OSDN Git Service

Passing uniform buffers to the vertex/pixel programs
[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                                 for(unsigned int q = 0; q < state.multiSample; q++)
602                                 {
603                                         Pointer<Byte> buffer = cBuffer[index] + q * *Pointer<Int>(data + OFFSET(DrawData, colorSliceB[index]));
604                                         Vector4f color = c[index];
605
606                                         if(state.multiSampleMask & (1 << q))
607                                         {
608                                                 alphaBlend(index, buffer, color, x);
609                                                 writeColor(index, buffer, x, color, sMask[q], zMask[q], cMask[q]);
610                                         }
611                                 }
612                                 break;
613                         default:
614                                 ASSERT(false);
615                         }
616                 }
617         }
618
619         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)
620         {
621                 if(sampler.type == Shader::PARAMETER_SAMPLER && sampler.rel.type == Shader::PARAMETER_VOID)
622                 {
623                         sampleTexture(c, sampler.index, u, v, w, q, dsx, dsy, project, bias, gradients, lodProvided);
624                 }
625                 else
626                 {
627                         Int index = As<Int>(Float(fetchRegisterF(sampler).x.x));
628
629                         for(int i = 0; i < TEXTURE_IMAGE_UNITS; i++)
630                         {
631                                 if(shader->usesSampler(i))
632                                 {
633                                         If(index == i)
634                                         {
635                                                 sampleTexture(c, i, u, v, w, q, dsx, dsy, project, bias, gradients, lodProvided);
636                                                 // FIXME: When the sampler states are the same, we could use one sampler and just index the texture
637                                         }
638                                 }
639                         }
640                 }
641         }
642
643         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)
644         {
645 #if PERF_PROFILE
646                 Long texTime = Ticks();
647 #endif
648
649                 Pointer<Byte> texture = data + OFFSET(DrawData, mipmap) + stage * sizeof(Texture);
650
651                 if(!project)
652                 {
653                         sampler[stage]->sampleTexture(texture, c, u, v, w, q, dsx, dsy, bias, gradients, lodProvided);
654                 }
655                 else
656                 {
657                         Float4 rq = reciprocal(q);
658
659                         Float4 u_q = u * rq;
660                         Float4 v_q = v * rq;
661                         Float4 w_q = w * rq;
662
663                         sampler[stage]->sampleTexture(texture, c, u_q, v_q, w_q, q, dsx, dsy, bias, gradients, lodProvided);
664                 }
665
666 #if PERF_PROFILE
667                 cycles[PERF_TEX] += Ticks() - texTime;
668 #endif
669         }
670
671         void PixelProgram::clampColor(Vector4f oC[RENDERTARGETS])
672         {
673                 for(int index = 0; index < RENDERTARGETS; index++)
674                 {
675                         if(!state.colorWriteActive(index) && !(index == 0 && state.alphaTestActive()))
676                         {
677                                 continue;
678                         }
679
680                         switch(state.targetFormat[index])
681                         {
682                         case FORMAT_NULL:
683                                 break;
684                         case FORMAT_R5G6B5:
685                         case FORMAT_A8R8G8B8:
686                         case FORMAT_A8B8G8R8:
687                         case FORMAT_X8R8G8B8:
688                         case FORMAT_X8B8G8R8:
689                         case FORMAT_A8:
690                         case FORMAT_G16R16:
691                         case FORMAT_A16B16G16R16:
692                                 oC[index].x = Max(oC[index].x, Float4(0.0f)); oC[index].x = Min(oC[index].x, Float4(1.0f));
693                                 oC[index].y = Max(oC[index].y, Float4(0.0f)); oC[index].y = Min(oC[index].y, Float4(1.0f));
694                                 oC[index].z = Max(oC[index].z, Float4(0.0f)); oC[index].z = Min(oC[index].z, Float4(1.0f));
695                                 oC[index].w = Max(oC[index].w, Float4(0.0f)); oC[index].w = Min(oC[index].w, Float4(1.0f));
696                                 break;
697                         case FORMAT_R32F:
698                         case FORMAT_G32R32F:
699                         case FORMAT_A32B32G32R32F:
700                                 break;
701                         default:
702                                 ASSERT(false);
703                         }
704                 }
705         }
706
707         Int4 PixelProgram::enableMask(const Shader::Instruction *instruction)
708         {
709                 Int4 enable = instruction->analysisBranch ? Int4(enableStack[enableIndex]) : Int4(0xFFFFFFFF);
710
711                 if(!whileTest)
712                 {
713                         if(shader->containsBreakInstruction() && instruction->analysisBreak)
714                         {
715                                 enable &= enableBreak;
716                         }
717
718                         if(shader->containsContinueInstruction() && instruction->analysisContinue)
719                         {
720                                 enable &= enableContinue;
721                         }
722
723                         if(shader->containsLeaveInstruction() && instruction->analysisLeave)
724                         {
725                                 enable &= enableLeave;
726                         }
727                 }
728
729                 return enable;
730         }
731
732         Vector4f PixelProgram::fetchRegisterF(const Src &src, unsigned int offset)
733         {
734                 Vector4f reg;
735                 unsigned int i = src.index + offset;
736
737                 switch(src.type)
738                 {
739                 case Shader::PARAMETER_TEMP:
740                         if(src.rel.type == Shader::PARAMETER_VOID)
741                         {
742                                 reg = r[i];
743                         }
744                         else
745                         {
746                                 Int a = relativeAddress(src, src.bufferIndex);
747
748                                 reg = r[i + a];
749                         }
750                         break;
751                 case Shader::PARAMETER_INPUT:
752                         {
753                                 if(src.rel.type == Shader::PARAMETER_VOID)   // Not relative
754                                 {
755                                         reg = v[i];
756                                 }
757                                 else
758                                 {
759                                         Int a = relativeAddress(src, src.bufferIndex);
760
761                                         reg = v[i + a];
762                                 }
763                         }
764                         break;
765                 case Shader::PARAMETER_CONST:
766                         reg = readConstant(src, offset);
767                         break;
768                 case Shader::PARAMETER_TEXTURE:
769                         reg = v[2 + i];
770                         break;
771                 case Shader::PARAMETER_MISCTYPE:
772                         if(src.index == 0) reg = vPos;
773                         if(src.index == 1) reg = vFace;
774                         break;
775                 case Shader::PARAMETER_SAMPLER:
776                         if(src.rel.type == Shader::PARAMETER_VOID)
777                         {
778                                 reg.x = As<Float4>(Int4(i));
779                         }
780                         else if(src.rel.type == Shader::PARAMETER_TEMP)
781                         {
782                                 reg.x = As<Float4>(Int4(i) + As<Int4>(r[src.rel.index].x));
783                         }
784                         return reg;
785                 case Shader::PARAMETER_PREDICATE:   return reg; // Dummy
786                 case Shader::PARAMETER_VOID:        return reg; // Dummy
787                 case Shader::PARAMETER_FLOAT4LITERAL:
788                         reg.x = Float4(src.value[0]);
789                         reg.y = Float4(src.value[1]);
790                         reg.z = Float4(src.value[2]);
791                         reg.w = Float4(src.value[3]);
792                         break;
793                 case Shader::PARAMETER_CONSTINT:    return reg; // Dummy
794                 case Shader::PARAMETER_CONSTBOOL:   return reg; // Dummy
795                 case Shader::PARAMETER_LOOP:        return reg; // Dummy
796                 case Shader::PARAMETER_COLOROUT:
797                         if(src.rel.type == Shader::PARAMETER_VOID)   // Not relative
798                         {
799                                 reg = oC[i];
800                         }
801                         else
802                         {
803                                 Int a = relativeAddress(src, src.bufferIndex);
804
805                                 reg = oC[i + a];
806                         }
807                         break;
808                 case Shader::PARAMETER_DEPTHOUT:
809                         reg.x = oDepth;
810                         break;
811                 default:
812                         ASSERT(false);
813                 }
814
815                 const Float4 &x = reg[(src.swizzle >> 0) & 0x3];
816                 const Float4 &y = reg[(src.swizzle >> 2) & 0x3];
817                 const Float4 &z = reg[(src.swizzle >> 4) & 0x3];
818                 const Float4 &w = reg[(src.swizzle >> 6) & 0x3];
819
820                 Vector4f mod;
821
822                 switch(src.modifier)
823                 {
824                 case Shader::MODIFIER_NONE:
825                         mod.x = x;
826                         mod.y = y;
827                         mod.z = z;
828                         mod.w = w;
829                         break;
830                 case Shader::MODIFIER_NEGATE:
831                         mod.x = -x;
832                         mod.y = -y;
833                         mod.z = -z;
834                         mod.w = -w;
835                         break;
836                 case Shader::MODIFIER_ABS:
837                         mod.x = Abs(x);
838                         mod.y = Abs(y);
839                         mod.z = Abs(z);
840                         mod.w = Abs(w);
841                         break;
842                 case Shader::MODIFIER_ABS_NEGATE:
843                         mod.x = -Abs(x);
844                         mod.y = -Abs(y);
845                         mod.z = -Abs(z);
846                         mod.w = -Abs(w);
847                         break;
848                 case Shader::MODIFIER_NOT:
849                         mod.x = As<Float4>(As<Int4>(x) ^ Int4(0xFFFFFFFF));
850                         mod.y = As<Float4>(As<Int4>(y) ^ Int4(0xFFFFFFFF));
851                         mod.z = As<Float4>(As<Int4>(z) ^ Int4(0xFFFFFFFF));
852                         mod.w = As<Float4>(As<Int4>(w) ^ Int4(0xFFFFFFFF));
853                         break;
854                 default:
855                         ASSERT(false);
856                 }
857
858                 return mod;
859         }
860
861         RValue<Pointer<Byte>> PixelProgram::uniformAddress(int bufferIndex, unsigned int index)
862         {
863                 if(bufferIndex == -1)
864                 {
865                         return data + OFFSET(DrawData, ps.c[index]);
866                 }
867                 else
868                 {
869                         return *Pointer<Pointer<Byte>>(data + OFFSET(DrawData, ps.u[bufferIndex])) + index;
870                 }
871         }
872
873         RValue<Pointer<Byte>> PixelProgram::uniformAddress(int bufferIndex, unsigned int index, Int& offset)
874         {
875                 return uniformAddress(bufferIndex, index) + offset * sizeof(float4);
876         }
877
878         Vector4f PixelProgram::readConstant(const Src &src, unsigned int offset)
879         {
880                 Vector4f c;
881                 unsigned int i = src.index + offset;
882
883                 if(src.rel.type == Shader::PARAMETER_VOID)   // Not relative
884                 {
885                         c.x = c.y = c.z = c.w = *Pointer<Float4>(uniformAddress(src.bufferIndex, i));
886
887                         c.x = c.x.xxxx;
888                         c.y = c.y.yyyy;
889                         c.z = c.z.zzzz;
890                         c.w = c.w.wwww;
891
892                         if(shader->containsDefineInstruction())   // Constant may be known at compile time
893                         {
894                                 for(size_t j = 0; j < shader->getLength(); j++)
895                                 {
896                                         const Shader::Instruction &instruction = *shader->getInstruction(j);
897
898                                         if(instruction.opcode == Shader::OPCODE_DEF)
899                                         {
900                                                 if(instruction.dst.index == i)
901                                                 {
902                                                         c.x = Float4(instruction.src[0].value[0]);
903                                                         c.y = Float4(instruction.src[0].value[1]);
904                                                         c.z = Float4(instruction.src[0].value[2]);
905                                                         c.w = Float4(instruction.src[0].value[3]);
906
907                                                         break;
908                                                 }
909                                         }
910                                 }
911                         }
912                 }
913                 else if(src.rel.type == Shader::PARAMETER_LOOP)
914                 {
915                         Int loopCounter = aL[loopDepth];
916
917                         c.x = c.y = c.z = c.w = *Pointer<Float4>(uniformAddress(src.bufferIndex, i, loopCounter));
918
919                         c.x = c.x.xxxx;
920                         c.y = c.y.yyyy;
921                         c.z = c.z.zzzz;
922                         c.w = c.w.wwww;
923                 }
924                 else
925                 {
926                         Int a = relativeAddress(src, src.bufferIndex);
927
928                         c.x = c.y = c.z = c.w = *Pointer<Float4>(uniformAddress(src.bufferIndex, i, a));
929
930                         c.x = c.x.xxxx;
931                         c.y = c.y.yyyy;
932                         c.z = c.z.zzzz;
933                         c.w = c.w.wwww;
934                 }
935
936                 return c;
937         }
938
939         Int PixelProgram::relativeAddress(const Shader::Parameter &var, int bufferIndex)
940         {
941                 ASSERT(var.rel.deterministic);
942
943                 if(var.rel.type == Shader::PARAMETER_TEMP)
944                 {
945                         return As<Int>(Extract(r[var.rel.index].x, 0)) * var.rel.scale;
946                 }
947                 else if(var.rel.type == Shader::PARAMETER_INPUT)
948                 {
949                         return As<Int>(Extract(v[var.rel.index].x, 0)) * var.rel.scale;
950                 }
951                 else if(var.rel.type == Shader::PARAMETER_OUTPUT)
952                 {
953                         return As<Int>(Extract(oC[var.rel.index].x, 0)) * var.rel.scale;
954                 }
955                 else if(var.rel.type == Shader::PARAMETER_CONST)
956                 {
957                         RValue<Int4> c = *Pointer<Int4>(uniformAddress(bufferIndex, var.rel.index));
958
959                         return Extract(c, 0) * var.rel.scale;
960                 }
961                 else if(var.rel.type == Shader::PARAMETER_LOOP)
962                 {
963                         return aL[loopDepth];
964                 }
965                 else ASSERT(false);
966
967                 return 0;
968         }
969
970         Float4 PixelProgram::linearToSRGB(const Float4 &x)   // Approximates x^(1.0/2.2)
971         {
972                 Float4 sqrtx = Rcp_pp(RcpSqrt_pp(x));
973                 Float4 sRGB = sqrtx * Float4(1.14f) - x * Float4(0.14f);
974
975                 return Min(Max(sRGB, Float4(0.0f)), Float4(1.0f));
976         }
977
978         void PixelProgram::M3X2(Vector4f &dst, Vector4f &src0, const Src &src1)
979         {
980                 Vector4f row0 = fetchRegisterF(src1, 0);
981                 Vector4f row1 = fetchRegisterF(src1, 1);
982
983                 dst.x = dot3(src0, row0);
984                 dst.y = dot3(src0, row1);
985         }
986
987         void PixelProgram::M3X3(Vector4f &dst, Vector4f &src0, const Src &src1)
988         {
989                 Vector4f row0 = fetchRegisterF(src1, 0);
990                 Vector4f row1 = fetchRegisterF(src1, 1);
991                 Vector4f row2 = fetchRegisterF(src1, 2);
992
993                 dst.x = dot3(src0, row0);
994                 dst.y = dot3(src0, row1);
995                 dst.z = dot3(src0, row2);
996         }
997
998         void PixelProgram::M3X4(Vector4f &dst, Vector4f &src0, const Src &src1)
999         {
1000                 Vector4f row0 = fetchRegisterF(src1, 0);
1001                 Vector4f row1 = fetchRegisterF(src1, 1);
1002                 Vector4f row2 = fetchRegisterF(src1, 2);
1003                 Vector4f row3 = fetchRegisterF(src1, 3);
1004
1005                 dst.x = dot3(src0, row0);
1006                 dst.y = dot3(src0, row1);
1007                 dst.z = dot3(src0, row2);
1008                 dst.w = dot3(src0, row3);
1009         }
1010
1011         void PixelProgram::M4X3(Vector4f &dst, Vector4f &src0, const Src &src1)
1012         {
1013                 Vector4f row0 = fetchRegisterF(src1, 0);
1014                 Vector4f row1 = fetchRegisterF(src1, 1);
1015                 Vector4f row2 = fetchRegisterF(src1, 2);
1016
1017                 dst.x = dot4(src0, row0);
1018                 dst.y = dot4(src0, row1);
1019                 dst.z = dot4(src0, row2);
1020         }
1021
1022         void PixelProgram::M4X4(Vector4f &dst, Vector4f &src0, const Src &src1)
1023         {
1024                 Vector4f row0 = fetchRegisterF(src1, 0);
1025                 Vector4f row1 = fetchRegisterF(src1, 1);
1026                 Vector4f row2 = fetchRegisterF(src1, 2);
1027                 Vector4f row3 = fetchRegisterF(src1, 3);
1028
1029                 dst.x = dot4(src0, row0);
1030                 dst.y = dot4(src0, row1);
1031                 dst.z = dot4(src0, row2);
1032                 dst.w = dot4(src0, row3);
1033         }
1034
1035         void PixelProgram::TEXLD(Vector4f &dst, Vector4f &src0, const Src &src1, bool project, bool bias)
1036         {
1037                 Vector4f tmp;
1038                 sampleTexture(tmp, src1, src0.x, src0.y, src0.z, src0.w, src0, src0, project, bias);
1039
1040                 dst.x = tmp[(src1.swizzle >> 0) & 0x3];
1041                 dst.y = tmp[(src1.swizzle >> 2) & 0x3];
1042                 dst.z = tmp[(src1.swizzle >> 4) & 0x3];
1043                 dst.w = tmp[(src1.swizzle >> 6) & 0x3];
1044         }
1045
1046         void PixelProgram::TEXOFFSET(Vector4f &dst, Vector4f &src0, const Src& src1, Vector4f &src2, Vector4f &src3, bool project, bool bias)
1047         {
1048                 UNIMPLEMENTED();
1049         }
1050
1051         void PixelProgram::TEXLDL(Vector4f &dst, Vector4f &src0, const Src &src1, Vector4f &offset, bool project, bool bias)
1052         {
1053                 UNIMPLEMENTED();
1054         }
1055
1056         void PixelProgram::TEXELFETCH(Vector4f &dst, Vector4f &src0, const Src& src1, Vector4f &src2)
1057         {
1058                 UNIMPLEMENTED();
1059         }
1060
1061         void PixelProgram::TEXELFETCH(Vector4f &dst, Vector4f &src0, const Src& src1, Vector4f &src2, Vector4f &offset)
1062         {
1063                 UNIMPLEMENTED();
1064         }
1065
1066         void PixelProgram::TEXGRAD(Vector4f &dst, Vector4f &src0, const Src& src1, Vector4f &src2, Vector4f &src3)
1067         {
1068                 UNIMPLEMENTED();
1069         }
1070
1071         void PixelProgram::TEXGRAD(Vector4f &dst, Vector4f &src0, const Src& src1, Vector4f &src2, Vector4f &src3, Vector4f &offset)
1072         {
1073                 UNIMPLEMENTED();
1074         }
1075
1076         void PixelProgram::TEXLDD(Vector4f &dst, Vector4f &src0, const Src &src1, Vector4f &src2, Vector4f &src3, bool project, bool bias)
1077         {
1078                 Vector4f tmp;
1079                 sampleTexture(tmp, src1, src0.x, src0.y, src0.z, src0.w, src2, src3, project, bias, true);
1080
1081                 dst.x = tmp[(src1.swizzle >> 0) & 0x3];
1082                 dst.y = tmp[(src1.swizzle >> 2) & 0x3];
1083                 dst.z = tmp[(src1.swizzle >> 4) & 0x3];
1084                 dst.w = tmp[(src1.swizzle >> 6) & 0x3];
1085         }
1086
1087         void PixelProgram::TEXLDL(Vector4f &dst, Vector4f &src0, const Src &src1, bool project, bool bias)
1088         {
1089                 Vector4f tmp;
1090                 sampleTexture(tmp, src1, src0.x, src0.y, src0.z, src0.w, src0, src0, project, bias, false, true);
1091
1092                 dst.x = tmp[(src1.swizzle >> 0) & 0x3];
1093                 dst.y = tmp[(src1.swizzle >> 2) & 0x3];
1094                 dst.z = tmp[(src1.swizzle >> 4) & 0x3];
1095                 dst.w = tmp[(src1.swizzle >> 6) & 0x3];
1096         }
1097
1098         void PixelProgram::TEXSIZE(Vector4f &dst, Float4 &lod, const Src &src1)
1099         {
1100                 Pointer<Byte> textureMipmap = data + OFFSET(DrawData, mipmap) + src1.index * sizeof(Texture) + OFFSET(Texture, mipmap);
1101                 for(int i = 0; i < 4; ++i)
1102                 {
1103                         Pointer<Byte> mipmap = textureMipmap + (As<Int>(Extract(lod, i)) + Int(1)) * sizeof(Mipmap);
1104                         dst.x = Insert(dst.x, As<Float>(Int(*Pointer<Short>(mipmap + OFFSET(Mipmap, width)))), i);
1105                         dst.y = Insert(dst.y, As<Float>(Int(*Pointer<Short>(mipmap + OFFSET(Mipmap, height)))), i);
1106                         dst.z = Insert(dst.z, As<Float>(Int(*Pointer<Short>(mipmap + OFFSET(Mipmap, depth)))), i);
1107                 }
1108         }
1109
1110         void PixelProgram::TEXKILL(Int cMask[4], Vector4f &src, unsigned char mask)
1111         {
1112                 Int kill = -1;
1113
1114                 if(mask & 0x1) kill &= SignMask(CmpNLT(src.x, Float4(0.0f)));
1115                 if(mask & 0x2) kill &= SignMask(CmpNLT(src.y, Float4(0.0f)));
1116                 if(mask & 0x4) kill &= SignMask(CmpNLT(src.z, Float4(0.0f)));
1117                 if(mask & 0x8) kill &= SignMask(CmpNLT(src.w, Float4(0.0f)));
1118
1119                 // FIXME: Dynamic branching affects TEXKILL?
1120                 //      if(shader->containsDynamicBranching())
1121                 //      {
1122                 //              kill = ~SignMask(enableMask());
1123                 //      }
1124
1125                 for(unsigned int q = 0; q < state.multiSample; q++)
1126                 {
1127                         cMask[q] &= kill;
1128                 }
1129
1130                 // FIXME: Branch to end of shader if all killed?
1131         }
1132
1133         void PixelProgram::DISCARD(Int cMask[4], const Shader::Instruction *instruction)
1134         {
1135                 Int kill = 0;
1136
1137                 if(shader->containsDynamicBranching())
1138                 {
1139                         kill = ~SignMask(enableMask(instruction));
1140                 }
1141
1142                 for(unsigned int q = 0; q < state.multiSample; q++)
1143                 {
1144                         cMask[q] &= kill;
1145                 }
1146
1147                 // FIXME: Branch to end of shader if all killed?
1148         }
1149
1150         void PixelProgram::DFDX(Vector4f &dst, Vector4f &src)
1151         {
1152                 dst.x = src.x.yyww - src.x.xxzz;
1153                 dst.y = src.y.yyww - src.y.xxzz;
1154                 dst.z = src.z.yyww - src.z.xxzz;
1155                 dst.w = src.w.yyww - src.w.xxzz;
1156         }
1157
1158         void PixelProgram::DFDY(Vector4f &dst, Vector4f &src)
1159         {
1160                 dst.x = src.x.zwzw - src.x.xyxy;
1161                 dst.y = src.y.zwzw - src.y.xyxy;
1162                 dst.z = src.z.zwzw - src.z.xyxy;
1163                 dst.w = src.w.zwzw - src.w.xyxy;
1164         }
1165
1166         void PixelProgram::FWIDTH(Vector4f &dst, Vector4f &src)
1167         {
1168                 // abs(dFdx(src)) + abs(dFdy(src));
1169                 dst.x = Abs(src.x.yyww - src.x.xxzz) + Abs(src.x.zwzw - src.x.xyxy);
1170                 dst.y = Abs(src.y.yyww - src.y.xxzz) + Abs(src.y.zwzw - src.y.xyxy);
1171                 dst.z = Abs(src.z.yyww - src.z.xxzz) + Abs(src.z.zwzw - src.z.xyxy);
1172                 dst.w = Abs(src.w.yyww - src.w.xxzz) + Abs(src.w.zwzw - src.w.xyxy);
1173         }
1174
1175         void PixelProgram::BREAK()
1176         {
1177                 llvm::BasicBlock *deadBlock = Nucleus::createBasicBlock();
1178                 llvm::BasicBlock *endBlock = loopRepEndBlock[loopRepDepth - 1];
1179
1180                 if(breakDepth == 0)
1181                 {
1182                         enableIndex = enableIndex - breakDepth;
1183                         Nucleus::createBr(endBlock);
1184                 }
1185                 else
1186                 {
1187                         enableBreak = enableBreak & ~enableStack[enableIndex];
1188                         Bool allBreak = SignMask(enableBreak) == 0x0;
1189
1190                         enableIndex = enableIndex - breakDepth;
1191                         branch(allBreak, endBlock, deadBlock);
1192                 }
1193
1194                 Nucleus::setInsertBlock(deadBlock);
1195                 enableIndex = enableIndex + breakDepth;
1196         }
1197
1198         void PixelProgram::BREAKC(Vector4f &src0, Vector4f &src1, Control control)
1199         {
1200                 Int4 condition;
1201
1202                 switch(control)
1203                 {
1204                 case Shader::CONTROL_GT: condition = CmpNLE(src0.x, src1.x); break;
1205                 case Shader::CONTROL_EQ: condition = CmpEQ(src0.x, src1.x);  break;
1206                 case Shader::CONTROL_GE: condition = CmpNLT(src0.x, src1.x); break;
1207                 case Shader::CONTROL_LT: condition = CmpLT(src0.x, src1.x);  break;
1208                 case Shader::CONTROL_NE: condition = CmpNEQ(src0.x, src1.x); break;
1209                 case Shader::CONTROL_LE: condition = CmpLE(src0.x, src1.x);  break;
1210                 default:
1211                         ASSERT(false);
1212                 }
1213
1214                 BREAK(condition);
1215         }
1216
1217         void PixelProgram::BREAKP(const Src &predicateRegister)   // FIXME: Factor out parts common with BREAKC
1218         {
1219                 Int4 condition = As<Int4>(p0[predicateRegister.swizzle & 0x3]);
1220
1221                 if(predicateRegister.modifier == Shader::MODIFIER_NOT)
1222                 {
1223                         condition = ~condition;
1224                 }
1225
1226                 BREAK(condition);
1227         }
1228
1229         void PixelProgram::BREAK(Int4 &condition)
1230         {
1231                 condition &= enableStack[enableIndex];
1232
1233                 llvm::BasicBlock *continueBlock = Nucleus::createBasicBlock();
1234                 llvm::BasicBlock *endBlock = loopRepEndBlock[loopRepDepth - 1];
1235
1236                 enableBreak = enableBreak & ~condition;
1237                 Bool allBreak = SignMask(enableBreak) == 0x0;
1238
1239                 enableIndex = enableIndex - breakDepth;
1240                 branch(allBreak, endBlock, continueBlock);
1241
1242                 Nucleus::setInsertBlock(continueBlock);
1243                 enableIndex = enableIndex + breakDepth;
1244         }
1245
1246         void PixelProgram::CONTINUE()
1247         {
1248                 enableContinue = enableContinue & ~enableStack[enableIndex];
1249         }
1250
1251         void PixelProgram::TEST()
1252         {
1253                 whileTest = true;
1254         }
1255
1256         void PixelProgram::CALL(int labelIndex, int callSiteIndex)
1257         {
1258                 if(!labelBlock[labelIndex])
1259                 {
1260                         labelBlock[labelIndex] = Nucleus::createBasicBlock();
1261                 }
1262
1263                 if(callRetBlock[labelIndex].size() > 1)
1264                 {
1265                         callStack[stackIndex++] = UInt(callSiteIndex);
1266                 }
1267
1268                 Int4 restoreLeave = enableLeave;
1269
1270                 Nucleus::createBr(labelBlock[labelIndex]);
1271                 Nucleus::setInsertBlock(callRetBlock[labelIndex][callSiteIndex]);
1272
1273                 enableLeave = restoreLeave;
1274         }
1275
1276         void PixelProgram::CALLNZ(int labelIndex, int callSiteIndex, const Src &src)
1277         {
1278                 if(src.type == Shader::PARAMETER_CONSTBOOL)
1279                 {
1280                         CALLNZb(labelIndex, callSiteIndex, src);
1281                 }
1282                 else if(src.type == Shader::PARAMETER_PREDICATE)
1283                 {
1284                         CALLNZp(labelIndex, callSiteIndex, src);
1285                 }
1286                 else ASSERT(false);
1287         }
1288
1289         void PixelProgram::CALLNZb(int labelIndex, int callSiteIndex, const Src &boolRegister)
1290         {
1291                 Bool condition = (*Pointer<Byte>(data + OFFSET(DrawData, ps.b[boolRegister.index])) != Byte(0));   // FIXME
1292
1293                 if(boolRegister.modifier == Shader::MODIFIER_NOT)
1294                 {
1295                         condition = !condition;
1296                 }
1297
1298                 if(!labelBlock[labelIndex])
1299                 {
1300                         labelBlock[labelIndex] = Nucleus::createBasicBlock();
1301                 }
1302
1303                 if(callRetBlock[labelIndex].size() > 1)
1304                 {
1305                         callStack[stackIndex++] = UInt(callSiteIndex);
1306                 }
1307
1308                 Int4 restoreLeave = enableLeave;
1309
1310                 branch(condition, labelBlock[labelIndex], callRetBlock[labelIndex][callSiteIndex]);
1311                 Nucleus::setInsertBlock(callRetBlock[labelIndex][callSiteIndex]);
1312
1313                 enableLeave = restoreLeave;
1314         }
1315
1316         void PixelProgram::CALLNZp(int labelIndex, int callSiteIndex, const Src &predicateRegister)
1317         {
1318                 Int4 condition = As<Int4>(p0[predicateRegister.swizzle & 0x3]);
1319
1320                 if(predicateRegister.modifier == Shader::MODIFIER_NOT)
1321                 {
1322                         condition = ~condition;
1323                 }
1324
1325                 condition &= enableStack[enableIndex];
1326
1327                 if(!labelBlock[labelIndex])
1328                 {
1329                         labelBlock[labelIndex] = Nucleus::createBasicBlock();
1330                 }
1331
1332                 if(callRetBlock[labelIndex].size() > 1)
1333                 {
1334                         callStack[stackIndex++] = UInt(callSiteIndex);
1335                 }
1336
1337                 enableIndex++;
1338                 enableStack[enableIndex] = condition;
1339                 Int4 restoreLeave = enableLeave;
1340
1341                 Bool notAllFalse = SignMask(condition) != 0;
1342                 branch(notAllFalse, labelBlock[labelIndex], callRetBlock[labelIndex][callSiteIndex]);
1343                 Nucleus::setInsertBlock(callRetBlock[labelIndex][callSiteIndex]);
1344
1345                 enableIndex--;
1346                 enableLeave = restoreLeave;
1347         }
1348
1349         void PixelProgram::ELSE()
1350         {
1351                 ifDepth--;
1352
1353                 llvm::BasicBlock *falseBlock = ifFalseBlock[ifDepth];
1354                 llvm::BasicBlock *endBlock = Nucleus::createBasicBlock();
1355
1356                 if(isConditionalIf[ifDepth])
1357                 {
1358                         Int4 condition = ~enableStack[enableIndex] & enableStack[enableIndex - 1];
1359                         Bool notAllFalse = SignMask(condition) != 0;
1360
1361                         branch(notAllFalse, falseBlock, endBlock);
1362
1363                         enableStack[enableIndex] = ~enableStack[enableIndex] & enableStack[enableIndex - 1];
1364                 }
1365                 else
1366                 {
1367                         Nucleus::createBr(endBlock);
1368                         Nucleus::setInsertBlock(falseBlock);
1369                 }
1370
1371                 ifFalseBlock[ifDepth] = endBlock;
1372
1373                 ifDepth++;
1374         }
1375
1376         void PixelProgram::ENDIF()
1377         {
1378                 ifDepth--;
1379
1380                 llvm::BasicBlock *endBlock = ifFalseBlock[ifDepth];
1381
1382                 Nucleus::createBr(endBlock);
1383                 Nucleus::setInsertBlock(endBlock);
1384
1385                 if(isConditionalIf[ifDepth])
1386                 {
1387                         breakDepth--;
1388                         enableIndex--;
1389                 }
1390         }
1391
1392         void PixelProgram::ENDLOOP()
1393         {
1394                 loopRepDepth--;
1395
1396                 aL[loopDepth] = aL[loopDepth] + increment[loopDepth];   // FIXME: +=
1397
1398                 llvm::BasicBlock *testBlock = loopRepTestBlock[loopRepDepth];
1399                 llvm::BasicBlock *endBlock = loopRepEndBlock[loopRepDepth];
1400
1401                 Nucleus::createBr(testBlock);
1402                 Nucleus::setInsertBlock(endBlock);
1403
1404                 loopDepth--;
1405                 enableBreak = Int4(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
1406         }
1407
1408         void PixelProgram::ENDREP()
1409         {
1410                 loopRepDepth--;
1411
1412                 llvm::BasicBlock *testBlock = loopRepTestBlock[loopRepDepth];
1413                 llvm::BasicBlock *endBlock = loopRepEndBlock[loopRepDepth];
1414
1415                 Nucleus::createBr(testBlock);
1416                 Nucleus::setInsertBlock(endBlock);
1417
1418                 loopDepth--;
1419                 enableBreak = Int4(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
1420         }
1421
1422         void PixelProgram::ENDWHILE()
1423         {
1424                 loopRepDepth--;
1425
1426                 llvm::BasicBlock *testBlock = loopRepTestBlock[loopRepDepth];
1427                 llvm::BasicBlock *endBlock = loopRepEndBlock[loopRepDepth];
1428
1429                 Nucleus::createBr(testBlock);
1430                 Nucleus::setInsertBlock(endBlock);
1431
1432                 enableIndex--;
1433                 enableBreak = Int4(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
1434                 whileTest = false;
1435         }
1436
1437         void PixelProgram::IF(const Src &src)
1438         {
1439                 if(src.type == Shader::PARAMETER_CONSTBOOL)
1440                 {
1441                         IFb(src);
1442                 }
1443                 else if(src.type == Shader::PARAMETER_PREDICATE)
1444                 {
1445                         IFp(src);
1446                 }
1447                 else
1448                 {
1449                         Int4 condition = As<Int4>(fetchRegisterF(src).x);
1450                         IF(condition);
1451                 }
1452         }
1453
1454         void PixelProgram::IFb(const Src &boolRegister)
1455         {
1456                 ASSERT(ifDepth < 24 + 4);
1457
1458                 Bool condition = (*Pointer<Byte>(data + OFFSET(DrawData, ps.b[boolRegister.index])) != Byte(0));   // FIXME
1459
1460                 if(boolRegister.modifier == Shader::MODIFIER_NOT)
1461                 {
1462                         condition = !condition;
1463                 }
1464
1465                 llvm::BasicBlock *trueBlock = Nucleus::createBasicBlock();
1466                 llvm::BasicBlock *falseBlock = Nucleus::createBasicBlock();
1467
1468                 branch(condition, trueBlock, falseBlock);
1469
1470                 isConditionalIf[ifDepth] = false;
1471                 ifFalseBlock[ifDepth] = falseBlock;
1472
1473                 ifDepth++;
1474         }
1475
1476         void PixelProgram::IFp(const Src &predicateRegister)
1477         {
1478                 Int4 condition = As<Int4>(p0[predicateRegister.swizzle & 0x3]);
1479
1480                 if(predicateRegister.modifier == Shader::MODIFIER_NOT)
1481                 {
1482                         condition = ~condition;
1483                 }
1484
1485                 IF(condition);
1486         }
1487
1488         void PixelProgram::IFC(Vector4f &src0, Vector4f &src1, Control control)
1489         {
1490                 Int4 condition;
1491
1492                 switch(control)
1493                 {
1494                 case Shader::CONTROL_GT: condition = CmpNLE(src0.x, src1.x); break;
1495                 case Shader::CONTROL_EQ: condition = CmpEQ(src0.x, src1.x);  break;
1496                 case Shader::CONTROL_GE: condition = CmpNLT(src0.x, src1.x); break;
1497                 case Shader::CONTROL_LT: condition = CmpLT(src0.x, src1.x);  break;
1498                 case Shader::CONTROL_NE: condition = CmpNEQ(src0.x, src1.x); break;
1499                 case Shader::CONTROL_LE: condition = CmpLE(src0.x, src1.x);  break;
1500                 default:
1501                         ASSERT(false);
1502                 }
1503
1504                 IF(condition);
1505         }
1506
1507         void PixelProgram::IF(Int4 &condition)
1508         {
1509                 condition &= enableStack[enableIndex];
1510
1511                 enableIndex++;
1512                 enableStack[enableIndex] = condition;
1513
1514                 llvm::BasicBlock *trueBlock = Nucleus::createBasicBlock();
1515                 llvm::BasicBlock *falseBlock = Nucleus::createBasicBlock();
1516
1517                 Bool notAllFalse = SignMask(condition) != 0;
1518
1519                 branch(notAllFalse, trueBlock, falseBlock);
1520
1521                 isConditionalIf[ifDepth] = true;
1522                 ifFalseBlock[ifDepth] = falseBlock;
1523
1524                 ifDepth++;
1525                 breakDepth++;
1526         }
1527
1528         void PixelProgram::LABEL(int labelIndex)
1529         {
1530                 if(!labelBlock[labelIndex])
1531                 {
1532                         labelBlock[labelIndex] = Nucleus::createBasicBlock();
1533                 }
1534
1535                 Nucleus::setInsertBlock(labelBlock[labelIndex]);
1536                 currentLabel = labelIndex;
1537         }
1538
1539         void PixelProgram::LOOP(const Src &integerRegister)
1540         {
1541                 loopDepth++;
1542
1543                 iteration[loopDepth] = *Pointer<Int>(data + OFFSET(DrawData, ps.i[integerRegister.index][0]));
1544                 aL[loopDepth] = *Pointer<Int>(data + OFFSET(DrawData, ps.i[integerRegister.index][1]));
1545                 increment[loopDepth] = *Pointer<Int>(data + OFFSET(DrawData, ps.i[integerRegister.index][2]));
1546
1547                 //      If(increment[loopDepth] == 0)
1548                 //      {
1549                 //              increment[loopDepth] = 1;
1550                 //      }
1551
1552                 llvm::BasicBlock *loopBlock = Nucleus::createBasicBlock();
1553                 llvm::BasicBlock *testBlock = Nucleus::createBasicBlock();
1554                 llvm::BasicBlock *endBlock = Nucleus::createBasicBlock();
1555
1556                 loopRepTestBlock[loopRepDepth] = testBlock;
1557                 loopRepEndBlock[loopRepDepth] = endBlock;
1558
1559                 // FIXME: jump(testBlock)
1560                 Nucleus::createBr(testBlock);
1561                 Nucleus::setInsertBlock(testBlock);
1562
1563                 branch(iteration[loopDepth] > 0, loopBlock, endBlock);
1564                 Nucleus::setInsertBlock(loopBlock);
1565
1566                 iteration[loopDepth] = iteration[loopDepth] - 1;   // FIXME: --
1567
1568                 loopRepDepth++;
1569                 breakDepth = 0;
1570         }
1571
1572         void PixelProgram::REP(const Src &integerRegister)
1573         {
1574                 loopDepth++;
1575
1576                 iteration[loopDepth] = *Pointer<Int>(data + OFFSET(DrawData, ps.i[integerRegister.index][0]));
1577                 aL[loopDepth] = aL[loopDepth - 1];
1578
1579                 llvm::BasicBlock *loopBlock = Nucleus::createBasicBlock();
1580                 llvm::BasicBlock *testBlock = Nucleus::createBasicBlock();
1581                 llvm::BasicBlock *endBlock = Nucleus::createBasicBlock();
1582
1583                 loopRepTestBlock[loopRepDepth] = testBlock;
1584                 loopRepEndBlock[loopRepDepth] = endBlock;
1585
1586                 // FIXME: jump(testBlock)
1587                 Nucleus::createBr(testBlock);
1588                 Nucleus::setInsertBlock(testBlock);
1589
1590                 branch(iteration[loopDepth] > 0, loopBlock, endBlock);
1591                 Nucleus::setInsertBlock(loopBlock);
1592
1593                 iteration[loopDepth] = iteration[loopDepth] - 1;   // FIXME: --
1594
1595                 loopRepDepth++;
1596                 breakDepth = 0;
1597         }
1598
1599         void PixelProgram::WHILE(const Src &temporaryRegister)
1600         {
1601                 enableIndex++;
1602
1603                 llvm::BasicBlock *loopBlock = Nucleus::createBasicBlock();
1604                 llvm::BasicBlock *testBlock = Nucleus::createBasicBlock();
1605                 llvm::BasicBlock *endBlock = Nucleus::createBasicBlock();
1606
1607                 loopRepTestBlock[loopRepDepth] = testBlock;
1608                 loopRepEndBlock[loopRepDepth] = endBlock;
1609
1610                 Int4 restoreBreak = enableBreak;
1611                 Int4 restoreContinue = enableContinue;
1612
1613                 // FIXME: jump(testBlock)
1614                 Nucleus::createBr(testBlock);
1615                 Nucleus::setInsertBlock(testBlock);
1616                 enableContinue = restoreContinue;
1617
1618                 const Vector4f &src = fetchRegisterF(temporaryRegister);
1619                 Int4 condition = As<Int4>(src.x);
1620                 condition &= enableStack[enableIndex - 1];
1621                 enableStack[enableIndex] = condition;
1622
1623                 Bool notAllFalse = SignMask(condition) != 0;
1624                 branch(notAllFalse, loopBlock, endBlock);
1625
1626                 Nucleus::setInsertBlock(endBlock);
1627                 enableBreak = restoreBreak;
1628
1629                 Nucleus::setInsertBlock(loopBlock);
1630
1631                 loopRepDepth++;
1632                 breakDepth = 0;
1633         }
1634
1635         void PixelProgram::RET()
1636         {
1637                 if(currentLabel == -1)
1638                 {
1639                         returnBlock = Nucleus::createBasicBlock();
1640                         Nucleus::createBr(returnBlock);
1641                 }
1642                 else
1643                 {
1644                         llvm::BasicBlock *unreachableBlock = Nucleus::createBasicBlock();
1645
1646                         if(callRetBlock[currentLabel].size() > 1)   // Pop the return destination from the call stack
1647                         {
1648                                 // FIXME: Encapsulate
1649                                 UInt index = callStack[--stackIndex];
1650
1651                                 llvm::Value *value = index.loadValue();
1652                                 llvm::Value *switchInst = Nucleus::createSwitch(value, unreachableBlock, (int)callRetBlock[currentLabel].size());
1653
1654                                 for(unsigned int i = 0; i < callRetBlock[currentLabel].size(); i++)
1655                                 {
1656                                         Nucleus::addSwitchCase(switchInst, i, callRetBlock[currentLabel][i]);
1657                                 }
1658                         }
1659                         else if(callRetBlock[currentLabel].size() == 1)   // Jump directly to the unique return destination
1660                         {
1661                                 Nucleus::createBr(callRetBlock[currentLabel][0]);
1662                         }
1663                         else   // Function isn't called
1664                         {
1665                                 Nucleus::createBr(unreachableBlock);
1666                         }
1667
1668                         Nucleus::setInsertBlock(unreachableBlock);
1669                         Nucleus::createUnreachable();
1670                 }
1671         }
1672
1673         void PixelProgram::LEAVE()
1674         {
1675                 enableLeave = enableLeave & ~enableStack[enableIndex];
1676
1677                 // FIXME: Return from function if all instances left
1678                 // FIXME: Use enableLeave in other control-flow constructs
1679         }
1680 }