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
, int sCount
= 3);
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
,
368 Modifier mod
, int sCount
)
377 for (int s
= 0; s
< sCount
&& 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 bool neg1
= (i
->src(0).mod
^ i
->src(1).mod
).neg();
492 if (isLIMM(i
->src(1), TYPE_F32
)) {
493 assert(i
->getDef(0)->reg
.data
.id
== i
->getSrc(2)->reg
.data
.id
);
495 // last source is dst, so force 2 sources
496 emitForm_L(i
, 0x600, 0x0, 0, 2);
498 if (i
->flagsDef
>= 0)
508 emitForm_21(i
, 0x0c0, 0x940);
528 CodeEmitterGK110::emitDMAD(const Instruction
*i
)
530 assert(!i
->saturate
);
533 emitForm_21(i
, 0x1b8, 0xb38);
538 bool neg1
= (i
->src(0).mod
^ i
->src(1).mod
).neg();
550 CodeEmitterGK110::emitMADSP(const Instruction
*i
)
552 emitForm_21(i
, 0x140, 0xa40);
554 if (i
->subOp
== NV50_IR_SUBOP_MADSP_SD
) {
555 code
[1] |= 0x00c00000;
557 code
[1] |= (i
->subOp
& 0x00f) << 19; // imadp1
558 code
[1] |= (i
->subOp
& 0x0f0) << 20; // imadp2
559 code
[1] |= (i
->subOp
& 0x100) << 11; // imadp3
560 code
[1] |= (i
->subOp
& 0x200) << 15; // imadp3
561 code
[1] |= (i
->subOp
& 0xc00) << 12; // imadp3
564 if (i
->flagsDef
>= 0)
569 CodeEmitterGK110::emitFMUL(const Instruction
*i
)
571 bool neg
= (i
->src(0).mod
^ i
->src(1).mod
).neg();
573 assert(i
->postFactor
>= -3 && i
->postFactor
<= 3);
575 if (isLIMM(i
->src(1), TYPE_F32
)) {
576 emitForm_L(i
, 0x200, 0x2, Modifier(0));
584 assert(i
->postFactor
== 0);
586 emitForm_21(i
, 0x234, 0xc34);
587 code
[1] |= ((i
->postFactor
> 0) ?
588 (7 - i
->postFactor
) : (0 - i
->postFactor
)) << 12;
606 CodeEmitterGK110::emitDMUL(const Instruction
*i
)
608 bool neg
= (i
->src(0).mod
^ i
->src(1).mod
).neg();
610 assert(!i
->postFactor
);
611 assert(!i
->saturate
);
615 emitForm_21(i
, 0x240, 0xc40);
629 CodeEmitterGK110::emitIMUL(const Instruction
*i
)
631 assert(!i
->src(0).mod
.neg() && !i
->src(1).mod
.neg());
632 assert(!i
->src(0).mod
.abs() && !i
->src(1).mod
.abs());
634 if (i
->src(1).getFile() == FILE_IMMEDIATE
) {
635 emitForm_L(i
, 0x280, 2, Modifier(0));
637 if (i
->subOp
== NV50_IR_SUBOP_MUL_HIGH
)
639 if (i
->sType
== TYPE_S32
)
642 emitForm_21(i
, 0x21c, 0xc1c);
644 if (i
->subOp
== NV50_IR_SUBOP_MUL_HIGH
)
646 if (i
->sType
== TYPE_S32
)
652 CodeEmitterGK110::emitFADD(const Instruction
*i
)
654 if (isLIMM(i
->src(1), TYPE_F32
)) {
655 assert(i
->rnd
== ROUND_N
);
656 assert(!i
->saturate
);
658 Modifier mod
= i
->src(1).mod
^
659 Modifier(i
->op
== OP_SUB
? NV50_IR_MOD_NEG
: 0);
661 emitForm_L(i
, 0x400, 0, mod
);
667 emitForm_21(i
, 0x22c, 0xc2c);
676 modNegAbsF32_3b(i
, 1);
677 if (i
->op
== OP_SUB
) code
[1] ^= 1 << 27;
681 if (i
->op
== OP_SUB
) code
[1] ^= 1 << 16;
687 CodeEmitterGK110::emitDADD(const Instruction
*i
)
689 assert(!i
->saturate
);
692 emitForm_21(i
, 0x238, 0xc38);
697 modNegAbsF32_3b(i
, 1);
698 if (i
->op
== OP_SUB
) code
[1] ^= 1 << 27;
702 if (i
->op
== OP_SUB
) code
[1] ^= 1 << 16;
707 CodeEmitterGK110::emitUADD(const Instruction
*i
)
709 uint8_t addOp
= (i
->src(0).mod
.neg() << 1) | i
->src(1).mod
.neg();
714 assert(!i
->src(0).mod
.abs() && !i
->src(1).mod
.abs());
716 if (isLIMM(i
->src(1), TYPE_S32
)) {
717 emitForm_L(i
, 0x400, 1, Modifier((addOp
& 1) ? NV50_IR_MOD_NEG
: 0));
722 assert(i
->flagsDef
< 0);
723 assert(i
->flagsSrc
< 0);
727 emitForm_21(i
, 0x208, 0xc08);
729 assert(addOp
!= 3); // would be add-plus-one
731 code
[1] |= addOp
<< 19;
733 if (i
->flagsDef
>= 0)
734 code
[1] |= 1 << 18; // write carry
735 if (i
->flagsSrc
>= 0)
736 code
[1] |= 1 << 14; // add carry
743 CodeEmitterGK110::emitIMAD(const Instruction
*i
)
746 i
->src(2).mod
.neg() | ((i
->src(0).mod
.neg() ^ i
->src(1).mod
.neg()) << 1);
748 emitForm_21(i
, 0x100, 0xa00);
751 code
[1] |= addOp
<< 26;
753 if (i
->sType
== TYPE_S32
)
754 code
[1] |= (1 << 19) | (1 << 24);
756 if (i
->subOp
== NV50_IR_SUBOP_MUL_HIGH
)
759 if (i
->flagsDef
>= 0) code
[1] |= 1 << 18;
760 if (i
->flagsSrc
>= 0) code
[1] |= 1 << 20;
766 CodeEmitterGK110::emitISAD(const Instruction
*i
)
768 assert(i
->dType
== TYPE_S32
|| i
->dType
== TYPE_U32
);
770 emitForm_21(i
, 0x1f4, 0xb74);
772 if (i
->dType
== TYPE_S32
)
777 CodeEmitterGK110::emitSHLADD(const Instruction
*i
)
779 uint8_t addOp
= (i
->src(0).mod
.neg() << 1) | i
->src(2).mod
.neg();
780 const ImmediateValue
*imm
= i
->src(1).get()->asImm();
783 if (i
->src(2).getFile() == FILE_IMMEDIATE
) {
785 code
[1] = 0xc0c << 20;
788 code
[1] = 0x20c << 20;
790 code
[1] |= addOp
<< 19;
795 srcId(i
->src(0), 10);
797 if (i
->flagsDef
>= 0)
800 assert(!(imm
->reg
.data
.u32
& 0xffffffe0));
801 code
[1] |= imm
->reg
.data
.u32
<< 10;
803 switch (i
->src(2).getFile()) {
805 assert(code
[0] & 0x2);
806 code
[1] |= 0xc << 28;
807 srcId(i
->src(2), 23);
809 case FILE_MEMORY_CONST
:
810 assert(code
[0] & 0x2);
811 code
[1] |= 0x4 << 28;
812 setCAddress14(i
->src(2));
815 assert(code
[0] & 0x1);
816 setShortImmediate(i
, 2);
819 assert(!"bad src2 file");
825 CodeEmitterGK110::emitNOT(const Instruction
*i
)
827 code
[0] = 0x0003fc02; // logop(mov2) dst, 0, not src
828 code
[1] = 0x22003800;
834 switch (i
->src(0).getFile()) {
836 code
[1] |= 0xc << 28;
837 srcId(i
->src(0), 23);
839 case FILE_MEMORY_CONST
:
840 code
[1] |= 0x4 << 28;
841 setCAddress14(i
->src(0));
850 CodeEmitterGK110::emitLogicOp(const Instruction
*i
, uint8_t subOp
)
852 if (i
->def(0).getFile() == FILE_PREDICATE
) {
853 code
[0] = 0x00000002 | (subOp
<< 27);
854 code
[1] = 0x84800000;
859 srcId(i
->src(0), 14);
860 if (i
->src(0).mod
== Modifier(NV50_IR_MOD_NOT
)) code
[0] |= 1 << 17;
861 srcId(i
->src(1), 32);
862 if (i
->src(1).mod
== Modifier(NV50_IR_MOD_NOT
)) code
[1] |= 1 << 3;
864 if (i
->defExists(1)) {
870 if (i
->predSrc
!= 2 && i
->srcExists(2)) {
871 code
[1] |= subOp
<< 16;
872 srcId(i
->src(2), 42);
873 if (i
->src(2).mod
== Modifier(NV50_IR_MOD_NOT
)) code
[1] |= 1 << 13;
878 if (isLIMM(i
->src(1), TYPE_S32
)) {
879 emitForm_L(i
, 0x200, 0, i
->src(1).mod
);
880 code
[1] |= subOp
<< 24;
883 emitForm_21(i
, 0x220, 0xc20);
884 code
[1] |= subOp
<< 12;
891 CodeEmitterGK110::emitPOPC(const Instruction
*i
)
893 assert(!isLIMM(i
->src(1), TYPE_S32
, true));
895 emitForm_21(i
, 0x204, 0xc04);
898 if (!(code
[0] & 0x1))
903 CodeEmitterGK110::emitINSBF(const Instruction
*i
)
905 emitForm_21(i
, 0x1f8, 0xb78);
909 CodeEmitterGK110::emitEXTBF(const Instruction
*i
)
911 emitForm_21(i
, 0x600, 0xc00);
913 if (i
->dType
== TYPE_S32
)
915 if (i
->subOp
== NV50_IR_SUBOP_EXTBF_REV
)
920 CodeEmitterGK110::emitBFIND(const Instruction
*i
)
922 emitForm_C(i
, 0x218, 0x2);
924 if (i
->dType
== TYPE_S32
)
926 if (i
->src(0).mod
== Modifier(NV50_IR_MOD_NOT
))
928 if (i
->subOp
== NV50_IR_SUBOP_BFIND_SAMT
)
933 CodeEmitterGK110::emitPERMT(const Instruction
*i
)
935 emitForm_21(i
, 0x1e0, 0xb60);
937 code
[1] |= i
->subOp
<< 19;
941 CodeEmitterGK110::emitShift(const Instruction
*i
)
943 if (i
->op
== OP_SHR
) {
944 emitForm_21(i
, 0x214, 0xc14);
945 if (isSignedType(i
->dType
))
948 emitForm_21(i
, 0x224, 0xc24);
951 if (i
->subOp
== NV50_IR_SUBOP_SHIFT_WRAP
)
956 CodeEmitterGK110::emitShift64(const Instruction
*i
)
958 if (i
->op
== OP_SHR
) {
959 emitForm_21(i
, 0x27c, 0xc7c);
960 if (isSignedType(i
->sType
))
962 if (i
->subOp
& NV50_IR_SUBOP_SHIFT_HIGH
)
965 emitForm_21(i
, 0xdfc, 0xf7c);
969 if (i
->subOp
& NV50_IR_SUBOP_SHIFT_WRAP
)
974 CodeEmitterGK110::emitPreOp(const Instruction
*i
)
976 emitForm_C(i
, 0x248, 0x2);
978 if (i
->op
== OP_PREEX2
)
986 CodeEmitterGK110::emitSFnOp(const Instruction
*i
, uint8_t subOp
)
988 code
[0] = 0x00000002 | (subOp
<< 23);
989 code
[1] = 0x84000000;
994 srcId(i
->src(0), 10);
1002 CodeEmitterGK110::emitMINMAX(const Instruction
*i
)
1026 emitForm_21(i
, op2
, op1
);
1028 if (i
->dType
== TYPE_S32
)
1030 code
[1] |= (i
->op
== OP_MIN
) ? 0x1c00 : 0x3c00; // [!]pt
1031 code
[1] |= i
->subOp
<< 14;
1032 if (i
->flagsDef
>= 0)
1033 code
[1] |= i
->subOp
<< 18;
1038 if (code
[0] & 0x1) {
1039 modNegAbsF32_3b(i
, 1);
1047 CodeEmitterGK110::emitCVT(const Instruction
*i
)
1049 const bool f2f
= isFloatType(i
->dType
) && isFloatType(i
->sType
);
1050 const bool f2i
= !isFloatType(i
->dType
) && isFloatType(i
->sType
);
1051 const bool i2f
= isFloatType(i
->dType
) && !isFloatType(i
->sType
);
1053 bool sat
= i
->saturate
;
1054 bool abs
= i
->src(0).mod
.abs();
1055 bool neg
= i
->src(0).mod
.neg();
1057 RoundMode rnd
= i
->rnd
;
1060 case OP_CEIL
: rnd
= f2f
? ROUND_PI
: ROUND_P
; break;
1061 case OP_FLOOR
: rnd
= f2f
? ROUND_MI
: ROUND_M
; break;
1062 case OP_TRUNC
: rnd
= f2f
? ROUND_ZI
: ROUND_Z
; break;
1063 case OP_SAT
: sat
= true; break;
1064 case OP_NEG
: neg
= !neg
; break;
1065 case OP_ABS
: abs
= true; neg
= false; break;
1072 if (i
->op
== OP_NEG
&& i
->dType
== TYPE_U32
)
1080 if (f2f
) op
= 0x254;
1081 else if (f2i
) op
= 0x258;
1082 else if (i2f
) op
= 0x25c;
1085 emitForm_C(i
, op
, 0x2);
1088 if (neg
) code
[1] |= 1 << 16;
1089 if (abs
) code
[1] |= 1 << 20;
1090 if (sat
) code
[1] |= 1 << 21;
1092 emitRoundMode(rnd
, 32 + 10, f2f
? (32 + 13) : -1);
1094 code
[0] |= typeSizeofLog2(dType
) << 10;
1095 code
[0] |= typeSizeofLog2(i
->sType
) << 12;
1096 code
[1] |= i
->subOp
<< 12;
1098 if (isSignedIntType(dType
))
1100 if (isSignedIntType(i
->sType
))
1105 CodeEmitterGK110::emitSET(const CmpInstruction
*i
)
1109 if (i
->def(0).getFile() == FILE_PREDICATE
) {
1111 case TYPE_F32
: op2
= 0x1d8; op1
= 0xb58; break;
1112 case TYPE_F64
: op2
= 0x1c0; op1
= 0xb40; break;
1118 emitForm_21(i
, op2
, op1
);
1122 if (!(code
[0] & 0x1)) {
1126 modNegAbsF32_3b(i
, 1);
1130 // normal DST field is negated predicate result
1131 code
[0] = (code
[0] & ~0xfc) | ((code
[0] << 3) & 0xe0);
1132 if (i
->defExists(1))
1133 defId(i
->def(1), 2);
1138 case TYPE_F32
: op2
= 0x000; op1
= 0x800; break;
1139 case TYPE_F64
: op2
= 0x080; op1
= 0x900; break;
1145 emitForm_21(i
, op2
, op1
);
1149 if (!(code
[0] & 0x1)) {
1153 modNegAbsF32_3b(i
, 1);
1157 if (i
->dType
== TYPE_F32
) {
1158 if (isFloatType(i
->sType
))
1164 if (i
->sType
== TYPE_S32
)
1167 if (i
->op
!= OP_SET
) {
1169 case OP_SET_AND
: code
[1] |= 0x0 << 16; break;
1170 case OP_SET_OR
: code
[1] |= 0x1 << 16; break;
1171 case OP_SET_XOR
: code
[1] |= 0x2 << 16; break;
1176 srcId(i
->src(2), 0x2a);
1178 code
[1] |= 0x7 << 10;
1180 if (i
->flagsSrc
>= 0)
1182 emitCondCode(i
->setCond
,
1183 isFloatType(i
->sType
) ? 0x33 : 0x34,
1184 isFloatType(i
->sType
) ? 0xf : 0x7);
1188 CodeEmitterGK110::emitSLCT(const CmpInstruction
*i
)
1190 CondCode cc
= i
->setCond
;
1191 if (i
->src(2).mod
.neg())
1192 cc
= reverseCondCode(cc
);
1194 if (i
->dType
== TYPE_F32
) {
1195 emitForm_21(i
, 0x1d0, 0xb50);
1197 emitCondCode(cc
, 0x33, 0xf);
1199 emitForm_21(i
, 0x1a0, 0xb20);
1200 emitCondCode(cc
, 0x34, 0x7);
1201 if (i
->dType
== TYPE_S32
)
1207 selpFlip(const FixupEntry
*entry
, uint32_t *code
, const FixupData
& data
)
1209 int loc
= entry
->loc
;
1210 if (data
.force_persample_interp
)
1211 code
[loc
+ 1] |= 1 << 13;
1213 code
[loc
+ 1] &= ~(1 << 13);
1216 void CodeEmitterGK110::emitSELP(const Instruction
*i
)
1218 emitForm_21(i
, 0x250, 0x050);
1220 if (i
->src(2).mod
& Modifier(NV50_IR_MOD_NOT
))
1223 if (i
->subOp
== 1) {
1224 addInterp(0, 0, selpFlip
);
1228 void CodeEmitterGK110::emitTEXBAR(const Instruction
*i
)
1230 code
[0] = 0x0000003e | (i
->subOp
<< 23);
1231 code
[1] = 0x77000000;
1236 void CodeEmitterGK110::emitTEXCSAA(const TexInstruction
*i
)
1238 code
[0] = 0x00000002;
1239 code
[1] = 0x76c00000;
1241 code
[1] |= i
->tex
.r
<< 9;
1242 // code[1] |= i->tex.s << (9 + 8);
1244 if (i
->tex
.liveOnly
)
1245 code
[0] |= 0x80000000;
1247 defId(i
->def(0), 2);
1248 srcId(i
->src(0), 10);
1252 isNextIndependentTex(const TexInstruction
*i
)
1254 if (!i
->next
|| !isTextureOp(i
->next
->op
))
1256 if (i
->getDef(0)->interfers(i
->next
->getSrc(0)))
1258 return !i
->next
->srcExists(1) || !i
->getDef(0)->interfers(i
->next
->getSrc(1));
1262 CodeEmitterGK110::emitTEX(const TexInstruction
*i
)
1264 const bool ind
= i
->tex
.rIndirectSrc
>= 0;
1267 code
[0] = 0x00000002;
1270 code
[1] = 0x7e000000;
1273 code
[1] = 0x7e800000;
1276 code
[1] = 0x78000000;
1279 code
[1] = 0x7dc00000;
1282 code
[1] = 0x7d800000;
1288 code
[0] = 0x00000002;
1289 code
[1] = 0x76000000;
1290 code
[1] |= i
->tex
.r
<< 9;
1293 code
[0] = 0x00000002;
1294 code
[1] = 0x76800000;
1295 code
[1] |= i
->tex
.r
<< 9;
1298 code
[0] = 0x00000002;
1299 code
[1] = 0x70000000;
1300 code
[1] |= i
->tex
.r
<< 13;
1303 code
[0] = 0x00000001;
1304 code
[1] = 0x70000000;
1305 code
[1] |= i
->tex
.r
<< 15;
1308 code
[0] = 0x00000001;
1309 code
[1] = 0x60000000;
1310 code
[1] |= i
->tex
.r
<< 15;
1315 code
[1] |= isNextIndependentTex(i
) ? 0x1 : 0x2; // t : p mode
1317 if (i
->tex
.liveOnly
)
1318 code
[0] |= 0x80000000;
1322 case OP_TXB
: code
[1] |= 0x2000; break;
1323 case OP_TXL
: code
[1] |= 0x3000; break;
1327 case OP_TXLQ
: break;
1329 assert(!"invalid texture op");
1333 if (i
->op
== OP_TXF
) {
1334 if (!i
->tex
.levelZero
)
1337 if (i
->tex
.levelZero
) {
1341 if (i
->op
!= OP_TXD
&& i
->tex
.derivAll
)
1346 code
[1] |= i
->tex
.mask
<< 2;
1348 const int src1
= (i
->predSrc
== 1) ? 2 : 1; // if predSrc == 1, !srcExists(2)
1350 defId(i
->def(0), 2);
1351 srcId(i
->src(0), 10);
1354 if (i
->op
== OP_TXG
) code
[1] |= i
->tex
.gatherComp
<< 13;
1357 code
[1] |= (i
->tex
.target
.isCube() ? 3 : (i
->tex
.target
.getDim() - 1)) << 7;
1358 if (i
->tex
.target
.isArray())
1360 if (i
->tex
.target
.isShadow())
1362 if (i
->tex
.target
== TEX_TARGET_2D_MS
||
1363 i
->tex
.target
== TEX_TARGET_2D_MS_ARRAY
)
1366 if (i
->srcExists(src1
) && i
->src(src1
).getFile() == FILE_IMMEDIATE
) {
1370 if (i
->tex
.useOffsets
== 1) {
1372 case OP_TXF
: code
[1] |= 0x200; break;
1373 case OP_TXD
: code
[1] |= 0x00400000; break;
1374 default: code
[1] |= 0x800; break;
1377 if (i
->tex
.useOffsets
== 4)
1382 CodeEmitterGK110::emitTXQ(const TexInstruction
*i
)
1384 code
[0] = 0x00000002;
1385 code
[1] = 0x75400001;
1387 switch (i
->tex
.query
) {
1388 case TXQ_DIMS
: code
[0] |= 0x01 << 25; break;
1389 case TXQ_TYPE
: code
[0] |= 0x02 << 25; break;
1390 case TXQ_SAMPLE_POSITION
: code
[0] |= 0x05 << 25; break;
1391 case TXQ_FILTER
: code
[0] |= 0x10 << 25; break;
1392 case TXQ_LOD
: code
[0] |= 0x12 << 25; break;
1393 case TXQ_BORDER_COLOUR
: code
[0] |= 0x16 << 25; break;
1395 assert(!"invalid texture query");
1399 code
[1] |= i
->tex
.mask
<< 2;
1400 code
[1] |= i
->tex
.r
<< 9;
1401 if (/*i->tex.sIndirectSrc >= 0 || */i
->tex
.rIndirectSrc
>= 0)
1402 code
[1] |= 0x08000000;
1404 defId(i
->def(0), 2);
1405 srcId(i
->src(0), 10);
1411 CodeEmitterGK110::emitQUADOP(const Instruction
*i
, uint8_t qOp
, uint8_t laneMask
)
1413 code
[0] = 0x00000002 | ((qOp
& 1) << 31);
1414 code
[1] = 0x7fc00200 | (qOp
>> 1) | (laneMask
<< 12); // dall
1416 defId(i
->def(0), 2);
1417 srcId(i
->src(0), 10);
1418 srcId((i
->srcExists(1) && i
->predSrc
!= 1) ? i
->src(1) : i
->src(0), 23);
1424 CodeEmitterGK110::emitPIXLD(const Instruction
*i
)
1426 emitForm_L(i
, 0x7f4, 2, Modifier(0));
1427 code
[1] |= i
->subOp
<< 2;
1428 code
[1] |= 0x00070000;
1432 CodeEmitterGK110::emitBAR(const Instruction
*i
)
1434 code
[0] = 0x00000002;
1435 code
[1] = 0x85400000;
1438 case NV50_IR_SUBOP_BAR_ARRIVE
: code
[1] |= 0x08; break;
1439 case NV50_IR_SUBOP_BAR_RED_AND
: code
[1] |= 0x50; break;
1440 case NV50_IR_SUBOP_BAR_RED_OR
: code
[1] |= 0x90; break;
1441 case NV50_IR_SUBOP_BAR_RED_POPC
: code
[1] |= 0x10; break;
1443 assert(i
->subOp
== NV50_IR_SUBOP_BAR_SYNC
);
1450 if (i
->src(0).getFile() == FILE_GPR
) {
1451 srcId(i
->src(0), 10);
1453 ImmediateValue
*imm
= i
->getSrc(0)->asImm();
1455 code
[0] |= imm
->reg
.data
.u32
<< 10;
1460 if (i
->src(1).getFile() == FILE_GPR
) {
1461 srcId(i
->src(1), 23);
1463 ImmediateValue
*imm
= i
->getSrc(0)->asImm();
1465 assert(imm
->reg
.data
.u32
<= 0xfff);
1466 code
[0] |= imm
->reg
.data
.u32
<< 23;
1467 code
[1] |= imm
->reg
.data
.u32
>> 9;
1471 if (i
->srcExists(2) && (i
->predSrc
!= 2)) {
1472 srcId(i
->src(2), 32 + 10);
1473 if (i
->src(2).mod
== Modifier(NV50_IR_MOD_NOT
))
1480 void CodeEmitterGK110::emitMEMBAR(const Instruction
*i
)
1482 code
[0] = 0x00000002 | NV50_IR_SUBOP_MEMBAR_SCOPE(i
->subOp
) << 8;
1483 code
[1] = 0x7cc00000;
1489 CodeEmitterGK110::emitFlow(const Instruction
*i
)
1491 const FlowInstruction
*f
= i
->asFlow();
1493 unsigned mask
; // bit 0: predicate, bit 1: target
1495 code
[0] = 0x00000000;
1499 code
[1] = f
->absolute
? 0x10800000 : 0x12000000;
1500 if (i
->srcExists(0) && i
->src(0).getFile() == FILE_MEMORY_CONST
)
1505 code
[1] = f
->absolute
? 0x11000000 : 0x13000000;
1506 if (i
->srcExists(0) && i
->src(0).getFile() == FILE_MEMORY_CONST
)
1511 case OP_EXIT
: code
[1] = 0x18000000; mask
= 1; break;
1512 case OP_RET
: code
[1] = 0x19000000; mask
= 1; break;
1513 case OP_DISCARD
: code
[1] = 0x19800000; mask
= 1; break;
1514 case OP_BREAK
: code
[1] = 0x1a000000; mask
= 1; break;
1515 case OP_CONT
: code
[1] = 0x1a800000; mask
= 1; break;
1517 case OP_JOINAT
: code
[1] = 0x14800000; mask
= 2; break;
1518 case OP_PREBREAK
: code
[1] = 0x15000000; mask
= 2; break;
1519 case OP_PRECONT
: code
[1] = 0x15800000; mask
= 2; break;
1520 case OP_PRERET
: code
[1] = 0x13800000; mask
= 2; break;
1522 case OP_QUADON
: code
[1] = 0x1b800000; mask
= 0; break;
1523 case OP_QUADPOP
: code
[1] = 0x1c000000; mask
= 0; break;
1524 case OP_BRKPT
: code
[1] = 0x00000000; mask
= 0; break;
1526 assert(!"invalid flow operation");
1532 if (i
->flagsSrc
< 0)
1544 if (f
->op
== OP_CALL
) {
1546 assert(f
->absolute
);
1547 uint32_t pcAbs
= targNVC0
->getBuiltinOffset(f
->target
.builtin
);
1548 addReloc(RelocEntry::TYPE_BUILTIN
, 0, pcAbs
, 0xff800000, 23);
1549 addReloc(RelocEntry::TYPE_BUILTIN
, 1, pcAbs
, 0x007fffff, -9);
1551 assert(!f
->absolute
);
1552 int32_t pcRel
= f
->target
.fn
->binPos
- (codeSize
+ 8);
1553 code
[0] |= (pcRel
& 0x1ff) << 23;
1554 code
[1] |= (pcRel
>> 9) & 0x7fff;
1558 int32_t pcRel
= f
->target
.bb
->binPos
- (codeSize
+ 8);
1559 if (writeIssueDelays
&& !(f
->target
.bb
->binPos
& 0x3f))
1561 // currently we don't want absolute branches
1562 assert(!f
->absolute
);
1563 code
[0] |= (pcRel
& 0x1ff) << 23;
1564 code
[1] |= (pcRel
>> 9) & 0x7fff;
1569 CodeEmitterGK110::emitVOTE(const Instruction
*i
)
1571 assert(i
->src(0).getFile() == FILE_PREDICATE
);
1573 code
[0] = 0x00000002;
1574 code
[1] = 0x86c00000 | (i
->subOp
<< 19);
1579 for (int d
= 0; i
->defExists(d
); d
++) {
1580 if (i
->def(d
).getFile() == FILE_PREDICATE
) {
1583 defId(i
->def(d
), 48);
1584 } else if (i
->def(d
).getFile() == FILE_GPR
) {
1587 defId(i
->def(d
), 2);
1589 assert(!"Unhandled def");
1593 code
[0] |= 255 << 2;
1596 if (i
->src(0).mod
== Modifier(NV50_IR_MOD_NOT
))
1598 srcId(i
->src(0), 42);
1602 CodeEmitterGK110::emitSUGType(DataType ty
, const int pos
)
1607 case TYPE_S32
: n
= 1; break;
1608 case TYPE_U8
: n
= 2; break;
1609 case TYPE_S8
: n
= 3; break;
1611 assert(ty
== TYPE_U32
);
1614 code
[pos
/ 32] |= n
<< (pos
% 32);
1618 CodeEmitterGK110::emitSUCachingMode(CacheMode c
)
1638 assert(!"invalid caching mode");
1641 code
[0] |= (n
& 1) << 31;
1642 code
[1] |= (n
& 2) >> 1;
1646 CodeEmitterGK110::setSUConst16(const Instruction
*i
, const int s
)
1648 const uint32_t offset
= i
->getSrc(s
)->reg
.data
.offset
;
1650 assert(offset
== (offset
& 0xfffc));
1652 code
[0] |= offset
<< 21;
1653 code
[1] |= offset
>> 11;
1654 code
[1] |= i
->getSrc(s
)->reg
.fileIndex
<< 5;
1658 CodeEmitterGK110::emitSULDGB(const TexInstruction
*i
)
1660 code
[0] = 0x00000002;
1661 code
[1] = 0x30000000 | (i
->subOp
<< 14);
1663 if (i
->src(1).getFile() == FILE_MEMORY_CONST
) {
1664 emitLoadStoreType(i
->dType
, 0x38);
1665 emitCachingMode(i
->cache
, 0x36);
1670 assert(i
->src(1).getFile() == FILE_GPR
);
1671 code
[1] |= 0x49800000;
1673 emitLoadStoreType(i
->dType
, 0x21);
1674 emitSUCachingMode(i
->cache
);
1676 srcId(i
->src(1), 23);
1679 emitSUGType(i
->sType
, 0x34);
1682 defId(i
->def(0), 2); // destination
1683 srcId(i
->src(0), 10); // address
1685 // surface predicate
1686 if (!i
->srcExists(2) || (i
->predSrc
== 2)) {
1687 code
[1] |= 0x7 << 10;
1689 if (i
->src(2).mod
== Modifier(NV50_IR_MOD_NOT
))
1691 srcId(i
->src(2), 32 + 10);
1696 CodeEmitterGK110::emitSUSTGx(const TexInstruction
*i
)
1698 assert(i
->op
== OP_SUSTP
);
1700 code
[0] = 0x00000002;
1701 code
[1] = 0x38000000;
1703 if (i
->src(1).getFile() == FILE_MEMORY_CONST
) {
1704 code
[0] |= i
->subOp
<< 2;
1706 if (i
->op
== OP_SUSTP
)
1707 code
[0] |= i
->tex
.mask
<< 4;
1709 emitSUGType(i
->sType
, 0x8);
1710 emitCachingMode(i
->cache
, 0x36);
1715 assert(i
->src(1).getFile() == FILE_GPR
);
1717 code
[0] |= i
->subOp
<< 23;
1718 code
[1] |= 0x41c00000;
1720 if (i
->op
== OP_SUSTP
)
1721 code
[0] |= i
->tex
.mask
<< 25;
1723 emitSUGType(i
->sType
, 0x1d);
1724 emitSUCachingMode(i
->cache
);
1726 srcId(i
->src(1), 2);
1730 srcId(i
->src(0), 10); // address
1731 srcId(i
->src(3), 42); // values
1733 // surface predicate
1734 if (!i
->srcExists(2) || (i
->predSrc
== 2)) {
1735 code
[1] |= 0x7 << 18;
1737 if (i
->src(2).mod
== Modifier(NV50_IR_MOD_NOT
))
1739 srcId(i
->src(2), 32 + 18);
1744 CodeEmitterGK110::emitSUCLAMPMode(uint16_t subOp
)
1747 switch (subOp
& ~NV50_IR_SUBOP_SUCLAMP_2D
) {
1748 case NV50_IR_SUBOP_SUCLAMP_SD(0, 1): m
= 0; break;
1749 case NV50_IR_SUBOP_SUCLAMP_SD(1, 1): m
= 1; break;
1750 case NV50_IR_SUBOP_SUCLAMP_SD(2, 1): m
= 2; break;
1751 case NV50_IR_SUBOP_SUCLAMP_SD(3, 1): m
= 3; break;
1752 case NV50_IR_SUBOP_SUCLAMP_SD(4, 1): m
= 4; break;
1753 case NV50_IR_SUBOP_SUCLAMP_PL(0, 1): m
= 5; break;
1754 case NV50_IR_SUBOP_SUCLAMP_PL(1, 1): m
= 6; break;
1755 case NV50_IR_SUBOP_SUCLAMP_PL(2, 1): m
= 7; break;
1756 case NV50_IR_SUBOP_SUCLAMP_PL(3, 1): m
= 8; break;
1757 case NV50_IR_SUBOP_SUCLAMP_PL(4, 1): m
= 9; break;
1758 case NV50_IR_SUBOP_SUCLAMP_BL(0, 1): m
= 10; break;
1759 case NV50_IR_SUBOP_SUCLAMP_BL(1, 1): m
= 11; break;
1760 case NV50_IR_SUBOP_SUCLAMP_BL(2, 1): m
= 12; break;
1761 case NV50_IR_SUBOP_SUCLAMP_BL(3, 1): m
= 13; break;
1762 case NV50_IR_SUBOP_SUCLAMP_BL(4, 1): m
= 14; break;
1767 if (subOp
& NV50_IR_SUBOP_SUCLAMP_2D
)
1772 CodeEmitterGK110::emitSUCalc(Instruction
*i
)
1774 ImmediateValue
*imm
= NULL
;
1775 uint64_t opc1
, opc2
;
1777 if (i
->srcExists(2)) {
1778 imm
= i
->getSrc(2)->asImm();
1780 i
->setSrc(2, NULL
); // special case, make emitForm_21 not assert
1784 case OP_SUCLAMP
: opc1
= 0xb00; opc2
= 0x580; break;
1785 case OP_SUBFM
: opc1
= 0xb68; opc2
= 0x1e8; break;
1786 case OP_SUEAU
: opc1
= 0xb6c; opc2
= 0x1ec; break;
1791 emitForm_21(i
, opc2
, opc1
);
1793 if (i
->op
== OP_SUCLAMP
) {
1794 if (i
->dType
== TYPE_S32
)
1796 emitSUCLAMPMode(i
->subOp
);
1799 if (i
->op
== OP_SUBFM
&& i
->subOp
== NV50_IR_SUBOP_SUBFM_3D
)
1802 if (i
->op
!= OP_SUEAU
) {
1803 const uint8_t pos
= i
->op
== OP_SUBFM
? 19 : 16;
1804 if (i
->def(0).getFile() == FILE_PREDICATE
) { // p, #
1805 code
[0] |= 255 << 2;
1806 code
[1] |= i
->getDef(1)->reg
.data
.id
<< pos
;
1808 if (i
->defExists(1)) { // r, p
1809 assert(i
->def(1).getFile() == FILE_PREDICATE
);
1810 code
[1] |= i
->getDef(1)->reg
.data
.id
<< pos
;
1812 code
[1] |= 7 << pos
;
1817 assert(i
->op
== OP_SUCLAMP
);
1819 code
[1] |= (imm
->reg
.data
.u32
& 0x3f) << 10; // sint6
1825 CodeEmitterGK110::emitVectorSubOp(const Instruction
*i
)
1827 switch (NV50_IR_SUBOP_Vn(i
->subOp
)) {
1829 code
[1] |= (i
->subOp
& 0x000f) << 7; // vsrc1
1830 code
[1] |= (i
->subOp
& 0x00e0) >> 6; // vsrc2
1831 code
[1] |= (i
->subOp
& 0x0100) << 13; // vsrc2
1832 code
[1] |= (i
->subOp
& 0x3c00) << 12; // vdst
1841 CodeEmitterGK110::emitVSHL(const Instruction
*i
)
1843 code
[0] = 0x00000002;
1844 code
[1] = 0xb8000000;
1846 assert(NV50_IR_SUBOP_Vn(i
->subOp
) == 0);
1848 if (isSignedType(i
->dType
)) code
[1] |= 1 << 25;
1849 if (isSignedType(i
->sType
)) code
[1] |= 1 << 19;
1854 defId(i
->def(0), 2);
1855 srcId(i
->src(0), 10);
1857 if (i
->getSrc(1)->reg
.file
== FILE_IMMEDIATE
) {
1858 ImmediateValue
*imm
= i
->getSrc(1)->asImm();
1860 code
[0] |= (imm
->reg
.data
.u32
& 0x01ff) << 23;
1861 code
[1] |= (imm
->reg
.data
.u32
& 0xfe00) >> 9;
1863 assert(i
->getSrc(1)->reg
.file
== FILE_GPR
);
1865 srcId(i
->src(1), 23);
1867 srcId(i
->src(2), 42);
1871 if (i
->flagsDef
>= 0)
1876 CodeEmitterGK110::emitAFETCH(const Instruction
*i
)
1878 uint32_t offset
= i
->src(0).get()->reg
.data
.offset
& 0x7ff;
1880 code
[0] = 0x00000002 | (offset
<< 23);
1881 code
[1] = 0x7d000000 | (offset
>> 9);
1883 if (i
->getSrc(0)->reg
.file
== FILE_SHADER_OUTPUT
)
1888 defId(i
->def(0), 2);
1889 srcId(i
->src(0).getIndirect(0), 10);
1893 CodeEmitterGK110::emitPFETCH(const Instruction
*i
)
1895 uint32_t prim
= i
->src(0).get()->reg
.data
.u32
;
1897 code
[0] = 0x00000002 | ((prim
& 0xff) << 23);
1898 code
[1] = 0x7f800000;
1902 const int src1
= (i
->predSrc
== 1) ? 2 : 1; // if predSrc == 1, !srcExists(2)
1904 defId(i
->def(0), 2);
1909 CodeEmitterGK110::emitVFETCH(const Instruction
*i
)
1911 unsigned int size
= typeSizeof(i
->dType
);
1912 uint32_t offset
= i
->src(0).get()->reg
.data
.offset
;
1914 code
[0] = 0x00000002 | (offset
<< 23);
1915 code
[1] = 0x7ec00000 | (offset
>> 9);
1916 code
[1] |= (size
/ 4 - 1) << 18;
1920 if (i
->getSrc(0)->reg
.file
== FILE_SHADER_OUTPUT
)
1921 code
[1] |= 0x8; // yes, TCPs can read from *outputs* of other threads
1925 defId(i
->def(0), 2);
1926 srcId(i
->src(0).getIndirect(0), 10);
1927 srcId(i
->src(0).getIndirect(1), 32 + 10); // vertex address
1931 CodeEmitterGK110::emitEXPORT(const Instruction
*i
)
1933 unsigned int size
= typeSizeof(i
->dType
);
1934 uint32_t offset
= i
->src(0).get()->reg
.data
.offset
;
1936 code
[0] = 0x00000002 | (offset
<< 23);
1937 code
[1] = 0x7f000000 | (offset
>> 9);
1938 code
[1] |= (size
/ 4 - 1) << 18;
1945 assert(i
->src(1).getFile() == FILE_GPR
);
1947 srcId(i
->src(0).getIndirect(0), 10);
1948 srcId(i
->src(0).getIndirect(1), 32 + 10); // vertex base address
1949 srcId(i
->src(1), 2);
1953 CodeEmitterGK110::emitOUT(const Instruction
*i
)
1955 assert(i
->src(0).getFile() == FILE_GPR
);
1957 emitForm_21(i
, 0x1f0, 0xb70);
1959 if (i
->op
== OP_EMIT
)
1961 if (i
->op
== OP_RESTART
|| i
->subOp
== NV50_IR_SUBOP_EMIT_RESTART
)
1966 CodeEmitterGK110::emitInterpMode(const Instruction
*i
)
1968 code
[1] |= (i
->ipa
& 0x3) << 21; // TODO: INTERP_SAMPLEID
1969 code
[1] |= (i
->ipa
& 0xc) << (19 - 2);
1973 interpApply(const FixupEntry
*entry
, uint32_t *code
, const FixupData
& data
)
1975 int ipa
= entry
->ipa
;
1976 int reg
= entry
->reg
;
1977 int loc
= entry
->loc
;
1979 if (data
.flatshade
&&
1980 (ipa
& NV50_IR_INTERP_MODE_MASK
) == NV50_IR_INTERP_SC
) {
1981 ipa
= NV50_IR_INTERP_FLAT
;
1983 } else if (data
.force_persample_interp
&&
1984 (ipa
& NV50_IR_INTERP_SAMPLE_MASK
) == NV50_IR_INTERP_DEFAULT
&&
1985 (ipa
& NV50_IR_INTERP_MODE_MASK
) != NV50_IR_INTERP_FLAT
) {
1986 ipa
|= NV50_IR_INTERP_CENTROID
;
1988 code
[loc
+ 1] &= ~(0xf << 19);
1989 code
[loc
+ 1] |= (ipa
& 0x3) << 21;
1990 code
[loc
+ 1] |= (ipa
& 0xc) << (19 - 2);
1991 code
[loc
+ 0] &= ~(0xff << 23);
1992 code
[loc
+ 0] |= reg
<< 23;
1996 CodeEmitterGK110::emitINTERP(const Instruction
*i
)
1998 const uint32_t base
= i
->getSrc(0)->reg
.data
.offset
;
2000 code
[0] = 0x00000002 | (base
<< 31);
2001 code
[1] = 0x74800000 | (base
>> 1);
2006 if (i
->op
== OP_PINTERP
) {
2007 srcId(i
->src(1), 23);
2008 addInterp(i
->ipa
, SDATA(i
->src(1)).id
, interpApply
);
2010 code
[0] |= 0xff << 23;
2011 addInterp(i
->ipa
, 0xff, interpApply
);
2014 srcId(i
->src(0).getIndirect(0), 10);
2018 defId(i
->def(0), 2);
2020 if (i
->getSampleMode() == NV50_IR_INTERP_OFFSET
)
2021 srcId(i
->src(i
->op
== OP_PINTERP
? 2 : 1), 32 + 10);
2023 code
[1] |= 0xff << 10;
2027 CodeEmitterGK110::emitLoadStoreType(DataType ty
, const int pos
)
2059 assert(!"invalid ld/st type");
2062 code
[pos
/ 32] |= n
<< (pos
% 32);
2066 CodeEmitterGK110::emitCachingMode(CacheMode c
, const int pos
)
2087 assert(!"invalid caching mode");
2090 code
[pos
/ 32] |= n
<< (pos
% 32);
2094 CodeEmitterGK110::emitSTORE(const Instruction
*i
)
2096 int32_t offset
= SDATA(i
->src(0)).offset
;
2098 switch (i
->src(0).getFile()) {
2099 case FILE_MEMORY_GLOBAL
: code
[1] = 0xe0000000; code
[0] = 0x00000000; break;
2100 case FILE_MEMORY_LOCAL
: code
[1] = 0x7a800000; code
[0] = 0x00000002; break;
2101 case FILE_MEMORY_SHARED
:
2102 code
[0] = 0x00000002;
2103 if (i
->subOp
== NV50_IR_SUBOP_STORE_UNLOCKED
)
2104 code
[1] = 0x78400000;
2106 code
[1] = 0x7ac00000;
2109 assert(!"invalid memory file");
2113 if (code
[0] & 0x2) {
2115 emitLoadStoreType(i
->dType
, 0x33);
2116 if (i
->src(0).getFile() == FILE_MEMORY_LOCAL
)
2117 emitCachingMode(i
->cache
, 0x2f);
2119 emitLoadStoreType(i
->dType
, 0x38);
2120 emitCachingMode(i
->cache
, 0x3b);
2122 code
[0] |= offset
<< 23;
2123 code
[1] |= offset
>> 9;
2125 // Unlocked store on shared memory can fail.
2126 if (i
->src(0).getFile() == FILE_MEMORY_SHARED
&&
2127 i
->subOp
== NV50_IR_SUBOP_STORE_UNLOCKED
) {
2128 assert(i
->defExists(0));
2129 defId(i
->def(0), 32 + 16);
2134 srcId(i
->src(1), 2);
2135 srcId(i
->src(0).getIndirect(0), 10);
2136 if (i
->src(0).getFile() == FILE_MEMORY_GLOBAL
&&
2137 i
->src(0).isIndirect(0) &&
2138 i
->getIndirect(0, 0)->reg
.size
== 8)
2143 CodeEmitterGK110::emitLOAD(const Instruction
*i
)
2145 int32_t offset
= SDATA(i
->src(0)).offset
;
2147 switch (i
->src(0).getFile()) {
2148 case FILE_MEMORY_GLOBAL
: code
[1] = 0xc0000000; code
[0] = 0x00000000; break;
2149 case FILE_MEMORY_LOCAL
: code
[1] = 0x7a000000; code
[0] = 0x00000002; break;
2150 case FILE_MEMORY_SHARED
:
2151 code
[0] = 0x00000002;
2152 if (i
->subOp
== NV50_IR_SUBOP_LOAD_LOCKED
)
2153 code
[1] = 0x77400000;
2155 code
[1] = 0x7a400000;
2157 case FILE_MEMORY_CONST
:
2158 if (!i
->src(0).isIndirect(0) && typeSizeof(i
->dType
) == 4) {
2163 code
[0] = 0x00000002;
2164 code
[1] = 0x7c800000 | (i
->src(0).get()->reg
.fileIndex
<< 7);
2165 code
[1] |= i
->subOp
<< 15;
2168 assert(!"invalid memory file");
2172 if (code
[0] & 0x2) {
2174 emitLoadStoreType(i
->dType
, 0x33);
2175 if (i
->src(0).getFile() == FILE_MEMORY_LOCAL
)
2176 emitCachingMode(i
->cache
, 0x2f);
2178 emitLoadStoreType(i
->dType
, 0x38);
2179 emitCachingMode(i
->cache
, 0x3b);
2181 code
[0] |= offset
<< 23;
2182 code
[1] |= offset
>> 9;
2184 // Locked store on shared memory can fail.
2186 if (i
->src(0).getFile() == FILE_MEMORY_SHARED
&&
2187 i
->subOp
== NV50_IR_SUBOP_LOAD_LOCKED
) {
2188 if (i
->def(0).getFile() == FILE_PREDICATE
) { // p, #
2191 } else if (i
->defExists(1)) { // r, p
2194 assert(!"Expected predicate dest for load locked");
2201 defId(i
->def(r
), 2);
2203 code
[0] |= 255 << 2;
2206 defId(i
->def(p
), 32 + 16);
2208 if (i
->getIndirect(0, 0)) {
2209 srcId(i
->src(0).getIndirect(0), 10);
2210 if (i
->getIndirect(0, 0)->reg
.size
== 8)
2213 code
[0] |= 255 << 10;
2218 CodeEmitterGK110::getSRegEncoding(const ValueRef
& ref
)
2220 switch (SDATA(ref
).sv
.sv
) {
2221 case SV_LANEID
: return 0x00;
2222 case SV_PHYSID
: return 0x03;
2223 case SV_VERTEX_COUNT
: return 0x10;
2224 case SV_INVOCATION_ID
: return 0x11;
2225 case SV_YDIR
: return 0x12;
2226 case SV_THREAD_KILL
: return 0x13;
2227 case SV_TID
: return 0x21 + SDATA(ref
).sv
.index
;
2228 case SV_CTAID
: return 0x25 + SDATA(ref
).sv
.index
;
2229 case SV_NTID
: return 0x29 + SDATA(ref
).sv
.index
;
2230 case SV_GRIDID
: return 0x2c;
2231 case SV_NCTAID
: return 0x2d + SDATA(ref
).sv
.index
;
2232 case SV_LBASE
: return 0x34;
2233 case SV_SBASE
: return 0x30;
2234 case SV_CLOCK
: return 0x50 + SDATA(ref
).sv
.index
;
2236 assert(!"no sreg for system value");
2242 CodeEmitterGK110::emitMOV(const Instruction
*i
)
2244 if (i
->def(0).getFile() == FILE_PREDICATE
) {
2245 if (i
->src(0).getFile() == FILE_GPR
) {
2246 // Use ISETP.NE.AND dst, PT, src, RZ, PT
2247 code
[0] = 0x00000002;
2248 code
[1] = 0xdb500000;
2250 code
[0] |= 0x7 << 2;
2251 code
[0] |= 0xff << 23;
2252 code
[1] |= 0x7 << 10;
2253 srcId(i
->src(0), 10);
2255 if (i
->src(0).getFile() == FILE_PREDICATE
) {
2256 // Use PSETP.AND.AND dst, PT, src, PT, PT
2257 code
[0] = 0x00000002;
2258 code
[1] = 0x84800000;
2260 code
[0] |= 0x7 << 2;
2261 code
[1] |= 0x7 << 0;
2262 code
[1] |= 0x7 << 10;
2264 srcId(i
->src(0), 14);
2266 assert(!"Unexpected source for predicate destination");
2270 defId(i
->def(0), 5);
2272 if (i
->src(0).getFile() == FILE_SYSTEM_VALUE
) {
2273 code
[0] = 0x00000002 | (getSRegEncoding(i
->src(0)) << 23);
2274 code
[1] = 0x86400000;
2276 defId(i
->def(0), 2);
2278 if (i
->src(0).getFile() == FILE_IMMEDIATE
) {
2279 code
[0] = 0x00000002 | (i
->lanes
<< 14);
2280 code
[1] = 0x74000000;
2282 defId(i
->def(0), 2);
2283 setImmediate32(i
, 0, Modifier(0));
2285 if (i
->src(0).getFile() == FILE_PREDICATE
) {
2286 code
[0] = 0x00000002;
2287 code
[1] = 0x84401c07;
2289 defId(i
->def(0), 2);
2290 srcId(i
->src(0), 14);
2292 emitForm_C(i
, 0x24c, 2);
2293 code
[1] |= i
->lanes
<< 10;
2298 uses64bitAddress(const Instruction
*ldst
)
2300 return ldst
->src(0).getFile() == FILE_MEMORY_GLOBAL
&&
2301 ldst
->src(0).isIndirect(0) &&
2302 ldst
->getIndirect(0, 0)->reg
.size
== 8;
2306 CodeEmitterGK110::emitATOM(const Instruction
*i
)
2308 const bool hasDst
= i
->defExists(0);
2309 const bool exch
= i
->subOp
== NV50_IR_SUBOP_ATOM_EXCH
;
2311 code
[0] = 0x00000002;
2312 if (i
->subOp
== NV50_IR_SUBOP_ATOM_CAS
)
2313 code
[1] = 0x77800000;
2315 code
[1] = 0x68000000;
2318 case NV50_IR_SUBOP_ATOM_CAS
: break;
2319 case NV50_IR_SUBOP_ATOM_EXCH
: code
[1] |= 0x04000000; break;
2320 default: code
[1] |= i
->subOp
<< 23; break;
2324 case TYPE_U32
: break;
2325 case TYPE_S32
: code
[1] |= 0x00100000; break;
2326 case TYPE_U64
: code
[1] |= 0x00200000; break;
2327 case TYPE_F32
: code
[1] |= 0x00300000; break;
2328 case TYPE_B128
: code
[1] |= 0x00400000; break; /* TODO: U128 */
2329 case TYPE_S64
: code
[1] |= 0x00500000; break;
2330 default: assert(!"unsupported type"); break;
2335 /* TODO: cas: check that src regs line up */
2336 /* TODO: cas: flip bits if $r255 is used */
2337 srcId(i
->src(1), 23);
2340 defId(i
->def(0), 2);
2343 code
[0] |= 255 << 2;
2346 if (hasDst
|| !exch
) {
2347 const int32_t offset
= SDATA(i
->src(0)).offset
;
2348 assert(offset
< 0x80000 && offset
>= -0x80000);
2349 code
[0] |= (offset
& 1) << 31;
2350 code
[1] |= (offset
& 0xffffe) >> 1;
2352 srcAddr32(i
->src(0), 31);
2355 if (i
->getIndirect(0, 0)) {
2356 srcId(i
->getIndirect(0, 0), 10);
2357 if (i
->getIndirect(0, 0)->reg
.size
== 8)
2360 code
[0] |= 255 << 10;
2365 CodeEmitterGK110::emitCCTL(const Instruction
*i
)
2367 int32_t offset
= SDATA(i
->src(0)).offset
;
2369 code
[0] = 0x00000002 | (i
->subOp
<< 2);
2371 if (i
->src(0).getFile() == FILE_MEMORY_GLOBAL
) {
2372 code
[1] = 0x7b000000;
2374 code
[1] = 0x7c000000;
2377 code
[0] |= offset
<< 23;
2378 code
[1] |= offset
>> 9;
2380 if (uses64bitAddress(i
))
2382 srcId(i
->src(0).getIndirect(0), 10);
2388 CodeEmitterGK110::emitInstruction(Instruction
*insn
)
2390 const unsigned int size
= (writeIssueDelays
&& !(codeSize
& 0x3f)) ? 16 : 8;
2392 if (insn
->encSize
!= 8) {
2393 ERROR("skipping unencodable instruction: ");
2397 if (codeSize
+ size
> codeSizeLimit
) {
2398 ERROR("code emitter output buffer too small\n");
2402 if (writeIssueDelays
) {
2403 int id
= (codeSize
& 0x3f) / 8 - 1;
2406 code
[0] = 0x00000000; // cf issue delay "instruction"
2407 code
[1] = 0x08000000;
2411 uint32_t *data
= code
- (id
* 2 + 2);
2414 case 0: data
[0] |= insn
->sched
<< 2; break;
2415 case 1: data
[0] |= insn
->sched
<< 10; break;
2416 case 2: data
[0] |= insn
->sched
<< 18; break;
2417 case 3: data
[0] |= insn
->sched
<< 26; data
[1] |= insn
->sched
>> 6; break;
2418 case 4: data
[1] |= insn
->sched
<< 2; break;
2419 case 5: data
[1] |= insn
->sched
<< 10; break;
2420 case 6: data
[1] |= insn
->sched
<< 18; break;
2427 // assert that instructions with multiple defs don't corrupt registers
2428 for (int d
= 0; insn
->defExists(d
); ++d
)
2429 assert(insn
->asTex() || insn
->def(d
).rep()->reg
.data
.id
>= 0);
2466 if (insn
->dType
== TYPE_F64
)
2468 else if (isFloatType(insn
->dType
))
2474 if (insn
->dType
== TYPE_F64
)
2476 else if (isFloatType(insn
->dType
))
2483 if (insn
->dType
== TYPE_F64
)
2485 else if (isFloatType(insn
->dType
))
2503 emitLogicOp(insn
, 0);
2506 emitLogicOp(insn
, 1);
2509 emitLogicOp(insn
, 2);
2513 if (typeSizeof(insn
->sType
) == 8)
2522 emitSET(insn
->asCmp());
2528 emitSLCT(insn
->asCmp());
2543 if (insn
->def(0).getFile() == FILE_PREDICATE
||
2544 insn
->src(0).getFile() == FILE_PREDICATE
)
2550 emitSFnOp(insn
, 5 + 2 * insn
->subOp
);
2553 emitSFnOp(insn
, 4 + 2 * insn
->subOp
);
2578 emitTEX(insn
->asTex());
2581 emitTXQ(insn
->asTex());
2606 emitQUADOP(insn
, insn
->subOp
, insn
->lanes
);
2609 emitQUADOP(insn
, insn
->src(0).mod
.neg() ? 0x66 : 0x99, 0x4);
2612 emitQUADOP(insn
, insn
->src(0).mod
.neg() ? 0x5a : 0xa5, 0x5);
2649 emitSULDGB(insn
->asTex());
2653 emitSUSTGx(insn
->asTex());
2666 ERROR("operation should have been eliminated");
2672 ERROR("operation should have been lowered\n");
2675 ERROR("unknown op: %u\n", insn
->op
);
2688 CodeEmitterGK110::getMinEncodingSize(const Instruction
*i
) const
2690 // No more short instruction encodings.
2695 CodeEmitterGK110::prepareEmission(Function
*func
)
2697 const Target
*targ
= func
->getProgram()->getTarget();
2699 CodeEmitter::prepareEmission(func
);
2701 if (targ
->hasSWSched
)
2702 calculateSchedDataNVC0(targ
, func
);
2705 CodeEmitterGK110::CodeEmitterGK110(const TargetNVC0
*target
)
2706 : CodeEmitter(target
),
2708 writeIssueDelays(target
->hasSWSched
)
2711 codeSize
= codeSizeLimit
= 0;
2716 TargetNVC0::createCodeEmitterGK110(Program::Type type
)
2718 CodeEmitterGK110
*emit
= new CodeEmitterGK110(this);
2719 emit
->setProgramType(type
);
2723 } // namespace nv50_ir