2 * Copyright 2011 Christoph Bumiller
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
23 #include "codegen/nv50_ir.h"
24 #include "codegen/nv50_ir_target_nv50.h"
28 #define NV50_OP_ENC_LONG 0
29 #define NV50_OP_ENC_SHORT 1
30 #define NV50_OP_ENC_IMM 2
31 #define NV50_OP_ENC_LONG_ALT 3
33 class CodeEmitterNV50 : public CodeEmitter
36 CodeEmitterNV50(const TargetNV50 *);
38 virtual bool emitInstruction(Instruction *);
40 virtual uint32_t getMinEncodingSize(const Instruction *) const;
42 inline void setProgramType(Program::Type pType) { progType = pType; }
44 virtual void prepareEmission(Function *);
47 Program::Type progType;
49 const TargetNV50 *targNV50;
52 inline void defId(const ValueDef&, const int pos);
53 inline void srcId(const ValueRef&, const int pos);
54 inline void srcId(const ValueRef *, const int pos);
56 inline void srcAddr16(const ValueRef&, bool adj, const int pos);
57 inline void srcAddr8(const ValueRef&, const int pos);
59 void emitFlagsRd(const Instruction *);
60 void emitFlagsWr(const Instruction *);
62 void emitCondCode(CondCode cc, DataType ty, int pos);
64 inline void setARegBits(unsigned int);
66 void setAReg16(const Instruction *, int s);
67 void setImmediate(const Instruction *, int s);
69 void setDst(const Value *);
70 void setDst(const Instruction *, int d);
71 void setSrcFileBits(const Instruction *, int enc);
72 void setSrc(const Instruction *, unsigned int s, int slot);
74 void emitForm_MAD(const Instruction *);
75 void emitForm_ADD(const Instruction *);
76 void emitForm_MUL(const Instruction *);
77 void emitForm_IMM(const Instruction *);
79 void emitLoadStoreSizeLG(DataType ty, int pos);
80 void emitLoadStoreSizeCS(DataType ty);
82 void roundMode_MAD(const Instruction *);
83 void roundMode_CVT(RoundMode);
85 void emitMNeg12(const Instruction *);
87 void emitLOAD(const Instruction *);
88 void emitSTORE(const Instruction *);
89 void emitMOV(const Instruction *);
90 void emitRDSV(const Instruction *);
92 void emitINTERP(const Instruction *);
93 void emitPFETCH(const Instruction *);
94 void emitOUT(const Instruction *);
96 void emitUADD(const Instruction *);
97 void emitAADD(const Instruction *);
98 void emitFADD(const Instruction *);
99 void emitIMUL(const Instruction *);
100 void emitFMUL(const Instruction *);
101 void emitFMAD(const Instruction *);
102 void emitIMAD(const Instruction *);
103 void emitISAD(const Instruction *);
105 void emitMINMAX(const Instruction *);
107 void emitPreOp(const Instruction *);
108 void emitSFnOp(const Instruction *, uint8_t subOp);
110 void emitShift(const Instruction *);
111 void emitARL(const Instruction *, unsigned int shl);
112 void emitLogicOp(const Instruction *);
113 void emitNOT(const Instruction *);
115 void emitCVT(const Instruction *);
116 void emitSET(const Instruction *);
118 void emitTEX(const TexInstruction *);
119 void emitTXQ(const TexInstruction *);
120 void emitTEXPREP(const TexInstruction *);
122 void emitQUADOP(const Instruction *, uint8_t lane, uint8_t quOp);
124 void emitFlow(const Instruction *, uint8_t flowOp);
125 void emitPRERETEmu(const FlowInstruction *);
126 void emitBAR(const Instruction *);
128 void emitATOM(const Instruction *);
131 #define SDATA(a) ((a).rep()->reg.data)
132 #define DDATA(a) ((a).rep()->reg.data)
134 void CodeEmitterNV50::srcId(const ValueRef& src, const int pos)
137 code[pos / 32] |= SDATA(src).id << (pos % 32);
140 void CodeEmitterNV50::srcId(const ValueRef *src, const int pos)
143 code[pos / 32] |= SDATA(*src).id << (pos % 32);
146 void CodeEmitterNV50::srcAddr16(const ValueRef& src, bool adj, const int pos)
150 int32_t offset = SDATA(src).offset;
152 assert(!adj || src.get()->reg.size <= 4);
154 offset /= src.get()->reg.size;
156 assert(offset <= 0x7fff && offset >= (int32_t)-0x8000 && (pos % 32) <= 16);
159 offset &= adj ? (0xffff >> (src.get()->reg.size >> 1)) : 0xffff;
161 code[pos / 32] |= offset << (pos % 32);
164 void CodeEmitterNV50::srcAddr8(const ValueRef& src, const int pos)
168 uint32_t offset = SDATA(src).offset;
170 assert((offset <= 0x1fc || offset == 0x3fc) && !(offset & 0x3));
172 code[pos / 32] |= (offset >> 2) << (pos % 32);
175 void CodeEmitterNV50::defId(const ValueDef& def, const int pos)
177 assert(def.get() && def.getFile() != FILE_SHADER_OUTPUT);
179 code[pos / 32] |= DDATA(def).id << (pos % 32);
183 CodeEmitterNV50::roundMode_MAD(const Instruction *insn)
186 case ROUND_M: code[1] |= 1 << 22; break;
187 case ROUND_P: code[1] |= 2 << 22; break;
188 case ROUND_Z: code[1] |= 3 << 22; break;
190 assert(insn->rnd == ROUND_N);
196 CodeEmitterNV50::emitMNeg12(const Instruction *i)
198 code[1] |= i->src(0).mod.neg() << 26;
199 code[1] |= i->src(1).mod.neg() << 27;
202 void CodeEmitterNV50::emitCondCode(CondCode cc, DataType ty, int pos)
206 assert(pos >= 32 || pos <= 27);
209 case CC_LT: enc = 0x1; break;
210 case CC_LTU: enc = 0x9; break;
211 case CC_EQ: enc = 0x2; break;
212 case CC_EQU: enc = 0xa; break;
213 case CC_LE: enc = 0x3; break;
214 case CC_LEU: enc = 0xb; break;
215 case CC_GT: enc = 0x4; break;
216 case CC_GTU: enc = 0xc; break;
217 case CC_NE: enc = 0x5; break;
218 case CC_NEU: enc = 0xd; break;
219 case CC_GE: enc = 0x6; break;
220 case CC_GEU: enc = 0xe; break;
221 case CC_TR: enc = 0xf; break;
222 case CC_FL: enc = 0x0; break;
224 case CC_O: enc = 0x10; break;
225 case CC_C: enc = 0x11; break;
226 case CC_A: enc = 0x12; break;
227 case CC_S: enc = 0x13; break;
228 case CC_NS: enc = 0x1c; break;
229 case CC_NA: enc = 0x1d; break;
230 case CC_NC: enc = 0x1e; break;
231 case CC_NO: enc = 0x1f; break;
235 assert(!"invalid condition code");
238 if (ty != TYPE_NONE && !isFloatType(ty))
239 enc &= ~0x8; // unordered only exists for float types
241 code[pos / 32] |= enc << (pos % 32);
245 CodeEmitterNV50::emitFlagsRd(const Instruction *i)
247 int s = (i->flagsSrc >= 0) ? i->flagsSrc : i->predSrc;
249 assert(!(code[1] & 0x00003f80));
252 assert(i->getSrc(s)->reg.file == FILE_FLAGS);
253 emitCondCode(i->cc, TYPE_NONE, 32 + 7);
254 srcId(i->src(s), 32 + 12);
261 CodeEmitterNV50::emitFlagsWr(const Instruction *i)
263 assert(!(code[1] & 0x70));
265 int flagsDef = i->flagsDef;
267 // find flags definition and check that it is the last def
269 for (int d = 0; i->defExists(d); ++d)
270 if (i->def(d).getFile() == FILE_FLAGS)
272 if (flagsDef >= 0 && 0) // TODO: enforce use of flagsDef at some point
273 WARN("Instruction::flagsDef was not set properly\n");
275 if (flagsDef == 0 && i->defExists(1))
276 WARN("flags def should not be the primary definition\n");
279 code[1] |= (DDATA(i->def(flagsDef)).id << 4) | 0x40;
284 CodeEmitterNV50::setARegBits(unsigned int u)
286 code[0] |= (u & 3) << 26;
291 CodeEmitterNV50::setAReg16(const Instruction *i, int s)
293 if (i->srcExists(s)) {
294 s = i->src(s).indirect[0];
296 setARegBits(SDATA(i->src(s)).id + 1);
301 CodeEmitterNV50::setImmediate(const Instruction *i, int s)
303 const ImmediateValue *imm = i->src(s).get()->asImm();
306 uint32_t u = imm->reg.data.u32;
308 if (i->src(s).mod & Modifier(NV50_IR_MOD_NOT))
312 code[0] |= (u & 0x3f) << 16;
313 code[1] |= (u >> 6) << 2;
317 CodeEmitterNV50::setDst(const Value *dst)
319 const Storage *reg = &dst->join->reg;
321 assert(reg->file != FILE_ADDRESS);
323 if (reg->data.id < 0 || reg->file == FILE_FLAGS) {
324 code[0] |= (127 << 2) | 1;
328 if (reg->file == FILE_SHADER_OUTPUT) {
330 id = reg->data.offset / 4;
339 CodeEmitterNV50::setDst(const Instruction *i, int d)
341 if (i->defExists(d)) {
342 setDst(i->getDef(d));
345 code[0] |= 0x01fc; // bit bucket
356 CodeEmitterNV50::setSrcFileBits(const Instruction *i, int enc)
360 for (unsigned int s = 0; s < Target::operationSrcNr[i->op]; ++s) {
361 switch (i->src(s).getFile()) {
364 case FILE_MEMORY_SHARED:
365 case FILE_SHADER_INPUT:
366 mode |= 1 << (s * 2);
368 case FILE_MEMORY_CONST:
369 mode |= 2 << (s * 2);
372 mode |= 3 << (s * 2);
375 ERROR("invalid file on source %i: %u\n", s, i->src(s).getFile());
383 case 0x01: // arr/grr
384 if (progType == Program::TYPE_GEOMETRY && i->src(0).isIndirect(0)) {
385 code[0] |= 0x01800000;
386 if (enc == NV50_OP_ENC_LONG || enc == NV50_OP_ENC_LONG_ALT)
387 code[1] |= 0x00200000;
389 if (enc == NV50_OP_ENC_SHORT)
390 code[0] |= 0x01000000;
392 code[1] |= 0x00200000;
396 assert(i->op == OP_MOV);
401 assert(progType == Program::TYPE_GEOMETRY ||
402 progType == Program::TYPE_COMPUTE);
403 code[0] |= 0x01000000;
404 if (progType == Program::TYPE_GEOMETRY && i->src(0).isIndirect(0)) {
405 int reg = i->src(0).getIndirect(0)->rep()->reg.data.id;
407 code[0] |= (reg + 1) << 26;
411 code[0] |= (enc == NV50_OP_ENC_LONG_ALT) ? 0x01000000 : 0x00800000;
412 code[1] |= (i->getSrc(1)->reg.fileIndex << 22);
414 case 0x09: // acr/gcr
415 if (progType == Program::TYPE_GEOMETRY && i->src(0).isIndirect(0)) {
416 code[0] |= 0x01800000;
418 code[0] |= (enc == NV50_OP_ENC_LONG_ALT) ? 0x01000000 : 0x00800000;
419 code[1] |= 0x00200000;
421 code[1] |= (i->getSrc(1)->reg.fileIndex << 22);
424 code[0] |= 0x01000000;
425 code[1] |= (i->getSrc(2)->reg.fileIndex << 22);
428 code[0] |= 0x01000000;
429 code[1] |= 0x00200000 | (i->getSrc(2)->reg.fileIndex << 22);
430 assert(progType != Program::TYPE_GEOMETRY);
433 ERROR("not encodable: %x\n", mode);
437 if (progType != Program::TYPE_COMPUTE)
440 if ((mode & 3) == 1) {
441 const int pos = i->src(1).getFile() == FILE_IMMEDIATE ? 13 : 14;
443 switch (i->getSrc(0)->reg.type) {
454 assert(i->getSrc(0)->reg.size == 4);
461 CodeEmitterNV50::setSrc(const Instruction *i, unsigned int s, int slot)
463 if (Target::operationSrcNr[i->op] <= s)
465 const Storage *reg = &i->src(s).rep()->reg;
467 unsigned int id = (reg->file == FILE_GPR) ?
469 reg->data.offset >> (reg->size >> 1); // no > 4 byte sources here
472 case 0: code[0] |= id << 9; break;
473 case 1: code[0] |= id << 16; break;
474 case 2: code[1] |= id << 14; break;
482 // - long instruction
483 // - 1 to 3 sources in slots 0, 1, 2 (rrr, arr, rcr, acr, rrc, arc, gcr, grr)
486 CodeEmitterNV50::emitForm_MAD(const Instruction *i)
488 assert(i->encSize == 8);
496 setSrcFileBits(i, NV50_OP_ENC_LONG);
501 if (i->getIndirect(0, 0)) {
502 assert(!i->srcExists(1) || !i->getIndirect(1, 0));
503 assert(!i->srcExists(2) || !i->getIndirect(2, 0));
505 } else if (i->srcExists(1) && i->getIndirect(1, 0)) {
506 assert(!i->srcExists(2) || !i->getIndirect(2, 0));
513 // like default form, but 2nd source in slot 2, and no 3rd source
515 CodeEmitterNV50::emitForm_ADD(const Instruction *i)
517 assert(i->encSize == 8);
525 setSrcFileBits(i, NV50_OP_ENC_LONG_ALT);
529 if (i->getIndirect(0, 0)) {
530 assert(!i->getIndirect(1, 0));
537 // default short form (rr, ar, rc, gr)
539 CodeEmitterNV50::emitForm_MUL(const Instruction *i)
541 assert(i->encSize == 4 && !(code[0] & 1));
542 assert(i->defExists(0));
543 assert(!i->getPredicate());
547 setSrcFileBits(i, NV50_OP_ENC_SHORT);
552 // usual immediate form
553 // - 1 to 3 sources where second is immediate (rir, gir)
554 // - no address or predicate possible
556 CodeEmitterNV50::emitForm_IMM(const Instruction *i)
558 assert(i->encSize == 8);
561 assert(i->defExists(0) && i->srcExists(0));
565 setSrcFileBits(i, NV50_OP_ENC_IMM);
566 if (Target::operationSrcNr[i->op] > 1) {
569 // If there is another source, it has to be the same as the dest reg.
576 CodeEmitterNV50::emitLoadStoreSizeLG(DataType ty, int pos)
581 case TYPE_F32: // fall through
582 case TYPE_S32: // fall through
583 case TYPE_U32: enc = 0x6; break;
584 case TYPE_B128: enc = 0x5; break;
585 case TYPE_F64: // fall through
586 case TYPE_S64: // fall through
587 case TYPE_U64: enc = 0x4; break;
588 case TYPE_S16: enc = 0x3; break;
589 case TYPE_U16: enc = 0x2; break;
590 case TYPE_S8: enc = 0x1; break;
591 case TYPE_U8: enc = 0x0; break;
594 assert(!"invalid load/store type");
597 code[pos / 32] |= enc << (pos % 32);
601 CodeEmitterNV50::emitLoadStoreSizeCS(DataType ty)
605 case TYPE_U16: code[1] |= 0x4000; break;
606 case TYPE_S16: code[1] |= 0x8000; break;
609 case TYPE_U32: code[1] |= 0xc000; break;
617 CodeEmitterNV50::emitLOAD(const Instruction *i)
619 DataFile sf = i->src(0).getFile();
620 int32_t offset = i->getSrc(0)->reg.data.offset;
623 case FILE_SHADER_INPUT:
624 if (progType == Program::TYPE_GEOMETRY && i->src(0).isIndirect(0))
625 code[0] = 0x11800001;
627 // use 'mov' where we can
628 code[0] = i->src(0).isIndirect(0) ? 0x00000001 : 0x10000001;
629 code[1] = 0x00200000 | (i->lanes << 14);
630 if (typeSizeof(i->dType) == 4)
631 code[1] |= 0x04000000;
633 case FILE_MEMORY_SHARED:
634 if (targ->getChipset() >= 0x84) {
635 assert(offset <= (int32_t)(0x3fff * typeSizeof(i->sType)));
636 code[0] = 0x10000001;
637 code[1] = 0x40000000;
639 if (typeSizeof(i->dType) == 4)
640 code[1] |= 0x04000000;
642 emitLoadStoreSizeCS(i->sType);
644 assert(offset <= (int32_t)(0x1f * typeSizeof(i->sType)));
645 code[0] = 0x10000001;
646 code[1] = 0x00200000 | (i->lanes << 14);
647 emitLoadStoreSizeCS(i->sType);
650 case FILE_MEMORY_CONST:
651 code[0] = 0x10000001;
652 code[1] = 0x20000000 | (i->getSrc(0)->reg.fileIndex << 22);
653 if (typeSizeof(i->dType) == 4)
654 code[1] |= 0x04000000;
655 emitLoadStoreSizeCS(i->sType);
657 case FILE_MEMORY_LOCAL:
658 code[0] = 0xd0000001;
659 code[1] = 0x40000000;
661 case FILE_MEMORY_GLOBAL:
662 code[0] = 0xd0000001 | (i->getSrc(0)->reg.fileIndex << 16);
663 code[1] = 0x80000000;
666 assert(!"invalid load source file");
669 if (sf == FILE_MEMORY_LOCAL ||
670 sf == FILE_MEMORY_GLOBAL)
671 emitLoadStoreSizeLG(i->sType, 21 + 32);
678 if (i->src(0).getFile() == FILE_MEMORY_GLOBAL) {
679 srcId(*i->src(0).getIndirect(0), 9);
682 srcAddr16(i->src(0), i->src(0).getFile() != FILE_MEMORY_LOCAL, 9);
687 CodeEmitterNV50::emitSTORE(const Instruction *i)
689 DataFile f = i->getSrc(0)->reg.file;
690 int32_t offset = i->getSrc(0)->reg.data.offset;
693 case FILE_SHADER_OUTPUT:
694 code[0] = 0x00000001 | ((offset >> 2) << 9);
695 code[1] = 0x80c00000;
696 srcId(i->src(1), 32 + 14);
698 case FILE_MEMORY_GLOBAL:
699 code[0] = 0xd0000001 | (i->getSrc(0)->reg.fileIndex << 16);
700 code[1] = 0xa0000000;
701 emitLoadStoreSizeLG(i->dType, 21 + 32);
704 case FILE_MEMORY_LOCAL:
705 code[0] = 0xd0000001;
706 code[1] = 0x60000000;
707 emitLoadStoreSizeLG(i->dType, 21 + 32);
710 case FILE_MEMORY_SHARED:
711 code[0] = 0x00000001;
712 code[1] = 0xe0000000;
713 switch (typeSizeof(i->dType)) {
715 code[0] |= offset << 9;
716 code[1] |= 0x00400000;
719 code[0] |= (offset >> 1) << 9;
722 code[0] |= (offset >> 2) << 9;
723 code[1] |= 0x04200000;
729 srcId(i->src(1), 32 + 14);
732 assert(!"invalid store destination file");
736 if (f == FILE_MEMORY_GLOBAL)
737 srcId(*i->src(0).getIndirect(0), 9);
741 if (f == FILE_MEMORY_LOCAL)
742 srcAddr16(i->src(0), false, 9);
748 CodeEmitterNV50::emitMOV(const Instruction *i)
750 DataFile sf = i->getSrc(0)->reg.file;
751 DataFile df = i->getDef(0)->reg.file;
753 assert(sf == FILE_GPR || df == FILE_GPR);
755 if (sf == FILE_FLAGS) {
756 code[0] = 0x00000001;
757 code[1] = 0x20000000;
759 srcId(i->src(0), 12);
762 if (sf == FILE_ADDRESS) {
763 code[0] = 0x00000001;
764 code[1] = 0x40000000;
766 setARegBits(SDATA(i->src(0)).id + 1);
769 if (df == FILE_FLAGS) {
770 code[0] = 0x00000001;
771 code[1] = 0xa0000000;
776 if (sf == FILE_IMMEDIATE) {
777 code[0] = 0x10008001;
778 code[1] = 0x00000003;
781 if (i->encSize == 4) {
782 code[0] = 0x10008000;
784 code[0] = 0x10000001;
785 code[1] = (typeSizeof(i->dType) == 2) ? 0 : 0x04000000;
786 code[1] |= (i->lanes << 14);
792 if (df == FILE_SHADER_OUTPUT) {
793 assert(i->encSize == 8);
798 static inline uint8_t getSRegEncoding(const ValueRef &ref)
800 switch (SDATA(ref).sv.sv) {
801 case SV_PHYSID: return 0;
802 case SV_CLOCK: return 1;
803 case SV_VERTEX_STRIDE: return 3;
804 // case SV_PM_COUNTER: return 4 + SDATA(ref).sv.index;
805 case SV_SAMPLE_INDEX: return 8;
807 assert(!"no sreg for system value");
813 CodeEmitterNV50::emitRDSV(const Instruction *i)
815 code[0] = 0x00000001;
816 code[1] = 0x60000000 | (getSRegEncoding(i->src(0)) << 14);
822 CodeEmitterNV50::emitNOP()
824 code[0] = 0xf0000001;
825 code[1] = 0xe0000000;
829 CodeEmitterNV50::emitQUADOP(const Instruction *i, uint8_t lane, uint8_t quOp)
831 code[0] = 0xc0000000 | (lane << 16);
832 code[1] = 0x80000000;
834 code[0] |= (quOp & 0x03) << 20;
835 code[1] |= (quOp & 0xfc) << 20;
839 if (!i->srcExists(1))
840 srcId(i->src(0), 32 + 14);
843 /* NOTE: This returns the base address of a vertex inside the primitive.
844 * src0 is an immediate, the index (not offset) of the vertex
845 * inside the primitive. XXX: signed or unsigned ?
846 * src1 (may be NULL) should use whatever units the hardware requires
847 * (on nv50 this is bytes, so, relative index * 4; signed 16 bit value).
850 CodeEmitterNV50::emitPFETCH(const Instruction *i)
852 const uint32_t prim = i->src(0).get()->reg.data.u32;
855 if (i->def(0).getFile() == FILE_ADDRESS) {
857 code[0] = 0x00000001 | ((DDATA(i->def(0)).id + 1) << 2);
858 code[1] = 0xc0200000;
859 code[0] |= prim << 9;
860 assert(!i->srcExists(1));
862 if (i->srcExists(1)) {
863 // ld b32 $rX a[$aX+base]
864 code[0] = 0x00000001;
865 code[1] = 0x04200000 | (0xf << 14);
867 code[0] |= prim << 9;
868 setARegBits(SDATA(i->src(1)).id + 1);
871 code[0] = 0x10000001;
872 code[1] = 0x04200000 | (0xf << 14);
874 code[0] |= prim << 9;
880 CodeEmitterNV50::emitINTERP(const Instruction *i)
882 code[0] = 0x80000000;
885 srcAddr8(i->src(0), 16);
887 if (i->encSize != 8 && i->getInterpMode() == NV50_IR_INTERP_FLAT) {
890 if (i->op == OP_PINTERP) {
894 if (i->getSampleMode() == NV50_IR_INTERP_CENTROID)
898 if (i->encSize == 8) {
899 if (i->getInterpMode() == NV50_IR_INTERP_FLAT)
902 code[1] = (code[0] & (3 << 24)) >> (24 - 16);
903 code[0] &= ~0x03000000;
910 CodeEmitterNV50::emitMINMAX(const Instruction *i)
912 if (i->dType == TYPE_F64) {
913 code[0] = 0xe0000000;
914 code[1] = (i->op == OP_MIN) ? 0xa0000000 : 0xc0000000;
916 code[0] = 0x30000000;
917 code[1] = 0x80000000;
919 code[1] |= 0x20000000;
922 case TYPE_F32: code[0] |= 0x80000000; break;
923 case TYPE_S32: code[1] |= 0x8c000000; break;
924 case TYPE_U32: code[1] |= 0x84000000; break;
925 case TYPE_S16: code[1] |= 0x80000000; break;
926 case TYPE_U16: break;
931 code[1] |= i->src(0).mod.abs() << 20;
932 code[1] |= i->src(0).mod.neg() << 26;
933 code[1] |= i->src(1).mod.abs() << 19;
934 code[1] |= i->src(1).mod.neg() << 27;
940 CodeEmitterNV50::emitFMAD(const Instruction *i)
942 const int neg_mul = i->src(0).mod.neg() ^ i->src(1).mod.neg();
943 const int neg_add = i->src(2).mod.neg();
945 code[0] = 0xe0000000;
947 if (i->src(1).getFile() == FILE_IMMEDIATE) {
950 code[0] |= neg_mul << 15;
951 code[0] |= neg_add << 22;
955 if (i->encSize == 4) {
957 code[0] |= neg_mul << 15;
958 code[0] |= neg_add << 22;
962 code[1] = neg_mul << 26;
963 code[1] |= neg_add << 27;
971 CodeEmitterNV50::emitFADD(const Instruction *i)
973 const int neg0 = i->src(0).mod.neg();
974 const int neg1 = i->src(1).mod.neg() ^ ((i->op == OP_SUB) ? 1 : 0);
976 code[0] = 0xb0000000;
978 assert(!(i->src(0).mod | i->src(1).mod).abs());
980 if (i->src(1).getFile() == FILE_IMMEDIATE) {
983 code[0] |= neg0 << 15;
984 code[0] |= neg1 << 22;
988 if (i->encSize == 8) {
991 code[1] |= neg0 << 26;
992 code[1] |= neg1 << 27;
997 code[0] |= neg0 << 15;
998 code[0] |= neg1 << 22;
1005 CodeEmitterNV50::emitUADD(const Instruction *i)
1007 const int neg0 = i->src(0).mod.neg();
1008 const int neg1 = i->src(1).mod.neg() ^ ((i->op == OP_SUB) ? 1 : 0);
1010 code[0] = 0x20008000;
1012 if (i->src(1).getFile() == FILE_IMMEDIATE) {
1016 if (i->encSize == 8) {
1017 code[0] = 0x20000000;
1018 code[1] = (typeSizeof(i->dType) == 2) ? 0 : 0x04000000;
1023 assert(!(neg0 && neg1));
1024 code[0] |= neg0 << 28;
1025 code[0] |= neg1 << 22;
1027 if (i->flagsSrc >= 0) {
1028 // addc == sub | subr
1029 assert(!(code[0] & 0x10400000) && !i->getPredicate());
1030 code[0] |= 0x10400000;
1031 srcId(i->src(i->flagsSrc), 32 + 12);
1036 CodeEmitterNV50::emitAADD(const Instruction *i)
1038 const int s = (i->op == OP_MOV) ? 0 : 1;
1040 code[0] = 0xd0000001 | (i->getSrc(s)->reg.data.u16 << 9);
1041 code[1] = 0x20000000;
1043 code[0] |= (DDATA(i->def(0)).id + 1) << 2;
1047 if (s && i->srcExists(0))
1048 setARegBits(SDATA(i->src(0)).id + 1);
1052 CodeEmitterNV50::emitIMUL(const Instruction *i)
1054 code[0] = 0x40000000;
1056 if (i->encSize == 8) {
1057 code[1] = (i->sType == TYPE_S16) ? (0x8000 | 0x4000) : 0x0000;
1060 if (i->sType == TYPE_S16)
1067 CodeEmitterNV50::emitFMUL(const Instruction *i)
1069 const int neg = (i->src(0).mod ^ i->src(1).mod).neg();
1071 code[0] = 0xc0000000;
1073 if (i->src(1).getFile() == FILE_IMMEDIATE) {
1081 if (i->encSize == 8) {
1082 code[1] = i->rnd == ROUND_Z ? 0x0000c000 : 0;
1084 code[1] |= 0x08000000;
1098 CodeEmitterNV50::emitIMAD(const Instruction *i)
1100 code[0] = 0x60000000;
1101 if (isSignedType(i->sType))
1102 code[1] = i->saturate ? 0x40000000 : 0x20000000;
1104 code[1] = 0x00000000;
1106 int neg1 = i->src(0).mod.neg() ^ i->src(1).mod.neg();
1107 int neg2 = i->src(2).mod.neg();
1109 assert(!(neg1 & neg2));
1110 code[1] |= neg1 << 27;
1111 code[1] |= neg2 << 26;
1115 if (i->flagsSrc >= 0) {
1116 // add with carry from $cX
1117 assert(!(code[1] & 0x0c000000) && !i->getPredicate());
1118 code[1] |= 0xc << 24;
1119 srcId(i->src(i->flagsSrc), 32 + 12);
1124 CodeEmitterNV50::emitISAD(const Instruction *i)
1126 if (i->encSize == 8) {
1127 code[0] = 0x50000000;
1129 case TYPE_U32: code[1] = 0x04000000; break;
1130 case TYPE_S32: code[1] = 0x0c000000; break;
1131 case TYPE_U16: code[1] = 0x00000000; break;
1132 case TYPE_S16: code[1] = 0x08000000; break;
1140 case TYPE_U32: code[0] = 0x50008000; break;
1141 case TYPE_S32: code[0] = 0x50008100; break;
1142 case TYPE_U16: code[0] = 0x50000000; break;
1143 case TYPE_S16: code[0] = 0x50000100; break;
1153 CodeEmitterNV50::emitSET(const Instruction *i)
1155 code[0] = 0x30000000;
1156 code[1] = 0x60000000;
1158 emitCondCode(i->asCmp()->setCond, i->sType, 32 + 14);
1161 case TYPE_F32: code[0] |= 0x80000000; break;
1162 case TYPE_S32: code[1] |= 0x0c000000; break;
1163 case TYPE_U32: code[1] |= 0x04000000; break;
1164 case TYPE_S16: code[1] |= 0x08000000; break;
1165 case TYPE_U16: break;
1170 if (i->src(0).mod.neg()) code[1] |= 0x04000000;
1171 if (i->src(1).mod.neg()) code[1] |= 0x08000000;
1172 if (i->src(0).mod.abs()) code[1] |= 0x00100000;
1173 if (i->src(1).mod.abs()) code[1] |= 0x00080000;
1179 CodeEmitterNV50::roundMode_CVT(RoundMode rnd)
1182 case ROUND_NI: code[1] |= 0x08000000; break;
1183 case ROUND_M: code[1] |= 0x00020000; break;
1184 case ROUND_MI: code[1] |= 0x08020000; break;
1185 case ROUND_P: code[1] |= 0x00040000; break;
1186 case ROUND_PI: code[1] |= 0x08040000; break;
1187 case ROUND_Z: code[1] |= 0x00060000; break;
1188 case ROUND_ZI: code[1] |= 0x08060000; break;
1190 assert(rnd == ROUND_N);
1196 CodeEmitterNV50::emitCVT(const Instruction *i)
1198 const bool f2f = isFloatType(i->dType) && isFloatType(i->sType);
1203 case OP_CEIL: rnd = f2f ? ROUND_PI : ROUND_P; break;
1204 case OP_FLOOR: rnd = f2f ? ROUND_MI : ROUND_M; break;
1205 case OP_TRUNC: rnd = f2f ? ROUND_ZI : ROUND_Z; break;
1211 if (i->op == OP_NEG && i->dType == TYPE_U32)
1216 code[0] = 0xa0000000;
1221 case TYPE_F64: code[1] = 0xc4404000; break;
1222 case TYPE_S64: code[1] = 0x44414000; break;
1223 case TYPE_U64: code[1] = 0x44404000; break;
1224 case TYPE_F32: code[1] = 0xc4400000; break;
1225 case TYPE_S32: code[1] = 0x44410000; break;
1226 case TYPE_U32: code[1] = 0x44400000; break;
1234 case TYPE_F64: code[1] = 0x8c404000; break;
1235 case TYPE_F32: code[1] = 0x8c400000; break;
1243 case TYPE_F64: code[1] = 0x84404000; break;
1244 case TYPE_F32: code[1] = 0x84400000; break;
1252 case TYPE_F64: code[1] = 0xc0404000; break;
1253 case TYPE_S64: code[1] = 0x40414000; break;
1254 case TYPE_U64: code[1] = 0x40404000; break;
1255 case TYPE_F32: code[1] = 0xc4004000; break;
1256 case TYPE_S32: code[1] = 0x44014000; break;
1257 case TYPE_U32: code[1] = 0x44004000; break;
1258 case TYPE_F16: code[1] = 0xc4000000; break;
1259 case TYPE_U16: code[1] = 0x44000000; break;
1267 case TYPE_F64: code[1] = 0x88404000; break;
1268 case TYPE_F32: code[1] = 0x8c004000; break;
1269 case TYPE_S32: code[1] = 0x0c014000; break;
1270 case TYPE_U32: code[1] = 0x0c004000; break;
1271 case TYPE_F16: code[1] = 0x8c000000; break;
1272 case TYPE_S16: code[1] = 0x0c010000; break;
1273 case TYPE_U16: code[1] = 0x0c000000; break;
1274 case TYPE_S8: code[1] = 0x0c018000; break;
1275 case TYPE_U8: code[1] = 0x0c008000; break;
1283 case TYPE_F64: code[1] = 0x80404000; break;
1284 case TYPE_F32: code[1] = 0x84004000; break;
1285 case TYPE_S32: code[1] = 0x04014000; break;
1286 case TYPE_U32: code[1] = 0x04004000; break;
1287 case TYPE_F16: code[1] = 0x84000000; break;
1288 case TYPE_S16: code[1] = 0x04010000; break;
1289 case TYPE_U16: code[1] = 0x04000000; break;
1290 case TYPE_S8: code[1] = 0x04018000; break;
1291 case TYPE_U8: code[1] = 0x04008000; break;
1305 if (typeSizeof(i->sType) == 1 && i->getSrc(0)->reg.size == 4)
1306 code[1] |= 0x00004000;
1311 case OP_ABS: code[1] |= 1 << 20; break;
1312 case OP_SAT: code[1] |= 1 << 19; break;
1313 case OP_NEG: code[1] |= 1 << 29; break;
1317 code[1] ^= i->src(0).mod.neg() << 29;
1318 code[1] |= i->src(0).mod.abs() << 20;
1322 assert(i->op != OP_ABS || !i->src(0).mod.neg());
1328 CodeEmitterNV50::emitPreOp(const Instruction *i)
1330 code[0] = 0xb0000000;
1331 code[1] = (i->op == OP_PREEX2) ? 0xc0004000 : 0xc0000000;
1333 code[1] |= i->src(0).mod.abs() << 20;
1334 code[1] |= i->src(0).mod.neg() << 26;
1340 CodeEmitterNV50::emitSFnOp(const Instruction *i, uint8_t subOp)
1342 code[0] = 0x90000000;
1344 if (i->encSize == 4) {
1345 assert(i->op == OP_RCP);
1346 code[0] |= i->src(0).mod.abs() << 15;
1347 code[0] |= i->src(0).mod.neg() << 22;
1350 code[1] = subOp << 29;
1351 code[1] |= i->src(0).mod.abs() << 20;
1352 code[1] |= i->src(0).mod.neg() << 26;
1358 CodeEmitterNV50::emitNOT(const Instruction *i)
1360 code[0] = 0xd0000000;
1361 code[1] = 0x0002c000;
1366 code[1] |= 0x04000000;
1376 CodeEmitterNV50::emitLogicOp(const Instruction *i)
1378 code[0] = 0xd0000000;
1381 if (i->src(1).getFile() == FILE_IMMEDIATE) {
1383 case OP_OR: code[0] |= 0x0100; break;
1384 case OP_XOR: code[0] |= 0x8000; break;
1386 assert(i->op == OP_AND);
1389 if (i->src(0).mod & Modifier(NV50_IR_MOD_NOT))
1395 case OP_AND: code[1] = 0x04000000; break;
1396 case OP_OR: code[1] = 0x04004000; break;
1397 case OP_XOR: code[1] = 0x04008000; break;
1402 if (i->src(0).mod & Modifier(NV50_IR_MOD_NOT))
1404 if (i->src(1).mod & Modifier(NV50_IR_MOD_NOT))
1412 CodeEmitterNV50::emitARL(const Instruction *i, unsigned int shl)
1414 code[0] = 0x00000001 | (shl << 16);
1415 code[1] = 0xc0000000;
1417 code[0] |= (DDATA(i->def(0)).id + 1) << 2;
1419 setSrcFileBits(i, NV50_OP_ENC_IMM);
1425 CodeEmitterNV50::emitShift(const Instruction *i)
1427 if (i->def(0).getFile() == FILE_ADDRESS) {
1428 assert(i->srcExists(1) && i->src(1).getFile() == FILE_IMMEDIATE);
1429 emitARL(i, i->getSrc(1)->reg.data.u32 & 0x3f);
1431 code[0] = 0x30000001;
1432 code[1] = (i->op == OP_SHR) ? 0xe4000000 : 0xc4000000;
1433 if (i->op == OP_SHR && isSignedType(i->sType))
1436 if (i->src(1).getFile() == FILE_IMMEDIATE) {
1438 code[0] |= (i->getSrc(1)->reg.data.u32 & 0x7f) << 16;
1439 defId(i->def(0), 2);
1440 srcId(i->src(0), 9);
1449 CodeEmitterNV50::emitOUT(const Instruction *i)
1451 code[0] = (i->op == OP_EMIT) ? 0xf0000201 : 0xf0000401;
1452 code[1] = 0xc0000000;
1458 CodeEmitterNV50::emitTEX(const TexInstruction *i)
1460 code[0] = 0xf0000001;
1461 code[1] = 0x00000000;
1465 code[1] = 0x20000000;
1468 code[1] = 0x40000000;
1471 code[0] |= 0x01000000;
1474 code[0] |= 0x01000000;
1475 code[1] = 0x80000000;
1478 code[1] = 0x60020000;
1481 assert(i->op == OP_TEX);
1485 code[0] |= i->tex.r << 9;
1486 code[0] |= i->tex.s << 17;
1488 int argc = i->tex.target.getArgCount();
1490 if (i->op == OP_TXB || i->op == OP_TXL || i->op == OP_TXF)
1492 if (i->tex.target.isShadow())
1496 code[0] |= (argc - 1) << 22;
1498 if (i->tex.target.isCube()) {
1499 code[0] |= 0x08000000;
1501 if (i->tex.useOffsets) {
1502 code[1] |= (i->tex.offset[0] & 0xf) << 24;
1503 code[1] |= (i->tex.offset[1] & 0xf) << 20;
1504 code[1] |= (i->tex.offset[2] & 0xf) << 16;
1507 code[0] |= (i->tex.mask & 0x3) << 25;
1508 code[1] |= (i->tex.mask & 0xc) << 12;
1510 if (i->tex.liveOnly)
1513 defId(i->def(0), 2);
1519 CodeEmitterNV50::emitTXQ(const TexInstruction *i)
1521 assert(i->tex.query == TXQ_DIMS);
1523 code[0] = 0xf0000001;
1524 code[1] = 0x60000000;
1526 code[0] |= i->tex.r << 9;
1527 code[0] |= i->tex.s << 17;
1529 code[0] |= (i->tex.mask & 0x3) << 25;
1530 code[1] |= (i->tex.mask & 0xc) << 12;
1532 defId(i->def(0), 2);
1538 CodeEmitterNV50::emitTEXPREP(const TexInstruction *i)
1540 code[0] = 0xf8000001 | (3 << 22) | (i->tex.s << 17) | (i->tex.r << 9);
1541 code[1] = 0x60010000;
1543 code[0] |= (i->tex.mask & 0x3) << 25;
1544 code[1] |= (i->tex.mask & 0xc) << 12;
1545 defId(i->def(0), 2);
1551 CodeEmitterNV50::emitPRERETEmu(const FlowInstruction *i)
1553 uint32_t pos = i->target.bb->binPos + 8; // +8 to skip an op */
1555 code[0] = 0x10000003; // bra
1556 code[1] = 0x00000780; // always
1559 case NV50_IR_SUBOP_EMU_PRERET + 0: // bra to the call
1561 case NV50_IR_SUBOP_EMU_PRERET + 1: // bra to skip the call
1565 assert(i->subOp == (NV50_IR_SUBOP_EMU_PRERET + 2));
1566 code[0] = 0x20000003; // call
1567 code[1] = 0x00000000; // no predicate
1570 addReloc(RelocEntry::TYPE_CODE, 0, pos, 0x07fff800, 9);
1571 addReloc(RelocEntry::TYPE_CODE, 1, pos, 0x000fc000, -4);
1575 CodeEmitterNV50::emitFlow(const Instruction *i, uint8_t flowOp)
1577 const FlowInstruction *f = i->asFlow();
1578 bool hasPred = false;
1579 bool hasTarg = false;
1581 code[0] = 0x00000003 | (flowOp << 28);
1582 code[1] = 0x00000000;
1602 if (i->subOp >= NV50_IR_SUBOP_EMU_PRERET) {
1617 if (f->op == OP_CALL) {
1619 pos = targNV50->getBuiltinOffset(f->target.builtin);
1621 pos = f->target.fn->binPos;
1624 pos = f->target.bb->binPos;
1627 code[0] |= ((pos >> 2) & 0xffff) << 11;
1628 code[1] |= ((pos >> 18) & 0x003f) << 14;
1630 RelocEntry::Type relocTy;
1632 relocTy = f->builtin ? RelocEntry::TYPE_BUILTIN : RelocEntry::TYPE_CODE;
1634 addReloc(relocTy, 0, pos, 0x07fff800, 9);
1635 addReloc(relocTy, 1, pos, 0x000fc000, -4);
1640 CodeEmitterNV50::emitBAR(const Instruction *i)
1642 ImmediateValue *barId = i->getSrc(0)->asImm();
1645 code[0] = 0x82000003 | (barId->reg.data.u32 << 21);
1646 code[1] = 0x00004000;
1648 if (i->subOp == NV50_IR_SUBOP_BAR_SYNC)
1653 CodeEmitterNV50::emitATOM(const Instruction *i)
1657 case NV50_IR_SUBOP_ATOM_ADD: subOp = 0x0; break;
1658 case NV50_IR_SUBOP_ATOM_MIN: subOp = 0x7; break;
1659 case NV50_IR_SUBOP_ATOM_MAX: subOp = 0x6; break;
1660 case NV50_IR_SUBOP_ATOM_INC: subOp = 0x4; break;
1661 case NV50_IR_SUBOP_ATOM_DEC: subOp = 0x5; break;
1662 case NV50_IR_SUBOP_ATOM_AND: subOp = 0xa; break;
1663 case NV50_IR_SUBOP_ATOM_OR: subOp = 0xb; break;
1664 case NV50_IR_SUBOP_ATOM_XOR: subOp = 0xc; break;
1665 case NV50_IR_SUBOP_ATOM_CAS: subOp = 0x2; break;
1666 case NV50_IR_SUBOP_ATOM_EXCH: subOp = 0x1; break;
1668 assert(!"invalid subop");
1671 code[0] = 0xd0000001;
1672 code[1] = 0xe0c00000 | (subOp << 2);
1673 if (isSignedType(i->dType))
1680 if (i->subOp == NV50_IR_SUBOP_ATOM_CAS)
1684 code[0] |= i->getSrc(0)->reg.fileIndex << 23;
1685 srcId(i->getIndirect(0, 0), 9);
1689 CodeEmitterNV50::emitInstruction(Instruction *insn)
1691 if (!insn->encSize) {
1692 ERROR("skipping unencodable instruction: "); insn->print();
1695 if (codeSize + insn->encSize > codeSizeLimit) {
1696 ERROR("code emitter output buffer too small\n");
1700 if (insn->bb->getProgram()->dbgFlags & NV50_IR_DEBUG_BASIC) {
1701 INFO("EMIT: "); insn->print();
1733 if (isFloatType(insn->dType))
1735 else if (insn->getDef(0)->reg.file == FILE_ADDRESS)
1741 if (isFloatType(insn->dType))
1748 if (isFloatType(insn->dType))
1784 if (insn->def(0).getFile() == FILE_ADDRESS)
1787 if (insn->def(0).getFile() == FILE_FLAGS ||
1788 insn->src(0).getFile() == FILE_FLAGS ||
1789 insn->src(0).getFile() == FILE_ADDRESS)
1822 emitTEX(insn->asTex());
1825 emitTXQ(insn->asTex());
1828 emitTEXPREP(insn->asTex());
1835 emitFlow(insn, 0x0);
1838 emitFlow(insn, 0x1);
1841 emitFlow(insn, 0x2);
1844 emitFlow(insn, 0x3);
1847 emitFlow(insn, 0x4);
1850 emitFlow(insn, 0x5);
1853 emitFlow(insn, 0x6);
1856 emitFlow(insn, 0x7);
1859 emitFlow(insn, 0xa);
1862 emitFlow(insn, 0xd);
1865 emitQUADOP(insn, insn->lanes, insn->subOp);
1868 emitQUADOP(insn, 4, insn->src(0).mod.neg() ? 0x66 : 0x99);
1871 emitQUADOP(insn, 5, insn->src(0).mod.neg() ? 0x5a : 0xa5);
1882 ERROR("operation should have been eliminated\n");
1896 ERROR("operation should have been lowered\n");
1899 ERROR("unknown op: %u\n", insn->op);
1902 if (insn->join || insn->op == OP_JOIN)
1905 if (insn->exit || insn->op == OP_EXIT)
1908 assert((insn->encSize == 8) == (code[0] & 1));
1910 code += insn->encSize / 4;
1911 codeSize += insn->encSize;
1916 CodeEmitterNV50::getMinEncodingSize(const Instruction *i) const
1918 const Target::OpInfo &info = targ->getOpInfo(i);
1920 if (info.minEncSize > 4)
1923 // check constraints on dst and src operands
1924 for (int d = 0; i->defExists(d); ++d) {
1925 if (i->def(d).rep()->reg.data.id > 63 ||
1926 i->def(d).rep()->reg.file != FILE_GPR)
1930 for (int s = 0; i->srcExists(s); ++s) {
1931 DataFile sf = i->src(s).getFile();
1933 if (sf != FILE_SHADER_INPUT || progType != Program::TYPE_FRAGMENT)
1935 if (i->src(s).rep()->reg.data.id > 63)
1939 // check modifiers & rounding
1940 if (i->join || i->lanes != 0xf || i->exit)
1942 if (i->op == OP_MUL && i->rnd != ROUND_N)
1946 return 8; // TODO: short tex encoding
1948 // check constraints on short MAD
1949 if (info.srcNr >= 2 && i->srcExists(2)) {
1950 if (!i->defExists(0) || !isFloatType(i->dType) ||
1951 i->def(0).rep()->reg.data.id != i->src(2).rep()->reg.data.id)
1955 return info.minEncSize;
1958 // Change the encoding size of an instruction after BBs have been scheduled.
1960 makeInstructionLong(Instruction *insn)
1962 if (insn->encSize == 8)
1964 Function *fn = insn->bb->getFunction();
1968 for (Instruction *i = insn->next; i && i->encSize == 4; ++n, i = i->next);
1972 insn->next->encSize = 8;
1974 if (insn->prev && insn->prev->encSize == 4) {
1976 insn->prev->encSize = 8;
1980 for (int i = fn->bbCount - 1; i >= 0 && fn->bbArray[i] != insn->bb; --i) {
1981 fn->bbArray[i]->binPos += 4;
1984 insn->bb->binSize += adj;
1988 trySetExitModifier(Instruction *insn)
1990 if (insn->op == OP_DISCARD ||
1991 insn->op == OP_QUADON ||
1992 insn->op == OP_QUADPOP)
1994 for (int s = 0; insn->srcExists(s); ++s)
1995 if (insn->src(s).getFile() == FILE_IMMEDIATE)
1997 if (insn->asFlow()) {
1998 if (insn->op == OP_CALL) // side effects !
2000 if (insn->getPredicate()) // cannot do conditional exit (or can we ?)
2005 makeInstructionLong(insn);
2010 replaceExitWithModifier(Function *func)
2012 BasicBlock *epilogue = BasicBlock::get(func->cfgExit);
2014 if (!epilogue->getExit() ||
2015 epilogue->getExit()->op != OP_EXIT) // only main will use OP_EXIT
2018 if (epilogue->getEntry()->op != OP_EXIT) {
2019 Instruction *insn = epilogue->getExit()->prev;
2020 if (!insn || !trySetExitModifier(insn))
2024 for (Graph::EdgeIterator ei = func->cfgExit->incident();
2025 !ei.end(); ei.next()) {
2026 BasicBlock *bb = BasicBlock::get(ei.getNode());
2027 Instruction *i = bb->getExit();
2029 if (!i || !trySetExitModifier(i))
2033 epilogue->binSize -= 8;
2035 delete_Instruction(func->getProgram(), epilogue->getExit());
2039 CodeEmitterNV50::prepareEmission(Function *func)
2041 CodeEmitter::prepareEmission(func);
2043 replaceExitWithModifier(func);
2046 CodeEmitterNV50::CodeEmitterNV50(const TargetNV50 *target) :
2047 CodeEmitter(target), targNV50(target)
2049 targ = target; // specialized
2051 codeSize = codeSizeLimit = 0;
2056 TargetNV50::getCodeEmitter(Program::Type type)
2058 CodeEmitterNV50 *emit = new CodeEmitterNV50(this);
2059 emit->setProgramType(type);
2063 } // namespace nv50_ir