2 * Copyright 2012 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_target_nvc0.h"
25 // CodeEmitter for GK110 encoding of the Fermi/Kepler ISA.
29 class CodeEmitterGK110
: public CodeEmitter
32 CodeEmitterGK110(const TargetNVC0
*);
34 virtual bool emitInstruction(Instruction
*);
35 virtual uint32_t getMinEncodingSize(const Instruction
*) const;
36 virtual void prepareEmission(Function
*);
38 inline void setProgramType(Program::Type pType
) { progType
= pType
; }
41 const TargetNVC0
*targNVC0
;
43 Program::Type progType
;
45 const bool writeIssueDelays
;
48 void emitForm_21(const Instruction
*, uint32_t opc2
, uint32_t opc1
);
49 void emitForm_C(const Instruction
*, uint32_t opc
, uint8_t ctg
);
50 void emitForm_L(const Instruction
*, uint32_t opc
, uint8_t ctg
, Modifier
);
52 void emitPredicate(const Instruction
*);
54 void setCAddress14(const ValueRef
&);
55 void setShortImmediate(const Instruction
*, const int s
);
56 void setImmediate32(const Instruction
*, const int s
, Modifier
);
57 void setSUConst16(const Instruction
*, const int s
);
59 void modNegAbsF32_3b(const Instruction
*, const int s
);
61 void emitCondCode(CondCode cc
, int pos
, uint8_t mask
);
62 void emitInterpMode(const Instruction
*);
63 void emitLoadStoreType(DataType ty
, const int pos
);
64 void emitCachingMode(CacheMode c
, const int pos
);
65 void emitSUGType(DataType
, const int pos
);
66 void emitSUCachingMode(CacheMode c
);
68 inline uint8_t getSRegEncoding(const ValueRef
&);
70 void emitRoundMode(RoundMode
, const int pos
, const int rintPos
);
71 void emitRoundModeF(RoundMode
, const int pos
);
72 void emitRoundModeI(RoundMode
, const int pos
);
74 void emitNegAbs12(const Instruction
*);
76 void emitNOP(const Instruction
*);
78 void emitLOAD(const Instruction
*);
79 void emitSTORE(const Instruction
*);
80 void emitMOV(const Instruction
*);
81 void emitATOM(const Instruction
*);
82 void emitCCTL(const Instruction
*);
84 void emitINTERP(const Instruction
*);
85 void emitAFETCH(const Instruction
*);
86 void emitPFETCH(const Instruction
*);
87 void emitVFETCH(const Instruction
*);
88 void emitEXPORT(const Instruction
*);
89 void emitOUT(const Instruction
*);
91 void emitUADD(const Instruction
*);
92 void emitFADD(const Instruction
*);
93 void emitDADD(const Instruction
*);
94 void emitIMUL(const Instruction
*);
95 void emitFMUL(const Instruction
*);
96 void emitDMUL(const Instruction
*);
97 void emitIMAD(const Instruction
*);
98 void emitISAD(const Instruction
*);
99 void emitSHLADD(const Instruction
*);
100 void emitFMAD(const Instruction
*);
101 void emitDMAD(const Instruction
*);
102 void emitMADSP(const Instruction
*i
);
104 void emitNOT(const Instruction
*);
105 void emitLogicOp(const Instruction
*, uint8_t subOp
);
106 void emitPOPC(const Instruction
*);
107 void emitINSBF(const Instruction
*);
108 void emitEXTBF(const Instruction
*);
109 void emitBFIND(const Instruction
*);
110 void emitPERMT(const Instruction
*);
111 void emitShift(const Instruction
*);
113 void emitSFnOp(const Instruction
*, uint8_t subOp
);
115 void emitCVT(const Instruction
*);
116 void emitMINMAX(const Instruction
*);
117 void emitPreOp(const Instruction
*);
119 void emitSET(const CmpInstruction
*);
120 void emitSLCT(const CmpInstruction
*);
121 void emitSELP(const Instruction
*);
123 void emitTEXBAR(const Instruction
*);
124 void emitTEX(const TexInstruction
*);
125 void emitTEXCSAA(const TexInstruction
*);
126 void emitTXQ(const TexInstruction
*);
128 void emitQUADOP(const Instruction
*, uint8_t qOp
, uint8_t laneMask
);
130 void emitPIXLD(const Instruction
*);
132 void emitBAR(const Instruction
*);
133 void emitMEMBAR(const Instruction
*);
135 void emitFlow(const Instruction
*);
137 void emitVOTE(const Instruction
*);
139 void emitSULDGB(const TexInstruction
*);
140 void emitSUSTGx(const TexInstruction
*);
141 void emitSUCLAMPMode(uint16_t);
142 void emitSUCalc(Instruction
*);
144 void emitVSHL(const Instruction
*);
145 void emitVectorSubOp(const Instruction
*);
147 inline void defId(const ValueDef
&, const int pos
);
148 inline void srcId(const ValueRef
&, const int pos
);
149 inline void srcId(const ValueRef
*, const int pos
);
150 inline void srcId(const Instruction
*, int s
, const int pos
);
152 inline void srcAddr32(const ValueRef
&, const int pos
); // address / 4
154 inline bool isLIMM(const ValueRef
&, DataType ty
, bool mod
= false);
157 #define GK110_GPR_ZERO 255
160 if (i->src(s).mod.neg()) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
162 if (i->src(s).mod.abs()) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
164 #define NOT_(b, s) if (i->src(s).mod & Modifier(NV50_IR_MOD_NOT)) \
165 code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
167 #define FTZ_(b) if (i->ftz) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
168 #define DNZ_(b) if (i->dnz) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
170 #define SAT_(b) if (i->saturate) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
172 #define RND_(b, t) emitRoundMode##t(i->rnd, 0x##b)
174 #define SDATA(a) ((a).rep()->reg.data)
175 #define DDATA(a) ((a).rep()->reg.data)
177 void CodeEmitterGK110::srcId(const ValueRef
& src
, const int pos
)
179 code
[pos
/ 32] |= (src
.get() ? SDATA(src
).id
: GK110_GPR_ZERO
) << (pos
% 32);
182 void CodeEmitterGK110::srcId(const ValueRef
*src
, const int pos
)
184 code
[pos
/ 32] |= (src
? SDATA(*src
).id
: GK110_GPR_ZERO
) << (pos
% 32);
187 void CodeEmitterGK110::srcId(const Instruction
*insn
, int s
, int pos
)
189 int r
= insn
->srcExists(s
) ? SDATA(insn
->src(s
)).id
: GK110_GPR_ZERO
;
190 code
[pos
/ 32] |= r
<< (pos
% 32);
193 void CodeEmitterGK110::srcAddr32(const ValueRef
& src
, const int pos
)
195 code
[pos
/ 32] |= (SDATA(src
).offset
>> 2) << (pos
% 32);
198 void CodeEmitterGK110::defId(const ValueDef
& def
, const int pos
)
200 code
[pos
/ 32] |= (def
.get() ? DDATA(def
).id
: GK110_GPR_ZERO
) << (pos
% 32);
203 bool CodeEmitterGK110::isLIMM(const ValueRef
& ref
, DataType ty
, bool mod
)
205 const ImmediateValue
*imm
= ref
.get()->asImm();
207 return imm
&& (imm
->reg
.data
.u32
& ((ty
== TYPE_F32
) ? 0xfff : 0xfff00000));
211 CodeEmitterGK110::emitRoundMode(RoundMode rnd
, const int pos
, const int rintPos
)
217 case ROUND_MI
: rint
= true; /* fall through */ case ROUND_M
: n
= 1; break;
218 case ROUND_PI
: rint
= true; /* fall through */ case ROUND_P
: n
= 2; break;
219 case ROUND_ZI
: rint
= true; /* fall through */ case ROUND_Z
: n
= 3; break;
221 rint
= rnd
== ROUND_NI
;
223 assert(rnd
== ROUND_N
|| rnd
== ROUND_NI
);
226 code
[pos
/ 32] |= n
<< (pos
% 32);
227 if (rint
&& rintPos
>= 0)
228 code
[rintPos
/ 32] |= 1 << (rintPos
% 32);
232 CodeEmitterGK110::emitRoundModeF(RoundMode rnd
, const int pos
)
237 case ROUND_M
: n
= 1; break;
238 case ROUND_P
: n
= 2; break;
239 case ROUND_Z
: n
= 3; break;
242 assert(rnd
== ROUND_N
);
245 code
[pos
/ 32] |= n
<< (pos
% 32);
249 CodeEmitterGK110::emitRoundModeI(RoundMode rnd
, const int pos
)
254 case ROUND_MI
: n
= 1; break;
255 case ROUND_PI
: n
= 2; break;
256 case ROUND_ZI
: n
= 3; break;
259 assert(rnd
== ROUND_NI
);
262 code
[pos
/ 32] |= n
<< (pos
% 32);
265 void CodeEmitterGK110::emitCondCode(CondCode cc
, int pos
, uint8_t mask
)
270 case CC_FL
: n
= 0x00; break;
271 case CC_LT
: n
= 0x01; break;
272 case CC_EQ
: n
= 0x02; break;
273 case CC_LE
: n
= 0x03; break;
274 case CC_GT
: n
= 0x04; break;
275 case CC_NE
: n
= 0x05; break;
276 case CC_GE
: n
= 0x06; break;
277 case CC_LTU
: n
= 0x09; break;
278 case CC_EQU
: n
= 0x0a; break;
279 case CC_LEU
: n
= 0x0b; break;
280 case CC_GTU
: n
= 0x0c; break;
281 case CC_NEU
: n
= 0x0d; break;
282 case CC_GEU
: n
= 0x0e; break;
283 case CC_TR
: n
= 0x0f; break;
284 case CC_NO
: n
= 0x10; break;
285 case CC_NC
: n
= 0x11; break;
286 case CC_NS
: n
= 0x12; break;
287 case CC_NA
: n
= 0x13; break;
288 case CC_A
: n
= 0x14; break;
289 case CC_S
: n
= 0x15; break;
290 case CC_C
: n
= 0x16; break;
291 case CC_O
: n
= 0x17; break;
294 assert(!"invalid condition code");
297 code
[pos
/ 32] |= (n
& mask
) << (pos
% 32);
301 CodeEmitterGK110::emitPredicate(const Instruction
*i
)
303 if (i
->predSrc
>= 0) {
304 srcId(i
->src(i
->predSrc
), 18);
305 if (i
->cc
== CC_NOT_P
)
306 code
[0] |= 8 << 18; // negate
307 assert(i
->getPredicate()->reg
.file
== FILE_PREDICATE
);
314 CodeEmitterGK110::setCAddress14(const ValueRef
& src
)
316 const Storage
& res
= src
.get()->asSym()->reg
;
317 const int32_t addr
= res
.data
.offset
/ 4;
319 code
[0] |= (addr
& 0x01ff) << 23;
320 code
[1] |= (addr
& 0x3e00) >> 9;
321 code
[1] |= res
.fileIndex
<< 5;
325 CodeEmitterGK110::setShortImmediate(const Instruction
*i
, const int s
)
327 const uint32_t u32
= i
->getSrc(s
)->asImm()->reg
.data
.u32
;
328 const uint64_t u64
= i
->getSrc(s
)->asImm()->reg
.data
.u64
;
330 if (i
->sType
== TYPE_F32
) {
331 assert(!(u32
& 0x00000fff));
332 code
[0] |= ((u32
& 0x001ff000) >> 12) << 23;
333 code
[1] |= ((u32
& 0x7fe00000) >> 21);
334 code
[1] |= ((u32
& 0x80000000) >> 4);
336 if (i
->sType
== TYPE_F64
) {
337 assert(!(u64
& 0x00000fffffffffffULL
));
338 code
[0] |= ((u64
& 0x001ff00000000000ULL
) >> 44) << 23;
339 code
[1] |= ((u64
& 0x7fe0000000000000ULL
) >> 53);
340 code
[1] |= ((u64
& 0x8000000000000000ULL
) >> 36);
342 assert((u32
& 0xfff00000) == 0 || (u32
& 0xfff00000) == 0xfff00000);
343 code
[0] |= (u32
& 0x001ff) << 23;
344 code
[1] |= (u32
& 0x7fe00) >> 9;
345 code
[1] |= (u32
& 0x80000) << 8;
350 CodeEmitterGK110::setImmediate32(const Instruction
*i
, const int s
,
353 uint32_t u32
= i
->getSrc(s
)->asImm()->reg
.data
.u32
;
356 ImmediateValue
imm(i
->getSrc(s
)->asImm(), i
->sType
);
358 u32
= imm
.reg
.data
.u32
;
361 code
[0] |= u32
<< 23;
366 CodeEmitterGK110::emitForm_L(const Instruction
*i
, uint32_t opc
, uint8_t ctg
,
376 for (int s
= 0; s
< 3 && i
->srcExists(s
); ++s
) {
377 switch (i
->src(s
).getFile()) {
379 srcId(i
->src(s
), s
? 42 : 10);
382 setImmediate32(i
, s
, mod
);
392 CodeEmitterGK110::emitForm_C(const Instruction
*i
, uint32_t opc
, uint8_t ctg
)
401 switch (i
->src(0).getFile()) {
402 case FILE_MEMORY_CONST
:
403 code
[1] |= 0x4 << 28;
404 setCAddress14(i
->src(0));
407 code
[1] |= 0xc << 28;
408 srcId(i
->src(0), 23);
416 // 0x2 for GPR, c[] and 0x1 for short immediate
418 CodeEmitterGK110::emitForm_21(const Instruction
*i
, uint32_t opc2
,
421 const bool imm
= i
->srcExists(1) && i
->src(1).getFile() == FILE_IMMEDIATE
;
424 if (i
->srcExists(2) && i
->src(2).getFile() == FILE_MEMORY_CONST
)
429 code
[1] = opc1
<< 20;
432 code
[1] = (0xc << 28) | (opc2
<< 20);
439 for (int s
= 0; s
< 3 && i
->srcExists(s
); ++s
) {
440 switch (i
->src(s
).getFile()) {
441 case FILE_MEMORY_CONST
:
442 code
[1] &= (s
== 2) ? ~(0x4 << 28) : ~(0x8 << 28);
443 setCAddress14(i
->src(s
));
446 setShortImmediate(i
, s
);
449 srcId(i
->src(s
), s
? ((s
== 2) ? 42 : s1
) : 10);
452 if (i
->op
== OP_SELP
) {
453 assert(s
== 2 && i
->src(s
).getFile() == FILE_PREDICATE
);
454 srcId(i
->src(s
), 42);
456 // ignore here, can be predicate or flags, but must not be address
464 assert(imm
|| (code
[1] & (0xc << 28)));
468 CodeEmitterGK110::modNegAbsF32_3b(const Instruction
*i
, const int s
)
470 if (i
->src(s
).mod
.abs()) code
[1] &= ~(1 << 27);
471 if (i
->src(s
).mod
.neg()) code
[1] ^= (1 << 27);
475 CodeEmitterGK110::emitNOP(const Instruction
*i
)
477 code
[0] = 0x00003c02;
478 code
[1] = 0x85800000;
483 code
[0] = 0x001c3c02;
487 CodeEmitterGK110::emitFMAD(const Instruction
*i
)
489 assert(!isLIMM(i
->src(1), TYPE_F32
));
491 emitForm_21(i
, 0x0c0, 0x940);
499 bool neg1
= (i
->src(0).mod
^ i
->src(1).mod
).neg();
511 CodeEmitterGK110::emitDMAD(const Instruction
*i
)
513 assert(!i
->saturate
);
516 emitForm_21(i
, 0x1b8, 0xb38);
521 bool neg1
= (i
->src(0).mod
^ i
->src(1).mod
).neg();
533 CodeEmitterGK110::emitMADSP(const Instruction
*i
)
535 emitForm_21(i
, 0x140, 0xa40);
537 if (i
->subOp
== NV50_IR_SUBOP_MADSP_SD
) {
538 code
[1] |= 0x00c00000;
540 code
[1] |= (i
->subOp
& 0x00f) << 19; // imadp1
541 code
[1] |= (i
->subOp
& 0x0f0) << 20; // imadp2
542 code
[1] |= (i
->subOp
& 0x100) << 11; // imadp3
543 code
[1] |= (i
->subOp
& 0x200) << 15; // imadp3
544 code
[1] |= (i
->subOp
& 0xc00) << 12; // imadp3
547 if (i
->flagsDef
>= 0)
552 CodeEmitterGK110::emitFMUL(const Instruction
*i
)
554 bool neg
= (i
->src(0).mod
^ i
->src(1).mod
).neg();
556 assert(i
->postFactor
>= -3 && i
->postFactor
<= 3);
558 if (isLIMM(i
->src(1), TYPE_F32
)) {
559 emitForm_L(i
, 0x200, 0x2, Modifier(0));
567 assert(i
->postFactor
== 0);
569 emitForm_21(i
, 0x234, 0xc34);
570 code
[1] |= ((i
->postFactor
> 0) ?
571 (7 - i
->postFactor
) : (0 - i
->postFactor
)) << 12;
589 CodeEmitterGK110::emitDMUL(const Instruction
*i
)
591 bool neg
= (i
->src(0).mod
^ i
->src(1).mod
).neg();
593 assert(!i
->postFactor
);
594 assert(!i
->saturate
);
598 emitForm_21(i
, 0x240, 0xc40);
612 CodeEmitterGK110::emitIMUL(const Instruction
*i
)
614 assert(!i
->src(0).mod
.neg() && !i
->src(1).mod
.neg());
615 assert(!i
->src(0).mod
.abs() && !i
->src(1).mod
.abs());
617 if (i
->src(1).getFile() == FILE_IMMEDIATE
) {
618 emitForm_L(i
, 0x280, 2, Modifier(0));
620 if (i
->subOp
== NV50_IR_SUBOP_MUL_HIGH
)
622 if (i
->sType
== TYPE_S32
)
625 emitForm_21(i
, 0x21c, 0xc1c);
627 if (i
->subOp
== NV50_IR_SUBOP_MUL_HIGH
)
629 if (i
->sType
== TYPE_S32
)
635 CodeEmitterGK110::emitFADD(const Instruction
*i
)
637 if (isLIMM(i
->src(1), TYPE_F32
)) {
638 assert(i
->rnd
== ROUND_N
);
639 assert(!i
->saturate
);
641 Modifier mod
= i
->src(1).mod
^
642 Modifier(i
->op
== OP_SUB
? NV50_IR_MOD_NEG
: 0);
644 emitForm_L(i
, 0x400, 0, mod
);
650 emitForm_21(i
, 0x22c, 0xc2c);
659 modNegAbsF32_3b(i
, 1);
660 if (i
->op
== OP_SUB
) code
[1] ^= 1 << 27;
664 if (i
->op
== OP_SUB
) code
[1] ^= 1 << 16;
670 CodeEmitterGK110::emitDADD(const Instruction
*i
)
672 assert(!i
->saturate
);
675 emitForm_21(i
, 0x238, 0xc38);
680 modNegAbsF32_3b(i
, 1);
681 if (i
->op
== OP_SUB
) code
[1] ^= 1 << 27;
685 if (i
->op
== OP_SUB
) code
[1] ^= 1 << 16;
690 CodeEmitterGK110::emitUADD(const Instruction
*i
)
692 uint8_t addOp
= (i
->src(0).mod
.neg() << 1) | i
->src(1).mod
.neg();
697 assert(!i
->src(0).mod
.abs() && !i
->src(1).mod
.abs());
699 if (isLIMM(i
->src(1), TYPE_S32
)) {
700 emitForm_L(i
, 0x400, 1, Modifier((addOp
& 1) ? NV50_IR_MOD_NEG
: 0));
705 assert(!i
->defExists(1));
706 assert(i
->flagsSrc
< 0);
710 emitForm_21(i
, 0x208, 0xc08);
712 assert(addOp
!= 3); // would be add-plus-one
714 code
[1] |= addOp
<< 19;
717 code
[1] |= 1 << 18; // write carry
718 if (i
->flagsSrc
>= 0)
719 code
[1] |= 1 << 14; // add carry
726 CodeEmitterGK110::emitIMAD(const Instruction
*i
)
729 i
->src(2).mod
.neg() | ((i
->src(0).mod
.neg() ^ i
->src(1).mod
.neg()) << 1);
731 emitForm_21(i
, 0x100, 0xa00);
734 code
[1] |= addOp
<< 26;
736 if (i
->sType
== TYPE_S32
)
737 code
[1] |= (1 << 19) | (1 << 24);
739 if (i
->subOp
== NV50_IR_SUBOP_MUL_HIGH
)
742 if (i
->flagsDef
>= 0) code
[1] |= 1 << 18;
743 if (i
->flagsSrc
>= 0) code
[1] |= 1 << 20;
749 CodeEmitterGK110::emitISAD(const Instruction
*i
)
751 assert(i
->dType
== TYPE_S32
|| i
->dType
== TYPE_U32
);
753 emitForm_21(i
, 0x1f4, 0xb74);
755 if (i
->dType
== TYPE_S32
)
760 CodeEmitterGK110::emitSHLADD(const Instruction
*i
)
762 uint8_t addOp
= (i
->src(0).mod
.neg() << 1) | i
->src(2).mod
.neg();
763 const ImmediateValue
*imm
= i
->src(1).get()->asImm();
766 if (i
->src(2).getFile() == FILE_IMMEDIATE
) {
768 code
[1] = 0xc0c << 20;
771 code
[1] = 0x20c << 20;
773 code
[1] |= addOp
<< 19;
778 srcId(i
->src(0), 10);
780 if (i
->flagsDef
>= 0)
783 assert(!(imm
->reg
.data
.u32
& 0xffffffe0));
784 code
[1] |= imm
->reg
.data
.u32
<< 10;
786 switch (i
->src(2).getFile()) {
788 assert(code
[0] & 0x2);
789 code
[1] |= 0xc << 28;
790 srcId(i
->src(2), 23);
792 case FILE_MEMORY_CONST
:
793 assert(code
[0] & 0x2);
794 code
[1] |= 0x4 << 28;
795 setCAddress14(i
->src(2));
798 assert(code
[0] & 0x1);
799 setShortImmediate(i
, 2);
802 assert(!"bad src2 file");
808 CodeEmitterGK110::emitNOT(const Instruction
*i
)
810 code
[0] = 0x0003fc02; // logop(mov2) dst, 0, not src
811 code
[1] = 0x22003800;
817 switch (i
->src(0).getFile()) {
819 code
[1] |= 0xc << 28;
820 srcId(i
->src(0), 23);
822 case FILE_MEMORY_CONST
:
823 code
[1] |= 0x4 << 28;
824 setCAddress14(i
->src(0));
833 CodeEmitterGK110::emitLogicOp(const Instruction
*i
, uint8_t subOp
)
835 if (i
->def(0).getFile() == FILE_PREDICATE
) {
836 code
[0] = 0x00000002 | (subOp
<< 27);
837 code
[1] = 0x84800000;
842 srcId(i
->src(0), 14);
843 if (i
->src(0).mod
== Modifier(NV50_IR_MOD_NOT
)) code
[0] |= 1 << 17;
844 srcId(i
->src(1), 32);
845 if (i
->src(1).mod
== Modifier(NV50_IR_MOD_NOT
)) code
[1] |= 1 << 3;
847 if (i
->defExists(1)) {
853 if (i
->predSrc
!= 2 && i
->srcExists(2)) {
854 code
[1] |= subOp
<< 16;
855 srcId(i
->src(2), 42);
856 if (i
->src(2).mod
== Modifier(NV50_IR_MOD_NOT
)) code
[1] |= 1 << 13;
861 if (isLIMM(i
->src(1), TYPE_S32
)) {
862 emitForm_L(i
, 0x200, 0, i
->src(1).mod
);
863 code
[1] |= subOp
<< 24;
866 emitForm_21(i
, 0x220, 0xc20);
867 code
[1] |= subOp
<< 12;
874 CodeEmitterGK110::emitPOPC(const Instruction
*i
)
876 assert(!isLIMM(i
->src(1), TYPE_S32
, true));
878 emitForm_21(i
, 0x204, 0xc04);
881 if (!(code
[0] & 0x1))
886 CodeEmitterGK110::emitINSBF(const Instruction
*i
)
888 emitForm_21(i
, 0x1f8, 0xb78);
892 CodeEmitterGK110::emitEXTBF(const Instruction
*i
)
894 emitForm_21(i
, 0x600, 0xc00);
896 if (i
->dType
== TYPE_S32
)
898 if (i
->subOp
== NV50_IR_SUBOP_EXTBF_REV
)
903 CodeEmitterGK110::emitBFIND(const Instruction
*i
)
905 emitForm_C(i
, 0x218, 0x2);
907 if (i
->dType
== TYPE_S32
)
909 if (i
->src(0).mod
== Modifier(NV50_IR_MOD_NOT
))
911 if (i
->subOp
== NV50_IR_SUBOP_BFIND_SAMT
)
916 CodeEmitterGK110::emitPERMT(const Instruction
*i
)
918 emitForm_21(i
, 0x1e0, 0xb60);
920 code
[1] |= i
->subOp
<< 19;
924 CodeEmitterGK110::emitShift(const Instruction
*i
)
926 if (i
->op
== OP_SHR
) {
927 emitForm_21(i
, 0x214, 0xc14);
928 if (isSignedType(i
->dType
))
931 emitForm_21(i
, 0x224, 0xc24);
934 if (i
->subOp
== NV50_IR_SUBOP_SHIFT_WRAP
)
939 CodeEmitterGK110::emitPreOp(const Instruction
*i
)
941 emitForm_C(i
, 0x248, 0x2);
943 if (i
->op
== OP_PREEX2
)
951 CodeEmitterGK110::emitSFnOp(const Instruction
*i
, uint8_t subOp
)
953 code
[0] = 0x00000002 | (subOp
<< 23);
954 code
[1] = 0x84000000;
959 srcId(i
->src(0), 10);
967 CodeEmitterGK110::emitMINMAX(const Instruction
*i
)
991 emitForm_21(i
, op2
, op1
);
993 if (i
->dType
== TYPE_S32
)
995 code
[1] |= (i
->op
== OP_MIN
) ? 0x1c00 : 0x3c00; // [!]pt
996 code
[1] |= i
->subOp
<< 14;
997 if (i
->flagsDef
>= 0)
998 code
[1] |= i
->subOp
<< 18;
1003 if (code
[0] & 0x1) {
1004 modNegAbsF32_3b(i
, 1);
1012 CodeEmitterGK110::emitCVT(const Instruction
*i
)
1014 const bool f2f
= isFloatType(i
->dType
) && isFloatType(i
->sType
);
1015 const bool f2i
= !isFloatType(i
->dType
) && isFloatType(i
->sType
);
1016 const bool i2f
= isFloatType(i
->dType
) && !isFloatType(i
->sType
);
1018 bool sat
= i
->saturate
;
1019 bool abs
= i
->src(0).mod
.abs();
1020 bool neg
= i
->src(0).mod
.neg();
1022 RoundMode rnd
= i
->rnd
;
1025 case OP_CEIL
: rnd
= f2f
? ROUND_PI
: ROUND_P
; break;
1026 case OP_FLOOR
: rnd
= f2f
? ROUND_MI
: ROUND_M
; break;
1027 case OP_TRUNC
: rnd
= f2f
? ROUND_ZI
: ROUND_Z
; break;
1028 case OP_SAT
: sat
= true; break;
1029 case OP_NEG
: neg
= !neg
; break;
1030 case OP_ABS
: abs
= true; neg
= false; break;
1037 if (i
->op
== OP_NEG
&& i
->dType
== TYPE_U32
)
1045 if (f2f
) op
= 0x254;
1046 else if (f2i
) op
= 0x258;
1047 else if (i2f
) op
= 0x25c;
1050 emitForm_C(i
, op
, 0x2);
1053 if (neg
) code
[1] |= 1 << 16;
1054 if (abs
) code
[1] |= 1 << 20;
1055 if (sat
) code
[1] |= 1 << 21;
1057 emitRoundMode(rnd
, 32 + 10, f2f
? (32 + 13) : -1);
1059 code
[0] |= typeSizeofLog2(dType
) << 10;
1060 code
[0] |= typeSizeofLog2(i
->sType
) << 12;
1061 code
[1] |= i
->subOp
<< 12;
1063 if (isSignedIntType(dType
))
1065 if (isSignedIntType(i
->sType
))
1070 CodeEmitterGK110::emitSET(const CmpInstruction
*i
)
1074 if (i
->def(0).getFile() == FILE_PREDICATE
) {
1076 case TYPE_F32
: op2
= 0x1d8; op1
= 0xb58; break;
1077 case TYPE_F64
: op2
= 0x1c0; op1
= 0xb40; break;
1083 emitForm_21(i
, op2
, op1
);
1087 if (!(code
[0] & 0x1)) {
1091 modNegAbsF32_3b(i
, 1);
1095 // normal DST field is negated predicate result
1096 code
[0] = (code
[0] & ~0xfc) | ((code
[0] << 3) & 0xe0);
1097 if (i
->defExists(1))
1098 defId(i
->def(1), 2);
1103 case TYPE_F32
: op2
= 0x000; op1
= 0x800; break;
1104 case TYPE_F64
: op2
= 0x080; op1
= 0x900; break;
1110 emitForm_21(i
, op2
, op1
);
1114 if (!(code
[0] & 0x1)) {
1118 modNegAbsF32_3b(i
, 1);
1122 if (i
->dType
== TYPE_F32
) {
1123 if (isFloatType(i
->sType
))
1129 if (i
->sType
== TYPE_S32
)
1132 if (i
->op
!= OP_SET
) {
1134 case OP_SET_AND
: code
[1] |= 0x0 << 16; break;
1135 case OP_SET_OR
: code
[1] |= 0x1 << 16; break;
1136 case OP_SET_XOR
: code
[1] |= 0x2 << 16; break;
1141 srcId(i
->src(2), 0x2a);
1143 code
[1] |= 0x7 << 10;
1145 emitCondCode(i
->setCond
,
1146 isFloatType(i
->sType
) ? 0x33 : 0x34,
1147 isFloatType(i
->sType
) ? 0xf : 0x7);
1151 CodeEmitterGK110::emitSLCT(const CmpInstruction
*i
)
1153 CondCode cc
= i
->setCond
;
1154 if (i
->src(2).mod
.neg())
1155 cc
= reverseCondCode(cc
);
1157 if (i
->dType
== TYPE_F32
) {
1158 emitForm_21(i
, 0x1d0, 0xb50);
1160 emitCondCode(cc
, 0x33, 0xf);
1162 emitForm_21(i
, 0x1a0, 0xb20);
1163 emitCondCode(cc
, 0x34, 0x7);
1168 selpFlip(const FixupEntry
*entry
, uint32_t *code
, const FixupData
& data
)
1170 int loc
= entry
->loc
;
1171 if (data
.force_persample_interp
)
1172 code
[loc
+ 1] |= 1 << 13;
1174 code
[loc
+ 1] &= ~(1 << 13);
1177 void CodeEmitterGK110::emitSELP(const Instruction
*i
)
1179 emitForm_21(i
, 0x250, 0x050);
1181 if (i
->src(2).mod
& Modifier(NV50_IR_MOD_NOT
))
1184 if (i
->subOp
== 1) {
1185 addInterp(0, 0, selpFlip
);
1189 void CodeEmitterGK110::emitTEXBAR(const Instruction
*i
)
1191 code
[0] = 0x0000003e | (i
->subOp
<< 23);
1192 code
[1] = 0x77000000;
1197 void CodeEmitterGK110::emitTEXCSAA(const TexInstruction
*i
)
1199 code
[0] = 0x00000002;
1200 code
[1] = 0x76c00000;
1202 code
[1] |= i
->tex
.r
<< 9;
1203 // code[1] |= i->tex.s << (9 + 8);
1205 if (i
->tex
.liveOnly
)
1206 code
[0] |= 0x80000000;
1208 defId(i
->def(0), 2);
1209 srcId(i
->src(0), 10);
1213 isNextIndependentTex(const TexInstruction
*i
)
1215 if (!i
->next
|| !isTextureOp(i
->next
->op
))
1217 if (i
->getDef(0)->interfers(i
->next
->getSrc(0)))
1219 return !i
->next
->srcExists(1) || !i
->getDef(0)->interfers(i
->next
->getSrc(1));
1223 CodeEmitterGK110::emitTEX(const TexInstruction
*i
)
1225 const bool ind
= i
->tex
.rIndirectSrc
>= 0;
1228 code
[0] = 0x00000002;
1231 code
[1] = 0x7e000000;
1234 code
[1] = 0x7e800000;
1237 code
[1] = 0x78000000;
1240 code
[1] = 0x7dc00000;
1243 code
[1] = 0x7d800000;
1249 code
[0] = 0x00000002;
1250 code
[1] = 0x76000000;
1251 code
[1] |= i
->tex
.r
<< 9;
1254 code
[0] = 0x00000002;
1255 code
[1] = 0x76800000;
1256 code
[1] |= i
->tex
.r
<< 9;
1259 code
[0] = 0x00000002;
1260 code
[1] = 0x70000000;
1261 code
[1] |= i
->tex
.r
<< 13;
1264 code
[0] = 0x00000001;
1265 code
[1] = 0x70000000;
1266 code
[1] |= i
->tex
.r
<< 15;
1269 code
[0] = 0x00000001;
1270 code
[1] = 0x60000000;
1271 code
[1] |= i
->tex
.r
<< 15;
1276 code
[1] |= isNextIndependentTex(i
) ? 0x1 : 0x2; // t : p mode
1278 if (i
->tex
.liveOnly
)
1279 code
[0] |= 0x80000000;
1283 case OP_TXB
: code
[1] |= 0x2000; break;
1284 case OP_TXL
: code
[1] |= 0x3000; break;
1288 case OP_TXLQ
: break;
1290 assert(!"invalid texture op");
1294 if (i
->op
== OP_TXF
) {
1295 if (!i
->tex
.levelZero
)
1298 if (i
->tex
.levelZero
) {
1302 if (i
->op
!= OP_TXD
&& i
->tex
.derivAll
)
1307 code
[1] |= i
->tex
.mask
<< 2;
1309 const int src1
= (i
->predSrc
== 1) ? 2 : 1; // if predSrc == 1, !srcExists(2)
1311 defId(i
->def(0), 2);
1312 srcId(i
->src(0), 10);
1315 if (i
->op
== OP_TXG
) code
[1] |= i
->tex
.gatherComp
<< 13;
1318 code
[1] |= (i
->tex
.target
.isCube() ? 3 : (i
->tex
.target
.getDim() - 1)) << 7;
1319 if (i
->tex
.target
.isArray())
1321 if (i
->tex
.target
.isShadow())
1323 if (i
->tex
.target
== TEX_TARGET_2D_MS
||
1324 i
->tex
.target
== TEX_TARGET_2D_MS_ARRAY
)
1327 if (i
->srcExists(src1
) && i
->src(src1
).getFile() == FILE_IMMEDIATE
) {
1331 if (i
->tex
.useOffsets
== 1) {
1333 case OP_TXF
: code
[1] |= 0x200; break;
1334 case OP_TXD
: code
[1] |= 0x00400000; break;
1335 default: code
[1] |= 0x800; break;
1338 if (i
->tex
.useOffsets
== 4)
1343 CodeEmitterGK110::emitTXQ(const TexInstruction
*i
)
1345 code
[0] = 0x00000002;
1346 code
[1] = 0x75400001;
1348 switch (i
->tex
.query
) {
1349 case TXQ_DIMS
: code
[0] |= 0x01 << 25; break;
1350 case TXQ_TYPE
: code
[0] |= 0x02 << 25; break;
1351 case TXQ_SAMPLE_POSITION
: code
[0] |= 0x05 << 25; break;
1352 case TXQ_FILTER
: code
[0] |= 0x10 << 25; break;
1353 case TXQ_LOD
: code
[0] |= 0x12 << 25; break;
1354 case TXQ_BORDER_COLOUR
: code
[0] |= 0x16 << 25; break;
1356 assert(!"invalid texture query");
1360 code
[1] |= i
->tex
.mask
<< 2;
1361 code
[1] |= i
->tex
.r
<< 9;
1362 if (/*i->tex.sIndirectSrc >= 0 || */i
->tex
.rIndirectSrc
>= 0)
1363 code
[1] |= 0x08000000;
1365 defId(i
->def(0), 2);
1366 srcId(i
->src(0), 10);
1372 CodeEmitterGK110::emitQUADOP(const Instruction
*i
, uint8_t qOp
, uint8_t laneMask
)
1374 code
[0] = 0x00000002 | ((qOp
& 1) << 31);
1375 code
[1] = 0x7fc00200 | (qOp
>> 1) | (laneMask
<< 12); // dall
1377 defId(i
->def(0), 2);
1378 srcId(i
->src(0), 10);
1379 srcId((i
->srcExists(1) && i
->predSrc
!= 1) ? i
->src(1) : i
->src(0), 23);
1385 CodeEmitterGK110::emitPIXLD(const Instruction
*i
)
1387 emitForm_L(i
, 0x7f4, 2, Modifier(0));
1388 code
[1] |= i
->subOp
<< 2;
1389 code
[1] |= 0x00070000;
1393 CodeEmitterGK110::emitBAR(const Instruction
*i
)
1395 code
[0] = 0x00000002;
1396 code
[1] = 0x85400000;
1399 case NV50_IR_SUBOP_BAR_ARRIVE
: code
[1] |= 0x08; break;
1400 case NV50_IR_SUBOP_BAR_RED_AND
: code
[1] |= 0x50; break;
1401 case NV50_IR_SUBOP_BAR_RED_OR
: code
[1] |= 0x90; break;
1402 case NV50_IR_SUBOP_BAR_RED_POPC
: code
[1] |= 0x10; break;
1404 assert(i
->subOp
== NV50_IR_SUBOP_BAR_SYNC
);
1411 if (i
->src(0).getFile() == FILE_GPR
) {
1412 srcId(i
->src(0), 10);
1414 ImmediateValue
*imm
= i
->getSrc(0)->asImm();
1416 code
[0] |= imm
->reg
.data
.u32
<< 10;
1421 if (i
->src(1).getFile() == FILE_GPR
) {
1422 srcId(i
->src(1), 23);
1424 ImmediateValue
*imm
= i
->getSrc(0)->asImm();
1426 assert(imm
->reg
.data
.u32
<= 0xfff);
1427 code
[0] |= imm
->reg
.data
.u32
<< 23;
1428 code
[1] |= imm
->reg
.data
.u32
>> 9;
1432 if (i
->srcExists(2) && (i
->predSrc
!= 2)) {
1433 srcId(i
->src(2), 32 + 10);
1434 if (i
->src(2).mod
== Modifier(NV50_IR_MOD_NOT
))
1441 void CodeEmitterGK110::emitMEMBAR(const Instruction
*i
)
1443 code
[0] = 0x00000002 | NV50_IR_SUBOP_MEMBAR_SCOPE(i
->subOp
) << 8;
1444 code
[1] = 0x7cc00000;
1450 CodeEmitterGK110::emitFlow(const Instruction
*i
)
1452 const FlowInstruction
*f
= i
->asFlow();
1454 unsigned mask
; // bit 0: predicate, bit 1: target
1456 code
[0] = 0x00000000;
1460 code
[1] = f
->absolute
? 0x10800000 : 0x12000000;
1461 if (i
->srcExists(0) && i
->src(0).getFile() == FILE_MEMORY_CONST
)
1466 code
[1] = f
->absolute
? 0x11000000 : 0x13000000;
1467 if (i
->srcExists(0) && i
->src(0).getFile() == FILE_MEMORY_CONST
)
1472 case OP_EXIT
: code
[1] = 0x18000000; mask
= 1; break;
1473 case OP_RET
: code
[1] = 0x19000000; mask
= 1; break;
1474 case OP_DISCARD
: code
[1] = 0x19800000; mask
= 1; break;
1475 case OP_BREAK
: code
[1] = 0x1a000000; mask
= 1; break;
1476 case OP_CONT
: code
[1] = 0x1a800000; mask
= 1; break;
1478 case OP_JOINAT
: code
[1] = 0x14800000; mask
= 2; break;
1479 case OP_PREBREAK
: code
[1] = 0x15000000; mask
= 2; break;
1480 case OP_PRECONT
: code
[1] = 0x15800000; mask
= 2; break;
1481 case OP_PRERET
: code
[1] = 0x13800000; mask
= 2; break;
1483 case OP_QUADON
: code
[1] = 0x1b800000; mask
= 0; break;
1484 case OP_QUADPOP
: code
[1] = 0x1c000000; mask
= 0; break;
1485 case OP_BRKPT
: code
[1] = 0x00000000; mask
= 0; break;
1487 assert(!"invalid flow operation");
1493 if (i
->flagsSrc
< 0)
1505 if (f
->op
== OP_CALL
) {
1507 assert(f
->absolute
);
1508 uint32_t pcAbs
= targNVC0
->getBuiltinOffset(f
->target
.builtin
);
1509 addReloc(RelocEntry::TYPE_BUILTIN
, 0, pcAbs
, 0xff800000, 23);
1510 addReloc(RelocEntry::TYPE_BUILTIN
, 1, pcAbs
, 0x007fffff, -9);
1512 assert(!f
->absolute
);
1513 int32_t pcRel
= f
->target
.fn
->binPos
- (codeSize
+ 8);
1514 code
[0] |= (pcRel
& 0x1ff) << 23;
1515 code
[1] |= (pcRel
>> 9) & 0x7fff;
1519 int32_t pcRel
= f
->target
.bb
->binPos
- (codeSize
+ 8);
1520 if (writeIssueDelays
&& !(f
->target
.bb
->binPos
& 0x3f))
1522 // currently we don't want absolute branches
1523 assert(!f
->absolute
);
1524 code
[0] |= (pcRel
& 0x1ff) << 23;
1525 code
[1] |= (pcRel
>> 9) & 0x7fff;
1530 CodeEmitterGK110::emitVOTE(const Instruction
*i
)
1532 assert(i
->src(0).getFile() == FILE_PREDICATE
);
1534 code
[0] = 0x00000002;
1535 code
[1] = 0x86c00000 | (i
->subOp
<< 19);
1540 for (int d
= 0; i
->defExists(d
); d
++) {
1541 if (i
->def(d
).getFile() == FILE_PREDICATE
) {
1544 defId(i
->def(d
), 48);
1545 } else if (i
->def(d
).getFile() == FILE_GPR
) {
1548 defId(i
->def(d
), 2);
1550 assert(!"Unhandled def");
1554 code
[0] |= 255 << 2;
1557 if (i
->src(0).mod
== Modifier(NV50_IR_MOD_NOT
))
1559 srcId(i
->src(0), 42);
1563 CodeEmitterGK110::emitSUGType(DataType ty
, const int pos
)
1568 case TYPE_S32
: n
= 1; break;
1569 case TYPE_U8
: n
= 2; break;
1570 case TYPE_S8
: n
= 3; break;
1572 assert(ty
== TYPE_U32
);
1575 code
[pos
/ 32] |= n
<< (pos
% 32);
1579 CodeEmitterGK110::emitSUCachingMode(CacheMode c
)
1599 assert(!"invalid caching mode");
1602 code
[0] |= (n
& 1) << 31;
1603 code
[1] |= (n
& 2) >> 1;
1607 CodeEmitterGK110::setSUConst16(const Instruction
*i
, const int s
)
1609 const uint32_t offset
= i
->getSrc(s
)->reg
.data
.offset
;
1611 assert(offset
== (offset
& 0xfffc));
1613 code
[0] |= offset
<< 21;
1614 code
[1] |= offset
>> 11;
1615 code
[1] |= i
->getSrc(s
)->reg
.fileIndex
<< 5;
1619 CodeEmitterGK110::emitSULDGB(const TexInstruction
*i
)
1621 code
[0] = 0x00000002;
1622 code
[1] = 0x30000000 | (i
->subOp
<< 14);
1624 if (i
->src(1).getFile() == FILE_MEMORY_CONST
) {
1625 emitLoadStoreType(i
->dType
, 0x38);
1626 emitCachingMode(i
->cache
, 0x36);
1631 assert(i
->src(1).getFile() == FILE_GPR
);
1632 code
[1] |= 0x49800000;
1634 emitLoadStoreType(i
->dType
, 0x21);
1635 emitSUCachingMode(i
->cache
);
1637 srcId(i
->src(1), 23);
1640 emitSUGType(i
->sType
, 0x34);
1643 defId(i
->def(0), 2); // destination
1644 srcId(i
->src(0), 10); // address
1646 // surface predicate
1647 if (!i
->srcExists(2) || (i
->predSrc
== 2)) {
1648 code
[1] |= 0x7 << 10;
1650 if (i
->src(2).mod
== Modifier(NV50_IR_MOD_NOT
))
1652 srcId(i
->src(2), 32 + 10);
1657 CodeEmitterGK110::emitSUSTGx(const TexInstruction
*i
)
1659 assert(i
->op
== OP_SUSTP
);
1661 code
[0] = 0x00000002;
1662 code
[1] = 0x38000000;
1664 if (i
->src(1).getFile() == FILE_MEMORY_CONST
) {
1665 code
[0] |= i
->subOp
<< 2;
1667 if (i
->op
== OP_SUSTP
)
1668 code
[0] |= i
->tex
.mask
<< 4;
1670 emitSUGType(i
->sType
, 0x8);
1671 emitCachingMode(i
->cache
, 0x36);
1676 assert(i
->src(1).getFile() == FILE_GPR
);
1678 code
[0] |= i
->subOp
<< 23;
1679 code
[1] |= 0x41c00000;
1681 if (i
->op
== OP_SUSTP
)
1682 code
[0] |= i
->tex
.mask
<< 25;
1684 emitSUGType(i
->sType
, 0x1d);
1685 emitSUCachingMode(i
->cache
);
1687 srcId(i
->src(1), 2);
1691 srcId(i
->src(0), 10); // address
1692 srcId(i
->src(3), 42); // values
1694 // surface predicate
1695 if (!i
->srcExists(2) || (i
->predSrc
== 2)) {
1696 code
[1] |= 0x7 << 18;
1698 if (i
->src(2).mod
== Modifier(NV50_IR_MOD_NOT
))
1700 srcId(i
->src(2), 32 + 18);
1705 CodeEmitterGK110::emitSUCLAMPMode(uint16_t subOp
)
1708 switch (subOp
& ~NV50_IR_SUBOP_SUCLAMP_2D
) {
1709 case NV50_IR_SUBOP_SUCLAMP_SD(0, 1): m
= 0; break;
1710 case NV50_IR_SUBOP_SUCLAMP_SD(1, 1): m
= 1; break;
1711 case NV50_IR_SUBOP_SUCLAMP_SD(2, 1): m
= 2; break;
1712 case NV50_IR_SUBOP_SUCLAMP_SD(3, 1): m
= 3; break;
1713 case NV50_IR_SUBOP_SUCLAMP_SD(4, 1): m
= 4; break;
1714 case NV50_IR_SUBOP_SUCLAMP_PL(0, 1): m
= 5; break;
1715 case NV50_IR_SUBOP_SUCLAMP_PL(1, 1): m
= 6; break;
1716 case NV50_IR_SUBOP_SUCLAMP_PL(2, 1): m
= 7; break;
1717 case NV50_IR_SUBOP_SUCLAMP_PL(3, 1): m
= 8; break;
1718 case NV50_IR_SUBOP_SUCLAMP_PL(4, 1): m
= 9; break;
1719 case NV50_IR_SUBOP_SUCLAMP_BL(0, 1): m
= 10; break;
1720 case NV50_IR_SUBOP_SUCLAMP_BL(1, 1): m
= 11; break;
1721 case NV50_IR_SUBOP_SUCLAMP_BL(2, 1): m
= 12; break;
1722 case NV50_IR_SUBOP_SUCLAMP_BL(3, 1): m
= 13; break;
1723 case NV50_IR_SUBOP_SUCLAMP_BL(4, 1): m
= 14; break;
1728 if (subOp
& NV50_IR_SUBOP_SUCLAMP_2D
)
1733 CodeEmitterGK110::emitSUCalc(Instruction
*i
)
1735 ImmediateValue
*imm
= NULL
;
1736 uint64_t opc1
, opc2
;
1738 if (i
->srcExists(2)) {
1739 imm
= i
->getSrc(2)->asImm();
1741 i
->setSrc(2, NULL
); // special case, make emitForm_21 not assert
1745 case OP_SUCLAMP
: opc1
= 0xb00; opc2
= 0x580; break;
1746 case OP_SUBFM
: opc1
= 0xb68; opc2
= 0x1e8; break;
1747 case OP_SUEAU
: opc1
= 0xb6c; opc2
= 0x1ec; break;
1752 emitForm_21(i
, opc2
, opc1
);
1754 if (i
->op
== OP_SUCLAMP
) {
1755 if (i
->dType
== TYPE_S32
)
1757 emitSUCLAMPMode(i
->subOp
);
1760 if (i
->op
== OP_SUBFM
&& i
->subOp
== NV50_IR_SUBOP_SUBFM_3D
)
1763 if (i
->op
!= OP_SUEAU
) {
1764 const uint8_t pos
= i
->op
== OP_SUBFM
? 19 : 16;
1765 if (i
->def(0).getFile() == FILE_PREDICATE
) { // p, #
1766 code
[0] |= 255 << 2;
1767 code
[1] |= i
->getDef(1)->reg
.data
.id
<< pos
;
1769 if (i
->defExists(1)) { // r, p
1770 assert(i
->def(1).getFile() == FILE_PREDICATE
);
1771 code
[1] |= i
->getDef(1)->reg
.data
.id
<< pos
;
1773 code
[1] |= 7 << pos
;
1778 assert(i
->op
== OP_SUCLAMP
);
1780 code
[1] |= (imm
->reg
.data
.u32
& 0x3f) << 10; // sint6
1786 CodeEmitterGK110::emitVectorSubOp(const Instruction
*i
)
1788 switch (NV50_IR_SUBOP_Vn(i
->subOp
)) {
1790 code
[1] |= (i
->subOp
& 0x000f) << 7; // vsrc1
1791 code
[1] |= (i
->subOp
& 0x00e0) >> 6; // vsrc2
1792 code
[1] |= (i
->subOp
& 0x0100) << 13; // vsrc2
1793 code
[1] |= (i
->subOp
& 0x3c00) << 12; // vdst
1802 CodeEmitterGK110::emitVSHL(const Instruction
*i
)
1804 code
[0] = 0x00000002;
1805 code
[1] = 0xb8000000;
1807 assert(NV50_IR_SUBOP_Vn(i
->subOp
) == 0);
1809 if (isSignedType(i
->dType
)) code
[1] |= 1 << 25;
1810 if (isSignedType(i
->sType
)) code
[1] |= 1 << 19;
1815 defId(i
->def(0), 2);
1816 srcId(i
->src(0), 10);
1818 if (i
->getSrc(1)->reg
.file
== FILE_IMMEDIATE
) {
1819 ImmediateValue
*imm
= i
->getSrc(1)->asImm();
1821 code
[0] |= (imm
->reg
.data
.u32
& 0x01ff) << 23;
1822 code
[1] |= (imm
->reg
.data
.u32
& 0xfe00) >> 9;
1824 assert(i
->getSrc(1)->reg
.file
== FILE_GPR
);
1826 srcId(i
->src(1), 23);
1828 srcId(i
->src(2), 42);
1832 if (i
->flagsDef
>= 0)
1837 CodeEmitterGK110::emitAFETCH(const Instruction
*i
)
1839 uint32_t offset
= i
->src(0).get()->reg
.data
.offset
& 0x7ff;
1841 code
[0] = 0x00000002 | (offset
<< 23);
1842 code
[1] = 0x7d000000 | (offset
>> 9);
1844 if (i
->getSrc(0)->reg
.file
== FILE_SHADER_OUTPUT
)
1849 defId(i
->def(0), 2);
1850 srcId(i
->src(0).getIndirect(0), 10);
1854 CodeEmitterGK110::emitPFETCH(const Instruction
*i
)
1856 uint32_t prim
= i
->src(0).get()->reg
.data
.u32
;
1858 code
[0] = 0x00000002 | ((prim
& 0xff) << 23);
1859 code
[1] = 0x7f800000;
1863 const int src1
= (i
->predSrc
== 1) ? 2 : 1; // if predSrc == 1, !srcExists(2)
1865 defId(i
->def(0), 2);
1870 CodeEmitterGK110::emitVFETCH(const Instruction
*i
)
1872 unsigned int size
= typeSizeof(i
->dType
);
1873 uint32_t offset
= i
->src(0).get()->reg
.data
.offset
;
1875 code
[0] = 0x00000002 | (offset
<< 23);
1876 code
[1] = 0x7ec00000 | (offset
>> 9);
1877 code
[1] |= (size
/ 4 - 1) << 18;
1881 if (i
->getSrc(0)->reg
.file
== FILE_SHADER_OUTPUT
)
1882 code
[1] |= 0x8; // yes, TCPs can read from *outputs* of other threads
1886 defId(i
->def(0), 2);
1887 srcId(i
->src(0).getIndirect(0), 10);
1888 srcId(i
->src(0).getIndirect(1), 32 + 10); // vertex address
1892 CodeEmitterGK110::emitEXPORT(const Instruction
*i
)
1894 unsigned int size
= typeSizeof(i
->dType
);
1895 uint32_t offset
= i
->src(0).get()->reg
.data
.offset
;
1897 code
[0] = 0x00000002 | (offset
<< 23);
1898 code
[1] = 0x7f000000 | (offset
>> 9);
1899 code
[1] |= (size
/ 4 - 1) << 18;
1906 assert(i
->src(1).getFile() == FILE_GPR
);
1908 srcId(i
->src(0).getIndirect(0), 10);
1909 srcId(i
->src(0).getIndirect(1), 32 + 10); // vertex base address
1910 srcId(i
->src(1), 2);
1914 CodeEmitterGK110::emitOUT(const Instruction
*i
)
1916 assert(i
->src(0).getFile() == FILE_GPR
);
1918 emitForm_21(i
, 0x1f0, 0xb70);
1920 if (i
->op
== OP_EMIT
)
1922 if (i
->op
== OP_RESTART
|| i
->subOp
== NV50_IR_SUBOP_EMIT_RESTART
)
1927 CodeEmitterGK110::emitInterpMode(const Instruction
*i
)
1929 code
[1] |= (i
->ipa
& 0x3) << 21; // TODO: INTERP_SAMPLEID
1930 code
[1] |= (i
->ipa
& 0xc) << (19 - 2);
1934 interpApply(const FixupEntry
*entry
, uint32_t *code
, const FixupData
& data
)
1936 int ipa
= entry
->ipa
;
1937 int reg
= entry
->reg
;
1938 int loc
= entry
->loc
;
1940 if (data
.flatshade
&&
1941 (ipa
& NV50_IR_INTERP_MODE_MASK
) == NV50_IR_INTERP_SC
) {
1942 ipa
= NV50_IR_INTERP_FLAT
;
1944 } else if (data
.force_persample_interp
&&
1945 (ipa
& NV50_IR_INTERP_SAMPLE_MASK
) == NV50_IR_INTERP_DEFAULT
&&
1946 (ipa
& NV50_IR_INTERP_MODE_MASK
) != NV50_IR_INTERP_FLAT
) {
1947 ipa
|= NV50_IR_INTERP_CENTROID
;
1949 code
[loc
+ 1] &= ~(0xf << 19);
1950 code
[loc
+ 1] |= (ipa
& 0x3) << 21;
1951 code
[loc
+ 1] |= (ipa
& 0xc) << (19 - 2);
1952 code
[loc
+ 0] &= ~(0xff << 23);
1953 code
[loc
+ 0] |= reg
<< 23;
1957 CodeEmitterGK110::emitINTERP(const Instruction
*i
)
1959 const uint32_t base
= i
->getSrc(0)->reg
.data
.offset
;
1961 code
[0] = 0x00000002 | (base
<< 31);
1962 code
[1] = 0x74800000 | (base
>> 1);
1967 if (i
->op
== OP_PINTERP
) {
1968 srcId(i
->src(1), 23);
1969 addInterp(i
->ipa
, SDATA(i
->src(1)).id
, interpApply
);
1971 code
[0] |= 0xff << 23;
1972 addInterp(i
->ipa
, 0xff, interpApply
);
1975 srcId(i
->src(0).getIndirect(0), 10);
1979 defId(i
->def(0), 2);
1981 if (i
->getSampleMode() == NV50_IR_INTERP_OFFSET
)
1982 srcId(i
->src(i
->op
== OP_PINTERP
? 2 : 1), 32 + 10);
1984 code
[1] |= 0xff << 10;
1988 CodeEmitterGK110::emitLoadStoreType(DataType ty
, const int pos
)
2020 assert(!"invalid ld/st type");
2023 code
[pos
/ 32] |= n
<< (pos
% 32);
2027 CodeEmitterGK110::emitCachingMode(CacheMode c
, const int pos
)
2048 assert(!"invalid caching mode");
2051 code
[pos
/ 32] |= n
<< (pos
% 32);
2055 CodeEmitterGK110::emitSTORE(const Instruction
*i
)
2057 int32_t offset
= SDATA(i
->src(0)).offset
;
2059 switch (i
->src(0).getFile()) {
2060 case FILE_MEMORY_GLOBAL
: code
[1] = 0xe0000000; code
[0] = 0x00000000; break;
2061 case FILE_MEMORY_LOCAL
: code
[1] = 0x7a800000; code
[0] = 0x00000002; break;
2062 case FILE_MEMORY_SHARED
:
2063 code
[0] = 0x00000002;
2064 if (i
->subOp
== NV50_IR_SUBOP_STORE_UNLOCKED
)
2065 code
[1] = 0x78400000;
2067 code
[1] = 0x7ac00000;
2070 assert(!"invalid memory file");
2074 if (code
[0] & 0x2) {
2076 emitLoadStoreType(i
->dType
, 0x33);
2077 if (i
->src(0).getFile() == FILE_MEMORY_LOCAL
)
2078 emitCachingMode(i
->cache
, 0x2f);
2080 emitLoadStoreType(i
->dType
, 0x38);
2081 emitCachingMode(i
->cache
, 0x3b);
2083 code
[0] |= offset
<< 23;
2084 code
[1] |= offset
>> 9;
2086 // Unlocked store on shared memory can fail.
2087 if (i
->src(0).getFile() == FILE_MEMORY_SHARED
&&
2088 i
->subOp
== NV50_IR_SUBOP_STORE_UNLOCKED
) {
2089 assert(i
->defExists(0));
2090 defId(i
->def(0), 32 + 16);
2095 srcId(i
->src(1), 2);
2096 srcId(i
->src(0).getIndirect(0), 10);
2097 if (i
->src(0).getFile() == FILE_MEMORY_GLOBAL
&&
2098 i
->src(0).isIndirect(0) &&
2099 i
->getIndirect(0, 0)->reg
.size
== 8)
2104 CodeEmitterGK110::emitLOAD(const Instruction
*i
)
2106 int32_t offset
= SDATA(i
->src(0)).offset
;
2108 switch (i
->src(0).getFile()) {
2109 case FILE_MEMORY_GLOBAL
: code
[1] = 0xc0000000; code
[0] = 0x00000000; break;
2110 case FILE_MEMORY_LOCAL
: code
[1] = 0x7a000000; code
[0] = 0x00000002; break;
2111 case FILE_MEMORY_SHARED
:
2112 code
[0] = 0x00000002;
2113 if (i
->subOp
== NV50_IR_SUBOP_LOAD_LOCKED
)
2114 code
[1] = 0x77400000;
2116 code
[1] = 0x7a400000;
2118 case FILE_MEMORY_CONST
:
2119 if (!i
->src(0).isIndirect(0) && typeSizeof(i
->dType
) == 4) {
2124 code
[0] = 0x00000002;
2125 code
[1] = 0x7c800000 | (i
->src(0).get()->reg
.fileIndex
<< 7);
2126 code
[1] |= i
->subOp
<< 15;
2129 assert(!"invalid memory file");
2133 if (code
[0] & 0x2) {
2135 emitLoadStoreType(i
->dType
, 0x33);
2136 if (i
->src(0).getFile() == FILE_MEMORY_LOCAL
)
2137 emitCachingMode(i
->cache
, 0x2f);
2139 emitLoadStoreType(i
->dType
, 0x38);
2140 emitCachingMode(i
->cache
, 0x3b);
2142 code
[0] |= offset
<< 23;
2143 code
[1] |= offset
>> 9;
2145 // Locked store on shared memory can fail.
2147 if (i
->src(0).getFile() == FILE_MEMORY_SHARED
&&
2148 i
->subOp
== NV50_IR_SUBOP_LOAD_LOCKED
) {
2149 if (i
->def(0).getFile() == FILE_PREDICATE
) { // p, #
2152 } else if (i
->defExists(1)) { // r, p
2155 assert(!"Expected predicate dest for load locked");
2162 defId(i
->def(r
), 2);
2164 code
[0] |= 255 << 2;
2167 defId(i
->def(p
), 32 + 16);
2169 if (i
->getIndirect(0, 0)) {
2170 srcId(i
->src(0).getIndirect(0), 10);
2171 if (i
->getIndirect(0, 0)->reg
.size
== 8)
2174 code
[0] |= 255 << 10;
2179 CodeEmitterGK110::getSRegEncoding(const ValueRef
& ref
)
2181 switch (SDATA(ref
).sv
.sv
) {
2182 case SV_LANEID
: return 0x00;
2183 case SV_PHYSID
: return 0x03;
2184 case SV_VERTEX_COUNT
: return 0x10;
2185 case SV_INVOCATION_ID
: return 0x11;
2186 case SV_YDIR
: return 0x12;
2187 case SV_THREAD_KILL
: return 0x13;
2188 case SV_TID
: return 0x21 + SDATA(ref
).sv
.index
;
2189 case SV_CTAID
: return 0x25 + SDATA(ref
).sv
.index
;
2190 case SV_NTID
: return 0x29 + SDATA(ref
).sv
.index
;
2191 case SV_GRIDID
: return 0x2c;
2192 case SV_NCTAID
: return 0x2d + SDATA(ref
).sv
.index
;
2193 case SV_LBASE
: return 0x34;
2194 case SV_SBASE
: return 0x30;
2195 case SV_CLOCK
: return 0x50 + SDATA(ref
).sv
.index
;
2197 assert(!"no sreg for system value");
2203 CodeEmitterGK110::emitMOV(const Instruction
*i
)
2205 if (i
->def(0).getFile() == FILE_PREDICATE
) {
2206 if (i
->src(0).getFile() == FILE_GPR
) {
2207 // Use ISETP.NE.AND dst, PT, src, RZ, PT
2208 code
[0] = 0x00000002;
2209 code
[1] = 0xdb500000;
2211 code
[0] |= 0x7 << 2;
2212 code
[0] |= 0xff << 23;
2213 code
[1] |= 0x7 << 10;
2214 srcId(i
->src(0), 10);
2216 if (i
->src(0).getFile() == FILE_PREDICATE
) {
2217 // Use PSETP.AND.AND dst, PT, src, PT, PT
2218 code
[0] = 0x00000002;
2219 code
[1] = 0x84800000;
2221 code
[0] |= 0x7 << 2;
2222 code
[1] |= 0x7 << 0;
2223 code
[1] |= 0x7 << 10;
2225 srcId(i
->src(0), 14);
2227 assert(!"Unexpected source for predicate destination");
2231 defId(i
->def(0), 5);
2233 if (i
->src(0).getFile() == FILE_SYSTEM_VALUE
) {
2234 code
[0] = 0x00000002 | (getSRegEncoding(i
->src(0)) << 23);
2235 code
[1] = 0x86400000;
2237 defId(i
->def(0), 2);
2239 if (i
->src(0).getFile() == FILE_IMMEDIATE
) {
2240 code
[0] = 0x00000002 | (i
->lanes
<< 14);
2241 code
[1] = 0x74000000;
2243 defId(i
->def(0), 2);
2244 setImmediate32(i
, 0, Modifier(0));
2246 if (i
->src(0).getFile() == FILE_PREDICATE
) {
2247 code
[0] = 0x00000002;
2248 code
[1] = 0x84401c07;
2250 defId(i
->def(0), 2);
2251 srcId(i
->src(0), 14);
2253 emitForm_C(i
, 0x24c, 2);
2254 code
[1] |= i
->lanes
<< 10;
2259 uses64bitAddress(const Instruction
*ldst
)
2261 return ldst
->src(0).getFile() == FILE_MEMORY_GLOBAL
&&
2262 ldst
->src(0).isIndirect(0) &&
2263 ldst
->getIndirect(0, 0)->reg
.size
== 8;
2267 CodeEmitterGK110::emitATOM(const Instruction
*i
)
2269 const bool hasDst
= i
->defExists(0);
2270 const bool exch
= i
->subOp
== NV50_IR_SUBOP_ATOM_EXCH
;
2272 code
[0] = 0x00000002;
2273 if (i
->subOp
== NV50_IR_SUBOP_ATOM_CAS
)
2274 code
[1] = 0x77800000;
2276 code
[1] = 0x68000000;
2279 case NV50_IR_SUBOP_ATOM_CAS
: break;
2280 case NV50_IR_SUBOP_ATOM_EXCH
: code
[1] |= 0x04000000; break;
2281 default: code
[1] |= i
->subOp
<< 23; break;
2285 case TYPE_U32
: break;
2286 case TYPE_S32
: code
[1] |= 0x00100000; break;
2287 case TYPE_U64
: code
[1] |= 0x00200000; break;
2288 case TYPE_F32
: code
[1] |= 0x00300000; break;
2289 case TYPE_B128
: code
[1] |= 0x00400000; break; /* TODO: U128 */
2290 case TYPE_S64
: code
[1] |= 0x00500000; break;
2291 default: assert(!"unsupported type"); break;
2296 /* TODO: cas: check that src regs line up */
2297 /* TODO: cas: flip bits if $r255 is used */
2298 srcId(i
->src(1), 23);
2301 defId(i
->def(0), 2);
2304 code
[0] |= 255 << 2;
2307 if (hasDst
|| !exch
) {
2308 const int32_t offset
= SDATA(i
->src(0)).offset
;
2309 assert(offset
< 0x80000 && offset
>= -0x80000);
2310 code
[0] |= (offset
& 1) << 31;
2311 code
[1] |= (offset
& 0xffffe) >> 1;
2313 srcAddr32(i
->src(0), 31);
2316 if (i
->getIndirect(0, 0)) {
2317 srcId(i
->getIndirect(0, 0), 10);
2318 if (i
->getIndirect(0, 0)->reg
.size
== 8)
2321 code
[0] |= 255 << 10;
2326 CodeEmitterGK110::emitCCTL(const Instruction
*i
)
2328 int32_t offset
= SDATA(i
->src(0)).offset
;
2330 code
[0] = 0x00000002 | (i
->subOp
<< 2);
2332 if (i
->src(0).getFile() == FILE_MEMORY_GLOBAL
) {
2333 code
[1] = 0x7b000000;
2335 code
[1] = 0x7c000000;
2338 code
[0] |= offset
<< 23;
2339 code
[1] |= offset
>> 9;
2341 if (uses64bitAddress(i
))
2343 srcId(i
->src(0).getIndirect(0), 10);
2349 CodeEmitterGK110::emitInstruction(Instruction
*insn
)
2351 const unsigned int size
= (writeIssueDelays
&& !(codeSize
& 0x3f)) ? 16 : 8;
2353 if (insn
->encSize
!= 8) {
2354 ERROR("skipping unencodable instruction: ");
2358 if (codeSize
+ size
> codeSizeLimit
) {
2359 ERROR("code emitter output buffer too small\n");
2363 if (writeIssueDelays
) {
2364 int id
= (codeSize
& 0x3f) / 8 - 1;
2367 code
[0] = 0x00000000; // cf issue delay "instruction"
2368 code
[1] = 0x08000000;
2372 uint32_t *data
= code
- (id
* 2 + 2);
2375 case 0: data
[0] |= insn
->sched
<< 2; break;
2376 case 1: data
[0] |= insn
->sched
<< 10; break;
2377 case 2: data
[0] |= insn
->sched
<< 18; break;
2378 case 3: data
[0] |= insn
->sched
<< 26; data
[1] |= insn
->sched
>> 6; break;
2379 case 4: data
[1] |= insn
->sched
<< 2; break;
2380 case 5: data
[1] |= insn
->sched
<< 10; break;
2381 case 6: data
[1] |= insn
->sched
<< 18; break;
2388 // assert that instructions with multiple defs don't corrupt registers
2389 for (int d
= 0; insn
->defExists(d
); ++d
)
2390 assert(insn
->asTex() || insn
->def(d
).rep()->reg
.data
.id
>= 0);
2427 if (insn
->dType
== TYPE_F64
)
2429 else if (isFloatType(insn
->dType
))
2435 if (insn
->dType
== TYPE_F64
)
2437 else if (isFloatType(insn
->dType
))
2444 if (insn
->dType
== TYPE_F64
)
2446 else if (isFloatType(insn
->dType
))
2464 emitLogicOp(insn
, 0);
2467 emitLogicOp(insn
, 1);
2470 emitLogicOp(insn
, 2);
2480 emitSET(insn
->asCmp());
2486 emitSLCT(insn
->asCmp());
2501 if (insn
->def(0).getFile() == FILE_PREDICATE
||
2502 insn
->src(0).getFile() == FILE_PREDICATE
)
2508 emitSFnOp(insn
, 5 + 2 * insn
->subOp
);
2511 emitSFnOp(insn
, 4 + 2 * insn
->subOp
);
2536 emitTEX(insn
->asTex());
2539 emitTXQ(insn
->asTex());
2564 emitQUADOP(insn
, insn
->subOp
, insn
->lanes
);
2567 emitQUADOP(insn
, insn
->src(0).mod
.neg() ? 0x66 : 0x99, 0x4);
2570 emitQUADOP(insn
, insn
->src(0).mod
.neg() ? 0x5a : 0xa5, 0x5);
2607 emitSULDGB(insn
->asTex());
2611 emitSUSTGx(insn
->asTex());
2624 ERROR("operation should have been eliminated");
2630 ERROR("operation should have been lowered\n");
2633 ERROR("unknown op: %u\n", insn
->op
);
2646 CodeEmitterGK110::getMinEncodingSize(const Instruction
*i
) const
2648 // No more short instruction encodings.
2653 CodeEmitterGK110::prepareEmission(Function
*func
)
2655 const Target
*targ
= func
->getProgram()->getTarget();
2657 CodeEmitter::prepareEmission(func
);
2659 if (targ
->hasSWSched
)
2660 calculateSchedDataNVC0(targ
, func
);
2663 CodeEmitterGK110::CodeEmitterGK110(const TargetNVC0
*target
)
2664 : CodeEmitter(target
),
2666 writeIssueDelays(target
->hasSWSched
)
2669 codeSize
= codeSizeLimit
= 0;
2674 TargetNVC0::createCodeEmitterGK110(Program::Type type
)
2676 CodeEmitterGK110
*emit
= new CodeEmitterGK110(this);
2677 emit
->setProgramType(type
);
2681 } // namespace nv50_ir