r);                                                break;
258                         case Shader::OPCODE_ENDIF:              ENDIF(r);                                               break;
259                         case Shader::OPCODE_ENDLOOP:    ENDLOOP(r);                                             break;
260                         case Shader::OPCODE_ENDREP:             ENDREP(r);                                              break;
261                         case Shader::OPCODE_ENDWHILE:   ENDWHILE(r);                                    break;
262                         case Shader::OPCODE_IF:                 IF(r, src0);                                    break;
263                         case Shader::OPCODE_IFC:                IFC(r, s0, s1, control);                break;
264                         case Shader::OPCODE_LABEL:              LABEL(dst.index);                               break;
265                         case Shader::OPCODE_LOOP:               LOOP(r, src1);                                  break;
266                         case Shader::OPCODE_REP:                REP(r, src0);                                   break;
267                         case Shader::OPCODE_WHILE:              WHILE(r, src0);                                 break;
268                         case Shader::OPCODE_RET:                RET(r);                                                 break;
269                         case Shader::OPCODE_LEAVE:              LEAVE(r);                                               break;
270                         case Shader::OPCODE_CMP:                cmp(d, s0, s1, control);                break;
271                         case Shader::OPCODE_ICMP:               icmp(d, s0, s1, control);               break;
272                         case Shader::OPCODE_UCMP:       ucmp(d, s0, s1, control);       break;
273                         case Shader::OPCODE_SELECT:             select(d, s0, s1, s2);                  break;
274                         case Shader::OPCODE_EXTRACT:    extract(d.x, s0, s1.x);                 break;
275                         case Shader::OPCODE_INSERT:             insert(d, s0, s1.x, s2.x);              break;
276                         case Shader::OPCODE_ALL:                all(d.x, s0);                                   break;
277                         case Shader::OPCODE_ANY:                any(d.x, s0);                                   break;
278                         case Shader::OPCODE_NOT:                not(d, s0);                                             break;
279                         case Shader::OPCODE_OR:         or(d, s0, s1);                  break;
280                         case Shader::OPCODE_XOR:        xor(d, s0, s1);                 break;
281                         case Shader::OPCODE_AND:        and(d, s0, s1);                 break;
282                         case Shader::OPCODE_EQ:         equal(d, s0, s1);               break;
283                         case Shader::OPCODE_NE:         notEqual(d, s0, s1);            break;
284                         case Shader::OPCODE_TEXLDL:             TEXLDL(r, d, s0, src1);                 break;
285                         case Shader::OPCODE_TEX:                TEX(r, d, s0, src1);                    break;
286                         case Shader::OPCODE_END:                                                                                break;
287                         default:
288                                 ASSERT(false);
289                         }
290
291                         if(dst.type != Shader::PARAMETER_VOID && dst.type != Shader::PARAMETER_LABEL && opcode != Shader::OPCODE_NOP)
292                         {
293                                 if(dst.integer)
294                                 {
295                                         switch(opcode)
296                                         {
297                                         case Shader::OPCODE_DIV:
298                                                 if(dst.x) d.x = Trunc(d.x);
299                                                 if(dst.y) d.y = Trunc(d.y);
300                                                 if(dst.z) d.z = Trunc(d.z);
301                                                 if(dst.w) d.w = Trunc(d.w);
302                                                 break;
303                                         default:
304                                                 break;   // No truncation to integer required when arguments are integer
305                                         }
306                                 }
307
308                                 if(dst.saturate)
309                                 {
310                                         if(dst.x) d.x = Max(d.x, Float4(0.0f));
311                                         if(dst.y) d.y = Max(d.y, Float4(0.0f));
312                                         if(dst.z) d.z = Max(d.z, Float4(0.0f));
313                                         if(dst.w) d.w = Max(d.w, Float4(0.0f));
314
315                                         if(dst.x) d.x = Min(d.x, Float4(1.0f));
316                                         if(dst.y) d.y = Min(d.y, Float4(1.0f));
317                                         if(dst.z) d.z = Min(d.z, Float4(1.0f));
318                                         if(dst.w) d.w = Min(d.w, Float4(1.0f));
319                                 }
320
321                                 if(instruction->isPredicated())
322                                 {
323                                         Vector4f pDst;   // FIXME: Rename
324
325                                         switch(dst.type)
326                                         {
327                                         case Shader::PARAMETER_VOID:                                                                                                                                            break;
328                                         case Shader::PARAMETER_TEMP:
329                                                 if(dst.rel.type == Shader::PARAMETER_VOID)
330                                                 {
331                                                         if(dst.x) pDst.x = r.r[dst.index].x;
332                                                         if(dst.y) pDst.y = r.r[dst.index].y;
333                                                         if(dst.z) pDst.z = r.r[dst.index].z;
334                                                         if(dst.w) pDst.w = r.r[dst.index].w;
335                                                 }
336                                                 else
337                                                 {
338                                                         Int a = relativeAddress(r, dst);
339
340                                                         if(dst.x) pDst.x = r.r[dst.index + a].x;
341                                                         if(dst.y) pDst.y = r.r[dst.index + a].y;
342                                                         if(dst.z) pDst.z = r.r[dst.index + a].z;
343                                                         if(dst.w) pDst.w = r.r[dst.index + a].w;
344                                                 }
345                                                 break;
346                                         case Shader::PARAMETER_ADDR:            pDst = r.a0;                                                                                                    break;
347                                         case Shader::PARAMETER_RASTOUT:
348                                                 switch(dst.index)
349                                                 {
350                                                 case 0:
351                                                         if(dst.x) pDst.x = r.o[Pos].x;
352                                                         if(dst.y) pDst.y = r.o[Pos].y;
353                                                         if(dst.z) pDst.z = r.o[Pos].z;
354                                                         if(dst.w) pDst.w = r.o[Pos].w;
355                                                         break;
356                                                 case 1:
357                                                         pDst.x = r.o[Fog].x;
358                                                         break;
359                                                 case 2:
360                                                         pDst.x = r.o[Pts].y;
361                                                         break;
362                                                 default:
363                                                         ASSERT(false);
364                                                 }
365                                                 break;
366                                         case Shader::PARAMETER_ATTROUT:
367                                                 if(dst.x) pDst.x = r.o[D0 + dst.index].x;
368                                                 if(dst.y) pDst.y = r.o[D0 + dst.index].y;
369                                                 if(dst.z) pDst.z = r.o[D0 + dst.index].z;
370                                                 if(dst.w) pDst.w = r.o[D0 + dst.index].w;
371                                                 break;
372                                         case Shader::PARAMETER_TEXCRDOUT:
373                                 //      case Shader::PARAMETER_OUTPUT:
374                                                 if(version < 0x0300)
375                                                 {
376                                                         if(dst.x) pDst.x = r.o[T0 + dst.index].x;
377                                                         if(dst.y) pDst.y = r.o[T0 + dst.index].y;
378                                                         if(dst.z) pDst.z = r.o[T0 + dst.index].z;
379                                                         if(dst.w) pDst.w = r.o[T0 + dst.index].w;
380                                                 }
381                                                 else
382                                                 {
383                                                         if(dst.rel.type == Shader::PARAMETER_VOID)   // Not relative
384                                                         {
385                                                                 if(dst.x) pDst.x = r.o[dst.index].x;
386                                                                 if(dst.y) pDst.y = r.o[dst.index].y;
387                                                                 if(dst.z) pDst.z = r.o[dst.index].z;
388                                                                 if(dst.w) pDst.w = r.o[dst.index].w;
389                                                         }
390                                                         else if(dst.rel.type == Shader::PARAMETER_LOOP)
391                                                         {
392                                                                 Int aL = r.aL[r.loopDepth];
393
394                                                                 if(dst.x) pDst.x = r.o[dst.index + aL].x;
395                                                                 if(dst.y) pDst.y = r.o[dst.index + aL].y;
396                                                                 if(dst.z) pDst.z = r.o[dst.index + aL].z;
397                                                                 if(dst.w) pDst.w = r.o[dst.index + aL].w;
398                                                         }
399                                                         else
400                                                         {
401                                                                 Int a = relativeAddress(r, dst);
402
403                                                                 if(dst.x) pDst.x = r.o[dst.index + a].x;
404                                                                 if(dst.y) pDst.y = r.o[dst.index + a].y;
405                                                                 if(dst.z) pDst.z = r.o[dst.index + a].z;
406                                                                 if(dst.w) pDst.w = r.o[dst.index + a].w;
407                                                         }
408                                                 }
409                                                 break;
410                                         case Shader::PARAMETER_LABEL:                                                                                                                                           break;
411                                         case Shader::PARAMETER_PREDICATE:       pDst = r.p0;                                                                                                    break;
412                                         case Shader::PARAMETER_INPUT:                                                                                                                                           break;
413                                         default:
414                                                 ASSERT(false);
415                                         }
416
417                                         Int4 enable = enableMask(r, instruction);
418
419                                         Int4 xEnable = enable;
420                                         Int4 yEnable = enable;
421                                         Int4 zEnable = enable;
422                                         Int4 wEnable = enable;
423
424                                         if(predicate)
425                                         {
426                                                 unsigned char pSwizzle = instruction->predicateSwizzle;
427
428                                                 Float4 xPredicate = r.p0[(pSwizzle >> 0) & 0x03];
429                                                 Float4 yPredicate = r.p0[(pSwizzle >> 2) & 0x03];
430                                                 Float4 zPredicate = r.p0[(pSwizzle >> 4) & 0x03];
431                                                 Float4 wPredicate = r.p0[(pSwizzle >> 6) & 0x03];
432
433                                                 if(!instruction->predicateNot)
434                                                 {
435                                                         if(dst.x) xEnable = xEnable & As<Int4>(xPredicate);
436                                                         if(dst.y) yEnable = yEnable & As<Int4>(yPredicate);
437                                                         if(dst.z) zEnable = zEnable & As<Int4>(zPredicate);
438                                                         if(dst.w) wEnable = wEnable & As<Int4>(wPredicate);
439                                                 }
440                                                 else
441                                                 {
442                                                         if(dst.x) xEnable = xEnable & ~As<Int4>(xPredicate);
443                                                         if(dst.y) yEnable = yEnable & ~As<Int4>(yPredicate);
444                                                         if(dst.z) zEnable = zEnable & ~As<Int4>(zPredicate);
445                                                         if(dst.w) wEnable = wEnable & ~As<Int4>(wPredicate);
446                                                 }
447                                         }
448
449                                         if(dst.x) d.x = As<Float4>(As<Int4>(d.x) & xEnable);
450                                         if(dst.y) d.y = As<Float4>(As<Int4>(d.y) & yEnable);
451                                         if(dst.z) d.z = As<Float4>(As<Int4>(d.z) & zEnable);
452                                         if(dst.w) d.w = As<Float4>(As<Int4>(d.w) & wEnable);
453
454                                         if(dst.x) d.x = As<Float4>(As<Int4>(d.x) | (As<Int4>(pDst.x) & ~xEnable));
455                                         if(dst.y) d.y = As<Float4>(As<Int4>(d.y) | (As<Int4>(pDst.y) & ~yEnable));
456                                         if(dst.z) d.z = As<Float4>(As<Int4>(d.z) | (As<Int4>(pDst.z) & ~zEnable));
457                                         if(dst.w) d.w = As<Float4>(As<Int4>(d.w) | (As<Int4>(pDst.w) & ~wEnable));
458                                 }
459
460                                 switch(dst.type)
461                                 {
462                                 case Shader::PARAMETER_VOID:
463                                         break;
464                                 case Shader::PARAMETER_TEMP:
465                                         if(dst.rel.type == Shader::PARAMETER_VOID)
466                                         {
467                                                 if(dst.x) r.r[dst.index].x = d.x;
468                                                 if(dst.y) r.r[dst.index].y = d.y;
469                                                 if(dst.z) r.r[dst.index].z = d.z;
470                                                 if(dst.w) r.r[dst.index].w = d.w;
471                                         }
472                                         else
473                                         {
474                                                 Int a = relativeAddress(r, dst);
475
476                                                 if(dst.x) r.r[dst.index + a].x = d.x;
477                                                 if(dst.y) r.r[dst.index + a].y = d.y;
478                                                 if(dst.z) r.r[dst.index + a].z = d.z;
479                                                 if(dst.w) r.r[dst.index + a].w = d.w;
480                                         }
481                                         break;
482                                 case Shader::PARAMETER_ADDR:
483                                         if(dst.x) r.a0.x = d.x;
484                                         if(dst.y) r.a0.y = d.y;
485                                         if(dst.z) r.a0.z = d.z;
486                                         if(dst.w) r.a0.w = d.w;
487                                         break;
488                                 case Shader::PARAMETER_RASTOUT:
489                                         switch(dst.index)
490                                         {
491                                         case 0:
492                                                 if(dst.x) r.o[Pos].x = d.x;
493                                                 if(dst.y) r.o[Pos].y = d.y;
494                                                 if(dst.z) r.o[Pos].z = d.z;
495                                                 if(dst.w) r.o[Pos].w = d.w;
496                                                 break;
497                                         case 1:
498                                                 r.o[Fog].x = d.x;
499                                                 break;
500                                         case 2:         
501                                                 r.o[Pts].y = d.x;
502                                                 break;
503                                         default:        ASSERT(false);
504                                         }
505                                         break;
506                                 case Shader::PARAMETER_ATTROUT: 
507                                         if(dst.x) r.o[D0 + dst.index].x = d.x;
508                                         if(dst.y) r.o[D0 + dst.index].y = d.y;
509                                         if(dst.z) r.o[D0 + dst.index].z = d.z;
510                                         if(dst.w) r.o[D0 + dst.index].w = d.w;
511                                         break;
512                                 case Shader::PARAMETER_TEXCRDOUT:
513                         //      case Shader::PARAMETER_OUTPUT:
514                                         if(version < 0x0300)
515                                         {
516                                                 if(dst.x) r.o[T0 + dst.index].x = d.x;
517                                                 if(dst.y) r.o[T0 + dst.index].y = d.y;
518                                                 if(dst.z) r.o[T0 + dst.index].z = d.z;
519                                                 if(dst.w) r.o[T0 + dst.index].w = d.w;
520                                         }
521                                         else
522                                         {
523                                                 if(dst.rel.type == Shader::PARAMETER_VOID)   // Not relative
524                                                 {
525                                                         if(dst.x) r.o[dst.index].x = d.x;
526                                                         if(dst.y) r.o[dst.index].y = d.y;
527                                                         if(dst.z) r.o[dst.index].z = d.z;
528                                                         if(dst.w) r.o[dst.index].w = d.w;
529                                                 }
530                                                 else if(dst.rel.type == Shader::PARAMETER_LOOP)
531                                                 {
532                                                         Int aL = r.aL[r.loopDepth];
533
534                                                         if(dst.x) r.o[dst.index + aL].x = d.x;
535                                                         if(dst.y) r.o[dst.index + aL].y = d.y;
536                                                         if(dst.z) r.o[dst.index + aL].z = d.z;
537                                                         if(dst.w) r.o[dst.index + aL].w = d.w;
538                                                 }
539                                                 else
540                                                 {
541                                                         Int a = relativeAddress(r, dst);
542
543                                                         if(dst.x) r.o[dst.index + a].x = d.x;
544                                                         if(dst.y) r.o[dst.index + a].y = d.y;
545                                                         if(dst.z) r.o[dst.index + a].z = d.z;
546                                                         if(dst.w) r.o[dst.index + a].w = d.w;
547                                                 }
548                                         }
549                                         break;
550                                 case Shader::PARAMETER_LABEL:                                                                                                                                           break;
551                                 case Shader::PARAMETER_PREDICATE:       r.p0 = d;                                                                                                               break;
552                                 case Shader::PARAMETER_INPUT:                                                                                                                                           break;
553                                 default:
554                                         ASSERT(false);
555                                 }
556                         }
557                 }
558
559                 if(currentLabel != -1)
560                 {
561                         Nucleus::setInsertBlock(returnBlock);
562                 }
563         }
564
565         void VertexProgram::passThrough(Registers &r)
566         {
567                 if(shader)
568                 {
569                         for(int i = 0; i < 12; i++)
570                         {
571                                 unsigned char usage = shader->output[i][0].usage;
572
573                                 switch(usage)
574                                 {
575                                 case 0xFF:
576                                         continue;
577                                 case Shader::USAGE_PSIZE:
578                                         r.o[i].y = r.v[i].x;
579                                         break;
580                                 case Shader::USAGE_TEXCOORD:
581                                         r.o[i].x = r.v[i].x;
582                                         r.o[i].y = r.v[i].y;
583                                         r.o[i].z = r.v[i].z;
584                                         r.o[i].w = r.v[i].w;
585                                         break;
586                                 case Shader::USAGE_POSITION:
587                                         r.o[i].x = r.v[i].x;
588                                         r.o[i].y = r.v[i].y;
589                                         r.o[i].z = r.v[i].z;
590                                         r.o[i].w = r.v[i].w;
591                                         break;
592                                 case Shader::USAGE_COLOR:
593                                         r.o[i].x = r.v[i].x;
594                                         r.o[i].y = r.v[i].y;
595                                         r.o[i].z = r.v[i].z;
596                                         r.o[i].w = r.v[i].w;
597                                         break;
598                                 case Shader::USAGE_FOG:
599                                         r.o[i].x = r.v[i].x;
600                                         break;
601                                 default:
602                                         ASSERT(false);
603                                 }
604                         }
605                 }
606                 else
607                 {
608                         r.o[Pos].x = r.v[PositionT].x;
609                         r.o[Pos].y = r.v[PositionT].y;
610                         r.o[Pos].z = r.v[PositionT].z;
611                         r.o[Pos].w = r.v[PositionT].w;
612
613                         for(int i = 0; i < 2; i++)
614                         {
615                                 r.o[D0 + i].x = r.v[Color0 + i].x;
616                                 r.o[D0 + i].y = r.v[Color0 + i].y;
617                                 r.o[D0 + i].z = r.v[Color0 + i].z;
618                                 r.o[D0 + i].w = r.v[Color0 + i].w;
619                         }
620
621                         for(int i = 0; i < 8; i++)
622                         {
623                                 r.o[T0 + i].x = r.v[TexCoord0 + i].x;
624                                 r.o[T0 + i].y = r.v[TexCoord0 + i].y;
625                                 r.o[T0 + i].z = r.v[TexCoord0 + i].z;
626                                 r.o[T0 + i].w = r.v[TexCoord0 + i].w;
627                         }
628
629                         r.o[Pts].y = r.v[PointSize].x;
630                 }
631         }
632
633         Vector4f VertexProgram::fetchRegisterF(Registers &r, const Src &src, int offset)
634         {
635                 int i = src.index + offset;
636
637                 Vector4f reg;
638
639                 switch(src.type)
640                 {
641                 case Shader::PARAMETER_TEMP:
642                         if(src.rel.type == Shader::PARAMETER_VOID)
643                         {
644                                 reg = r.r[i];
645                         }
646                         else
647                         {
648                                 reg = r.r[i + relativeAddress(r, src)];
649                         }
650                         break;
651                 case Shader::PARAMETER_CONST:
652                         reg = readConstant(r, src, offset);
653                         break;
654                 case Shader::PARAMETER_INPUT:
655             if(src.rel.type == Shader::PARAMETER_VOID)
656                         {
657                                 reg = r.v[i];
658                         }
659                         else
660                         {
661                                 reg = r.v[i + relativeAddress(r, src)];
662                         }
663             break;
664                 case Shader::PARAMETER_VOID:                    return r.r[0];   // Dummy
665                 case Shader::PARAMETER_FLOAT4LITERAL:
666                         reg.x = Float4(src.value[0]);
667                         reg.y = Float4(src.value[1]);
668                         reg.z = Float4(src.value[2]);
669                         reg.w = Float4(src.value[3]);
670                         break;
671                 case Shader::PARAMETER_ADDR:                    reg = r.a0;             break;
672                 case Shader::PARAMETER_CONSTBOOL:               return r.r[0];   // Dummy
673                 case Shader::PARAMETER_CONSTINT:                return r.r[0];   // Dummy
674                 case Shader::PARAMETER_LOOP:                    return r.r[0];   // Dummy
675                 case Shader::PARAMETER_PREDICATE:               return r.r[0];   // Dummy
676                 case Shader::PARAMETER_SAMPLER:
677                         if(src.rel.type == Shader::PARAMETER_VOID)
678                         {
679                                 reg.x = As<Float4>(Int4(i));
680                         }
681                         else if(src.rel.type == Shader::PARAMETER_TEMP)
682                         {
683                                 reg.x = As<Float4>(Int4(i) + RoundInt(r.r[src.rel.index].x));
684                         }
685                         return reg;
686                 case Shader::PARAMETER_OUTPUT:
687             if(src.rel.type == Shader::PARAMETER_VOID)
688                         {
689                                 reg = r.o[i];
690                         }
691                         else
692                         {
693                                 reg = r.o[i + relativeAddress(r, src)];
694                         }
695                         break;
696                 case Shader::PARAMETER_MISCTYPE:
697                         reg.x = Float(r.instanceID);
698                         return reg;
699                 default:
700                         ASSERT(false);
701                 }
702
703                 const Float4 &x = reg[(src.swizzle >> 0) & 0x3];
704                 const Float4 &y = reg[(src.swizzle >> 2) & 0x3];
705                 const Float4 &z = reg[(src.swizzle >> 4) & 0x3];
706                 const Float4 &w = reg[(src.swizzle >> 6) & 0x3];
707
708                 Vector4f mod;
709
710                 switch(src.modifier)
711                 {
712                 case Shader::MODIFIER_NONE:
713                         mod.x = x;
714                         mod.y = y;
715                         mod.z = z;
716                         mod.w = w;
717                         break;
718                 case Shader::MODIFIER_NEGATE:
719                         mod.x = -x;
720                         mod.y = -y;
721                         mod.z = -z;
722                         mod.w = -w;
723                         break;
724                 case Shader::MODIFIER_ABS:
725                         mod.x = Abs(x);
726                         mod.y = Abs(y);
727                         mod.z = Abs(z);
728                         mod.w = Abs(w);
729                         break;
730                 case Shader::MODIFIER_ABS_NEGATE:
731                         mod.x = -Abs(x);
732                         mod.y = -Abs(y);
733                         mod.z = -Abs(z);
734                         mod.w = -Abs(w);
735                         break;
736                 case Shader::MODIFIER_NOT:
737                         mod.x = As<Float4>(As<Int4>(x) ^ Int4(0xFFFFFFFF));
738                         mod.y = As<Float4>(As<Int4>(y) ^ Int4(0xFFFFFFFF));
739                         mod.z = As<Float4>(As<Int4>(z) ^ Int4(0xFFFFFFFF));
740                         mod.w = As<Float4>(As<Int4>(w) ^ Int4(0xFFFFFFFF));
741                         break;
742                 default:
743                         ASSERT(false);
744                 }
745
746                 return mod;
747         }
748
749         Vector4f VertexProgram::readConstant(Registers &r, const Src &src, int offset)
750         {
751                 Vector4f c;
752
753                 int i = src.index + offset;
754
755                 if(src.rel.type == Shader::PARAMETER_VOID)   // Not relative
756                 {
757                         c.x = c.y = c.z = c.w = *Pointer<Float4>(r.data + OFFSET(DrawData,vs.c[i]));
758
759                         c.x = c.x.xxxx;
760                         c.y = c.y.yyyy;
761                         c.z = c.z.zzzz;
762                         c.w = c.w.wwww;
763
764                         if(shader->containsDefineInstruction())   // Constant may be known at compile time
765                         {
766                                 for(size_t j = 0; j < shader->getLength(); j++)
767                                 {
768                                         const Shader::Instruction &instruction = *shader->getInstruction(j);
769
770                                         if(instruction.opcode == Shader::OPCODE_DEF)
771                                         {
772                                                 if(instruction.dst.index == i)
773                                                 {
774                                                         c.x = Float4(instruction.src[0].value[0]);
775                                                         c.y = Float4(instruction.src[0].value[1]);
776                                                         c.z = Float4(instruction.src[0].value[2]);
777                                                         c.w = Float4(instruction.src[0].value[3]);
778
779                                                         break;
780                                                 }
781                                         }
782                                 }
783                         }
784                 }
785                 else if(src.rel.type == Shader::PARAMETER_LOOP)
786                 {
787                         Int loopCounter = r.aL[r.loopDepth];
788
789                         c.x = c.y = c.z = c.w = *Pointer<Float4>(r.data + OFFSET(DrawData,vs.c[i]) + loopCounter * 16);
790
791                         c.x = c.x.xxxx;
792                         c.y = c.y.yyyy;
793                         c.z = c.z.zzzz;
794                         c.w = c.w.wwww;
795                 }
796                 else
797                 {
798                         if(src.rel.deterministic)
799                         {
800                                 Int a = relativeAddress(r, src);
801                         
802                                 c.x = c.y = c.z = c.w = *Pointer<Float4>(r.data + OFFSET(DrawData,vs.c[i]) + a * 16);
803
804                                 c.x = c.x.xxxx;
805                                 c.y = c.y.yyyy;
806                                 c.z = c.z.zzzz;
807                                 c.w = c.w.wwww;
808                         }
809                         else
810                         {
811                                 int component = src.rel.swizzle & 0x03;
812                                 Float4 a;
813
814                                 switch(src.rel.type)
815                                 {
816                                 case Shader::PARAMETER_ADDR:   a = r.a0[component]; break;
817                                 case Shader::PARAMETER_TEMP:   a = r.r[src.rel.index][component]; break;
818                                 case Shader::PARAMETER_INPUT:  a = r.v[src.rel.index][component]; break;
819                                 case Shader::PARAMETER_OUTPUT: a = r.o[src.rel.index][component]; break;
820                                 case Shader::PARAMETER_CONST:  a = *Pointer<Float>(r.data + OFFSET(DrawData,vs.c[src.rel.index][component])); break;
821                                 default: ASSERT(false);
822                                 }
823
824                                 Int4 index = Int4(i) + RoundInt(a) * Int4(src.rel.scale);
825
826                                 index = Min(As<UInt4>(index), UInt4(256));   // Clamp to constant register range, c[256] = {0, 0, 0, 0}
827                                 
828                                 Int index0 = Extract(index, 0);
829                                 Int index1 = Extract(index, 1);
830                                 Int index2 = Extract(index, 2);
831                                 Int index3 = Extract(index, 3);
832
833                                 c.x = *Pointer<Float4>(r.data + OFFSET(DrawData,vs.c) + index0 * 16, 16);
834                                 c.y = *Pointer<Float4>(r.data + OFFSET(DrawData,vs.c) + index1 * 16, 16);
835                                 c.z = *Pointer<Float4>(r.data + OFFSET(DrawData,vs.c) + index2 * 16, 16);
836                                 c.w = *Pointer<Float4>(r.data + OFFSET(DrawData,vs.c) + index3 * 16, 16);
837
838                                 transpose4x4(c.x, c.y, c.z, c.w);
839                         }
840                 }
841
842                 return c;
843         }
844
845         Int VertexProgram::relativeAddress(Registers &r, const Shader::Parameter &var)
846         {
847                 ASSERT(var.rel.deterministic);
848
849                 if(var.rel.type == Shader::PARAMETER_TEMP)
850                 {
851                         return RoundInt(Extract(r.r[var.rel.index].x, 0)) * var.rel.scale;
852                 }
853                 else if(var.rel.type == Shader::PARAMETER_INPUT)
854                 {
855                         return RoundInt(Extract(r.v[var.rel.index].x, 0)) * var.rel.scale;
856                 }
857                 else if(var.rel.type == Shader::PARAMETER_OUTPUT)
858                 {
859                         return RoundInt(Extract(r.o[var.rel.index].x, 0)) * var.rel.scale;
860                 }
861                 else if(var.rel.type == Shader::PARAMETER_CONST)
862                 {
863                         RValue<Float4> c = *Pointer<Float4>(r.data + OFFSET(DrawData,vs.c[var.rel.index]));
864
865                         return RoundInt(Extract(c, 0)) * var.rel.scale;
866                 }
867                 else ASSERT(false);
868
869                 return 0;
870         }
871
872         Int4 VertexProgram::enableMask(Registers &r, const Shader::Instruction *instruction)
873         {
874                 Int4 enable = instruction->analysisBranch ? Int4(r.enableStack[r.enableIndex]) : Int4(0xFFFFFFFF);
875                 
876                 if(!whileTest)
877                 {
878                         if(shader->containsBreakInstruction() && instruction->analysisBreak)
879                         {
880                                 enable &= r.enableBreak;
881                         }
882
883                         if(shader->containsContinueInstruction() && instruction->analysisContinue)
884                         {
885                                 enable &= r.enableContinue;
886                         }
887
888                         if(shader->containsLeaveInstruction() && instruction->analysisLeave)
889                         {
890                                 enable &= r.enableLeave;
891                         }
892                 }
893
894                 return enable;
895         }
896
897         void VertexProgram::M3X2(Registers &r, Vector4f &dst, Vector4f &src0, Src &src1)
898         {
899                 Vector4f row0 = fetchRegisterF(r, src1, 0);
900                 Vector4f row1 = fetchRegisterF(r, src1, 1);
901
902                 dst.x = dot3(src0, row0);
903                 dst.y = dot3(src0, row1);
904         }
905
906         void VertexProgram::M3X3(Registers &r, Vector4f &dst, Vector4f &src0, Src &src1)
907         {
908                 Vector4f row0 = fetchRegisterF(r, src1, 0);
909                 Vector4f row1 = fetchRegisterF(r, src1, 1);
910                 Vector4f row2 = fetchRegisterF(r, src1, 2);
911
912                 dst.x = dot3(src0, row0);
913                 dst.y = dot3(src0, row1);
914                 dst.z = dot3(src0, row2);
915         }
916
917         void VertexProgram::M3X4(Registers &r, Vector4f &dst, Vector4f &src0, Src &src1)
918         {
919                 Vector4f row0 = fetchRegisterF(r, src1, 0);
920                 Vector4f row1 = fetchRegisterF(r, src1, 1);
921                 Vector4f row2 = fetchRegisterF(r, src1, 2);
922                 Vector4f row3 = fetchRegisterF(r, src1, 3);
923
924                 dst.x = dot3(src0, row0);
925                 dst.y = dot3(src0, row1);
926                 dst.z = dot3(src0, row2);
927                 dst.w = dot3(src0, row3);
928         }
929
930         void VertexProgram::M4X3(Registers &r, Vector4f &dst, Vector4f &src0, Src &src1)
931         {
932                 Vector4f row0 = fetchRegisterF(r, src1, 0);
933                 Vector4f row1 = fetchRegisterF(r, src1, 1);
934                 Vector4f row2 = fetchRegisterF(r, src1, 2);
935
936                 dst.x = dot4(src0, row0);
937                 dst.y = dot4(src0, row1);
938                 dst.z = dot4(src0, row2);
939         }
940
941         void VertexProgram::M4X4(Registers &r, Vector4f &dst, Vector4f &src0, Src &src1)
942         {
943                 Vector4f row0 = fetchRegisterF(r, src1, 0);
944                 Vector4f row1 = fetchRegisterF(r, src1, 1);
945                 Vector4f row2 = fetchRegisterF(r, src1, 2);
946                 Vector4f row3 = fetchRegisterF(r, src1, 3);
947
948                 dst.x = dot4(src0, row0);
949                 dst.y = dot4(src0, row1);
950                 dst.z = dot4(src0, row2);
951                 dst.w = dot4(src0, row3);
952         }
953
954         void VertexProgram::BREAK(Registers &r)
955         {
956                 llvm::BasicBlock *deadBlock = Nucleus::createBasicBlock();
957                 llvm::BasicBlock *endBlock = loopRepEndBlock[loopRepDepth - 1];
958
959                 if(breakDepth == 0)
960                 {
961                         r.enableIndex = r.enableIndex - breakDepth;
962                         Nucleus::createBr(endBlock);
963                 }
964                 else
965                 {
966                         r.enableBreak = r.enableBreak & ~r.enableStack[r.enableIndex];
967                         Bool allBreak = SignMask(r.enableBreak) == 0x0;
968
969                         r.enableIndex = r.enableIndex - breakDepth;
970                         branch(allBreak, endBlock, deadBlock);
971                 }
972
973                 Nucleus::setInsertBlock(deadBlock);
974                 r.enableIndex = r.enableIndex + breakDepth;
975         }
976
977         void VertexProgram::BREAKC(Registers &r, Vector4f &src0, Vector4f &src1, Control control)
978         {
979                 Int4 condition;
980
981                 switch(control)
982                 {
983                 case Shader::CONTROL_GT: condition = CmpNLE(src0.x,  src1.x);   break;
984                 case Shader::CONTROL_EQ: condition = CmpEQ(src0.x, src1.x);             break;
985                 case Shader::CONTROL_GE: condition = CmpNLT(src0.x, src1.x);    break;
986                 case Shader::CONTROL_LT: condition = CmpLT(src0.x,  src1.x);    break;
987                 case Shader::CONTROL_NE: condition = CmpNEQ(src0.x, src1.x);    break;
988                 case Shader::CONTROL_LE: condition = CmpLE(src0.x, src1.x);             break;
989                 default:
990                         ASSERT(false);
991                 }
992
993                 BREAK(r, condition);
994         }
995
996         void VertexProgram::BREAKP(Registers &r, const Src &predicateRegister)   // FIXME: Factor out parts common with BREAKC
997         {
998                 Int4 condition = As<Int4>(r.p0[predicateRegister.swizzle & 0x3]);
999
1000                 if(predicateRegister.modifier == Shader::MODIFIER_NOT)
1001                 {
1002                         condition = ~condition;
1003                 }
1004
1005                 BREAK(r, condition);
1006         }
1007
1008         void VertexProgram::BREAK(Registers &r, Int4 &condition)
1009         {
1010                 condition &= r.enableStack[r.enableIndex];
1011
1012                 llvm::BasicBlock *continueBlock = Nucleus::createBasicBlock();
1013                 llvm::BasicBlock *endBlock = loopRepEndBlock[loopRepDepth - 1];
1014
1015                 r.enableBreak = r.enableBreak & ~condition;
1016                 Bool allBreak = SignMask(r.enableBreak) == 0x0;
1017
1018                 r.enableIndex = r.enableIndex - breakDepth;
1019                 branch(allBreak, endBlock, continueBlock);
1020
1021                 Nucleus::setInsertBlock(continueBlock);
1022                 r.enableIndex = r.enableIndex + breakDepth;
1023         }
1024
1025         void VertexProgram::CONTINUE(Registers &r)
1026         {
1027                 r.enableContinue = r.enableContinue & ~r.enableStack[r.enableIndex];
1028         }
1029
1030         void VertexProgram::TEST()
1031         {
1032                 whileTest = true;
1033         }
1034
1035         void VertexProgram::CALL(Registers &r, int labelIndex, int callSiteIndex)
1036         {
1037                 if(!labelBlock[labelIndex])
1038                 {
1039                         labelBlock[labelIndex] = Nucleus::createBasicBlock();
1040                 }
1041
1042                 if(callRetBlock[labelIndex].size() > 1)
1043                 {
1044                         r.callStack[r.stackIndex++] = UInt(callSiteIndex);
1045                 }
1046
1047                 Int4 restoreLeave = r.enableLeave;
1048
1049                 Nucleus::createBr(labelBlock[labelIndex]);
1050                 Nucleus::setInsertBlock(callRetBlock[labelIndex][callSiteIndex]);
1051
1052                 r.enableLeave = restoreLeave;
1053         }
1054
1055         void VertexProgram::CALLNZ(Registers &r, int labelIndex, int callSiteIndex, const Src &src)
1056         {
1057                 if(src.type == Shader::PARAMETER_CONSTBOOL)
1058                 {
1059                         CALLNZb(r, labelIndex, callSiteIndex, src);
1060                 }
1061                 else if(src.type == Shader::PARAMETER_PREDICATE)
1062                 {
1063                         CALLNZp(r, labelIndex, callSiteIndex, src);
1064                 }
1065                 else ASSERT(false);
1066         }
1067
1068         void VertexProgram::CALLNZb(Registers &r, int labelIndex, int callSiteIndex, const Src &boolRegister)
1069         {
1070                 Bool condition = (*Pointer<Byte>(r.data + OFFSET(DrawData,vs.b[boolRegister.index])) != Byte(0));   // FIXME
1071                 
1072                 if(boolRegister.modifier == Shader::MODIFIER_NOT)
1073                 {
1074                         condition = !condition; 
1075                 }
1076
1077                 if(!labelBlock[labelIndex])
1078                 {
1079                         labelBlock[labelIndex] = Nucleus::createBasicBlock();
1080                 }
1081
1082                 if(callRetBlock[labelIndex].size() > 1)
1083                 {
1084                         r.callStack[r.stackIndex++] = UInt(callSiteIndex);
1085                 }
1086
1087                 Int4 restoreLeave = r.enableLeave;
1088
1089                 branch(condition, labelBlock[labelIndex], callRetBlock[labelIndex][callSiteIndex]);
1090                 Nucleus::setInsertBlock(callRetBlock[labelIndex][callSiteIndex]);
1091
1092                 r.enableLeave = restoreLeave;
1093         }
1094
1095         void VertexProgram::CALLNZp(Registers &r, int labelIndex, int callSiteIndex, const Src &predicateRegister)
1096         {
1097                 Int4 condition = As<Int4>(r.p0[predicateRegister.swizzle & 0x3]);
1098
1099                 if(predicateRegister.modifier == Shader::MODIFIER_NOT)
1100                 {
1101                         condition = ~condition;
1102                 }
1103
1104                 condition &= r.enableStack[r.enableIndex];
1105
1106                 if(!labelBlock[labelIndex])
1107                 {
1108                         labelBlock[labelIndex] = Nucleus::createBasicBlock();
1109                 }
1110
1111                 if(callRetBlock[labelIndex].size() > 1)
1112                 {
1113                         r.callStack[r.stackIndex++] = UInt(callSiteIndex);
1114                 }
1115
1116                 r.enableIndex++;
1117                 r.enableStack[r.enableIndex] = condition;
1118                 Int4 restoreLeave = r.enableLeave;
1119
1120                 Bool notAllFalse = SignMask(condition) != 0;
1121                 branch(notAllFalse, labelBlock[labelIndex], callRetBlock[labelIndex][callSiteIndex]);
1122                 Nucleus::setInsertBlock(callRetBlock[labelIndex][callSiteIndex]);
1123
1124                 r.enableIndex--;
1125                 r.enableLeave = restoreLeave;
1126         }
1127
1128         void VertexProgram::ELSE(Registers &r)
1129         {
1130                 ifDepth--;
1131
1132                 llvm::BasicBlock *falseBlock = ifFalseBlock[ifDepth];
1133                 llvm::BasicBlock *endBlock = Nucleus::createBasicBlock();
1134
1135                 if(isConditionalIf[ifDepth])
1136                 {
1137                         Int4 condition = ~r.enableStack[r.enableIndex] & r.enableStack[r.enableIndex - 1];
1138                         Bool notAllFalse = SignMask(condition) != 0;
1139
1140                         branch(notAllFalse, falseBlock, endBlock);
1141
1142                         r.enableStack[r.enableIndex] = ~r.enableStack[r.enableIndex] & r.enableStack[r.enableIndex - 1];
1143                 }
1144                 else
1145                 {
1146                         Nucleus::createBr(endBlock);
1147                         Nucleus::setInsertBlock(falseBlock);
1148                 }
1149
1150                 ifFalseBlock[ifDepth] = endBlock;
1151
1152                 ifDepth++;
1153         }
1154
1155         void VertexProgram::ENDIF(Registers &r)
1156         {
1157                 ifDepth--;
1158
1159                 llvm::BasicBlock *endBlock = ifFalseBlock[ifDepth];
1160
1161                 Nucleus::createBr(endBlock);
1162                 Nucleus::setInsertBlock(endBlock);
1163
1164                 if(isConditionalIf[ifDepth])
1165                 {
1166                         breakDepth--;
1167                         r.enableIndex--;
1168                 }
1169         }
1170
1171         void VertexProgram::ENDLOOP(Registers &r)
1172         {
1173                 loopRepDepth--;
1174
1175                 r.aL[r.loopDepth] = r.aL[r.loopDepth] + r.increment[r.loopDepth];   // FIXME: +=
1176
1177                 llvm::BasicBlock *testBlock = loopRepTestBlock[loopRepDepth];
1178                 llvm::BasicBlock *endBlock = loopRepEndBlock[loopRepDepth];
1179
1180                 Nucleus::createBr(testBlock);
1181                 Nucleus::setInsertBlock(endBlock);
1182
1183                 r.loopDepth--;
1184                 r.enableBreak = Int4(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
1185         }
1186
1187         void VertexProgram::ENDREP(Registers &r)
1188         {
1189                 loopRepDepth--;
1190
1191                 llvm::BasicBlock *testBlock = loopRepTestBlock[loopRepDepth];
1192                 llvm::BasicBlock *endBlock = loopRepEndBlock[loopRepDepth];
1193
1194                 Nucleus::createBr(testBlock);
1195                 Nucleus::setInsertBlock(endBlock);
1196
1197                 r.loopDepth--;
1198                 r.enableBreak = Int4(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
1199         }
1200
1201         void VertexProgram::ENDWHILE(Registers &r)
1202         {
1203                 loopRepDepth--;
1204
1205                 llvm::BasicBlock *testBlock = loopRepTestBlock[loopRepDepth];
1206                 llvm::BasicBlock *endBlock = loopRepEndBlock[loopRepDepth];
1207
1208                 Nucleus::createBr(testBlock);
1209                 Nucleus::setInsertBlock(endBlock);
1210
1211                 r.enableIndex--;
1212                 r.enableBreak = Int4(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
1213                 whileTest = false;
1214         }
1215
1216         void VertexProgram::IF(Registers &r, const Src &src)
1217         {
1218                 if(src.type == Shader::PARAMETER_CONSTBOOL)
1219                 {
1220                         IFb(r, src);
1221                 }
1222                 else if(src.type == Shader::PARAMETER_PREDICATE)
1223                 {
1224                         IFp(r, src);
1225                 }
1226                 else
1227                 {
1228                         Int4 condition = As<Int4>(fetchRegisterF(r, src).x);
1229                         IF(r, condition);
1230                 }
1231         }
1232
1233         void VertexProgram::IFb(Registers &r, const Src &boolRegister)
1234         {
1235                 ASSERT(ifDepth < 24 + 4);
1236
1237                 Bool condition = (*Pointer<Byte>(r.data + OFFSET(DrawData,vs.b[boolRegister.index])) != Byte(0));   // FIXME
1238
1239                 if(boolRegister.modifier == Shader::MODIFIER_NOT)
1240                 {
1241                         condition = !condition;
1242                 }
1243
1244                 llvm::BasicBlock *trueBlock = Nucleus::createBasicBlock();
1245                 llvm::BasicBlock *falseBlock = Nucleus::createBasicBlock();
1246
1247                 branch(condition, trueBlock, falseBlock);
1248
1249                 isConditionalIf[ifDepth] = false;
1250                 ifFalseBlock[ifDepth] = falseBlock;
1251
1252                 ifDepth++;
1253         }
1254
1255         void VertexProgram::IFp(Registers &r, const Src &predicateRegister)
1256         {
1257                 Int4 condition = As<Int4>(r.p0[predicateRegister.swizzle & 0x3]);
1258
1259                 if(predicateRegister.modifier == Shader::MODIFIER_NOT)
1260                 {
1261                         condition = ~condition;
1262                 }
1263
1264                 IF(r, condition);
1265         }
1266
1267         void VertexProgram::IFC(Registers &r, Vector4f &src0, Vector4f &src1, Control control)
1268         {
1269                 Int4 condition;
1270
1271                 switch(control)
1272                 {
1273                 case Shader::CONTROL_GT: condition = CmpNLE(src0.x,  src1.x);   break;
1274                 case Shader::CONTROL_EQ: condition = CmpEQ(src0.x, src1.x);             break;
1275                 case Shader::CONTROL_GE: condition = CmpNLT(src0.x, src1.x);    break;
1276                 case Shader::CONTROL_LT: condition = CmpLT(src0.x,  src1.x);    break;
1277                 case Shader::CONTROL_NE: condition = CmpNEQ(src0.x, src1.x);    break;
1278                 case Shader::CONTROL_LE: condition = CmpLE(src0.x, src1.x);             break;
1279                 default:
1280                         ASSERT(false);
1281                 }
1282
1283                 IF(r, condition);
1284         }
1285
1286         void VertexProgram::IF(Registers &r, Int4 &condition)
1287         {
1288                 condition &= r.enableStack[r.enableIndex];
1289
1290                 r.enableIndex++;
1291                 r.enableStack[r.enableIndex] = condition;
1292
1293                 llvm::BasicBlock *trueBlock = Nucleus::createBasicBlock();
1294                 llvm::BasicBlock *falseBlock = Nucleus::createBasicBlock();
1295
1296                 Bool notAllFalse = SignMask(condition) != 0;
1297
1298                 branch(notAllFalse, trueBlock, falseBlock);
1299
1300                 isConditionalIf[ifDepth] = true;
1301                 ifFalseBlock[ifDepth] = falseBlock;
1302
1303                 ifDepth++;
1304                 breakDepth++;
1305         }
1306
1307         void VertexProgram::LABEL(int labelIndex)
1308         {
1309                 if(!labelBlock[labelIndex])
1310                 {
1311                         labelBlock[labelIndex] = Nucleus::createBasicBlock();
1312                 }
1313
1314                 Nucleus::setInsertBlock(labelBlock[labelIndex]);
1315                 currentLabel = labelIndex;
1316         }
1317
1318         void VertexProgram::LOOP(Registers &r, const Src &integerRegister)
1319         {
1320                 r.loopDepth++;
1321
1322                 r.iteration[r.loopDepth] = *Pointer<Int>(r.data + OFFSET(DrawData,vs.i[integerRegister.index][0]));
1323                 r.aL[r.loopDepth] = *Pointer<Int>(r.data + OFFSET(DrawData,vs.i[integerRegister.index][1]));
1324                 r.increment[r.loopDepth] = *Pointer<Int>(r.data + OFFSET(DrawData,vs.i[integerRegister.index][2]));
1325
1326                 // FIXME: Compiles to two instructions?
1327                 If(r.increment[r.loopDepth] == 0)
1328                 {
1329                         r.increment[r.loopDepth] = 1;
1330                 }
1331
1332                 llvm::BasicBlock *loopBlock = Nucleus::createBasicBlock();
1333                 llvm::BasicBlock *testBlock = Nucleus::createBasicBlock();
1334                 llvm::BasicBlock *endBlock = Nucleus::createBasicBlock();
1335
1336                 loopRepTestBlock[loopRepDepth] = testBlock;
1337                 loopRepEndBlock[loopRepDepth] = endBlock;
1338
1339                 // FIXME: jump(testBlock)
1340                 Nucleus::createBr(testBlock);
1341                 Nucleus::setInsertBlock(testBlock);
1342
1343                 branch(r.iteration[r.loopDepth] > 0, loopBlock, endBlock);
1344                 Nucleus::setInsertBlock(loopBlock);
1345
1346                 r.iteration[r.loopDepth] = r.iteration[r.loopDepth] - 1;   // FIXME: --
1347                 
1348                 loopRepDepth++;
1349                 breakDepth = 0;
1350         }
1351
1352         void VertexProgram::REP(Registers &r, const Src &integerRegister)
1353         {
1354                 r.loopDepth++;
1355
1356                 r.iteration[r.loopDepth] = *Pointer<Int>(r.data + OFFSET(DrawData,vs.i[integerRegister.index][0]));
1357                 r.aL[r.loopDepth] = r.aL[r.loopDepth - 1];
1358
1359                 llvm::BasicBlock *loopBlock = Nucleus::createBasicBlock();
1360                 llvm::BasicBlock *testBlock = Nucleus::createBasicBlock();
1361                 llvm::BasicBlock *endBlock = Nucleus::createBasicBlock();
1362
1363                 loopRepTestBlock[loopRepDepth] = testBlock;
1364                 loopRepEndBlock[loopRepDepth] = endBlock;
1365
1366                 // FIXME: jump(testBlock)
1367                 Nucleus::createBr(testBlock);
1368                 Nucleus::setInsertBlock(testBlock);
1369
1370                 branch(r.iteration[r.loopDepth] > 0, loopBlock, endBlock);
1371                 Nucleus::setInsertBlock(loopBlock);
1372
1373                 r.iteration[r.loopDepth] = r.iteration[r.loopDepth] - 1;   // FIXME: --
1374
1375                 loopRepDepth++;
1376                 breakDepth = 0;
1377         }
1378
1379         void VertexProgram::WHILE(Registers &r, const Src &temporaryRegister)
1380         {
1381                 r.enableIndex++;
1382
1383                 llvm::BasicBlock *loopBlock = Nucleus::createBasicBlock();
1384                 llvm::BasicBlock *testBlock = Nucleus::createBasicBlock();
1385                 llvm::BasicBlock *endBlock = Nucleus::createBasicBlock();
1386                 
1387                 loopRepTestBlock[loopRepDepth] = testBlock;
1388                 loopRepEndBlock[loopRepDepth] = endBlock;
1389
1390                 Int4 restoreBreak = r.enableBreak;
1391                 Int4 restoreContinue = r.enableContinue;
1392
1393                 // FIXME: jump(testBlock)
1394                 Nucleus::createBr(testBlock);
1395                 Nucleus::setInsertBlock(testBlock);
1396                 r.enableContinue = restoreContinue;
1397
1398                 const Vector4f &src = fetchRegisterF(r, temporaryRegister);
1399                 Int4 condition = As<Int4>(src.x);
1400                 condition &= r.enableStack[r.enableIndex - 1];
1401                 r.enableStack[r.enableIndex] = condition;
1402
1403                 Bool notAllFalse = SignMask(condition) != 0;
1404                 branch(notAllFalse, loopBlock, endBlock);
1405                 
1406                 Nucleus::setInsertBlock(endBlock);
1407                 r.enableBreak = restoreBreak;
1408                 
1409                 Nucleus::setInsertBlock(loopBlock);
1410
1411                 loopRepDepth++;
1412                 breakDepth = 0;
1413         }
1414
1415         void VertexProgram::RET(Registers &r)
1416         {
1417                 if(currentLabel == -1)
1418                 {
1419                         returnBlock = Nucleus::createBasicBlock();
1420                         Nucleus::createBr(returnBlock);
1421                 }
1422                 else
1423                 {
1424                         llvm::BasicBlock *unreachableBlock = Nucleus::createBasicBlock();
1425
1426                         if(callRetBlock[currentLabel].size() > 1)   // Pop the return destination from the call stack
1427                         {
1428                                 // FIXME: Encapsulate
1429                                 UInt index = r.callStack[--r.stackIndex];
1430  
1431                                 llvm::Value *value = index.loadValue();
1432                                 llvm::Value *switchInst = Nucleus::createSwitch(value, unreachableBlock, (int)callRetBlock[currentLabel].size());
1433
1434                                 for(unsigned int i = 0; i < callRetBlock[currentLabel].size(); i++)
1435                                 {
1436                                         Nucleus::addSwitchCase(switchInst, i, callRetBlock[currentLabel][i]);
1437                                 }
1438                         }
1439                         else if(callRetBlock[currentLabel].size() == 1)   // Jump directly to the unique return destination
1440                         {
1441                                 Nucleus::createBr(callRetBlock[currentLabel][0]);
1442                         }
1443                         else   // Function isn't called
1444                         {
1445                                 Nucleus::createBr(unreachableBlock);
1446                         }
1447
1448                         Nucleus::setInsertBlock(unreachableBlock);
1449                         Nucleus::createUnreachable();
1450                 }
1451         }
1452
1453         void VertexProgram::LEAVE(Registers &r)
1454         {
1455                 r.enableLeave = r.enableLeave & ~r.enableStack[r.enableIndex];
1456
1457                 // FIXME: Return from function if all instances left
1458                 // FIXME: Use enableLeave in other control-flow constructs
1459         }
1460
1461         void VertexProgram::TEXLDL(Registers &r, Vector4f &dst, Vector4f &src0, const Src &src1)
1462         {
1463                 Vector4f tmp;
1464                 sampleTexture(r, tmp, src1, src0.x, src0.y, src0.z, src0.w);
1465
1466                 dst.x = tmp[(src1.swizzle >> 0) & 0x3];
1467                 dst.y = tmp[(src1.swizzle >> 2) & 0x3];
1468                 dst.z = tmp[(src1.swizzle >> 4) & 0x3];
1469                 dst.w = tmp[(src1.swizzle >> 6) & 0x3];
1470         }
1471
1472         void VertexProgram::TEX(Registers &r, Vector4f &dst, Vector4f &src0, const Src &src1)
1473         {
1474                 Float4 lod = Float4(0.0f);
1475                 Vector4f tmp;
1476                 sampleTexture(r, tmp, src1, src0.x, src0.y, src0.z, lod);
1477
1478                 dst.x = tmp[(src1.swizzle >> 0) & 0x3];
1479                 dst.y = tmp[(src1.swizzle >> 2) & 0x3];
1480                 dst.z = tmp[(src1.swizzle >> 4) & 0x3];
1481                 dst.w = tmp[(src1.swizzle >> 6) & 0x3];
1482         }
1483
1484         void VertexProgram::sampleTexture(Registers &r, Vector4f &c, const Src &s, Float4 &u, Float4 &v, Float4 &w, Float4 &q)
1485         {
1486                 if(s.type == Shader::PARAMETER_SAMPLER && s.rel.type == Shader::PARAMETER_VOID)
1487                 {
1488                         Pointer<Byte> texture = r.data + OFFSET(DrawData,mipmap[16]) + s.index * sizeof(Texture);
1489                         sampler[s.index]->sampleTexture(texture, c, u, v, w, q, r.a0, r.a0, false, false, true);        
1490                 }
1491                 else
1492                 {
1493                         Int index = As<Int>(Float(fetchRegisterF(r, s).x.x));
1494
1495                         for(int i = 0; i < 16; i++)
1496                         {
1497                                 if(shader->usesSampler(i))
1498                                 {
1499                                         If(index == i)
1500                                         {
1501                                                 Pointer<Byte> texture = r.data + OFFSET(DrawData,mipmap[16]) + i * sizeof(Texture);
1502                                                 sampler[i]->sampleTexture(texture, c, u, v, w, q, r.a0, r.a0, false, false, true);
1503                                                 // FIXME: When the sampler states are the same, we could use one sampler and just index the texture
1504                                         }
1505                                 }
1506                         }
1507                 }
1508         }
1509 }