OSDN Git Service

Added 32 bit integer output types
[android-x86/external-swiftshader.git] / src / Shader / PixelProgram.cpp
1 // SwiftShader Software Renderer
2 //
3 // Copyright(c) 2015 Google 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 Google 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 "PixelProgram.hpp"
13 #include "Primitive.hpp"
14 #include "Renderer.hpp"
15 #include "SamplerCore.hpp"
16
17 namespace sw
18 {
19         extern bool postBlendSRGB;
20         extern bool booleanFaceRegister;
21         extern bool halfIntegerCoordinates;     // Pixel centers are not at integer coordinates
22         extern bool fullPixelPositionRegister;
23
24         void PixelProgram::setBuiltins(Int &x, Int &y, Float4(&z)[4], Float4 &w)
25         {
26                 if(shader->getVersion() >= 0x0300)
27                 {
28                         if(shader->vPosDeclared)
29                         {
30                                 if(!halfIntegerCoordinates)
31                                 {
32                                         vPos.x = Float4(Float(x)) + Float4(0, 1, 0, 1);
33                                         vPos.y = Float4(Float(y)) + Float4(0, 0, 1, 1);
34                                 }
35                                 else
36                                 {
37                                         vPos.x = Float4(Float(x)) + Float4(0.5f, 1.5f, 0.5f, 1.5f);
38                                         vPos.y = Float4(Float(y)) + Float4(0.5f, 0.5f, 1.5f, 1.5f);
39                                 }
40
41                                 if(fullPixelPositionRegister)
42                                 {
43                                         vPos.z = z[0]; // FIXME: Centroid?
44                                         vPos.w = w;    // FIXME: Centroid?
45                                 }
46                         }
47
48                         if(shader->vFaceDeclared)
49                         {
50                                 Float4 area = *Pointer<Float>(primitive + OFFSET(Primitive, area));
51                                 Float4 face = booleanFaceRegister ? Float4(As<Float4>(CmpNLT(area, Float4(0.0f)))) : area;
52
53                                 vFace.x = face;
54                                 vFace.y = face;
55                                 vFace.z = face;
56                                 vFace.w = face;
57                         }
58                 }
59         }
60
61         void PixelProgram::applyShader(Int cMask[4])
62         {
63                 enableIndex = 0;
64                 stackIndex = 0;
65
66                 if(shader->containsLeaveInstruction())
67                 {
68                         enableLeave = Int4(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
69                 }
70
71                 for(int i = 0; i < RENDERTARGETS; i++)
72                 {
73                         if(state.targetFormat[i] != FORMAT_NULL)
74                         {
75                                 oC[i] = Vector4f(0.0f, 0.0f, 0.0f, 0.0f);
76                         }
77                 }
78
79                 // Create all call site return blocks up front
80                 for(size_t i = 0; i < shader->getLength(); i++)
81                 {
82                         const Shader::Instruction *instruction = shader->getInstruction(i);
83                         Shader::Opcode opcode = instruction->opcode;
84
85                         if(opcode == Shader::OPCODE_CALL || opcode == Shader::OPCODE_CALLNZ)
86                         {
87                                 const Dst &dst = instruction->dst;
88
89                                 ASSERT(callRetBlock[dst.label].size() == dst.callSite);
90                                 callRetBlock[dst.label].push_back(Nucleus::createBasicBlock());
91                         }
92                 }
93
94                 for(size_t i = 0; i < shader->getLength(); i++)
95                 {
96                         const Shader::Instruction *instruction = shader->getInstruction(i);
97                         Shader::Opcode opcode = instruction->opcode;
98
99                         if(opcode == Shader::OPCODE_DCL || opcode == Shader::OPCODE_DEF || opcode == Shader::OPCODE_DEFI || opcode == Shader::OPCODE_DEFB)
100                         {
101                                 continue;
102                         }
103
104                         const Dst &dst = instruction->dst;
105                         const Src &src0 = instruction->src[0];
106                         const Src &src1 = instruction->src[1];
107                         const Src &src2 = instruction->src[2];
108                         const Src &src3 = instruction->src[3];
109                         const Src &src4 = instruction->src[4];
110
111                         bool predicate = instruction->predicate;
112                         Control control = instruction->control;
113                         bool pp = dst.partialPrecision;
114                         bool project = instruction->project;
115                         bool bias = instruction->bias;
116
117                         Vector4f d;
118                         Vector4f s0;
119                         Vector4f s1;
120                         Vector4f s2;
121                         Vector4f s3;
122                         Vector4f s4;
123
124                         if(opcode == Shader::OPCODE_TEXKILL)   // Takes destination as input
125                         {
126                                 if(dst.type == Shader::PARAMETER_TEXTURE)
127                                 {
128                                         d.x = v[2 + dst.index].x;
129                                         d.y = v[2 + dst.index].y;
130                                         d.z = v[2 + dst.index].z;
131                                         d.w = v[2 + dst.index].w;
132                                 }
133                                 else
134                                 {
135                                         d = r[dst.index];
136                                 }
137                         }
138
139                         if(src0.type != Shader::PARAMETER_VOID) s0 = fetchRegisterF(src0);
140                         if(src1.type != Shader::PARAMETER_VOID) s1 = fetchRegisterF(src1);
141                         if(src2.type != Shader::PARAMETER_VOID) s2 = fetchRegisterF(src2);
142                         if(src3.type != Shader::PARAMETER_VOID) s3 = fetchRegisterF(src3);
143                         if(src4.type != Shader::PARAMETER_VOID) s4 = fetchRegisterF(src4);
144
145                         switch(opcode)
146                         {
147                         case Shader::OPCODE_PS_2_0:                                                    break;
148                         case Shader::OPCODE_PS_2_x:                                                    break;
149                         case Shader::OPCODE_PS_3_0:                                                    break;
150                         case Shader::OPCODE_DEF:                                                       break;
151                         case Shader::OPCODE_DCL:                                                       break;
152                         case Shader::OPCODE_NOP:                                                       break;
153                         case Shader::OPCODE_MOV:        mov(d, s0);                                    break;
154                         case Shader::OPCODE_NEG:        neg(d, s0);                                    break;
155                         case Shader::OPCODE_INEG:       ineg(d, s0);                                   break;
156                         case Shader::OPCODE_F2B:        f2b(d, s0);                                    break;
157                         case Shader::OPCODE_B2F:        b2f(d, s0);                                    break;
158                         case Shader::OPCODE_F2I:        f2i(d, s0);                                    break;
159                         case Shader::OPCODE_I2F:        i2f(d, s0);                                    break;
160                         case Shader::OPCODE_F2U:        f2u(d, s0);                                    break;
161                         case Shader::OPCODE_U2F:        u2f(d, s0);                                    break;
162                         case Shader::OPCODE_I2B:        i2b(d, s0);                                    break;
163                         case Shader::OPCODE_B2I:        b2i(d, s0);                                    break;
164                         case Shader::OPCODE_ADD:        add(d, s0, s1);                                break;
165                         case Shader::OPCODE_IADD:       iadd(d, s0, s1);                               break;
166                         case Shader::OPCODE_SUB:        sub(d, s0, s1);                                break;
167                         case Shader::OPCODE_ISUB:       isub(d, s0, s1);                               break;
168                         case Shader::OPCODE_MUL:        mul(d, s0, s1);                                break;
169                         case Shader::OPCODE_IMUL:       imul(d, s0, s1);                               break;
170                         case Shader::OPCODE_MAD:        mad(d, s0, s1, s2);                            break;
171                         case Shader::OPCODE_IMAD:       imad(d, s0, s1, s2);                           break;
172                         case Shader::OPCODE_DP1:        dp1(d, s0, s1);                                break;
173                         case Shader::OPCODE_DP2:        dp2(d, s0, s1);                                break;
174                         case Shader::OPCODE_DP2ADD:     dp2add(d, s0, s1, s2);                         break;
175                         case Shader::OPCODE_DP3:        dp3(d, s0, s1);                                break;
176                         case Shader::OPCODE_DP4:        dp4(d, s0, s1);                                break;
177                         case Shader::OPCODE_DET2:       det2(d, s0, s1);                               break;
178                         case Shader::OPCODE_DET3:       det3(d, s0, s1, s2);                           break;
179                         case Shader::OPCODE_DET4:       det4(d, s0, s1, s2, s3);                       break;
180                         case Shader::OPCODE_CMP0:       cmp0(d, s0, s1, s2);                           break;
181                         case Shader::OPCODE_ICMP:       icmp(d, s0, s1, control);                      break;
182                         case Shader::OPCODE_UCMP:       ucmp(d, s0, s1, control);                      break;
183                         case Shader::OPCODE_SELECT:     select(d, s0, s1, s2);                         break;
184                         case Shader::OPCODE_EXTRACT:    extract(d.x, s0, s1.x);                        break;
185                         case Shader::OPCODE_INSERT:     insert(d, s0, s1.x, s2.x);                     break;
186                         case Shader::OPCODE_FRC:        frc(d, s0);                                    break;
187                         case Shader::OPCODE_TRUNC:      trunc(d, s0);                                  break;
188                         case Shader::OPCODE_FLOOR:      floor(d, s0);                                  break;
189                         case Shader::OPCODE_ROUND:      round(d, s0);                                  break;
190                         case Shader::OPCODE_ROUNDEVEN:  roundEven(d, s0);                              break;
191                         case Shader::OPCODE_CEIL:       ceil(d, s0);                                   break;
192                         case Shader::OPCODE_EXP2X:      exp2x(d, s0, pp);                              break;
193                         case Shader::OPCODE_EXP2:       exp2(d, s0, pp);                               break;
194                         case Shader::OPCODE_LOG2X:      log2x(d, s0, pp);                              break;
195                         case Shader::OPCODE_LOG2:       log2(d, s0, pp);                               break;
196                         case Shader::OPCODE_EXP:        exp(d, s0, pp);                                break;
197                         case Shader::OPCODE_LOG:        log(d, s0, pp);                                break;
198                         case Shader::OPCODE_RCPX:       rcpx(d, s0, pp);                               break;
199                         case Shader::OPCODE_DIV:        div(d, s0, s1);                                break;
200                         case Shader::OPCODE_IDIV:       idiv(d, s0, s1);                               break;
201                         case Shader::OPCODE_UDIV:       udiv(d, s0, s1);                               break;
202                         case Shader::OPCODE_MOD:        mod(d, s0, s1);                                break;
203                         case Shader::OPCODE_IMOD:       imod(d, s0, s1);                               break;
204                         case Shader::OPCODE_UMOD:       umod(d, s0, s1);                               break;
205                         case Shader::OPCODE_SHL:        shl(d, s0, s1);                                break;
206                         case Shader::OPCODE_ISHR:       ishr(d, s0, s1);                               break;
207                         case Shader::OPCODE_USHR:       ushr(d, s0, s1);                               break;
208                         case Shader::OPCODE_RSQX:       rsqx(d, s0, pp);                               break;
209                         case Shader::OPCODE_SQRT:       sqrt(d, s0, pp);                               break;
210                         case Shader::OPCODE_RSQ:        rsq(d, s0, pp);                                break;
211                         case Shader::OPCODE_LEN2:       len2(d.x, s0, pp);                             break;
212                         case Shader::OPCODE_LEN3:       len3(d.x, s0, pp);                             break;
213                         case Shader::OPCODE_LEN4:       len4(d.x, s0, pp);                             break;
214                         case Shader::OPCODE_DIST1:      dist1(d.x, s0, s1, pp);                        break;
215                         case Shader::OPCODE_DIST2:      dist2(d.x, s0, s1, pp);                        break;
216                         case Shader::OPCODE_DIST3:      dist3(d.x, s0, s1, pp);                        break;
217                         case Shader::OPCODE_DIST4:      dist4(d.x, s0, s1, pp);                        break;
218                         case Shader::OPCODE_MIN:        min(d, s0, s1);                                break;
219                         case Shader::OPCODE_IMIN:       imin(d, s0, s1);                               break;
220                         case Shader::OPCODE_UMIN:       umin(d, s0, s1);                               break;
221                         case Shader::OPCODE_MAX:        max(d, s0, s1);                                break;
222                         case Shader::OPCODE_IMAX:       imax(d, s0, s1);                               break;
223                         case Shader::OPCODE_UMAX:       umax(d, s0, s1);                               break;
224                         case Shader::OPCODE_LRP:        lrp(d, s0, s1, s2);                            break;
225                         case Shader::OPCODE_STEP:       step(d, s0, s1);                               break;
226                         case Shader::OPCODE_SMOOTH:     smooth(d, s0, s1, s2);                         break;
227                         case Shader::OPCODE_FLOATBITSTOINT:
228                         case Shader::OPCODE_FLOATBITSTOUINT:
229                         case Shader::OPCODE_INTBITSTOFLOAT:
230                         case Shader::OPCODE_UINTBITSTOFLOAT: d = s0;                                   break;
231                         case Shader::OPCODE_POWX:       powx(d, s0, s1, pp);                           break;
232                         case Shader::OPCODE_POW:        pow(d, s0, s1, pp);                            break;
233                         case Shader::OPCODE_SGN:        sgn(d, s0);                                    break;
234                         case Shader::OPCODE_CRS:        crs(d, s0, s1);                                break;
235                         case Shader::OPCODE_FORWARD1:   forward1(d, s0, s1, s2);                       break;
236                         case Shader::OPCODE_FORWARD2:   forward2(d, s0, s1, s2);                       break;
237                         case Shader::OPCODE_FORWARD3:   forward3(d, s0, s1, s2);                       break;
238                         case Shader::OPCODE_FORWARD4:   forward4(d, s0, s1, s2);                       break;
239                         case Shader::OPCODE_REFLECT1:   reflect1(d, s0, s1);                           break;
240                         case Shader::OPCODE_REFLECT2:   reflect2(d, s0, s1);                           break;
241                         case Shader::OPCODE_REFLECT3:   reflect3(d, s0, s1);                           break;
242                         case Shader::OPCODE_REFLECT4:   reflect4(d, s0, s1);                           break;
243                         case Shader::OPCODE_REFRACT1:   refract1(d, s0, s1, s2.x);                     break;
244                         case Shader::OPCODE_REFRACT2:   refract2(d, s0, s1, s2.x);                     break;
245                         case Shader::OPCODE_REFRACT3:   refract3(d, s0, s1, s2.x);                     break;
246                         case Shader::OPCODE_REFRACT4:   refract4(d, s0, s1, s2.x);                     break;
247                         case Shader::OPCODE_NRM2:       nrm2(d, s0, pp);                               break;
248                         case Shader::OPCODE_NRM3:       nrm3(d, s0, pp);                               break;
249                         case Shader::OPCODE_NRM4:       nrm4(d, s0, pp);                               break;
250                         case Shader::OPCODE_ABS:        abs(d, s0);                                    break;
251                         case Shader::OPCODE_SINCOS:     sincos(d, s0, pp);                             break;
252                         case Shader::OPCODE_COS:        cos(d, s0, pp);                                break;
253                         case Shader::OPCODE_SIN:        sin(d, s0, pp);                                break;
254                         case Shader::OPCODE_TAN:        tan(d, s0, pp);                                break;
255                         case Shader::OPCODE_ACOS:       acos(d, s0, pp);                               break;
256                         case Shader::OPCODE_ASIN:       asin(d, s0, pp);                               break;
257                         case Shader::OPCODE_ATAN:       atan(d, s0, pp);                               break;
258                         case Shader::OPCODE_ATAN2:      atan2(d, s0, s1, pp);                          break;
259                         case Shader::OPCODE_COSH:       cosh(d, s0, pp);                               break;
260                         case Shader::OPCODE_SINH:       sinh(d, s0, pp);                               break;
261                         case Shader::OPCODE_TANH:       tanh(d, s0, pp);                               break;
262                         case Shader::OPCODE_ACOSH:      acosh(d, s0, pp);                              break;
263                         case Shader::OPCODE_ASINH:      asinh(d, s0, pp);                              break;
264                         case Shader::OPCODE_ATANH:      atanh(d, s0, pp);                              break;
265                         case Shader::OPCODE_M4X4:       M4X4(d, s0, src1);                             break;
266                         case Shader::OPCODE_M4X3:       M4X3(d, s0, src1);                             break;
267                         case Shader::OPCODE_M3X4:       M3X4(d, s0, src1);                             break;
268                         case Shader::OPCODE_M3X3:       M3X3(d, s0, src1);                             break;
269                         case Shader::OPCODE_M3X2:       M3X2(d, s0, src1);                             break;
270                         case Shader::OPCODE_TEX:        TEXLD(d, s0, src1, project, bias);             break;
271                         case Shader::OPCODE_TEXLDD:     TEXLDD(d, s0, src1, s2, s3, project, bias);    break;
272                         case Shader::OPCODE_TEXLDL:     TEXLDL(d, s0, src1, project, bias);            break;
273                         case Shader::OPCODE_TEXSIZE:    TEXSIZE(d, s0.x, src1);                        break;
274                         case Shader::OPCODE_TEXKILL:    TEXKILL(cMask, d, dst.mask);                   break;
275                         case Shader::OPCODE_TEXOFFSET:  TEXOFFSET(d, s0, src1, s2, s3, project, bias); break;
276                         case Shader::OPCODE_TEXLDLOFFSET: TEXLDL(d, s0, src1, s2, project, bias);      break;
277                         case Shader::OPCODE_TEXELFETCH: TEXELFETCH(d, s0, src1, s2);                   break;
278                         case Shader::OPCODE_TEXELFETCHOFFSET: TEXELFETCH(d, s0, src1, s2, s3);         break;
279                         case Shader::OPCODE_TEXGRAD:    TEXGRAD(d, s0, src1, s2, s3);                  break;
280                         case Shader::OPCODE_TEXGRADOFFSET: TEXGRAD(d, s0, src1, s2, s3, s4);           break;
281                         case Shader::OPCODE_DISCARD:    DISCARD(cMask, instruction);                   break;
282                         case Shader::OPCODE_DFDX:       DFDX(d, s0);                                   break;
283                         case Shader::OPCODE_DFDY:       DFDY(d, s0);                                   break;
284                         case Shader::OPCODE_FWIDTH:     FWIDTH(d, s0);                                 break;
285                         case Shader::OPCODE_BREAK:      BREAK();                                       break;
286                         case Shader::OPCODE_BREAKC:     BREAKC(s0, s1, control);                       break;
287                         case Shader::OPCODE_BREAKP:     BREAKP(src0);                                  break;
288                         case Shader::OPCODE_CONTINUE:   CONTINUE();                                    break;
289                         case Shader::OPCODE_TEST:       TEST();                                        break;
290                         case Shader::OPCODE_CALL:       CALL(dst.label, dst.callSite);                 break;
291                         case Shader::OPCODE_CALLNZ:     CALLNZ(dst.label, dst.callSite, src0);         break;
292                         case Shader::OPCODE_ELSE:       ELSE();                                        break;
293                         case Shader::OPCODE_ENDIF:      ENDIF();                                       break;
294                         case Shader::OPCODE_ENDLOOP:    ENDLOOP();                                     break;
295                         case Shader::OPCODE_ENDREP:     ENDREP();                                      break;
296                         case Shader::OPCODE_ENDWHILE:   ENDWHILE();                                    break;
297                         case Shader::OPCODE_IF:         IF(src0);                                      break;
298                         case Shader::OPCODE_IFC:        IFC(s0, s1, control);                          break;
299                         case Shader::OPCODE_LABEL:      LABEL(dst.index);                              break;
300                         case Shader::OPCODE_LOOP:       LOOP(src1);                                    break;
301                         case Shader::OPCODE_REP:        REP(src0);                                     break;
302                         case Shader::OPCODE_WHILE:      WHILE(src0);                                   break;
303                         case Shader::OPCODE_RET:        RET();                                         break;
304                         case Shader::OPCODE_LEAVE:      LEAVE();                                       break;
305                         case Shader::OPCODE_CMP:        cmp(d, s0, s1, control);                       break;
306                         case Shader::OPCODE_ALL:        all(d.x, s0);                                  break;
307                         case Shader::OPCODE_ANY:        any(d.x, s0);                                  break;
308                         case Shader::OPCODE_NOT:        not(d, s0);                                    break;
309                         case Shader::OPCODE_OR:         or(d, s0, s1);                                 break;
310                         case Shader::OPCODE_XOR:        xor(d, s0, s1);                                break;
311                         case Shader::OPCODE_AND:        and(d, s0, s1);                                break;
312                         case Shader::OPCODE_EQ:         equal(d, s0, s1);                              break;
313                         case Shader::OPCODE_NE:         notEqual(d, s0, s1);                           break;
314                         case Shader::OPCODE_END:                                                       break;
315                         default:
316                                 ASSERT(false);
317                         }
318
319                         if(dst.type != Shader::PARAMETER_VOID && dst.type != Shader::PARAMETER_LABEL && opcode != Shader::OPCODE_TEXKILL && opcode != Shader::OPCODE_NOP)
320                         {
321                                 if(dst.integer)
322                                 {
323                                         switch(opcode)
324                                         {
325                                         case Shader::OPCODE_DIV:
326                                                 if(dst.x) d.x = Trunc(d.x);
327                                                 if(dst.y) d.y = Trunc(d.y);
328                                                 if(dst.z) d.z = Trunc(d.z);
329                                                 if(dst.w) d.w = Trunc(d.w);
330                                                 break;
331                                         default:
332                                                 break;   // No truncation to integer required when arguments are integer
333                                         }
334                                 }
335
336                                 if(dst.saturate)
337                                 {
338                                         if(dst.x) d.x = Max(d.x, Float4(0.0f));
339                                         if(dst.y) d.y = Max(d.y, Float4(0.0f));
340                                         if(dst.z) d.z = Max(d.z, Float4(0.0f));
341                                         if(dst.w) d.w = Max(d.w, Float4(0.0f));
342
343                                         if(dst.x) d.x = Min(d.x, Float4(1.0f));
344                                         if(dst.y) d.y = Min(d.y, Float4(1.0f));
345                                         if(dst.z) d.z = Min(d.z, Float4(1.0f));
346                                         if(dst.w) d.w = Min(d.w, Float4(1.0f));
347                                 }
348
349                                 if(instruction->isPredicated())
350                                 {
351                                         Vector4f pDst;   // FIXME: Rename
352
353                                         switch(dst.type)
354                                         {
355                                         case Shader::PARAMETER_TEMP:
356                                                 if(dst.rel.type == Shader::PARAMETER_VOID)
357                                                 {
358                                                         if(dst.x) pDst.x = r[dst.index].x;
359                                                         if(dst.y) pDst.y = r[dst.index].y;
360                                                         if(dst.z) pDst.z = r[dst.index].z;
361                                                         if(dst.w) pDst.w = r[dst.index].w;
362                                                 }
363                                                 else
364                                                 {
365                                                         Int a = relativeAddress(dst);
366
367                                                         if(dst.x) pDst.x = r[dst.index + a].x;
368                                                         if(dst.y) pDst.y = r[dst.index + a].y;
369                                                         if(dst.z) pDst.z = r[dst.index + a].z;
370                                                         if(dst.w) pDst.w = r[dst.index + a].w;
371                                                 }
372                                                 break;
373                                         case Shader::PARAMETER_COLOROUT:
374                                                 if(dst.rel.type == Shader::PARAMETER_VOID)
375                                                 {
376                                                         if(dst.x) pDst.x = oC[dst.index].x;
377                                                         if(dst.y) pDst.y = oC[dst.index].y;
378                                                         if(dst.z) pDst.z = oC[dst.index].z;
379                                                         if(dst.w) pDst.w = oC[dst.index].w;
380                                                 }
381                                                 else
382                                                 {
383                                                         Int a = relativeAddress(dst) + dst.index;
384
385                                                         if(dst.x) pDst.x = oC[a].x;
386                                                         if(dst.y) pDst.y = oC[a].y;
387                                                         if(dst.z) pDst.z = oC[a].z;
388                                                         if(dst.w) pDst.w = oC[a].w;
389                                                 }
390                                                 break;
391                                         case Shader::PARAMETER_PREDICATE:
392                                                 if(dst.x) pDst.x = p0.x;
393                                                 if(dst.y) pDst.y = p0.y;
394                                                 if(dst.z) pDst.z = p0.z;
395                                                 if(dst.w) pDst.w = p0.w;
396                                                 break;
397                                         case Shader::PARAMETER_DEPTHOUT:
398                                                 pDst.x = oDepth;
399                                                 break;
400                                         default:
401                                                 ASSERT(false);
402                                         }
403
404                                         Int4 enable = enableMask(instruction);
405
406                                         Int4 xEnable = enable;
407                                         Int4 yEnable = enable;
408                                         Int4 zEnable = enable;
409                                         Int4 wEnable = enable;
410
411                                         if(predicate)
412                                         {
413                                                 unsigned char pSwizzle = instruction->predicateSwizzle;
414
415                                                 Float4 xPredicate = p0[(pSwizzle >> 0) & 0x03];
416                                                 Float4 yPredicate = p0[(pSwizzle >> 2) & 0x03];
417                                                 Float4 zPredicate = p0[(pSwizzle >> 4) & 0x03];
418                                                 Float4 wPredicate = p0[(pSwizzle >> 6) & 0x03];
419
420                                                 if(!instruction->predicateNot)
421                                                 {
422                                                         if(dst.x) xEnable = xEnable & As<Int4>(xPredicate);
423                                                         if(dst.y) yEnable = yEnable & As<Int4>(yPredicate);
424                                                         if(dst.z) zEnable = zEnable & As<Int4>(zPredicate);
425                                                         if(dst.w) wEnable = wEnable & As<Int4>(wPredicate);
426                                                 }
427                                                 else
428                                                 {
429                                                         if(dst.x) xEnable = xEnable & ~As<Int4>(xPredicate);
430                                                         if(dst.y) yEnable = yEnable & ~As<Int4>(yPredicate);
431                                                         if(dst.z) zEnable = zEnable & ~As<Int4>(zPredicate);
432                                                         if(dst.w) wEnable = wEnable & ~As<Int4>(wPredicate);
433                                                 }
434                                         }
435
436                                         if(dst.x) d.x = As<Float4>(As<Int4>(d.x) & xEnable);
437                                         if(dst.y) d.y = As<Float4>(As<Int4>(d.y) & yEnable);
438                                         if(dst.z) d.z = As<Float4>(As<Int4>(d.z) & zEnable);
439                                         if(dst.w) d.w = As<Float4>(As<Int4>(d.w) & wEnable);
440
441                                         if(dst.x) d.x = As<Float4>(As<Int4>(d.x) | (As<Int4>(pDst.x) & ~xEnable));
442                                         if(dst.y) d.y = As<Float4>(As<Int4>(d.y) | (As<Int4>(pDst.y) & ~yEnable));
443                                         if(dst.z) d.z = As<Float4>(As<Int4>(d.z) | (As<Int4>(pDst.z) & ~zEnable));
444                                         if(dst.w) d.w = As<Float4>(As<Int4>(d.w) | (As<Int4>(pDst.w) & ~wEnable));
445                                 }
446
447                                 switch(dst.type)
448                                 {
449                                 case Shader::PARAMETER_TEMP:
450                                         if(dst.rel.type == Shader::PARAMETER_VOID)
451                                         {
452                                                 if(dst.x) r[dst.index].x = d.x;
453                                                 if(dst.y) r[dst.index].y = d.y;
454                                                 if(dst.z) r[dst.index].z = d.z;
455                                                 if(dst.w) r[dst.index].w = d.w;
456                                         }
457                                         else
458                                         {
459                                                 Int a = relativeAddress(dst);
460
461                                                 if(dst.x) r[dst.index + a].x = d.x;
462                                                 if(dst.y) r[dst.index + a].y = d.y;
463                                                 if(dst.z) r[dst.index + a].z = d.z;
464                                                 if(dst.w) r[dst.index + a].w = d.w;
465                                         }
466                                         break;
467                                 case Shader::PARAMETER_COLOROUT:
468                                         if(dst.rel.type == Shader::PARAMETER_VOID)
469                                         {
470                                                 if(dst.x) { oC[dst.index].x = d.x; }
471                                                 if(dst.y) { oC[dst.index].y = d.y; }
472                                                 if(dst.z) { oC[dst.index].z = d.z; }
473                                                 if(dst.w) { oC[dst.index].w = d.w; }
474                                         }
475                                         else
476                                         {
477                                                 Int a = relativeAddress(dst) + dst.index;
478
479                                                 if(dst.x) { oC[a].x = d.x; }
480                                                 if(dst.y) { oC[a].y = d.y; }
481                                                 if(dst.z) { oC[a].z = d.z; }
482                                                 if(dst.w) { oC[a].w = d.w; }
483                                         }
484                                         break;
485                                 case Shader::PARAMETER_PREDICATE:
486                                         if(dst.x) p0.x = d.x;
487                                         if(dst.y) p0.y = d.y;
488                                         if(dst.z) p0.z = d.z;
489                                         if(dst.w) p0.w = d.w;
490                                         break;
491                                 case Shader::PARAMETER_DEPTHOUT:
492                                         oDepth = d.x;
493                                         break;
494                                 default:
495                                         ASSERT(false);
496                                 }
497                         }
498                 }
499
500                 if(currentLabel != -1)
501                 {
502                         Nucleus::setInsertBlock(returnBlock);
503                 }
504
505                 for(int i = 0; i < RENDERTARGETS; i++)
506                 {
507                         c[i] = oC[i];
508                 }
509         }
510
511         Bool PixelProgram::alphaTest(Int cMask[4])
512         {
513                 clampColor(c);
514
515                 if(!state.alphaTestActive())
516                 {
517                         return true;
518                 }
519
520                 Int aMask;
521
522                 if(state.transparencyAntialiasing == TRANSPARENCY_NONE)
523                 {
524                         Short4 alpha = RoundShort4(c[0].w * Float4(0x1000));
525
526                         PixelRoutine::alphaTest(aMask, alpha);
527
528                         for(unsigned int q = 0; q < state.multiSample; q++)
529                         {
530                                 cMask[q] &= aMask;
531                         }
532                 }
533                 else if(state.transparencyAntialiasing == TRANSPARENCY_ALPHA_TO_COVERAGE)
534                 {
535                         alphaToCoverage(cMask, c[0].w);
536                 }
537                 else ASSERT(false);
538
539                 Int pass = cMask[0];
540
541                 for(unsigned int q = 1; q < state.multiSample; q++)
542                 {
543                         pass = pass | cMask[q];
544                 }
545
546                 return pass != 0x0;
547         }
548
549         void PixelProgram::rasterOperation(Float4 &fog, Pointer<Byte> cBuffer[4], Int &x, Int sMask[4], Int zMask[4], Int cMask[4])
550         {
551                 for(int index = 0; index < RENDERTARGETS; index++)
552                 {
553                         if(!state.colorWriteActive(index))
554                         {
555                                 continue;
556                         }
557
558                         if(!postBlendSRGB && state.writeSRGB)
559                         {
560                                 c[index].x = linearToSRGB(c[index].x);
561                                 c[index].y = linearToSRGB(c[index].y);
562                                 c[index].z = linearToSRGB(c[index].z);
563                         }
564
565                         if(index == 0)
566                         {
567                                 fogBlend(c[index], fog);
568                         }
569
570                         switch(state.targetFormat[index])
571                         {
572                         case FORMAT_R5G6B5:
573                         case FORMAT_X8R8G8B8:
574                         case FORMAT_X8B8G8R8:
575                         case FORMAT_A8R8G8B8:
576                         case FORMAT_A8B8G8R8:
577                         case FORMAT_A8:
578                         case FORMAT_G16R16:
579                         case FORMAT_A16B16G16R16:
580                                 for(unsigned int q = 0; q < state.multiSample; q++)
581                                 {
582                                         Pointer<Byte> buffer = cBuffer[index] + q * *Pointer<Int>(data + OFFSET(DrawData, colorSliceB[index]));
583                                         Vector4s color;
584
585                                         color.x = convertFixed16(c[index].x, false);
586                                         color.y = convertFixed16(c[index].y, false);
587                                         color.z = convertFixed16(c[index].z, false);
588                                         color.w = convertFixed16(c[index].w, false);
589
590                                         if(state.multiSampleMask & (1 << q))
591                                         {
592                                                 alphaBlend(index, buffer, color, x);
593                                                 logicOperation(index, buffer, color, x);
594                                                 writeColor(index, buffer, x, color, sMask[q], zMask[q], cMask[q]);
595                                         }
596                                 }
597                                 break;
598                         case FORMAT_R32F:
599                         case FORMAT_G32R32F:
600                         case FORMAT_A32B32G32R32F:
601                         case FORMAT_R32I:
602                         case FORMAT_G32R32I:
603                         case FORMAT_A32B32G32R32I:
604                         case FORMAT_R32UI:
605                         case FORMAT_G32R32UI:
606                         case FORMAT_A32B32G32R32UI:
607                                 for(unsigned int q = 0; q < state.multiSample; q++)
608                                 {
609                                         Pointer<Byte> buffer = cBuffer[index] + q * *Pointer<Int>(data + OFFSET(DrawData, colorSliceB[index]));
610                                         Vector4f color = c[index];
611
612                                         if(state.multiSampleMask & (1 << q))
613                                         {
614                                                 alphaBlend(index, buffer, color, x);
615                                                 writeColor(index, buffer, x, color, sMask[q], zMask[q], cMask[q]);
616                                         }
617                                 }
618                                 break;
619                         default:
620                                 ASSERT(false);
621                         }
622                 }
623         }
624
625         void PixelProgram::sampleTexture(Vector4f &c, const Src &sampler, Float4 &u, Float4 &v, Float4 &w, Float4 &q, Vector4f &dsx, Vector4f &dsy, bool project, bool bias, bool gradients, bool lodProvided)
626         {
627                 if(sampler.type == Shader::PARAMETER_SAMPLER && sampler.rel.type == Shader::PARAMETER_VOID)
628                 {
629                         sampleTexture(c, sampler.index, u, v, w, q, dsx, dsy, project, bias, gradients, lodProvided);
630                 }
631                 else
632                 {
633                         Int index = As<Int>(Float(fetchRegisterF(sampler).x.x));
634
635                         for(int i = 0; i < TEXTURE_IMAGE_UNITS; i++)
636                         {
637                                 if(shader->usesSampler(i))
638                                 {
639                                         If(index == i)
640                                         {
641                                                 sampleTexture(c, i, u, v, w, q, dsx, dsy, project, bias, gradients, lodProvided);
642                                                 // FIXME: When the sampler states are the same, we could use one sampler and just index the texture
643                                         }
644                                 }
645                         }
646                 }
647         }
648
649         void PixelProgram::sampleTexture(Vector4f &c, int stage, Float4 &u, Float4 &v, Float4 &w, Float4 &q, Vector4f &dsx, Vector4f &dsy, bool project, bool bias, bool gradients, bool lodProvided)
650         {
651 #if PERF_PROFILE
652                 Long texTime = Ticks();
653 #endif
654
655                 Pointer<Byte> texture = data + OFFSET(DrawData, mipmap) + stage * sizeof(Texture);
656
657                 if(!project)
658                 {
659                         sampler[stage]->sampleTexture(texture, c, u, v, w, q, dsx, dsy, bias, gradients, lodProvided);
660                 }
661                 else
662                 {
663                         Float4 rq = reciprocal(q);
664
665                         Float4 u_q = u * rq;
666                         Float4 v_q = v * rq;
667                         Float4 w_q = w * rq;
668
669                         sampler[stage]->sampleTexture(texture, c, u_q, v_q, w_q, q, dsx, dsy, bias, gradients, lodProvided);
670                 }
671
672 #if PERF_PROFILE
673                 cycles[PERF_TEX] += Ticks() - texTime;
674 #endif
675         }
676
677         void PixelProgram::clampColor(Vector4f oC[RENDERTARGETS])
678         {
679                 for(int index = 0; index < RENDERTARGETS; index++)
680                 {
681                         if(!state.colorWriteActive(index) && !(index == 0 && state.alphaTestActive()))
682                         {
683                                 continue;
684                         }
685
686                         switch(state.targetFormat[index])
687                         {
688                         case FORMAT_NULL:
689                                 break;
690                         case FORMAT_R5G6B5:
691                         case FORMAT_A8R8G8B8:
692                         case FORMAT_A8B8G8R8:
693                         case FORMAT_X8R8G8B8:
694                         case FORMAT_X8B8G8R8:
695                         case FORMAT_A8:
696                         case FORMAT_G16R16:
697                         case FORMAT_A16B16G16R16:
698                                 oC[index].x = Max(oC[index].x, Float4(0.0f)); oC[index].x = Min(oC[index].x, Float4(1.0f));
699                                 oC[index].y = Max(oC[index].y, Float4(0.0f)); oC[index].y = Min(oC[index].y, Float4(1.0f));
700                                 oC[index].z = Max(oC[index].z, Float4(0.0f)); oC[index].z = Min(oC[index].z, Float4(1.0f));
701                                 oC[index].w = Max(oC[index].w, Float4(0.0f)); oC[index].w = Min(oC[index].w, Float4(1.0f));
702                                 break;
703                         case FORMAT_R32F:
704                         case FORMAT_G32R32F:
705                         case FORMAT_A32B32G32R32F:
706                         case FORMAT_R32I:
707                         case FORMAT_G32R32I:
708                         case FORMAT_A32B32G32R32I:
709                         case FORMAT_R32UI:
710                         case FORMAT_G32R32UI:
711                         case FORMAT_A32B32G32R32UI:
712                                 break;
713                         default:
714                                 ASSERT(false);
715                         }
716                 }
717         }
718
719         Int4 PixelProgram::enableMask(const Shader::Instruction *instruction)
720         {
721                 Int4 enable = instruction->analysisBranch ? Int4(enableStack[enableIndex]) : Int4(0xFFFFFFFF);
722
723                 if(!whileTest)
724                 {
725                         if(shader->containsBreakInstruction() && instruction->analysisBreak)
726                         {
727                                 enable &= enableBreak;
728                         }
729
730                         if(shader->containsContinueInstruction() && instruction->analysisContinue)
731                         {
732                                 enable &= enableContinue;
733                         }
734
735                         if(shader->containsLeaveInstruction() && instruction->analysisLeave)
736                         {
737                                 enable &= enableLeave;
738                         }
739                 }
740
741                 return enable;
742         }
743
744         Vector4f PixelProgram::fetchRegisterF(const Src &src, unsigned int offset)
745         {
746                 Vector4f reg;
747                 unsigned int i = src.index + offset;
748
749                 switch(src.type)
750                 {
751                 case Shader::PARAMETER_TEMP:
752                         if(src.rel.type == Shader::PARAMETER_VOID)
753                         {
754                                 reg = r[i];
755                         }
756                         else
757                         {
758                                 Int a = relativeAddress(src, src.bufferIndex);
759
760                                 reg = r[i + a];
761                         }
762                         break;
763                 case Shader::PARAMETER_INPUT:
764                         {
765                                 if(src.rel.type == Shader::PARAMETER_VOID)   // Not relative
766                                 {
767                                         reg = v[i];
768                                 }
769                                 else
770                                 {
771                                         Int a = relativeAddress(src, src.bufferIndex);
772
773                                         reg = v[i + a];
774                                 }
775                         }
776                         break;
777                 case Shader::PARAMETER_CONST:
778                         reg = readConstant(src, offset);
779                         break;
780                 case Shader::PARAMETER_TEXTURE:
781                         reg = v[2 + i];
782                         break;
783                 case Shader::PARAMETER_MISCTYPE:
784                         if(src.index == 0) reg = vPos;
785                         if(src.index == 1) reg = vFace;
786                         break;
787                 case Shader::PARAMETER_SAMPLER:
788                         if(src.rel.type == Shader::PARAMETER_VOID)
789                         {
790                                 reg.x = As<Float4>(Int4(i));
791                         }
792                         else if(src.rel.type == Shader::PARAMETER_TEMP)
793                         {
794                                 reg.x = As<Float4>(Int4(i) + As<Int4>(r[src.rel.index].x));
795                         }
796                         return reg;
797                 case Shader::PARAMETER_PREDICATE:   return reg; // Dummy
798                 case Shader::PARAMETER_VOID:        return reg; // Dummy
799                 case Shader::PARAMETER_FLOAT4LITERAL:
800                         reg.x = Float4(src.value[0]);
801                         reg.y = Float4(src.value[1]);
802                         reg.z = Float4(src.value[2]);
803                         reg.w = Float4(src.value[3]);
804                         break;
805                 case Shader::PARAMETER_CONSTINT:    return reg; // Dummy
806                 case Shader::PARAMETER_CONSTBOOL:   return reg; // Dummy
807                 case Shader::PARAMETER_LOOP:        return reg; // Dummy
808                 case Shader::PARAMETER_COLOROUT:
809                         if(src.rel.type == Shader::PARAMETER_VOID)   // Not relative
810                         {
811                                 reg = oC[i];
812                         }
813                         else
814                         {
815                                 Int a = relativeAddress(src, src.bufferIndex);
816
817                                 reg = oC[i + a];
818                         }
819                         break;
820                 case Shader::PARAMETER_DEPTHOUT:
821                         reg.x = oDepth;
822                         break;
823                 default:
824                         ASSERT(false);
825                 }
826
827                 const Float4 &x = reg[(src.swizzle >> 0) & 0x3];
828                 const Float4 &y = reg[(src.swizzle >> 2) & 0x3];
829                 const Float4 &z = reg[(src.swizzle >> 4) & 0x3];
830                 const Float4 &w = reg[(src.swizzle >> 6) & 0x3];
831
832                 Vector4f mod;
833
834                 switch(src.modifier)
835                 {
836                 case Shader::MODIFIER_NONE:
837                         mod.x = x;
838                         mod.y = y;
839                         mod.z = z;
840                         mod.w = w;
841                         break;
842                 case Shader::MODIFIER_NEGATE:
843                         mod.x = -x;
844                         mod.y = -y;
845                         mod.z = -z;
846                         mod.w = -w;
847                         break;
848                 case Shader::MODIFIER_ABS:
849                         mod.x = Abs(x);
850                         mod.y = Abs(y);
851                         mod.z = Abs(z);
852                         mod.w = Abs(w);
853                         break;
854                 case Shader::MODIFIER_ABS_NEGATE:
855                         mod.x = -Abs(x);
856                         mod.y = -Abs(y);
857                         mod.z = -Abs(z);
858                         mod.w = -Abs(w);
859                         break;
860                 case Shader::MODIFIER_NOT:
861                         mod.x = As<Float4>(As<Int4>(x) ^ Int4(0xFFFFFFFF));
862                         mod.y = As<Float4>(As<Int4>(y) ^ Int4(0xFFFFFFFF));
863                         mod.z = As<Float4>(As<Int4>(z) ^ Int4(0xFFFFFFFF));
864                         mod.w = As<Float4>(As<Int4>(w) ^ Int4(0xFFFFFFFF));
865                         break;
866                 default:
867                         ASSERT(false);
868                 }
869
870                 return mod;
871         }
872
873         RValue<Pointer<Byte>> PixelProgram::uniformAddress(int bufferIndex, unsigned int index)
874         {
875                 if(bufferIndex == -1)
876                 {
877                         return data + OFFSET(DrawData, ps.c[index]);
878                 }
879                 else
880                 {
881                         return *Pointer<Pointer<Byte>>(data + OFFSET(DrawData, ps.u[bufferIndex])) + index;
882                 }
883         }
884
885         RValue<Pointer<Byte>> PixelProgram::uniformAddress(int bufferIndex, unsigned int index, Int& offset)
886         {
887                 return uniformAddress(bufferIndex, index) + offset * sizeof(float4);
888         }
889
890         Vector4f PixelProgram::readConstant(const Src &src, unsigned int offset)
891         {
892                 Vector4f c;
893                 unsigned int i = src.index + offset;
894
895                 if(src.rel.type == Shader::PARAMETER_VOID)   // Not relative
896                 {
897                         c.x = c.y = c.z = c.w = *Pointer<Float4>(uniformAddress(src.bufferIndex, i));
898
899                         c.x = c.x.xxxx;
900                         c.y = c.y.yyyy;
901                         c.z = c.z.zzzz;
902                         c.w = c.w.wwww;
903
904                         if(shader->containsDefineInstruction())   // Constant may be known at compile time
905                         {
906                                 for(size_t j = 0; j < shader->getLength(); j++)
907                                 {
908                                         const Shader::Instruction &instruction = *shader->getInstruction(j);
909
910                                         if(instruction.opcode == Shader::OPCODE_DEF)
911                                         {
912                                                 if(instruction.dst.index == i)
913                                                 {
914                                                         c.x = Float4(instruction.src[0].value[0]);
915                                                         c.y = Float4(instruction.src[0].value[1]);
916                                                         c.z = Float4(instruction.src[0].value[2]);
917                                                         c.w = Float4(instruction.src[0].value[3]);
918
919                                                         break;
920                                                 }
921                                         }
922                                 }
923                         }
924                 }
925                 else if(src.rel.type == Shader::PARAMETER_LOOP)
926                 {
927                         Int loopCounter = aL[loopDepth];
928
929                         c.x = c.y = c.z = c.w = *Pointer<Float4>(uniformAddress(src.bufferIndex, i, loopCounter));
930
931                         c.x = c.x.xxxx;
932                         c.y = c.y.yyyy;
933                         c.z = c.z.zzzz;
934                         c.w = c.w.wwww;
935                 }
936                 else
937                 {
938                         Int a = relativeAddress(src, src.bufferIndex);
939
940                         c.x = c.y = c.z = c.w = *Pointer<Float4>(uniformAddress(src.bufferIndex, i, a));
941
942                         c.x = c.x.xxxx;
943                         c.y = c.y.yyyy;
944                         c.z = c.z.zzzz;
945                         c.w = c.w.wwww;
946                 }
947
948                 return c;
949         }
950
951         Int PixelProgram::relativeAddress(const Shader::Parameter &var, int bufferIndex)
952         {
953                 ASSERT(var.rel.deterministic);
954
955                 if(var.rel.type == Shader::PARAMETER_TEMP)
956                 {
957                         return As<Int>(Extract(r[var.rel.index].x, 0)) * var.rel.scale;
958                 }
959                 else if(var.rel.type == Shader::PARAMETER_INPUT)
960                 {
961                         return As<Int>(Extract(v[var.rel.index].x, 0)) * var.rel.scale;
962                 }
963                 else if(var.rel.type == Shader::PARAMETER_OUTPUT)
964                 {
965                         return As<Int>(Extract(oC[var.rel.index].x, 0)) * var.rel.scale;
966                 }
967                 else if(var.rel.type == Shader::PARAMETER_CONST)
968                 {
969                         return *Pointer<Int>(uniformAddress(bufferIndex, var.rel.index)) * var.rel.scale;
970                 }
971                 else if(var.rel.type == Shader::PARAMETER_LOOP)
972                 {
973                         return aL[loopDepth];
974                 }
975                 else ASSERT(false);
976
977                 return 0;
978         }
979
980         Float4 PixelProgram::linearToSRGB(const Float4 &x)   // Approximates x^(1.0/2.2)
981         {
982                 Float4 sqrtx = Rcp_pp(RcpSqrt_pp(x));
983                 Float4 sRGB = sqrtx * Float4(1.14f) - x * Float4(0.14f);
984
985                 return Min(Max(sRGB, Float4(0.0f)), Float4(1.0f));
986         }
987
988         void PixelProgram::M3X2(Vector4f &dst, Vector4f &src0, const Src &src1)
989         {
990                 Vector4f row0 = fetchRegisterF(src1, 0);
991                 Vector4f row1 = fetchRegisterF(src1, 1);
992
993                 dst.x = dot3(src0, row0);
994                 dst.y = dot3(src0, row1);
995         }
996
997         void PixelProgram::M3X3(Vector4f &dst, Vector4f &src0, const Src &src1)
998         {
999                 Vector4f row0 = fetchRegisterF(src1, 0);
1000                 Vector4f row1 = fetchRegisterF(src1, 1);
1001                 Vector4f row2 = fetchRegisterF(src1, 2);
1002
1003                 dst.x = dot3(src0, row0);
1004                 dst.y = dot3(src0, row1);
1005                 dst.z = dot3(src0, row2);
1006         }
1007
1008         void PixelProgram::M3X4(Vector4f &dst, Vector4f &src0, const Src &src1)
1009         {
1010                 Vector4f row0 = fetchRegisterF(src1, 0);
1011                 Vector4f row1 = fetchRegisterF(src1, 1);
1012                 Vector4f row2 = fetchRegisterF(src1, 2);
1013                 Vector4f row3 = fetchRegisterF(src1, 3);
1014
1015                 dst.x = dot3(src0, row0);
1016                 dst.y = dot3(src0, row1);
1017                 dst.z = dot3(src0, row2);
1018                 dst.w = dot3(src0, row3);
1019         }
1020
1021         void PixelProgram::M4X3(Vector4f &dst, Vector4f &src0, const Src &src1)
1022         {
1023                 Vector4f row0 = fetchRegisterF(src1, 0);
1024                 Vector4f row1 = fetchRegisterF(src1, 1);
1025                 Vector4f row2 = fetchRegisterF(src1, 2);
1026
1027                 dst.x = dot4(src0, row0);
1028                 dst.y = dot4(src0, row1);
1029                 dst.z = dot4(src0, row2);
1030         }
1031
1032         void PixelProgram::M4X4(Vector4f &dst, Vector4f &src0, const Src &src1)
1033         {
1034                 Vector4f row0 = fetchRegisterF(src1, 0);
1035                 Vector4f row1 = fetchRegisterF(src1, 1);
1036                 Vector4f row2 = fetchRegisterF(src1, 2);
1037                 Vector4f row3 = fetchRegisterF(src1, 3);
1038
1039                 dst.x = dot4(src0, row0);
1040                 dst.y = dot4(src0, row1);
1041                 dst.z = dot4(src0, row2);
1042                 dst.w = dot4(src0, row3);
1043         }
1044
1045         void PixelProgram::TEXLD(Vector4f &dst, Vector4f &src0, const Src &src1, bool project, bool bias)
1046         {
1047                 Vector4f tmp;
1048                 sampleTexture(tmp, src1, src0.x, src0.y, src0.z, src0.w, src0, src0, project, bias);
1049
1050                 dst.x = tmp[(src1.swizzle >> 0) & 0x3];
1051                 dst.y = tmp[(src1.swizzle >> 2) & 0x3];
1052                 dst.z = tmp[(src1.swizzle >> 4) & 0x3];
1053                 dst.w = tmp[(src1.swizzle >> 6) & 0x3];
1054         }
1055
1056         void PixelProgram::TEXOFFSET(Vector4f &dst, Vector4f &src0, const Src& src1, Vector4f &src2, Vector4f &src3, bool project, bool bias)
1057         {
1058                 UNIMPLEMENTED();
1059         }
1060
1061         void PixelProgram::TEXLDL(Vector4f &dst, Vector4f &src0, const Src &src1, Vector4f &offset, bool project, bool bias)
1062         {
1063                 UNIMPLEMENTED();
1064         }
1065
1066         void PixelProgram::TEXELFETCH(Vector4f &dst, Vector4f &src0, const Src& src1, Vector4f &src2)
1067         {
1068                 UNIMPLEMENTED();
1069         }
1070
1071         void PixelProgram::TEXELFETCH(Vector4f &dst, Vector4f &src0, const Src& src1, Vector4f &src2, Vector4f &offset)
1072         {
1073                 UNIMPLEMENTED();
1074         }
1075
1076         void PixelProgram::TEXGRAD(Vector4f &dst, Vector4f &src0, const Src& src1, Vector4f &src2, Vector4f &src3)
1077         {
1078                 UNIMPLEMENTED();
1079         }
1080
1081         void PixelProgram::TEXGRAD(Vector4f &dst, Vector4f &src0, const Src& src1, Vector4f &src2, Vector4f &src3, Vector4f &offset)
1082         {
1083                 UNIMPLEMENTED();
1084         }
1085
1086         void PixelProgram::TEXLDD(Vector4f &dst, Vector4f &src0, const Src &src1, Vector4f &src2, Vector4f &src3, bool project, bool bias)
1087         {
1088                 Vector4f tmp;
1089                 sampleTexture(tmp, src1, src0.x, src0.y, src0.z, src0.w, src2, src3, project, bias, true);
1090
1091                 dst.x = tmp[(src1.swizzle >> 0) & 0x3];
1092                 dst.y = tmp[(src1.swizzle >> 2) & 0x3];
1093                 dst.z = tmp[(src1.swizzle >> 4) & 0x3];
1094                 dst.w = tmp[(src1.swizzle >> 6) & 0x3];
1095         }
1096
1097         void PixelProgram::TEXLDL(Vector4f &dst, Vector4f &src0, const Src &src1, bool project, bool bias)
1098         {
1099                 Vector4f tmp;
1100                 sampleTexture(tmp, src1, src0.x, src0.y, src0.z, src0.w, src0, src0, project, bias, false, true);
1101
1102                 dst.x = tmp[(src1.swizzle >> 0) & 0x3];
1103                 dst.y = tmp[(src1.swizzle >> 2) & 0x3];
1104                 dst.z = tmp[(src1.swizzle >> 4) & 0x3];
1105                 dst.w = tmp[(src1.swizzle >> 6) & 0x3];
1106         }
1107
1108         void PixelProgram::TEXSIZE(Vector4f &dst, Float4 &lod, const Src &src1)
1109         {
1110                 Pointer<Byte> textureMipmap = data + OFFSET(DrawData, mipmap) + src1.index * sizeof(Texture) + OFFSET(Texture, mipmap);
1111                 for(int i = 0; i < 4; ++i)
1112                 {
1113                         Pointer<Byte> mipmap = textureMipmap + (As<Int>(Extract(lod, i)) + Int(1)) * sizeof(Mipmap);
1114                         dst.x = Insert(dst.x, As<Float>(Int(*Pointer<Short>(mipmap + OFFSET(Mipmap, width)))), i);
1115                         dst.y = Insert(dst.y, As<Float>(Int(*Pointer<Short>(mipmap + OFFSET(Mipmap, height)))), i);
1116                         dst.z = Insert(dst.z, As<Float>(Int(*Pointer<Short>(mipmap + OFFSET(Mipmap, depth)))), i);
1117                 }
1118         }
1119
1120         void PixelProgram::TEXKILL(Int cMask[4], Vector4f &src, unsigned char mask)
1121         {
1122                 Int kill = -1;
1123
1124                 if(mask & 0x1) kill &= SignMask(CmpNLT(src.x, Float4(0.0f)));
1125                 if(mask & 0x2) kill &= SignMask(CmpNLT(src.y, Float4(0.0f)));
1126                 if(mask & 0x4) kill &= SignMask(CmpNLT(src.z, Float4(0.0f)));
1127                 if(mask & 0x8) kill &= SignMask(CmpNLT(src.w, Float4(0.0f)));
1128
1129                 // FIXME: Dynamic branching affects TEXKILL?
1130                 //      if(shader->containsDynamicBranching())
1131                 //      {
1132                 //              kill = ~SignMask(enableMask());
1133                 //      }
1134
1135                 for(unsigned int q = 0; q < state.multiSample; q++)
1136                 {
1137                         cMask[q] &= kill;
1138                 }
1139
1140                 // FIXME: Branch to end of shader if all killed?
1141         }
1142
1143         void PixelProgram::DISCARD(Int cMask[4], const Shader::Instruction *instruction)
1144         {
1145                 Int kill = 0;
1146
1147                 if(shader->containsDynamicBranching())
1148                 {
1149                         kill = ~SignMask(enableMask(instruction));
1150                 }
1151
1152                 for(unsigned int q = 0; q < state.multiSample; q++)
1153                 {
1154                         cMask[q] &= kill;
1155                 }
1156
1157                 // FIXME: Branch to end of shader if all killed?
1158         }
1159
1160         void PixelProgram::DFDX(Vector4f &dst, Vector4f &src)
1161         {
1162                 dst.x = src.x.yyww - src.x.xxzz;
1163                 dst.y = src.y.yyww - src.y.xxzz;
1164                 dst.z = src.z.yyww - src.z.xxzz;
1165                 dst.w = src.w.yyww - src.w.xxzz;
1166         }
1167
1168         void PixelProgram::DFDY(Vector4f &dst, Vector4f &src)
1169         {
1170                 dst.x = src.x.zwzw - src.x.xyxy;
1171                 dst.y = src.y.zwzw - src.y.xyxy;
1172                 dst.z = src.z.zwzw - src.z.xyxy;
1173                 dst.w = src.w.zwzw - src.w.xyxy;
1174         }
1175
1176         void PixelProgram::FWIDTH(Vector4f &dst, Vector4f &src)
1177         {
1178                 // abs(dFdx(src)) + abs(dFdy(src));
1179                 dst.x = Abs(src.x.yyww - src.x.xxzz) + Abs(src.x.zwzw - src.x.xyxy);
1180                 dst.y = Abs(src.y.yyww - src.y.xxzz) + Abs(src.y.zwzw - src.y.xyxy);
1181                 dst.z = Abs(src.z.yyww - src.z.xxzz) + Abs(src.z.zwzw - src.z.xyxy);
1182                 dst.w = Abs(src.w.yyww - src.w.xxzz) + Abs(src.w.zwzw - src.w.xyxy);
1183         }
1184
1185         void PixelProgram::BREAK()
1186         {
1187                 llvm::BasicBlock *deadBlock = Nucleus::createBasicBlock();
1188                 llvm::BasicBlock *endBlock = loopRepEndBlock[loopRepDepth - 1];
1189
1190                 if(breakDepth == 0)
1191                 {
1192                         enableIndex = enableIndex - breakDepth;
1193                         Nucleus::createBr(endBlock);
1194                 }
1195                 else
1196                 {
1197                         enableBreak = enableBreak & ~enableStack[enableIndex];
1198                         Bool allBreak = SignMask(enableBreak) == 0x0;
1199
1200                         enableIndex = enableIndex - breakDepth;
1201                         branch(allBreak, endBlock, deadBlock);
1202                 }
1203
1204                 Nucleus::setInsertBlock(deadBlock);
1205                 enableIndex = enableIndex + breakDepth;
1206         }
1207
1208         void PixelProgram::BREAKC(Vector4f &src0, Vector4f &src1, Control control)
1209         {
1210                 Int4 condition;
1211
1212                 switch(control)
1213                 {
1214                 case Shader::CONTROL_GT: condition = CmpNLE(src0.x, src1.x); break;
1215                 case Shader::CONTROL_EQ: condition = CmpEQ(src0.x, src1.x);  break;
1216                 case Shader::CONTROL_GE: condition = CmpNLT(src0.x, src1.x); break;
1217                 case Shader::CONTROL_LT: condition = CmpLT(src0.x, src1.x);  break;
1218                 case Shader::CONTROL_NE: condition = CmpNEQ(src0.x, src1.x); break;
1219                 case Shader::CONTROL_LE: condition = CmpLE(src0.x, src1.x);  break;
1220                 default:
1221                         ASSERT(false);
1222                 }
1223
1224                 BREAK(condition);
1225         }
1226
1227         void PixelProgram::BREAKP(const Src &predicateRegister)   // FIXME: Factor out parts common with BREAKC
1228         {
1229                 Int4 condition = As<Int4>(p0[predicateRegister.swizzle & 0x3]);
1230
1231                 if(predicateRegister.modifier == Shader::MODIFIER_NOT)
1232                 {
1233                         condition = ~condition;
1234                 }
1235
1236                 BREAK(condition);
1237         }
1238
1239         void PixelProgram::BREAK(Int4 &condition)
1240         {
1241                 condition &= enableStack[enableIndex];
1242
1243                 llvm::BasicBlock *continueBlock = Nucleus::createBasicBlock();
1244                 llvm::BasicBlock *endBlock = loopRepEndBlock[loopRepDepth - 1];
1245
1246                 enableBreak = enableBreak & ~condition;
1247                 Bool allBreak = SignMask(enableBreak) == 0x0;
1248
1249                 enableIndex = enableIndex - breakDepth;
1250                 branch(allBreak, endBlock, continueBlock);
1251
1252                 Nucleus::setInsertBlock(continueBlock);
1253                 enableIndex = enableIndex + breakDepth;
1254         }
1255
1256         void PixelProgram::CONTINUE()
1257         {
1258                 enableContinue = enableContinue & ~enableStack[enableIndex];
1259         }
1260
1261         void PixelProgram::TEST()
1262         {
1263                 whileTest = true;
1264         }
1265
1266         void PixelProgram::CALL(int labelIndex, int callSiteIndex)
1267         {
1268                 if(!labelBlock[labelIndex])
1269                 {
1270                         labelBlock[labelIndex] = Nucleus::createBasicBlock();
1271                 }
1272
1273                 if(callRetBlock[labelIndex].size() > 1)
1274                 {
1275                         callStack[stackIndex++] = UInt(callSiteIndex);
1276                 }
1277
1278                 Int4 restoreLeave = enableLeave;
1279
1280                 Nucleus::createBr(labelBlock[labelIndex]);
1281                 Nucleus::setInsertBlock(callRetBlock[labelIndex][callSiteIndex]);
1282
1283                 enableLeave = restoreLeave;
1284         }
1285
1286         void PixelProgram::CALLNZ(int labelIndex, int callSiteIndex, const Src &src)
1287         {
1288                 if(src.type == Shader::PARAMETER_CONSTBOOL)
1289                 {
1290                         CALLNZb(labelIndex, callSiteIndex, src);
1291                 }
1292                 else if(src.type == Shader::PARAMETER_PREDICATE)
1293                 {
1294                         CALLNZp(labelIndex, callSiteIndex, src);
1295                 }
1296                 else ASSERT(false);
1297         }
1298
1299         void PixelProgram::CALLNZb(int labelIndex, int callSiteIndex, const Src &boolRegister)
1300         {
1301                 Bool condition = (*Pointer<Byte>(data + OFFSET(DrawData, ps.b[boolRegister.index])) != Byte(0));   // FIXME
1302
1303                 if(boolRegister.modifier == Shader::MODIFIER_NOT)
1304                 {
1305                         condition = !condition;
1306                 }
1307
1308                 if(!labelBlock[labelIndex])
1309                 {
1310                         labelBlock[labelIndex] = Nucleus::createBasicBlock();
1311                 }
1312
1313                 if(callRetBlock[labelIndex].size() > 1)
1314                 {
1315                         callStack[stackIndex++] = UInt(callSiteIndex);
1316                 }
1317
1318                 Int4 restoreLeave = enableLeave;
1319
1320                 branch(condition, labelBlock[labelIndex], callRetBlock[labelIndex][callSiteIndex]);
1321                 Nucleus::setInsertBlock(callRetBlock[labelIndex][callSiteIndex]);
1322
1323                 enableLeave = restoreLeave;
1324         }
1325
1326         void PixelProgram::CALLNZp(int labelIndex, int callSiteIndex, const Src &predicateRegister)
1327         {
1328                 Int4 condition = As<Int4>(p0[predicateRegister.swizzle & 0x3]);
1329
1330                 if(predicateRegister.modifier == Shader::MODIFIER_NOT)
1331                 {
1332                         condition = ~condition;
1333                 }
1334
1335                 condition &= enableStack[enableIndex];
1336
1337                 if(!labelBlock[labelIndex])
1338                 {
1339                         labelBlock[labelIndex] = Nucleus::createBasicBlock();
1340                 }
1341
1342                 if(callRetBlock[labelIndex].size() > 1)
1343                 {
1344                         callStack[stackIndex++] = UInt(callSiteIndex);
1345                 }
1346
1347                 enableIndex++;
1348                 enableStack[enableIndex] = condition;
1349                 Int4 restoreLeave = enableLeave;
1350
1351                 Bool notAllFalse = SignMask(condition) != 0;
1352                 branch(notAllFalse, labelBlock[labelIndex], callRetBlock[labelIndex][callSiteIndex]);
1353                 Nucleus::setInsertBlock(callRetBlock[labelIndex][callSiteIndex]);
1354
1355                 enableIndex--;
1356                 enableLeave = restoreLeave;
1357         }
1358
1359         void PixelProgram::ELSE()
1360         {
1361                 ifDepth--;
1362
1363                 llvm::BasicBlock *falseBlock = ifFalseBlock[ifDepth];
1364                 llvm::BasicBlock *endBlock = Nucleus::createBasicBlock();
1365
1366                 if(isConditionalIf[ifDepth])
1367                 {
1368                         Int4 condition = ~enableStack[enableIndex] & enableStack[enableIndex - 1];
1369                         Bool notAllFalse = SignMask(condition) != 0;
1370
1371                         branch(notAllFalse, falseBlock, endBlock);
1372
1373                         enableStack[enableIndex] = ~enableStack[enableIndex] & enableStack[enableIndex - 1];
1374                 }
1375                 else
1376                 {
1377                         Nucleus::createBr(endBlock);
1378                         Nucleus::setInsertBlock(falseBlock);
1379                 }
1380
1381                 ifFalseBlock[ifDepth] = endBlock;
1382
1383                 ifDepth++;
1384         }
1385
1386         void PixelProgram::ENDIF()
1387         {
1388                 ifDepth--;
1389
1390                 llvm::BasicBlock *endBlock = ifFalseBlock[ifDepth];
1391
1392                 Nucleus::createBr(endBlock);
1393                 Nucleus::setInsertBlock(endBlock);
1394
1395                 if(isConditionalIf[ifDepth])
1396                 {
1397                         breakDepth--;
1398                         enableIndex--;
1399                 }
1400         }
1401
1402         void PixelProgram::ENDLOOP()
1403         {
1404                 loopRepDepth--;
1405
1406                 aL[loopDepth] = aL[loopDepth] + increment[loopDepth];   // FIXME: +=
1407
1408                 llvm::BasicBlock *testBlock = loopRepTestBlock[loopRepDepth];
1409                 llvm::BasicBlock *endBlock = loopRepEndBlock[loopRepDepth];
1410
1411                 Nucleus::createBr(testBlock);
1412                 Nucleus::setInsertBlock(endBlock);
1413
1414                 loopDepth--;
1415                 enableBreak = Int4(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
1416         }
1417
1418         void PixelProgram::ENDREP()
1419         {
1420                 loopRepDepth--;
1421
1422                 llvm::BasicBlock *testBlock = loopRepTestBlock[loopRepDepth];
1423                 llvm::BasicBlock *endBlock = loopRepEndBlock[loopRepDepth];
1424
1425                 Nucleus::createBr(testBlock);
1426                 Nucleus::setInsertBlock(endBlock);
1427
1428                 loopDepth--;
1429                 enableBreak = Int4(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
1430         }
1431
1432         void PixelProgram::ENDWHILE()
1433         {
1434                 loopRepDepth--;
1435
1436                 llvm::BasicBlock *testBlock = loopRepTestBlock[loopRepDepth];
1437                 llvm::BasicBlock *endBlock = loopRepEndBlock[loopRepDepth];
1438
1439                 Nucleus::createBr(testBlock);
1440                 Nucleus::setInsertBlock(endBlock);
1441
1442                 enableIndex--;
1443                 enableBreak = Int4(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
1444                 whileTest = false;
1445         }
1446
1447         void PixelProgram::IF(const Src &src)
1448         {
1449                 if(src.type == Shader::PARAMETER_CONSTBOOL)
1450                 {
1451                         IFb(src);
1452                 }
1453                 else if(src.type == Shader::PARAMETER_PREDICATE)
1454                 {
1455                         IFp(src);
1456                 }
1457                 else
1458                 {
1459                         Int4 condition = As<Int4>(fetchRegisterF(src).x);
1460                         IF(condition);
1461                 }
1462         }
1463
1464         void PixelProgram::IFb(const Src &boolRegister)
1465         {
1466                 ASSERT(ifDepth < 24 + 4);
1467
1468                 Bool condition = (*Pointer<Byte>(data + OFFSET(DrawData, ps.b[boolRegister.index])) != Byte(0));   // FIXME
1469
1470                 if(boolRegister.modifier == Shader::MODIFIER_NOT)
1471                 {
1472                         condition = !condition;
1473                 }
1474
1475                 llvm::BasicBlock *trueBlock = Nucleus::createBasicBlock();
1476                 llvm::BasicBlock *falseBlock = Nucleus::createBasicBlock();
1477
1478                 branch(condition, trueBlock, falseBlock);
1479
1480                 isConditionalIf[ifDepth] = false;
1481                 ifFalseBlock[ifDepth] = falseBlock;
1482
1483                 ifDepth++;
1484         }
1485
1486         void PixelProgram::IFp(const Src &predicateRegister)
1487         {
1488                 Int4 condition = As<Int4>(p0[predicateRegister.swizzle & 0x3]);
1489
1490                 if(predicateRegister.modifier == Shader::MODIFIER_NOT)
1491                 {
1492                         condition = ~condition;
1493                 }
1494
1495                 IF(condition);
1496         }
1497
1498         void PixelProgram::IFC(Vector4f &src0, Vector4f &src1, Control control)
1499         {
1500                 Int4 condition;
1501
1502                 switch(control)
1503                 {
1504                 case Shader::CONTROL_GT: condition = CmpNLE(src0.x, src1.x); break;
1505                 case Shader::CONTROL_EQ: condition = CmpEQ(src0.x, src1.x);  break;
1506                 case Shader::CONTROL_GE: condition = CmpNLT(src0.x, src1.x); break;
1507                 case Shader::CONTROL_LT: condition = CmpLT(src0.x, src1.x);  break;
1508                 case Shader::CONTROL_NE: condition = CmpNEQ(src0.x, src1.x); break;
1509                 case Shader::CONTROL_LE: condition = CmpLE(src0.x, src1.x);  break;
1510                 default:
1511                         ASSERT(false);
1512                 }
1513
1514                 IF(condition);
1515         }
1516
1517         void PixelProgram::IF(Int4 &condition)
1518         {
1519                 condition &= enableStack[enableIndex];
1520
1521                 enableIndex++;
1522                 enableStack[enableIndex] = condition;
1523
1524                 llvm::BasicBlock *trueBlock = Nucleus::createBasicBlock();
1525                 llvm::BasicBlock *falseBlock = Nucleus::createBasicBlock();
1526
1527                 Bool notAllFalse = SignMask(condition) != 0;
1528
1529                 branch(notAllFalse, trueBlock, falseBlock);
1530
1531                 isConditionalIf[ifDepth] = true;
1532                 ifFalseBlock[ifDepth] = falseBlock;
1533
1534                 ifDepth++;
1535                 breakDepth++;
1536         }
1537
1538         void PixelProgram::LABEL(int labelIndex)
1539         {
1540                 if(!labelBlock[labelIndex])
1541                 {
1542                         labelBlock[labelIndex] = Nucleus::createBasicBlock();
1543                 }
1544
1545                 Nucleus::setInsertBlock(labelBlock[labelIndex]);
1546                 currentLabel = labelIndex;
1547         }
1548
1549         void PixelProgram::LOOP(const Src &integerRegister)
1550         {
1551                 loopDepth++;
1552
1553                 iteration[loopDepth] = *Pointer<Int>(data + OFFSET(DrawData, ps.i[integerRegister.index][0]));
1554                 aL[loopDepth] = *Pointer<Int>(data + OFFSET(DrawData, ps.i[integerRegister.index][1]));
1555                 increment[loopDepth] = *Pointer<Int>(data + OFFSET(DrawData, ps.i[integerRegister.index][2]));
1556
1557                 //      If(increment[loopDepth] == 0)
1558                 //      {
1559                 //              increment[loopDepth] = 1;
1560                 //      }
1561
1562                 llvm::BasicBlock *loopBlock = Nucleus::createBasicBlock();
1563                 llvm::BasicBlock *testBlock = Nucleus::createBasicBlock();
1564                 llvm::BasicBlock *endBlock = Nucleus::createBasicBlock();
1565
1566                 loopRepTestBlock[loopRepDepth] = testBlock;
1567                 loopRepEndBlock[loopRepDepth] = endBlock;
1568
1569                 // FIXME: jump(testBlock)
1570                 Nucleus::createBr(testBlock);
1571                 Nucleus::setInsertBlock(testBlock);
1572
1573                 branch(iteration[loopDepth] > 0, loopBlock, endBlock);
1574                 Nucleus::setInsertBlock(loopBlock);
1575
1576                 iteration[loopDepth] = iteration[loopDepth] - 1;   // FIXME: --
1577
1578                 loopRepDepth++;
1579                 breakDepth = 0;
1580         }
1581
1582         void PixelProgram::REP(const Src &integerRegister)
1583         {
1584                 loopDepth++;
1585
1586                 iteration[loopDepth] = *Pointer<Int>(data + OFFSET(DrawData, ps.i[integerRegister.index][0]));
1587                 aL[loopDepth] = aL[loopDepth - 1];
1588
1589                 llvm::BasicBlock *loopBlock = Nucleus::createBasicBlock();
1590                 llvm::BasicBlock *testBlock = Nucleus::createBasicBlock();
1591                 llvm::BasicBlock *endBlock = Nucleus::createBasicBlock();
1592
1593                 loopRepTestBlock[loopRepDepth] = testBlock;
1594                 loopRepEndBlock[loopRepDepth] = endBlock;
1595
1596                 // FIXME: jump(testBlock)
1597                 Nucleus::createBr(testBlock);
1598                 Nucleus::setInsertBlock(testBlock);
1599
1600                 branch(iteration[loopDepth] > 0, loopBlock, endBlock);
1601                 Nucleus::setInsertBlock(loopBlock);
1602
1603                 iteration[loopDepth] = iteration[loopDepth] - 1;   // FIXME: --
1604
1605                 loopRepDepth++;
1606                 breakDepth = 0;
1607         }
1608
1609         void PixelProgram::WHILE(const Src &temporaryRegister)
1610         {
1611                 enableIndex++;
1612
1613                 llvm::BasicBlock *loopBlock = Nucleus::createBasicBlock();
1614                 llvm::BasicBlock *testBlock = Nucleus::createBasicBlock();
1615                 llvm::BasicBlock *endBlock = Nucleus::createBasicBlock();
1616
1617                 loopRepTestBlock[loopRepDepth] = testBlock;
1618                 loopRepEndBlock[loopRepDepth] = endBlock;
1619
1620                 Int4 restoreBreak = enableBreak;
1621                 Int4 restoreContinue = enableContinue;
1622
1623                 // FIXME: jump(testBlock)
1624                 Nucleus::createBr(testBlock);
1625                 Nucleus::setInsertBlock(testBlock);
1626                 enableContinue = restoreContinue;
1627
1628                 const Vector4f &src = fetchRegisterF(temporaryRegister);
1629                 Int4 condition = As<Int4>(src.x);
1630                 condition &= enableStack[enableIndex - 1];
1631                 enableStack[enableIndex] = condition;
1632
1633                 Bool notAllFalse = SignMask(condition) != 0;
1634                 branch(notAllFalse, loopBlock, endBlock);
1635
1636                 Nucleus::setInsertBlock(endBlock);
1637                 enableBreak = restoreBreak;
1638
1639                 Nucleus::setInsertBlock(loopBlock);
1640
1641                 loopRepDepth++;
1642                 breakDepth = 0;
1643         }
1644
1645         void PixelProgram::RET()
1646         {
1647                 if(currentLabel == -1)
1648                 {
1649                         returnBlock = Nucleus::createBasicBlock();
1650                         Nucleus::createBr(returnBlock);
1651                 }
1652                 else
1653                 {
1654                         llvm::BasicBlock *unreachableBlock = Nucleus::createBasicBlock();
1655
1656                         if(callRetBlock[currentLabel].size() > 1)   // Pop the return destination from the call stack
1657                         {
1658                                 // FIXME: Encapsulate
1659                                 UInt index = callStack[--stackIndex];
1660
1661                                 llvm::Value *value = index.loadValue();
1662                                 llvm::Value *switchInst = Nucleus::createSwitch(value, unreachableBlock, (int)callRetBlock[currentLabel].size());
1663
1664                                 for(unsigned int i = 0; i < callRetBlock[currentLabel].size(); i++)
1665                                 {
1666                                         Nucleus::addSwitchCase(switchInst, i, callRetBlock[currentLabel][i]);
1667                                 }
1668                         }
1669                         else if(callRetBlock[currentLabel].size() == 1)   // Jump directly to the unique return destination
1670                         {
1671                                 Nucleus::createBr(callRetBlock[currentLabel][0]);
1672                         }
1673                         else   // Function isn't called
1674                         {
1675                                 Nucleus::createBr(unreachableBlock);
1676                         }
1677
1678                         Nucleus::setInsertBlock(unreachableBlock);
1679                         Nucleus::createUnreachable();
1680                 }
1681         }
1682
1683         void PixelProgram::LEAVE()
1684         {
1685                 enableLeave = enableLeave & ~enableStack[enableIndex];
1686
1687                 // FIXME: Return from function if all instances left
1688                 // FIXME: Use enableLeave in other control-flow constructs
1689         }
1690 }