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
*);
112 void emitShift64(const Instruction
*);
114 void emitSFnOp(const Instruction
*, uint8_t subOp
);
116 void emitCVT(const Instruction
*);
117 void emitMINMAX(const Instruction
*);
118 void emitPreOp(const Instruction
*);
120 void emitSET(const CmpInstruction
*);
121 void emitSLCT(const CmpInstruction
*);
122 void emitSELP(const Instruction
*);
124 void emitTEXBAR(const Instruction
*);
125 void emitTEX(const TexInstruction
*);
126 void emitTEXCSAA(const TexInstruction
*);
127 void emitTXQ(const TexInstruction
*);
129 void emitQUADOP(const Instruction
*, uint8_t qOp
, uint8_t laneMask
);
131 void emitPIXLD(const Instruction
*);
133 void emitBAR(const Instruction
*);
134 void emitMEMBAR(const Instruction
*);
136 void emitFlow(const Instruction
*);
138 void emitVOTE(const Instruction
*);
140 void emitSULDGB(const TexInstruction
*);
141 void emitSUSTGx(const TexInstruction
*);
142 void emitSUCLAMPMode(uint16_t);
143 void emitSUCalc(Instruction
*);
145 void emitVSHL(const Instruction
*);
146 void emitVectorSubOp(const Instruction
*);
148 inline void defId(const ValueDef
&, const int pos
);
149 inline void srcId(const ValueRef
&, const int pos
);
150 inline void srcId(const ValueRef
*, const int pos
);
151 inline void srcId(const Instruction
*, int s
, const int pos
);
153 inline void srcAddr32(const ValueRef
&, const int pos
); // address / 4
155 inline bool isLIMM(const ValueRef
&, DataType ty
, bool mod
= false);
158 #define GK110_GPR_ZERO 255
161 if (i->src(s).mod.neg()) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
163 if (i->src(s).mod.abs()) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
165 #define NOT_(b, s) if (i->src(s).mod & Modifier(NV50_IR_MOD_NOT)) \
166 code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
168 #define FTZ_(b) if (i->ftz) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
169 #define DNZ_(b) if (i->dnz) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
171 #define SAT_(b) if (i->saturate) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
173 #define RND_(b, t) emitRoundMode##t(i->rnd, 0x##b)
175 #define SDATA(a) ((a).rep()->reg.data)
176 #define DDATA(a) ((a).rep()->reg.data)
178 void CodeEmitterGK110::srcId(const ValueRef
& src
, const int pos
)
180 code
[pos
/ 32] |= (src
.get() ? SDATA(src
).id
: GK110_GPR_ZERO
) << (pos
% 32);
183 void CodeEmitterGK110::srcId(const ValueRef
*src
, const int pos
)
185 code
[pos
/ 32] |= (src
? SDATA(*src
).id
: GK110_GPR_ZERO
) << (pos
% 32);
188 void CodeEmitterGK110::srcId(const Instruction
*insn
, int s
, int pos
)
190 int r
= insn
->srcExists(s
) ? SDATA(insn
->src(s
)).id
: GK110_GPR_ZERO
;
191 code
[pos
/ 32] |= r
<< (pos
% 32);
194 void CodeEmitterGK110::srcAddr32(const ValueRef
& src
, const int pos
)
196 code
[pos
/ 32] |= (SDATA(src
).offset
>> 2) << (pos
% 32);
199 void CodeEmitterGK110::defId(const ValueDef
& def
, const int pos
)
201 code
[pos
/ 32] |= (def
.get() && def
.getFile() != FILE_FLAGS
? DDATA(def
).id
: GK110_GPR_ZERO
) << (pos
% 32);
204 bool CodeEmitterGK110::isLIMM(const ValueRef
& ref
, DataType ty
, bool mod
)
206 const ImmediateValue
*imm
= ref
.get()->asImm();
208 return imm
&& (imm
->reg
.data
.u32
& ((ty
== TYPE_F32
) ? 0xfff : 0xfff00000));
212 CodeEmitterGK110::emitRoundMode(RoundMode rnd
, const int pos
, const int rintPos
)
218 case ROUND_MI
: rint
= true; /* fall through */ case ROUND_M
: n
= 1; break;
219 case ROUND_PI
: rint
= true; /* fall through */ case ROUND_P
: n
= 2; break;
220 case ROUND_ZI
: rint
= true; /* fall through */ case ROUND_Z
: n
= 3; break;
222 rint
= rnd
== ROUND_NI
;
224 assert(rnd
== ROUND_N
|| rnd
== ROUND_NI
);
227 code
[pos
/ 32] |= n
<< (pos
% 32);
228 if (rint
&& rintPos
>= 0)
229 code
[rintPos
/ 32] |= 1 << (rintPos
% 32);
233 CodeEmitterGK110::emitRoundModeF(RoundMode rnd
, const int pos
)
238 case ROUND_M
: n
= 1; break;
239 case ROUND_P
: n
= 2; break;
240 case ROUND_Z
: n
= 3; break;
243 assert(rnd
== ROUND_N
);
246 code
[pos
/ 32] |= n
<< (pos
% 32);
250 CodeEmitterGK110::emitRoundModeI(RoundMode rnd
, const int pos
)
255 case ROUND_MI
: n
= 1; break;
256 case ROUND_PI
: n
= 2; break;
257 case ROUND_ZI
: n
= 3; break;
260 assert(rnd
== ROUND_NI
);
263 code
[pos
/ 32] |= n
<< (pos
% 32);
266 void CodeEmitterGK110::emitCondCode(CondCode cc
, int pos
, uint8_t mask
)
271 case CC_FL
: n
= 0x00; break;
272 case CC_LT
: n
= 0x01; break;
273 case CC_EQ
: n
= 0x02; break;
274 case CC_LE
: n
= 0x03; break;
275 case CC_GT
: n
= 0x04; break;
276 case CC_NE
: n
= 0x05; break;
277 case CC_GE
: n
= 0x06; break;
278 case CC_LTU
: n
= 0x09; break;
279 case CC_EQU
: n
= 0x0a; break;
280 case CC_LEU
: n
= 0x0b; break;
281 case CC_GTU
: n
= 0x0c; break;
282 case CC_NEU
: n
= 0x0d; break;
283 case CC_GEU
: n
= 0x0e; break;
284 case CC_TR
: n
= 0x0f; break;
285 case CC_NO
: n
= 0x10; break;
286 case CC_NC
: n
= 0x11; break;
287 case CC_NS
: n
= 0x12; break;
288 case CC_NA
: n
= 0x13; break;
289 case CC_A
: n
= 0x14; break;
290 case CC_S
: n
= 0x15; break;
291 case CC_C
: n
= 0x16; break;
292 case CC_O
: n
= 0x17; break;
295 assert(!"invalid condition code");
298 code
[pos
/ 32] |= (n
& mask
) << (pos
% 32);
302 CodeEmitterGK110::emitPredicate(const Instruction
*i
)
304 if (i
->predSrc
>= 0) {
305 srcId(i
->src(i
->predSrc
), 18);
306 if (i
->cc
== CC_NOT_P
)
307 code
[0] |= 8 << 18; // negate
308 assert(i
->getPredicate()->reg
.file
== FILE_PREDICATE
);
315 CodeEmitterGK110::setCAddress14(const ValueRef
& src
)
317 const Storage
& res
= src
.get()->asSym()->reg
;
318 const int32_t addr
= res
.data
.offset
/ 4;
320 code
[0] |= (addr
& 0x01ff) << 23;
321 code
[1] |= (addr
& 0x3e00) >> 9;
322 code
[1] |= res
.fileIndex
<< 5;
326 CodeEmitterGK110::setShortImmediate(const Instruction
*i
, const int s
)
328 const uint32_t u32
= i
->getSrc(s
)->asImm()->reg
.data
.u32
;
329 const uint64_t u64
= i
->getSrc(s
)->asImm()->reg
.data
.u64
;
331 if (i
->sType
== TYPE_F32
) {
332 assert(!(u32
& 0x00000fff));
333 code
[0] |= ((u32
& 0x001ff000) >> 12) << 23;
334 code
[1] |= ((u32
& 0x7fe00000) >> 21);
335 code
[1] |= ((u32
& 0x80000000) >> 4);
337 if (i
->sType
== TYPE_F64
) {
338 assert(!(u64
& 0x00000fffffffffffULL
));
339 code
[0] |= ((u64
& 0x001ff00000000000ULL
) >> 44) << 23;
340 code
[1] |= ((u64
& 0x7fe0000000000000ULL
) >> 53);
341 code
[1] |= ((u64
& 0x8000000000000000ULL
) >> 36);
343 assert((u32
& 0xfff00000) == 0 || (u32
& 0xfff00000) == 0xfff00000);
344 code
[0] |= (u32
& 0x001ff) << 23;
345 code
[1] |= (u32
& 0x7fe00) >> 9;
346 code
[1] |= (u32
& 0x80000) << 8;
351 CodeEmitterGK110::setImmediate32(const Instruction
*i
, const int s
,
354 uint32_t u32
= i
->getSrc(s
)->asImm()->reg
.data
.u32
;
357 ImmediateValue
imm(i
->getSrc(s
)->asImm(), i
->sType
);
359 u32
= imm
.reg
.data
.u32
;
362 code
[0] |= u32
<< 23;
367 CodeEmitterGK110::emitForm_L(const Instruction
*i
, uint32_t opc
, uint8_t ctg
,
377 for (int s
= 0; s
< 3 && i
->srcExists(s
); ++s
) {
378 switch (i
->src(s
).getFile()) {
380 srcId(i
->src(s
), s
? 42 : 10);
383 setImmediate32(i
, s
, mod
);
393 CodeEmitterGK110::emitForm_C(const Instruction
*i
, uint32_t opc
, uint8_t ctg
)
402 switch (i
->src(0).getFile()) {
403 case FILE_MEMORY_CONST
:
404 code
[1] |= 0x4 << 28;
405 setCAddress14(i
->src(0));
408 code
[1] |= 0xc << 28;
409 srcId(i
->src(0), 23);
417 // 0x2 for GPR, c[] and 0x1 for short immediate
419 CodeEmitterGK110::emitForm_21(const Instruction
*i
, uint32_t opc2
,
422 const bool imm
= i
->srcExists(1) && i
->src(1).getFile() == FILE_IMMEDIATE
;
425 if (i
->srcExists(2) && i
->src(2).getFile() == FILE_MEMORY_CONST
)
430 code
[1] = opc1
<< 20;
433 code
[1] = (0xc << 28) | (opc2
<< 20);
440 for (int s
= 0; s
< 3 && i
->srcExists(s
); ++s
) {
441 switch (i
->src(s
).getFile()) {
442 case FILE_MEMORY_CONST
:
443 code
[1] &= (s
== 2) ? ~(0x4 << 28) : ~(0x8 << 28);
444 setCAddress14(i
->src(s
));
447 setShortImmediate(i
, s
);
450 srcId(i
->src(s
), s
? ((s
== 2) ? 42 : s1
) : 10);
453 if (i
->op
== OP_SELP
) {
454 assert(s
== 2 && i
->src(s
).getFile() == FILE_PREDICATE
);
455 srcId(i
->src(s
), 42);
457 // ignore here, can be predicate or flags, but must not be address
465 assert(imm
|| (code
[1] & (0xc << 28)));
469 CodeEmitterGK110::modNegAbsF32_3b(const Instruction
*i
, const int s
)
471 if (i
->src(s
).mod
.abs()) code
[1] &= ~(1 << 27);
472 if (i
->src(s
).mod
.neg()) code
[1] ^= (1 << 27);
476 CodeEmitterGK110::emitNOP(const Instruction
*i
)
478 code
[0] = 0x00003c02;
479 code
[1] = 0x85800000;
484 code
[0] = 0x001c3c02;
488 CodeEmitterGK110::emitFMAD(const Instruction
*i
)
490 assert(!isLIMM(i
->src(1), TYPE_F32
));
492 emitForm_21(i
, 0x0c0, 0x940);
500 bool neg1
= (i
->src(0).mod
^ i
->src(1).mod
).neg();
512 CodeEmitterGK110::emitDMAD(const Instruction
*i
)
514 assert(!i
->saturate
);
517 emitForm_21(i
, 0x1b8, 0xb38);
522 bool neg1
= (i
->src(0).mod
^ i
->src(1).mod
).neg();
534 CodeEmitterGK110::emitMADSP(const Instruction
*i
)
536 emitForm_21(i
, 0x140, 0xa40);
538 if (i
->subOp
== NV50_IR_SUBOP_MADSP_SD
) {
539 code
[1] |= 0x00c00000;
541 code
[1] |= (i
->subOp
& 0x00f) << 19; // imadp1
542 code
[1] |= (i
->subOp
& 0x0f0) << 20; // imadp2
543 code
[1] |= (i
->subOp
& 0x100) << 11; // imadp3
544 code
[1] |= (i
->subOp
& 0x200) << 15; // imadp3
545 code
[1] |= (i
->subOp
& 0xc00) << 12; // imadp3
548 if (i
->flagsDef
>= 0)
553 CodeEmitterGK110::emitFMUL(const Instruction
*i
)
555 bool neg
= (i
->src(0).mod
^ i
->src(1).mod
).neg();
557 assert(i
->postFactor
>= -3 && i
->postFactor
<= 3);
559 if (isLIMM(i
->src(1), TYPE_F32
)) {
560 emitForm_L(i
, 0x200, 0x2, Modifier(0));
568 assert(i
->postFactor
== 0);
570 emitForm_21(i
, 0x234, 0xc34);
571 code
[1] |= ((i
->postFactor
> 0) ?
572 (7 - i
->postFactor
) : (0 - i
->postFactor
)) << 12;
590 CodeEmitterGK110::emitDMUL(const Instruction
*i
)
592 bool neg
= (i
->src(0).mod
^ i
->src(1).mod
).neg();
594 assert(!i
->postFactor
);
595 assert(!i
->saturate
);
599 emitForm_21(i
, 0x240, 0xc40);
613 CodeEmitterGK110::emitIMUL(const Instruction
*i
)
615 assert(!i
->src(0).mod
.neg() && !i
->src(1).mod
.neg());
616 assert(!i
->src(0).mod
.abs() && !i
->src(1).mod
.abs());
618 if (i
->src(1).getFile() == FILE_IMMEDIATE
) {
619 emitForm_L(i
, 0x280, 2, Modifier(0));
621 if (i
->subOp
== NV50_IR_SUBOP_MUL_HIGH
)
623 if (i
->sType
== TYPE_S32
)
626 emitForm_21(i
, 0x21c, 0xc1c);
628 if (i
->subOp
== NV50_IR_SUBOP_MUL_HIGH
)
630 if (i
->sType
== TYPE_S32
)
636 CodeEmitterGK110::emitFADD(const Instruction
*i
)
638 if (isLIMM(i
->src(1), TYPE_F32
)) {
639 assert(i
->rnd
== ROUND_N
);
640 assert(!i
->saturate
);
642 Modifier mod
= i
->src(1).mod
^
643 Modifier(i
->op
== OP_SUB
? NV50_IR_MOD_NEG
: 0);
645 emitForm_L(i
, 0x400, 0, mod
);
651 emitForm_21(i
, 0x22c, 0xc2c);
660 modNegAbsF32_3b(i
, 1);
661 if (i
->op
== OP_SUB
) code
[1] ^= 1 << 27;
665 if (i
->op
== OP_SUB
) code
[1] ^= 1 << 16;
671 CodeEmitterGK110::emitDADD(const Instruction
*i
)
673 assert(!i
->saturate
);
676 emitForm_21(i
, 0x238, 0xc38);
681 modNegAbsF32_3b(i
, 1);
682 if (i
->op
== OP_SUB
) code
[1] ^= 1 << 27;
686 if (i
->op
== OP_SUB
) code
[1] ^= 1 << 16;
691 CodeEmitterGK110::emitUADD(const Instruction
*i
)
693 uint8_t addOp
= (i
->src(0).mod
.neg() << 1) | i
->src(1).mod
.neg();
698 assert(!i
->src(0).mod
.abs() && !i
->src(1).mod
.abs());
700 if (isLIMM(i
->src(1), TYPE_S32
)) {
701 emitForm_L(i
, 0x400, 1, Modifier((addOp
& 1) ? NV50_IR_MOD_NEG
: 0));
706 assert(i
->flagsDef
< 0);
707 assert(i
->flagsSrc
< 0);
711 emitForm_21(i
, 0x208, 0xc08);
713 assert(addOp
!= 3); // would be add-plus-one
715 code
[1] |= addOp
<< 19;
717 if (i
->flagsDef
>= 0)
718 code
[1] |= 1 << 18; // write carry
719 if (i
->flagsSrc
>= 0)
720 code
[1] |= 1 << 14; // add carry
727 CodeEmitterGK110::emitIMAD(const Instruction
*i
)
730 i
->src(2).mod
.neg() | ((i
->src(0).mod
.neg() ^ i
->src(1).mod
.neg()) << 1);
732 emitForm_21(i
, 0x100, 0xa00);
735 code
[1] |= addOp
<< 26;
737 if (i
->sType
== TYPE_S32
)
738 code
[1] |= (1 << 19) | (1 << 24);
740 if (i
->subOp
== NV50_IR_SUBOP_MUL_HIGH
)
743 if (i
->flagsDef
>= 0) code
[1] |= 1 << 18;
744 if (i
->flagsSrc
>= 0) code
[1] |= 1 << 20;
750 CodeEmitterGK110::emitISAD(const Instruction
*i
)
752 assert(i
->dType
== TYPE_S32
|| i
->dType
== TYPE_U32
);
754 emitForm_21(i
, 0x1f4, 0xb74);
756 if (i
->dType
== TYPE_S32
)
761 CodeEmitterGK110::emitSHLADD(const Instruction
*i
)
763 uint8_t addOp
= (i
->src(0).mod
.neg() << 1) | i
->src(2).mod
.neg();
764 const ImmediateValue
*imm
= i
->src(1).get()->asImm();
767 if (i
->src(2).getFile() == FILE_IMMEDIATE
) {
769 code
[1] = 0xc0c << 20;
772 code
[1] = 0x20c << 20;
774 code
[1] |= addOp
<< 19;
779 srcId(i
->src(0), 10);
781 if (i
->flagsDef
>= 0)
784 assert(!(imm
->reg
.data
.u32
& 0xffffffe0));
785 code
[1] |= imm
->reg
.data
.u32
<< 10;
787 switch (i
->src(2).getFile()) {
789 assert(code
[0] & 0x2);
790 code
[1] |= 0xc << 28;
791 srcId(i
->src(2), 23);
793 case FILE_MEMORY_CONST
:
794 assert(code
[0] & 0x2);
795 code
[1] |= 0x4 << 28;
796 setCAddress14(i
->src(2));
799 assert(code
[0] & 0x1);
800 setShortImmediate(i
, 2);
803 assert(!"bad src2 file");
809 CodeEmitterGK110::emitNOT(const Instruction
*i
)
811 code
[0] = 0x0003fc02; // logop(mov2) dst, 0, not src
812 code
[1] = 0x22003800;
818 switch (i
->src(0).getFile()) {
820 code
[1] |= 0xc << 28;
821 srcId(i
->src(0), 23);
823 case FILE_MEMORY_CONST
:
824 code
[1] |= 0x4 << 28;
825 setCAddress14(i
->src(0));
834 CodeEmitterGK110::emitLogicOp(const Instruction
*i
, uint8_t subOp
)
836 if (i
->def(0).getFile() == FILE_PREDICATE
) {
837 code
[0] = 0x00000002 | (subOp
<< 27);
838 code
[1] = 0x84800000;
843 srcId(i
->src(0), 14);
844 if (i
->src(0).mod
== Modifier(NV50_IR_MOD_NOT
)) code
[0] |= 1 << 17;
845 srcId(i
->src(1), 32);
846 if (i
->src(1).mod
== Modifier(NV50_IR_MOD_NOT
)) code
[1] |= 1 << 3;
848 if (i
->defExists(1)) {
854 if (i
->predSrc
!= 2 && i
->srcExists(2)) {
855 code
[1] |= subOp
<< 16;
856 srcId(i
->src(2), 42);
857 if (i
->src(2).mod
== Modifier(NV50_IR_MOD_NOT
)) code
[1] |= 1 << 13;
862 if (isLIMM(i
->src(1), TYPE_S32
)) {
863 emitForm_L(i
, 0x200, 0, i
->src(1).mod
);
864 code
[1] |= subOp
<< 24;
867 emitForm_21(i
, 0x220, 0xc20);
868 code
[1] |= subOp
<< 12;
875 CodeEmitterGK110::emitPOPC(const Instruction
*i
)
877 assert(!isLIMM(i
->src(1), TYPE_S32
, true));
879 emitForm_21(i
, 0x204, 0xc04);
882 if (!(code
[0] & 0x1))
887 CodeEmitterGK110::emitINSBF(const Instruction
*i
)
889 emitForm_21(i
, 0x1f8, 0xb78);
893 CodeEmitterGK110::emitEXTBF(const Instruction
*i
)
895 emitForm_21(i
, 0x600, 0xc00);
897 if (i
->dType
== TYPE_S32
)
899 if (i
->subOp
== NV50_IR_SUBOP_EXTBF_REV
)
904 CodeEmitterGK110::emitBFIND(const Instruction
*i
)
906 emitForm_C(i
, 0x218, 0x2);
908 if (i
->dType
== TYPE_S32
)
910 if (i
->src(0).mod
== Modifier(NV50_IR_MOD_NOT
))
912 if (i
->subOp
== NV50_IR_SUBOP_BFIND_SAMT
)
917 CodeEmitterGK110::emitPERMT(const Instruction
*i
)
919 emitForm_21(i
, 0x1e0, 0xb60);
921 code
[1] |= i
->subOp
<< 19;
925 CodeEmitterGK110::emitShift(const Instruction
*i
)
927 if (i
->op
== OP_SHR
) {
928 emitForm_21(i
, 0x214, 0xc14);
929 if (isSignedType(i
->dType
))
932 emitForm_21(i
, 0x224, 0xc24);
935 if (i
->subOp
== NV50_IR_SUBOP_SHIFT_WRAP
)
940 CodeEmitterGK110::emitShift64(const Instruction
*i
)
942 if (i
->op
== OP_SHR
) {
943 emitForm_21(i
, 0x27c, 0xc7c);
944 if (isSignedType(i
->sType
))
946 if (i
->subOp
& NV50_IR_SUBOP_SHIFT_HIGH
)
949 emitForm_21(i
, 0xdfc, 0xf7c);
953 if (i
->subOp
& NV50_IR_SUBOP_SHIFT_WRAP
)
958 CodeEmitterGK110::emitPreOp(const Instruction
*i
)
960 emitForm_C(i
, 0x248, 0x2);
962 if (i
->op
== OP_PREEX2
)
970 CodeEmitterGK110::emitSFnOp(const Instruction
*i
, uint8_t subOp
)
972 code
[0] = 0x00000002 | (subOp
<< 23);
973 code
[1] = 0x84000000;
978 srcId(i
->src(0), 10);
986 CodeEmitterGK110::emitMINMAX(const Instruction
*i
)
1010 emitForm_21(i
, op2
, op1
);
1012 if (i
->dType
== TYPE_S32
)
1014 code
[1] |= (i
->op
== OP_MIN
) ? 0x1c00 : 0x3c00; // [!]pt
1015 code
[1] |= i
->subOp
<< 14;
1016 if (i
->flagsDef
>= 0)
1017 code
[1] |= i
->subOp
<< 18;
1022 if (code
[0] & 0x1) {
1023 modNegAbsF32_3b(i
, 1);
1031 CodeEmitterGK110::emitCVT(const Instruction
*i
)
1033 const bool f2f
= isFloatType(i
->dType
) && isFloatType(i
->sType
);
1034 const bool f2i
= !isFloatType(i
->dType
) && isFloatType(i
->sType
);
1035 const bool i2f
= isFloatType(i
->dType
) && !isFloatType(i
->sType
);
1037 bool sat
= i
->saturate
;
1038 bool abs
= i
->src(0).mod
.abs();
1039 bool neg
= i
->src(0).mod
.neg();
1041 RoundMode rnd
= i
->rnd
;
1044 case OP_CEIL
: rnd
= f2f
? ROUND_PI
: ROUND_P
; break;
1045 case OP_FLOOR
: rnd
= f2f
? ROUND_MI
: ROUND_M
; break;
1046 case OP_TRUNC
: rnd
= f2f
? ROUND_ZI
: ROUND_Z
; break;
1047 case OP_SAT
: sat
= true; break;
1048 case OP_NEG
: neg
= !neg
; break;
1049 case OP_ABS
: abs
= true; neg
= false; break;
1056 if (i
->op
== OP_NEG
&& i
->dType
== TYPE_U32
)
1064 if (f2f
) op
= 0x254;
1065 else if (f2i
) op
= 0x258;
1066 else if (i2f
) op
= 0x25c;
1069 emitForm_C(i
, op
, 0x2);
1072 if (neg
) code
[1] |= 1 << 16;
1073 if (abs
) code
[1] |= 1 << 20;
1074 if (sat
) code
[1] |= 1 << 21;
1076 emitRoundMode(rnd
, 32 + 10, f2f
? (32 + 13) : -1);
1078 code
[0] |= typeSizeofLog2(dType
) << 10;
1079 code
[0] |= typeSizeofLog2(i
->sType
) << 12;
1080 code
[1] |= i
->subOp
<< 12;
1082 if (isSignedIntType(dType
))
1084 if (isSignedIntType(i
->sType
))
1089 CodeEmitterGK110::emitSET(const CmpInstruction
*i
)
1093 if (i
->def(0).getFile() == FILE_PREDICATE
) {
1095 case TYPE_F32
: op2
= 0x1d8; op1
= 0xb58; break;
1096 case TYPE_F64
: op2
= 0x1c0; op1
= 0xb40; break;
1102 emitForm_21(i
, op2
, op1
);
1106 if (!(code
[0] & 0x1)) {
1110 modNegAbsF32_3b(i
, 1);
1114 // normal DST field is negated predicate result
1115 code
[0] = (code
[0] & ~0xfc) | ((code
[0] << 3) & 0xe0);
1116 if (i
->defExists(1))
1117 defId(i
->def(1), 2);
1122 case TYPE_F32
: op2
= 0x000; op1
= 0x800; break;
1123 case TYPE_F64
: op2
= 0x080; op1
= 0x900; break;
1129 emitForm_21(i
, op2
, op1
);
1133 if (!(code
[0] & 0x1)) {
1137 modNegAbsF32_3b(i
, 1);
1141 if (i
->dType
== TYPE_F32
) {
1142 if (isFloatType(i
->sType
))
1148 if (i
->sType
== TYPE_S32
)
1151 if (i
->op
!= OP_SET
) {
1153 case OP_SET_AND
: code
[1] |= 0x0 << 16; break;
1154 case OP_SET_OR
: code
[1] |= 0x1 << 16; break;
1155 case OP_SET_XOR
: code
[1] |= 0x2 << 16; break;
1160 srcId(i
->src(2), 0x2a);
1162 code
[1] |= 0x7 << 10;
1164 if (i
->flagsSrc
>= 0)
1166 emitCondCode(i
->setCond
,
1167 isFloatType(i
->sType
) ? 0x33 : 0x34,
1168 isFloatType(i
->sType
) ? 0xf : 0x7);
1172 CodeEmitterGK110::emitSLCT(const CmpInstruction
*i
)
1174 CondCode cc
= i
->setCond
;
1175 if (i
->src(2).mod
.neg())
1176 cc
= reverseCondCode(cc
);
1178 if (i
->dType
== TYPE_F32
) {
1179 emitForm_21(i
, 0x1d0, 0xb50);
1181 emitCondCode(cc
, 0x33, 0xf);
1183 emitForm_21(i
, 0x1a0, 0xb20);
1184 emitCondCode(cc
, 0x34, 0x7);
1185 if (i
->dType
== TYPE_S32
)
1191 selpFlip(const FixupEntry
*entry
, uint32_t *code
, const FixupData
& data
)
1193 int loc
= entry
->loc
;
1194 if (data
.force_persample_interp
)
1195 code
[loc
+ 1] |= 1 << 13;
1197 code
[loc
+ 1] &= ~(1 << 13);
1200 void CodeEmitterGK110::emitSELP(const Instruction
*i
)
1202 emitForm_21(i
, 0x250, 0x050);
1204 if (i
->src(2).mod
& Modifier(NV50_IR_MOD_NOT
))
1207 if (i
->subOp
== 1) {
1208 addInterp(0, 0, selpFlip
);
1212 void CodeEmitterGK110::emitTEXBAR(const Instruction
*i
)
1214 code
[0] = 0x0000003e | (i
->subOp
<< 23);
1215 code
[1] = 0x77000000;
1220 void CodeEmitterGK110::emitTEXCSAA(const TexInstruction
*i
)
1222 code
[0] = 0x00000002;
1223 code
[1] = 0x76c00000;
1225 code
[1] |= i
->tex
.r
<< 9;
1226 // code[1] |= i->tex.s << (9 + 8);
1228 if (i
->tex
.liveOnly
)
1229 code
[0] |= 0x80000000;
1231 defId(i
->def(0), 2);
1232 srcId(i
->src(0), 10);
1236 isNextIndependentTex(const TexInstruction
*i
)
1238 if (!i
->next
|| !isTextureOp(i
->next
->op
))
1240 if (i
->getDef(0)->interfers(i
->next
->getSrc(0)))
1242 return !i
->next
->srcExists(1) || !i
->getDef(0)->interfers(i
->next
->getSrc(1));
1246 CodeEmitterGK110::emitTEX(const TexInstruction
*i
)
1248 const bool ind
= i
->tex
.rIndirectSrc
>= 0;
1251 code
[0] = 0x00000002;
1254 code
[1] = 0x7e000000;
1257 code
[1] = 0x7e800000;
1260 code
[1] = 0x78000000;
1263 code
[1] = 0x7dc00000;
1266 code
[1] = 0x7d800000;
1272 code
[0] = 0x00000002;
1273 code
[1] = 0x76000000;
1274 code
[1] |= i
->tex
.r
<< 9;
1277 code
[0] = 0x00000002;
1278 code
[1] = 0x76800000;
1279 code
[1] |= i
->tex
.r
<< 9;
1282 code
[0] = 0x00000002;
1283 code
[1] = 0x70000000;
1284 code
[1] |= i
->tex
.r
<< 13;
1287 code
[0] = 0x00000001;
1288 code
[1] = 0x70000000;
1289 code
[1] |= i
->tex
.r
<< 15;
1292 code
[0] = 0x00000001;
1293 code
[1] = 0x60000000;
1294 code
[1] |= i
->tex
.r
<< 15;
1299 code
[1] |= isNextIndependentTex(i
) ? 0x1 : 0x2; // t : p mode
1301 if (i
->tex
.liveOnly
)
1302 code
[0] |= 0x80000000;
1306 case OP_TXB
: code
[1] |= 0x2000; break;
1307 case OP_TXL
: code
[1] |= 0x3000; break;
1311 case OP_TXLQ
: break;
1313 assert(!"invalid texture op");
1317 if (i
->op
== OP_TXF
) {
1318 if (!i
->tex
.levelZero
)
1321 if (i
->tex
.levelZero
) {
1325 if (i
->op
!= OP_TXD
&& i
->tex
.derivAll
)
1330 code
[1] |= i
->tex
.mask
<< 2;
1332 const int src1
= (i
->predSrc
== 1) ? 2 : 1; // if predSrc == 1, !srcExists(2)
1334 defId(i
->def(0), 2);
1335 srcId(i
->src(0), 10);
1338 if (i
->op
== OP_TXG
) code
[1] |= i
->tex
.gatherComp
<< 13;
1341 code
[1] |= (i
->tex
.target
.isCube() ? 3 : (i
->tex
.target
.getDim() - 1)) << 7;
1342 if (i
->tex
.target
.isArray())
1344 if (i
->tex
.target
.isShadow())
1346 if (i
->tex
.target
== TEX_TARGET_2D_MS
||
1347 i
->tex
.target
== TEX_TARGET_2D_MS_ARRAY
)
1350 if (i
->srcExists(src1
) && i
->src(src1
).getFile() == FILE_IMMEDIATE
) {
1354 if (i
->tex
.useOffsets
== 1) {
1356 case OP_TXF
: code
[1] |= 0x200; break;
1357 case OP_TXD
: code
[1] |= 0x00400000; break;
1358 default: code
[1] |= 0x800; break;
1361 if (i
->tex
.useOffsets
== 4)
1366 CodeEmitterGK110::emitTXQ(const TexInstruction
*i
)
1368 code
[0] = 0x00000002;
1369 code
[1] = 0x75400001;
1371 switch (i
->tex
.query
) {
1372 case TXQ_DIMS
: code
[0] |= 0x01 << 25; break;
1373 case TXQ_TYPE
: code
[0] |= 0x02 << 25; break;
1374 case TXQ_SAMPLE_POSITION
: code
[0] |= 0x05 << 25; break;
1375 case TXQ_FILTER
: code
[0] |= 0x10 << 25; break;
1376 case TXQ_LOD
: code
[0] |= 0x12 << 25; break;
1377 case TXQ_BORDER_COLOUR
: code
[0] |= 0x16 << 25; break;
1379 assert(!"invalid texture query");
1383 code
[1] |= i
->tex
.mask
<< 2;
1384 code
[1] |= i
->tex
.r
<< 9;
1385 if (/*i->tex.sIndirectSrc >= 0 || */i
->tex
.rIndirectSrc
>= 0)
1386 code
[1] |= 0x08000000;
1388 defId(i
->def(0), 2);
1389 srcId(i
->src(0), 10);
1395 CodeEmitterGK110::emitQUADOP(const Instruction
*i
, uint8_t qOp
, uint8_t laneMask
)
1397 code
[0] = 0x00000002 | ((qOp
& 1) << 31);
1398 code
[1] = 0x7fc00200 | (qOp
>> 1) | (laneMask
<< 12); // dall
1400 defId(i
->def(0), 2);
1401 srcId(i
->src(0), 10);
1402 srcId((i
->srcExists(1) && i
->predSrc
!= 1) ? i
->src(1) : i
->src(0), 23);
1408 CodeEmitterGK110::emitPIXLD(const Instruction
*i
)
1410 emitForm_L(i
, 0x7f4, 2, Modifier(0));
1411 code
[1] |= i
->subOp
<< 2;
1412 code
[1] |= 0x00070000;
1416 CodeEmitterGK110::emitBAR(const Instruction
*i
)
1418 code
[0] = 0x00000002;
1419 code
[1] = 0x85400000;
1422 case NV50_IR_SUBOP_BAR_ARRIVE
: code
[1] |= 0x08; break;
1423 case NV50_IR_SUBOP_BAR_RED_AND
: code
[1] |= 0x50; break;
1424 case NV50_IR_SUBOP_BAR_RED_OR
: code
[1] |= 0x90; break;
1425 case NV50_IR_SUBOP_BAR_RED_POPC
: code
[1] |= 0x10; break;
1427 assert(i
->subOp
== NV50_IR_SUBOP_BAR_SYNC
);
1434 if (i
->src(0).getFile() == FILE_GPR
) {
1435 srcId(i
->src(0), 10);
1437 ImmediateValue
*imm
= i
->getSrc(0)->asImm();
1439 code
[0] |= imm
->reg
.data
.u32
<< 10;
1444 if (i
->src(1).getFile() == FILE_GPR
) {
1445 srcId(i
->src(1), 23);
1447 ImmediateValue
*imm
= i
->getSrc(0)->asImm();
1449 assert(imm
->reg
.data
.u32
<= 0xfff);
1450 code
[0] |= imm
->reg
.data
.u32
<< 23;
1451 code
[1] |= imm
->reg
.data
.u32
>> 9;
1455 if (i
->srcExists(2) && (i
->predSrc
!= 2)) {
1456 srcId(i
->src(2), 32 + 10);
1457 if (i
->src(2).mod
== Modifier(NV50_IR_MOD_NOT
))
1464 void CodeEmitterGK110::emitMEMBAR(const Instruction
*i
)
1466 code
[0] = 0x00000002 | NV50_IR_SUBOP_MEMBAR_SCOPE(i
->subOp
) << 8;
1467 code
[1] = 0x7cc00000;
1473 CodeEmitterGK110::emitFlow(const Instruction
*i
)
1475 const FlowInstruction
*f
= i
->asFlow();
1477 unsigned mask
; // bit 0: predicate, bit 1: target
1479 code
[0] = 0x00000000;
1483 code
[1] = f
->absolute
? 0x10800000 : 0x12000000;
1484 if (i
->srcExists(0) && i
->src(0).getFile() == FILE_MEMORY_CONST
)
1489 code
[1] = f
->absolute
? 0x11000000 : 0x13000000;
1490 if (i
->srcExists(0) && i
->src(0).getFile() == FILE_MEMORY_CONST
)
1495 case OP_EXIT
: code
[1] = 0x18000000; mask
= 1; break;
1496 case OP_RET
: code
[1] = 0x19000000; mask
= 1; break;
1497 case OP_DISCARD
: code
[1] = 0x19800000; mask
= 1; break;
1498 case OP_BREAK
: code
[1] = 0x1a000000; mask
= 1; break;
1499 case OP_CONT
: code
[1] = 0x1a800000; mask
= 1; break;
1501 case OP_JOINAT
: code
[1] = 0x14800000; mask
= 2; break;
1502 case OP_PREBREAK
: code
[1] = 0x15000000; mask
= 2; break;
1503 case OP_PRECONT
: code
[1] = 0x15800000; mask
= 2; break;
1504 case OP_PRERET
: code
[1] = 0x13800000; mask
= 2; break;
1506 case OP_QUADON
: code
[1] = 0x1b800000; mask
= 0; break;
1507 case OP_QUADPOP
: code
[1] = 0x1c000000; mask
= 0; break;
1508 case OP_BRKPT
: code
[1] = 0x00000000; mask
= 0; break;
1510 assert(!"invalid flow operation");
1516 if (i
->flagsSrc
< 0)
1528 if (f
->op
== OP_CALL
) {
1530 assert(f
->absolute
);
1531 uint32_t pcAbs
= targNVC0
->getBuiltinOffset(f
->target
.builtin
);
1532 addReloc(RelocEntry::TYPE_BUILTIN
, 0, pcAbs
, 0xff800000, 23);
1533 addReloc(RelocEntry::TYPE_BUILTIN
, 1, pcAbs
, 0x007fffff, -9);
1535 assert(!f
->absolute
);
1536 int32_t pcRel
= f
->target
.fn
->binPos
- (codeSize
+ 8);
1537 code
[0] |= (pcRel
& 0x1ff) << 23;
1538 code
[1] |= (pcRel
>> 9) & 0x7fff;
1542 int32_t pcRel
= f
->target
.bb
->binPos
- (codeSize
+ 8);
1543 if (writeIssueDelays
&& !(f
->target
.bb
->binPos
& 0x3f))
1545 // currently we don't want absolute branches
1546 assert(!f
->absolute
);
1547 code
[0] |= (pcRel
& 0x1ff) << 23;
1548 code
[1] |= (pcRel
>> 9) & 0x7fff;
1553 CodeEmitterGK110::emitVOTE(const Instruction
*i
)
1555 assert(i
->src(0).getFile() == FILE_PREDICATE
);
1557 code
[0] = 0x00000002;
1558 code
[1] = 0x86c00000 | (i
->subOp
<< 19);
1563 for (int d
= 0; i
->defExists(d
); d
++) {
1564 if (i
->def(d
).getFile() == FILE_PREDICATE
) {
1567 defId(i
->def(d
), 48);
1568 } else if (i
->def(d
).getFile() == FILE_GPR
) {
1571 defId(i
->def(d
), 2);
1573 assert(!"Unhandled def");
1577 code
[0] |= 255 << 2;
1580 if (i
->src(0).mod
== Modifier(NV50_IR_MOD_NOT
))
1582 srcId(i
->src(0), 42);
1586 CodeEmitterGK110::emitSUGType(DataType ty
, const int pos
)
1591 case TYPE_S32
: n
= 1; break;
1592 case TYPE_U8
: n
= 2; break;
1593 case TYPE_S8
: n
= 3; break;
1595 assert(ty
== TYPE_U32
);
1598 code
[pos
/ 32] |= n
<< (pos
% 32);
1602 CodeEmitterGK110::emitSUCachingMode(CacheMode c
)
1622 assert(!"invalid caching mode");
1625 code
[0] |= (n
& 1) << 31;
1626 code
[1] |= (n
& 2) >> 1;
1630 CodeEmitterGK110::setSUConst16(const Instruction
*i
, const int s
)
1632 const uint32_t offset
= i
->getSrc(s
)->reg
.data
.offset
;
1634 assert(offset
== (offset
& 0xfffc));
1636 code
[0] |= offset
<< 21;
1637 code
[1] |= offset
>> 11;
1638 code
[1] |= i
->getSrc(s
)->reg
.fileIndex
<< 5;
1642 CodeEmitterGK110::emitSULDGB(const TexInstruction
*i
)
1644 code
[0] = 0x00000002;
1645 code
[1] = 0x30000000 | (i
->subOp
<< 14);
1647 if (i
->src(1).getFile() == FILE_MEMORY_CONST
) {
1648 emitLoadStoreType(i
->dType
, 0x38);
1649 emitCachingMode(i
->cache
, 0x36);
1654 assert(i
->src(1).getFile() == FILE_GPR
);
1655 code
[1] |= 0x49800000;
1657 emitLoadStoreType(i
->dType
, 0x21);
1658 emitSUCachingMode(i
->cache
);
1660 srcId(i
->src(1), 23);
1663 emitSUGType(i
->sType
, 0x34);
1666 defId(i
->def(0), 2); // destination
1667 srcId(i
->src(0), 10); // address
1669 // surface predicate
1670 if (!i
->srcExists(2) || (i
->predSrc
== 2)) {
1671 code
[1] |= 0x7 << 10;
1673 if (i
->src(2).mod
== Modifier(NV50_IR_MOD_NOT
))
1675 srcId(i
->src(2), 32 + 10);
1680 CodeEmitterGK110::emitSUSTGx(const TexInstruction
*i
)
1682 assert(i
->op
== OP_SUSTP
);
1684 code
[0] = 0x00000002;
1685 code
[1] = 0x38000000;
1687 if (i
->src(1).getFile() == FILE_MEMORY_CONST
) {
1688 code
[0] |= i
->subOp
<< 2;
1690 if (i
->op
== OP_SUSTP
)
1691 code
[0] |= i
->tex
.mask
<< 4;
1693 emitSUGType(i
->sType
, 0x8);
1694 emitCachingMode(i
->cache
, 0x36);
1699 assert(i
->src(1).getFile() == FILE_GPR
);
1701 code
[0] |= i
->subOp
<< 23;
1702 code
[1] |= 0x41c00000;
1704 if (i
->op
== OP_SUSTP
)
1705 code
[0] |= i
->tex
.mask
<< 25;
1707 emitSUGType(i
->sType
, 0x1d);
1708 emitSUCachingMode(i
->cache
);
1710 srcId(i
->src(1), 2);
1714 srcId(i
->src(0), 10); // address
1715 srcId(i
->src(3), 42); // values
1717 // surface predicate
1718 if (!i
->srcExists(2) || (i
->predSrc
== 2)) {
1719 code
[1] |= 0x7 << 18;
1721 if (i
->src(2).mod
== Modifier(NV50_IR_MOD_NOT
))
1723 srcId(i
->src(2), 32 + 18);
1728 CodeEmitterGK110::emitSUCLAMPMode(uint16_t subOp
)
1731 switch (subOp
& ~NV50_IR_SUBOP_SUCLAMP_2D
) {
1732 case NV50_IR_SUBOP_SUCLAMP_SD(0, 1): m
= 0; break;
1733 case NV50_IR_SUBOP_SUCLAMP_SD(1, 1): m
= 1; break;
1734 case NV50_IR_SUBOP_SUCLAMP_SD(2, 1): m
= 2; break;
1735 case NV50_IR_SUBOP_SUCLAMP_SD(3, 1): m
= 3; break;
1736 case NV50_IR_SUBOP_SUCLAMP_SD(4, 1): m
= 4; break;
1737 case NV50_IR_SUBOP_SUCLAMP_PL(0, 1): m
= 5; break;
1738 case NV50_IR_SUBOP_SUCLAMP_PL(1, 1): m
= 6; break;
1739 case NV50_IR_SUBOP_SUCLAMP_PL(2, 1): m
= 7; break;
1740 case NV50_IR_SUBOP_SUCLAMP_PL(3, 1): m
= 8; break;
1741 case NV50_IR_SUBOP_SUCLAMP_PL(4, 1): m
= 9; break;
1742 case NV50_IR_SUBOP_SUCLAMP_BL(0, 1): m
= 10; break;
1743 case NV50_IR_SUBOP_SUCLAMP_BL(1, 1): m
= 11; break;
1744 case NV50_IR_SUBOP_SUCLAMP_BL(2, 1): m
= 12; break;
1745 case NV50_IR_SUBOP_SUCLAMP_BL(3, 1): m
= 13; break;
1746 case NV50_IR_SUBOP_SUCLAMP_BL(4, 1): m
= 14; break;
1751 if (subOp
& NV50_IR_SUBOP_SUCLAMP_2D
)
1756 CodeEmitterGK110::emitSUCalc(Instruction
*i
)
1758 ImmediateValue
*imm
= NULL
;
1759 uint64_t opc1
, opc2
;
1761 if (i
->srcExists(2)) {
1762 imm
= i
->getSrc(2)->asImm();
1764 i
->setSrc(2, NULL
); // special case, make emitForm_21 not assert
1768 case OP_SUCLAMP
: opc1
= 0xb00; opc2
= 0x580; break;
1769 case OP_SUBFM
: opc1
= 0xb68; opc2
= 0x1e8; break;
1770 case OP_SUEAU
: opc1
= 0xb6c; opc2
= 0x1ec; break;
1775 emitForm_21(i
, opc2
, opc1
);
1777 if (i
->op
== OP_SUCLAMP
) {
1778 if (i
->dType
== TYPE_S32
)
1780 emitSUCLAMPMode(i
->subOp
);
1783 if (i
->op
== OP_SUBFM
&& i
->subOp
== NV50_IR_SUBOP_SUBFM_3D
)
1786 if (i
->op
!= OP_SUEAU
) {
1787 const uint8_t pos
= i
->op
== OP_SUBFM
? 19 : 16;
1788 if (i
->def(0).getFile() == FILE_PREDICATE
) { // p, #
1789 code
[0] |= 255 << 2;
1790 code
[1] |= i
->getDef(1)->reg
.data
.id
<< pos
;
1792 if (i
->defExists(1)) { // r, p
1793 assert(i
->def(1).getFile() == FILE_PREDICATE
);
1794 code
[1] |= i
->getDef(1)->reg
.data
.id
<< pos
;
1796 code
[1] |= 7 << pos
;
1801 assert(i
->op
== OP_SUCLAMP
);
1803 code
[1] |= (imm
->reg
.data
.u32
& 0x3f) << 10; // sint6
1809 CodeEmitterGK110::emitVectorSubOp(const Instruction
*i
)
1811 switch (NV50_IR_SUBOP_Vn(i
->subOp
)) {
1813 code
[1] |= (i
->subOp
& 0x000f) << 7; // vsrc1
1814 code
[1] |= (i
->subOp
& 0x00e0) >> 6; // vsrc2
1815 code
[1] |= (i
->subOp
& 0x0100) << 13; // vsrc2
1816 code
[1] |= (i
->subOp
& 0x3c00) << 12; // vdst
1825 CodeEmitterGK110::emitVSHL(const Instruction
*i
)
1827 code
[0] = 0x00000002;
1828 code
[1] = 0xb8000000;
1830 assert(NV50_IR_SUBOP_Vn(i
->subOp
) == 0);
1832 if (isSignedType(i
->dType
)) code
[1] |= 1 << 25;
1833 if (isSignedType(i
->sType
)) code
[1] |= 1 << 19;
1838 defId(i
->def(0), 2);
1839 srcId(i
->src(0), 10);
1841 if (i
->getSrc(1)->reg
.file
== FILE_IMMEDIATE
) {
1842 ImmediateValue
*imm
= i
->getSrc(1)->asImm();
1844 code
[0] |= (imm
->reg
.data
.u32
& 0x01ff) << 23;
1845 code
[1] |= (imm
->reg
.data
.u32
& 0xfe00) >> 9;
1847 assert(i
->getSrc(1)->reg
.file
== FILE_GPR
);
1849 srcId(i
->src(1), 23);
1851 srcId(i
->src(2), 42);
1855 if (i
->flagsDef
>= 0)
1860 CodeEmitterGK110::emitAFETCH(const Instruction
*i
)
1862 uint32_t offset
= i
->src(0).get()->reg
.data
.offset
& 0x7ff;
1864 code
[0] = 0x00000002 | (offset
<< 23);
1865 code
[1] = 0x7d000000 | (offset
>> 9);
1867 if (i
->getSrc(0)->reg
.file
== FILE_SHADER_OUTPUT
)
1872 defId(i
->def(0), 2);
1873 srcId(i
->src(0).getIndirect(0), 10);
1877 CodeEmitterGK110::emitPFETCH(const Instruction
*i
)
1879 uint32_t prim
= i
->src(0).get()->reg
.data
.u32
;
1881 code
[0] = 0x00000002 | ((prim
& 0xff) << 23);
1882 code
[1] = 0x7f800000;
1886 const int src1
= (i
->predSrc
== 1) ? 2 : 1; // if predSrc == 1, !srcExists(2)
1888 defId(i
->def(0), 2);
1893 CodeEmitterGK110::emitVFETCH(const Instruction
*i
)
1895 unsigned int size
= typeSizeof(i
->dType
);
1896 uint32_t offset
= i
->src(0).get()->reg
.data
.offset
;
1898 code
[0] = 0x00000002 | (offset
<< 23);
1899 code
[1] = 0x7ec00000 | (offset
>> 9);
1900 code
[1] |= (size
/ 4 - 1) << 18;
1904 if (i
->getSrc(0)->reg
.file
== FILE_SHADER_OUTPUT
)
1905 code
[1] |= 0x8; // yes, TCPs can read from *outputs* of other threads
1909 defId(i
->def(0), 2);
1910 srcId(i
->src(0).getIndirect(0), 10);
1911 srcId(i
->src(0).getIndirect(1), 32 + 10); // vertex address
1915 CodeEmitterGK110::emitEXPORT(const Instruction
*i
)
1917 unsigned int size
= typeSizeof(i
->dType
);
1918 uint32_t offset
= i
->src(0).get()->reg
.data
.offset
;
1920 code
[0] = 0x00000002 | (offset
<< 23);
1921 code
[1] = 0x7f000000 | (offset
>> 9);
1922 code
[1] |= (size
/ 4 - 1) << 18;
1929 assert(i
->src(1).getFile() == FILE_GPR
);
1931 srcId(i
->src(0).getIndirect(0), 10);
1932 srcId(i
->src(0).getIndirect(1), 32 + 10); // vertex base address
1933 srcId(i
->src(1), 2);
1937 CodeEmitterGK110::emitOUT(const Instruction
*i
)
1939 assert(i
->src(0).getFile() == FILE_GPR
);
1941 emitForm_21(i
, 0x1f0, 0xb70);
1943 if (i
->op
== OP_EMIT
)
1945 if (i
->op
== OP_RESTART
|| i
->subOp
== NV50_IR_SUBOP_EMIT_RESTART
)
1950 CodeEmitterGK110::emitInterpMode(const Instruction
*i
)
1952 code
[1] |= (i
->ipa
& 0x3) << 21; // TODO: INTERP_SAMPLEID
1953 code
[1] |= (i
->ipa
& 0xc) << (19 - 2);
1957 interpApply(const FixupEntry
*entry
, uint32_t *code
, const FixupData
& data
)
1959 int ipa
= entry
->ipa
;
1960 int reg
= entry
->reg
;
1961 int loc
= entry
->loc
;
1963 if (data
.flatshade
&&
1964 (ipa
& NV50_IR_INTERP_MODE_MASK
) == NV50_IR_INTERP_SC
) {
1965 ipa
= NV50_IR_INTERP_FLAT
;
1967 } else if (data
.force_persample_interp
&&
1968 (ipa
& NV50_IR_INTERP_SAMPLE_MASK
) == NV50_IR_INTERP_DEFAULT
&&
1969 (ipa
& NV50_IR_INTERP_MODE_MASK
) != NV50_IR_INTERP_FLAT
) {
1970 ipa
|= NV50_IR_INTERP_CENTROID
;
1972 code
[loc
+ 1] &= ~(0xf << 19);
1973 code
[loc
+ 1] |= (ipa
& 0x3) << 21;
1974 code
[loc
+ 1] |= (ipa
& 0xc) << (19 - 2);
1975 code
[loc
+ 0] &= ~(0xff << 23);
1976 code
[loc
+ 0] |= reg
<< 23;
1980 CodeEmitterGK110::emitINTERP(const Instruction
*i
)
1982 const uint32_t base
= i
->getSrc(0)->reg
.data
.offset
;
1984 code
[0] = 0x00000002 | (base
<< 31);
1985 code
[1] = 0x74800000 | (base
>> 1);
1990 if (i
->op
== OP_PINTERP
) {
1991 srcId(i
->src(1), 23);
1992 addInterp(i
->ipa
, SDATA(i
->src(1)).id
, interpApply
);
1994 code
[0] |= 0xff << 23;
1995 addInterp(i
->ipa
, 0xff, interpApply
);
1998 srcId(i
->src(0).getIndirect(0), 10);
2002 defId(i
->def(0), 2);
2004 if (i
->getSampleMode() == NV50_IR_INTERP_OFFSET
)
2005 srcId(i
->src(i
->op
== OP_PINTERP
? 2 : 1), 32 + 10);
2007 code
[1] |= 0xff << 10;
2011 CodeEmitterGK110::emitLoadStoreType(DataType ty
, const int pos
)
2043 assert(!"invalid ld/st type");
2046 code
[pos
/ 32] |= n
<< (pos
% 32);
2050 CodeEmitterGK110::emitCachingMode(CacheMode c
, const int pos
)
2071 assert(!"invalid caching mode");
2074 code
[pos
/ 32] |= n
<< (pos
% 32);
2078 CodeEmitterGK110::emitSTORE(const Instruction
*i
)
2080 int32_t offset
= SDATA(i
->src(0)).offset
;
2082 switch (i
->src(0).getFile()) {
2083 case FILE_MEMORY_GLOBAL
: code
[1] = 0xe0000000; code
[0] = 0x00000000; break;
2084 case FILE_MEMORY_LOCAL
: code
[1] = 0x7a800000; code
[0] = 0x00000002; break;
2085 case FILE_MEMORY_SHARED
:
2086 code
[0] = 0x00000002;
2087 if (i
->subOp
== NV50_IR_SUBOP_STORE_UNLOCKED
)
2088 code
[1] = 0x78400000;
2090 code
[1] = 0x7ac00000;
2093 assert(!"invalid memory file");
2097 if (code
[0] & 0x2) {
2099 emitLoadStoreType(i
->dType
, 0x33);
2100 if (i
->src(0).getFile() == FILE_MEMORY_LOCAL
)
2101 emitCachingMode(i
->cache
, 0x2f);
2103 emitLoadStoreType(i
->dType
, 0x38);
2104 emitCachingMode(i
->cache
, 0x3b);
2106 code
[0] |= offset
<< 23;
2107 code
[1] |= offset
>> 9;
2109 // Unlocked store on shared memory can fail.
2110 if (i
->src(0).getFile() == FILE_MEMORY_SHARED
&&
2111 i
->subOp
== NV50_IR_SUBOP_STORE_UNLOCKED
) {
2112 assert(i
->defExists(0));
2113 defId(i
->def(0), 32 + 16);
2118 srcId(i
->src(1), 2);
2119 srcId(i
->src(0).getIndirect(0), 10);
2120 if (i
->src(0).getFile() == FILE_MEMORY_GLOBAL
&&
2121 i
->src(0).isIndirect(0) &&
2122 i
->getIndirect(0, 0)->reg
.size
== 8)
2127 CodeEmitterGK110::emitLOAD(const Instruction
*i
)
2129 int32_t offset
= SDATA(i
->src(0)).offset
;
2131 switch (i
->src(0).getFile()) {
2132 case FILE_MEMORY_GLOBAL
: code
[1] = 0xc0000000; code
[0] = 0x00000000; break;
2133 case FILE_MEMORY_LOCAL
: code
[1] = 0x7a000000; code
[0] = 0x00000002; break;
2134 case FILE_MEMORY_SHARED
:
2135 code
[0] = 0x00000002;
2136 if (i
->subOp
== NV50_IR_SUBOP_LOAD_LOCKED
)
2137 code
[1] = 0x77400000;
2139 code
[1] = 0x7a400000;
2141 case FILE_MEMORY_CONST
:
2142 if (!i
->src(0).isIndirect(0) && typeSizeof(i
->dType
) == 4) {
2147 code
[0] = 0x00000002;
2148 code
[1] = 0x7c800000 | (i
->src(0).get()->reg
.fileIndex
<< 7);
2149 code
[1] |= i
->subOp
<< 15;
2152 assert(!"invalid memory file");
2156 if (code
[0] & 0x2) {
2158 emitLoadStoreType(i
->dType
, 0x33);
2159 if (i
->src(0).getFile() == FILE_MEMORY_LOCAL
)
2160 emitCachingMode(i
->cache
, 0x2f);
2162 emitLoadStoreType(i
->dType
, 0x38);
2163 emitCachingMode(i
->cache
, 0x3b);
2165 code
[0] |= offset
<< 23;
2166 code
[1] |= offset
>> 9;
2168 // Locked store on shared memory can fail.
2170 if (i
->src(0).getFile() == FILE_MEMORY_SHARED
&&
2171 i
->subOp
== NV50_IR_SUBOP_LOAD_LOCKED
) {
2172 if (i
->def(0).getFile() == FILE_PREDICATE
) { // p, #
2175 } else if (i
->defExists(1)) { // r, p
2178 assert(!"Expected predicate dest for load locked");
2185 defId(i
->def(r
), 2);
2187 code
[0] |= 255 << 2;
2190 defId(i
->def(p
), 32 + 16);
2192 if (i
->getIndirect(0, 0)) {
2193 srcId(i
->src(0).getIndirect(0), 10);
2194 if (i
->getIndirect(0, 0)->reg
.size
== 8)
2197 code
[0] |= 255 << 10;
2202 CodeEmitterGK110::getSRegEncoding(const ValueRef
& ref
)
2204 switch (SDATA(ref
).sv
.sv
) {
2205 case SV_LANEID
: return 0x00;
2206 case SV_PHYSID
: return 0x03;
2207 case SV_VERTEX_COUNT
: return 0x10;
2208 case SV_INVOCATION_ID
: return 0x11;
2209 case SV_YDIR
: return 0x12;
2210 case SV_THREAD_KILL
: return 0x13;
2211 case SV_TID
: return 0x21 + SDATA(ref
).sv
.index
;
2212 case SV_CTAID
: return 0x25 + SDATA(ref
).sv
.index
;
2213 case SV_NTID
: return 0x29 + SDATA(ref
).sv
.index
;
2214 case SV_GRIDID
: return 0x2c;
2215 case SV_NCTAID
: return 0x2d + SDATA(ref
).sv
.index
;
2216 case SV_LBASE
: return 0x34;
2217 case SV_SBASE
: return 0x30;
2218 case SV_CLOCK
: return 0x50 + SDATA(ref
).sv
.index
;
2220 assert(!"no sreg for system value");
2226 CodeEmitterGK110::emitMOV(const Instruction
*i
)
2228 if (i
->def(0).getFile() == FILE_PREDICATE
) {
2229 if (i
->src(0).getFile() == FILE_GPR
) {
2230 // Use ISETP.NE.AND dst, PT, src, RZ, PT
2231 code
[0] = 0x00000002;
2232 code
[1] = 0xdb500000;
2234 code
[0] |= 0x7 << 2;
2235 code
[0] |= 0xff << 23;
2236 code
[1] |= 0x7 << 10;
2237 srcId(i
->src(0), 10);
2239 if (i
->src(0).getFile() == FILE_PREDICATE
) {
2240 // Use PSETP.AND.AND dst, PT, src, PT, PT
2241 code
[0] = 0x00000002;
2242 code
[1] = 0x84800000;
2244 code
[0] |= 0x7 << 2;
2245 code
[1] |= 0x7 << 0;
2246 code
[1] |= 0x7 << 10;
2248 srcId(i
->src(0), 14);
2250 assert(!"Unexpected source for predicate destination");
2254 defId(i
->def(0), 5);
2256 if (i
->src(0).getFile() == FILE_SYSTEM_VALUE
) {
2257 code
[0] = 0x00000002 | (getSRegEncoding(i
->src(0)) << 23);
2258 code
[1] = 0x86400000;
2260 defId(i
->def(0), 2);
2262 if (i
->src(0).getFile() == FILE_IMMEDIATE
) {
2263 code
[0] = 0x00000002 | (i
->lanes
<< 14);
2264 code
[1] = 0x74000000;
2266 defId(i
->def(0), 2);
2267 setImmediate32(i
, 0, Modifier(0));
2269 if (i
->src(0).getFile() == FILE_PREDICATE
) {
2270 code
[0] = 0x00000002;
2271 code
[1] = 0x84401c07;
2273 defId(i
->def(0), 2);
2274 srcId(i
->src(0), 14);
2276 emitForm_C(i
, 0x24c, 2);
2277 code
[1] |= i
->lanes
<< 10;
2282 uses64bitAddress(const Instruction
*ldst
)
2284 return ldst
->src(0).getFile() == FILE_MEMORY_GLOBAL
&&
2285 ldst
->src(0).isIndirect(0) &&
2286 ldst
->getIndirect(0, 0)->reg
.size
== 8;
2290 CodeEmitterGK110::emitATOM(const Instruction
*i
)
2292 const bool hasDst
= i
->defExists(0);
2293 const bool exch
= i
->subOp
== NV50_IR_SUBOP_ATOM_EXCH
;
2295 code
[0] = 0x00000002;
2296 if (i
->subOp
== NV50_IR_SUBOP_ATOM_CAS
)
2297 code
[1] = 0x77800000;
2299 code
[1] = 0x68000000;
2302 case NV50_IR_SUBOP_ATOM_CAS
: break;
2303 case NV50_IR_SUBOP_ATOM_EXCH
: code
[1] |= 0x04000000; break;
2304 default: code
[1] |= i
->subOp
<< 23; break;
2308 case TYPE_U32
: break;
2309 case TYPE_S32
: code
[1] |= 0x00100000; break;
2310 case TYPE_U64
: code
[1] |= 0x00200000; break;
2311 case TYPE_F32
: code
[1] |= 0x00300000; break;
2312 case TYPE_B128
: code
[1] |= 0x00400000; break; /* TODO: U128 */
2313 case TYPE_S64
: code
[1] |= 0x00500000; break;
2314 default: assert(!"unsupported type"); break;
2319 /* TODO: cas: check that src regs line up */
2320 /* TODO: cas: flip bits if $r255 is used */
2321 srcId(i
->src(1), 23);
2324 defId(i
->def(0), 2);
2327 code
[0] |= 255 << 2;
2330 if (hasDst
|| !exch
) {
2331 const int32_t offset
= SDATA(i
->src(0)).offset
;
2332 assert(offset
< 0x80000 && offset
>= -0x80000);
2333 code
[0] |= (offset
& 1) << 31;
2334 code
[1] |= (offset
& 0xffffe) >> 1;
2336 srcAddr32(i
->src(0), 31);
2339 if (i
->getIndirect(0, 0)) {
2340 srcId(i
->getIndirect(0, 0), 10);
2341 if (i
->getIndirect(0, 0)->reg
.size
== 8)
2344 code
[0] |= 255 << 10;
2349 CodeEmitterGK110::emitCCTL(const Instruction
*i
)
2351 int32_t offset
= SDATA(i
->src(0)).offset
;
2353 code
[0] = 0x00000002 | (i
->subOp
<< 2);
2355 if (i
->src(0).getFile() == FILE_MEMORY_GLOBAL
) {
2356 code
[1] = 0x7b000000;
2358 code
[1] = 0x7c000000;
2361 code
[0] |= offset
<< 23;
2362 code
[1] |= offset
>> 9;
2364 if (uses64bitAddress(i
))
2366 srcId(i
->src(0).getIndirect(0), 10);
2372 CodeEmitterGK110::emitInstruction(Instruction
*insn
)
2374 const unsigned int size
= (writeIssueDelays
&& !(codeSize
& 0x3f)) ? 16 : 8;
2376 if (insn
->encSize
!= 8) {
2377 ERROR("skipping unencodable instruction: ");
2381 if (codeSize
+ size
> codeSizeLimit
) {
2382 ERROR("code emitter output buffer too small\n");
2386 if (writeIssueDelays
) {
2387 int id
= (codeSize
& 0x3f) / 8 - 1;
2390 code
[0] = 0x00000000; // cf issue delay "instruction"
2391 code
[1] = 0x08000000;
2395 uint32_t *data
= code
- (id
* 2 + 2);
2398 case 0: data
[0] |= insn
->sched
<< 2; break;
2399 case 1: data
[0] |= insn
->sched
<< 10; break;
2400 case 2: data
[0] |= insn
->sched
<< 18; break;
2401 case 3: data
[0] |= insn
->sched
<< 26; data
[1] |= insn
->sched
>> 6; break;
2402 case 4: data
[1] |= insn
->sched
<< 2; break;
2403 case 5: data
[1] |= insn
->sched
<< 10; break;
2404 case 6: data
[1] |= insn
->sched
<< 18; break;
2411 // assert that instructions with multiple defs don't corrupt registers
2412 for (int d
= 0; insn
->defExists(d
); ++d
)
2413 assert(insn
->asTex() || insn
->def(d
).rep()->reg
.data
.id
>= 0);
2450 if (insn
->dType
== TYPE_F64
)
2452 else if (isFloatType(insn
->dType
))
2458 if (insn
->dType
== TYPE_F64
)
2460 else if (isFloatType(insn
->dType
))
2467 if (insn
->dType
== TYPE_F64
)
2469 else if (isFloatType(insn
->dType
))
2487 emitLogicOp(insn
, 0);
2490 emitLogicOp(insn
, 1);
2493 emitLogicOp(insn
, 2);
2497 if (typeSizeof(insn
->sType
) == 8)
2506 emitSET(insn
->asCmp());
2512 emitSLCT(insn
->asCmp());
2527 if (insn
->def(0).getFile() == FILE_PREDICATE
||
2528 insn
->src(0).getFile() == FILE_PREDICATE
)
2534 emitSFnOp(insn
, 5 + 2 * insn
->subOp
);
2537 emitSFnOp(insn
, 4 + 2 * insn
->subOp
);
2562 emitTEX(insn
->asTex());
2565 emitTXQ(insn
->asTex());
2590 emitQUADOP(insn
, insn
->subOp
, insn
->lanes
);
2593 emitQUADOP(insn
, insn
->src(0).mod
.neg() ? 0x66 : 0x99, 0x4);
2596 emitQUADOP(insn
, insn
->src(0).mod
.neg() ? 0x5a : 0xa5, 0x5);
2633 emitSULDGB(insn
->asTex());
2637 emitSUSTGx(insn
->asTex());
2650 ERROR("operation should have been eliminated");
2656 ERROR("operation should have been lowered\n");
2659 ERROR("unknown op: %u\n", insn
->op
);
2672 CodeEmitterGK110::getMinEncodingSize(const Instruction
*i
) const
2674 // No more short instruction encodings.
2679 CodeEmitterGK110::prepareEmission(Function
*func
)
2681 const Target
*targ
= func
->getProgram()->getTarget();
2683 CodeEmitter::prepareEmission(func
);
2685 if (targ
->hasSWSched
)
2686 calculateSchedDataNVC0(targ
, func
);
2689 CodeEmitterGK110::CodeEmitterGK110(const TargetNVC0
*target
)
2690 : CodeEmitter(target
),
2692 writeIssueDelays(target
->hasSWSched
)
2695 codeSize
= codeSizeLimit
= 0;
2700 TargetNVC0::createCodeEmitterGK110(Program::Type type
)
2702 CodeEmitterGK110
*emit
= new CodeEmitterGK110(this);
2703 emit
->setProgramType(type
);
2707 } // namespace nv50_ir