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
);
58 void modNegAbsF32_3b(const Instruction
*, const int s
);
60 void emitCondCode(CondCode cc
, int pos
, uint8_t mask
);
61 void emitInterpMode(const Instruction
*);
62 void emitLoadStoreType(DataType ty
, const int pos
);
63 void emitCachingMode(CacheMode c
, const int pos
);
65 inline uint8_t getSRegEncoding(const ValueRef
&);
67 void emitRoundMode(RoundMode
, const int pos
, const int rintPos
);
68 void emitRoundModeF(RoundMode
, const int pos
);
69 void emitRoundModeI(RoundMode
, const int pos
);
71 void emitNegAbs12(const Instruction
*);
73 void emitNOP(const Instruction
*);
75 void emitLOAD(const Instruction
*);
76 void emitSTORE(const Instruction
*);
77 void emitMOV(const Instruction
*);
79 void emitINTERP(const Instruction
*);
80 void emitPFETCH(const Instruction
*);
81 void emitVFETCH(const Instruction
*);
82 void emitEXPORT(const Instruction
*);
83 void emitOUT(const Instruction
*);
85 void emitUADD(const Instruction
*);
86 void emitFADD(const Instruction
*);
87 void emitIMUL(const Instruction
*);
88 void emitFMUL(const Instruction
*);
89 void emitIMAD(const Instruction
*);
90 void emitISAD(const Instruction
*);
91 void emitFMAD(const Instruction
*);
93 void emitNOT(const Instruction
*);
94 void emitLogicOp(const Instruction
*, uint8_t subOp
);
95 void emitPOPC(const Instruction
*);
96 void emitINSBF(const Instruction
*);
97 void emitEXTBF(const Instruction
*);
98 void emitBFIND(const Instruction
*);
99 void emitShift(const Instruction
*);
101 void emitSFnOp(const Instruction
*, uint8_t subOp
);
103 void emitCVT(const Instruction
*);
104 void emitMINMAX(const Instruction
*);
105 void emitPreOp(const Instruction
*);
107 void emitSET(const CmpInstruction
*);
108 void emitSLCT(const CmpInstruction
*);
109 void emitSELP(const Instruction
*);
111 void emitTEXBAR(const Instruction
*);
112 void emitTEX(const TexInstruction
*);
113 void emitTEXCSAA(const TexInstruction
*);
114 void emitTXQ(const TexInstruction
*);
116 void emitQUADOP(const Instruction
*, uint8_t qOp
, uint8_t laneMask
);
118 void emitPIXLD(const Instruction
*);
120 void emitFlow(const Instruction
*);
122 inline void defId(const ValueDef
&, const int pos
);
123 inline void srcId(const ValueRef
&, const int pos
);
124 inline void srcId(const ValueRef
*, const int pos
);
125 inline void srcId(const Instruction
*, int s
, const int pos
);
127 inline void srcAddr32(const ValueRef
&, const int pos
); // address / 4
129 inline bool isLIMM(const ValueRef
&, DataType ty
, bool mod
= false);
132 #define GK110_GPR_ZERO 255
135 if (i->src(s).mod.neg()) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
137 if (i->src(s).mod.abs()) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
139 #define NOT_(b, s) if (i->src(s).mod & Modifier(NV50_IR_MOD_NOT)) \
140 code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
142 #define FTZ_(b) if (i->ftz) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
143 #define DNZ_(b) if (i->dnz) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
145 #define SAT_(b) if (i->saturate) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
147 #define RND_(b, t) emitRoundMode##t(i->rnd, 0x##b)
149 #define SDATA(a) ((a).rep()->reg.data)
150 #define DDATA(a) ((a).rep()->reg.data)
152 void CodeEmitterGK110::srcId(const ValueRef
& src
, const int pos
)
154 code
[pos
/ 32] |= (src
.get() ? SDATA(src
).id
: GK110_GPR_ZERO
) << (pos
% 32);
157 void CodeEmitterGK110::srcId(const ValueRef
*src
, const int pos
)
159 code
[pos
/ 32] |= (src
? SDATA(*src
).id
: GK110_GPR_ZERO
) << (pos
% 32);
162 void CodeEmitterGK110::srcId(const Instruction
*insn
, int s
, int pos
)
164 int r
= insn
->srcExists(s
) ? SDATA(insn
->src(s
)).id
: GK110_GPR_ZERO
;
165 code
[pos
/ 32] |= r
<< (pos
% 32);
168 void CodeEmitterGK110::srcAddr32(const ValueRef
& src
, const int pos
)
170 code
[pos
/ 32] |= (SDATA(src
).offset
>> 2) << (pos
% 32);
173 void CodeEmitterGK110::defId(const ValueDef
& def
, const int pos
)
175 code
[pos
/ 32] |= (def
.get() ? DDATA(def
).id
: GK110_GPR_ZERO
) << (pos
% 32);
178 bool CodeEmitterGK110::isLIMM(const ValueRef
& ref
, DataType ty
, bool mod
)
180 const ImmediateValue
*imm
= ref
.get()->asImm();
182 return imm
&& (imm
->reg
.data
.u32
& ((ty
== TYPE_F32
) ? 0xfff : 0xfff00000));
186 CodeEmitterGK110::emitRoundMode(RoundMode rnd
, const int pos
, const int rintPos
)
192 case ROUND_MI
: rint
= true; /* fall through */ case ROUND_M
: n
= 1; break;
193 case ROUND_PI
: rint
= true; /* fall through */ case ROUND_P
: n
= 2; break;
194 case ROUND_ZI
: rint
= true; /* fall through */ case ROUND_Z
: n
= 3; break;
196 rint
= rnd
== ROUND_NI
;
198 assert(rnd
== ROUND_N
|| rnd
== ROUND_NI
);
201 code
[pos
/ 32] |= n
<< (pos
% 32);
202 if (rint
&& rintPos
>= 0)
203 code
[rintPos
/ 32] |= 1 << (rintPos
% 32);
207 CodeEmitterGK110::emitRoundModeF(RoundMode rnd
, const int pos
)
212 case ROUND_M
: n
= 1; break;
213 case ROUND_P
: n
= 2; break;
214 case ROUND_Z
: n
= 3; break;
217 assert(rnd
== ROUND_N
);
220 code
[pos
/ 32] |= n
<< (pos
% 32);
224 CodeEmitterGK110::emitRoundModeI(RoundMode rnd
, const int pos
)
229 case ROUND_MI
: n
= 1; break;
230 case ROUND_PI
: n
= 2; break;
231 case ROUND_ZI
: n
= 3; break;
234 assert(rnd
== ROUND_NI
);
237 code
[pos
/ 32] |= n
<< (pos
% 32);
240 void CodeEmitterGK110::emitCondCode(CondCode cc
, int pos
, uint8_t mask
)
245 case CC_FL
: n
= 0x00; break;
246 case CC_LT
: n
= 0x01; break;
247 case CC_EQ
: n
= 0x02; break;
248 case CC_LE
: n
= 0x03; break;
249 case CC_GT
: n
= 0x04; break;
250 case CC_NE
: n
= 0x05; break;
251 case CC_GE
: n
= 0x06; break;
252 case CC_LTU
: n
= 0x09; break;
253 case CC_EQU
: n
= 0x0a; break;
254 case CC_LEU
: n
= 0x0b; break;
255 case CC_GTU
: n
= 0x0c; break;
256 case CC_NEU
: n
= 0x0d; break;
257 case CC_GEU
: n
= 0x0e; break;
258 case CC_TR
: n
= 0x0f; break;
259 case CC_NO
: n
= 0x10; break;
260 case CC_NC
: n
= 0x11; break;
261 case CC_NS
: n
= 0x12; break;
262 case CC_NA
: n
= 0x13; break;
263 case CC_A
: n
= 0x14; break;
264 case CC_S
: n
= 0x15; break;
265 case CC_C
: n
= 0x16; break;
266 case CC_O
: n
= 0x17; break;
269 assert(!"invalid condition code");
272 code
[pos
/ 32] |= (n
& mask
) << (pos
% 32);
276 CodeEmitterGK110::emitPredicate(const Instruction
*i
)
278 if (i
->predSrc
>= 0) {
279 srcId(i
->src(i
->predSrc
), 18);
280 if (i
->cc
== CC_NOT_P
)
281 code
[0] |= 8 << 18; // negate
282 assert(i
->getPredicate()->reg
.file
== FILE_PREDICATE
);
289 CodeEmitterGK110::setCAddress14(const ValueRef
& src
)
291 const Storage
& res
= src
.get()->asSym()->reg
;
292 const int32_t addr
= res
.data
.offset
/ 4;
294 code
[0] |= (addr
& 0x01ff) << 23;
295 code
[1] |= (addr
& 0x3e00) >> 9;
296 code
[1] |= res
.fileIndex
<< 5;
300 CodeEmitterGK110::setShortImmediate(const Instruction
*i
, const int s
)
302 const uint32_t u32
= i
->getSrc(s
)->asImm()->reg
.data
.u32
;
303 const uint64_t u64
= i
->getSrc(s
)->asImm()->reg
.data
.u64
;
305 if (i
->sType
== TYPE_F32
) {
306 assert(!(u32
& 0x00000fff));
307 code
[0] |= ((u32
& 0x001ff000) >> 12) << 23;
308 code
[1] |= ((u32
& 0x7fe00000) >> 21);
309 code
[1] |= ((u32
& 0x80000000) >> 4);
311 if (i
->sType
== TYPE_F64
) {
312 assert(!(u64
& 0x00000fffffffffffULL
));
313 code
[0] |= ((u64
& 0x001ff00000000000ULL
) >> 44) << 23;
314 code
[1] |= ((u64
& 0x7fe0000000000000ULL
) >> 53);
315 code
[1] |= ((u64
& 0x8000000000000000ULL
) >> 36);
317 assert((u32
& 0xfff00000) == 0 || (u32
& 0xfff00000) == 0xfff00000);
318 code
[0] |= (u32
& 0x001ff) << 23;
319 code
[1] |= (u32
& 0x7fe00) >> 9;
320 code
[1] |= (u32
& 0x80000) << 8;
325 CodeEmitterGK110::setImmediate32(const Instruction
*i
, const int s
,
328 uint32_t u32
= i
->getSrc(s
)->asImm()->reg
.data
.u32
;
331 ImmediateValue
imm(i
->getSrc(s
)->asImm(), i
->sType
);
333 u32
= imm
.reg
.data
.u32
;
336 code
[0] |= u32
<< 23;
341 CodeEmitterGK110::emitForm_L(const Instruction
*i
, uint32_t opc
, uint8_t ctg
,
351 for (int s
= 0; s
< 3 && i
->srcExists(s
); ++s
) {
352 switch (i
->src(s
).getFile()) {
354 srcId(i
->src(s
), s
? 42 : 10);
357 setImmediate32(i
, s
, mod
);
367 CodeEmitterGK110::emitForm_C(const Instruction
*i
, uint32_t opc
, uint8_t ctg
)
376 switch (i
->src(0).getFile()) {
377 case FILE_MEMORY_CONST
:
378 code
[1] |= 0x4 << 28;
379 setCAddress14(i
->src(0));
382 code
[1] |= 0xc << 28;
383 srcId(i
->src(0), 23);
391 // 0x2 for GPR, c[] and 0x1 for short immediate
393 CodeEmitterGK110::emitForm_21(const Instruction
*i
, uint32_t opc2
,
396 const bool imm
= i
->srcExists(1) && i
->src(1).getFile() == FILE_IMMEDIATE
;
399 if (i
->srcExists(2) && i
->src(2).getFile() == FILE_MEMORY_CONST
)
404 code
[1] = opc1
<< 20;
407 code
[1] = (0xc << 28) | (opc2
<< 20);
414 for (int s
= 0; s
< 3 && i
->srcExists(s
); ++s
) {
415 switch (i
->src(s
).getFile()) {
416 case FILE_MEMORY_CONST
:
417 code
[1] &= (s
== 2) ? ~(0x4 << 28) : ~(0x8 << 28);
418 setCAddress14(i
->src(s
));
421 setShortImmediate(i
, s
);
424 srcId(i
->src(s
), s
? ((s
== 2) ? 42 : s1
) : 10);
427 // ignore here, can be predicate or flags, but must not be address
435 assert(imm
|| (code
[1] & (0xc << 28)));
439 CodeEmitterGK110::modNegAbsF32_3b(const Instruction
*i
, const int s
)
441 if (i
->src(s
).mod
.abs()) code
[1] &= ~(1 << 27);
442 if (i
->src(s
).mod
.neg()) code
[1] ^= (1 << 27);
446 CodeEmitterGK110::emitNOP(const Instruction
*i
)
448 code
[0] = 0x00003c02;
449 code
[1] = 0x85800000;
454 code
[0] = 0x001c3c02;
458 CodeEmitterGK110::emitFMAD(const Instruction
*i
)
460 assert(!isLIMM(i
->src(1), TYPE_F32
));
462 emitForm_21(i
, 0x0c0, 0x940);
470 bool neg1
= (i
->src(0).mod
^ i
->src(1).mod
).neg();
482 CodeEmitterGK110::emitFMUL(const Instruction
*i
)
484 bool neg
= (i
->src(0).mod
^ i
->src(1).mod
).neg();
486 assert(i
->postFactor
>= -3 && i
->postFactor
<= 3);
488 if (isLIMM(i
->src(1), TYPE_F32
)) {
489 emitForm_L(i
, 0x200, 0x2, Modifier(0));
497 assert(i
->postFactor
== 0);
499 emitForm_21(i
, 0x234, 0xc34);
500 code
[1] |= ((i
->postFactor
> 0) ?
501 (7 - i
->postFactor
) : (0 - i
->postFactor
)) << 12;
519 CodeEmitterGK110::emitIMUL(const Instruction
*i
)
521 assert(!i
->src(0).mod
.neg() && !i
->src(1).mod
.neg());
522 assert(!i
->src(0).mod
.abs() && !i
->src(1).mod
.abs());
524 if (isLIMM(i
->src(1), TYPE_S32
)) {
525 emitForm_L(i
, 0x280, 2, Modifier(0));
527 assert(i
->subOp
!= NV50_IR_SUBOP_MUL_HIGH
);
529 if (i
->sType
== TYPE_S32
)
532 emitForm_21(i
, 0x21c, 0xc1c);
534 if (i
->subOp
== NV50_IR_SUBOP_MUL_HIGH
)
536 if (i
->sType
== TYPE_S32
)
542 CodeEmitterGK110::emitFADD(const Instruction
*i
)
544 if (isLIMM(i
->src(1), TYPE_F32
)) {
545 assert(i
->rnd
== ROUND_N
);
546 assert(!i
->saturate
);
548 Modifier mod
= i
->src(1).mod
^
549 Modifier(i
->op
== OP_SUB
? NV50_IR_MOD_NEG
: 0);
551 emitForm_L(i
, 0x400, 0, mod
);
557 emitForm_21(i
, 0x22c, 0xc2c);
566 modNegAbsF32_3b(i
, 1);
567 if (i
->op
== OP_SUB
) code
[1] ^= 1 << 27;
571 if (i
->op
== OP_SUB
) code
[1] ^= 1 << 16;
577 CodeEmitterGK110::emitUADD(const Instruction
*i
)
579 uint8_t addOp
= (i
->src(0).mod
.neg() << 1) | i
->src(1).mod
.neg();
584 assert(!i
->src(0).mod
.abs() && !i
->src(1).mod
.abs());
586 if (isLIMM(i
->src(1), TYPE_S32
)) {
587 emitForm_L(i
, 0x400, 1, Modifier((addOp
& 1) ? NV50_IR_MOD_NEG
: 0));
592 assert(!i
->defExists(1));
593 assert(i
->flagsSrc
< 0);
597 emitForm_21(i
, 0x208, 0xc08);
599 assert(addOp
!= 3); // would be add-plus-one
601 code
[1] |= addOp
<< 19;
604 code
[1] |= 1 << 18; // write carry
605 if (i
->flagsSrc
>= 0)
606 code
[1] |= 1 << 14; // add carry
614 CodeEmitterGK110::emitIMAD(const Instruction
*i
)
617 (i
->src(2).mod
.neg() << 1) | (i
->src(0).mod
.neg() ^ i
->src(1).mod
.neg());
619 emitForm_21(i
, 0x100, 0xa00);
622 code
[1] |= addOp
<< 26;
624 if (i
->sType
== TYPE_S32
)
625 code
[1] |= (1 << 19) | (1 << 24);
631 if (i
->subOp
== NV50_IR_SUBOP_MUL_HIGH
)
638 CodeEmitterGK110::emitISAD(const Instruction
*i
)
640 assert(i
->dType
== TYPE_S32
|| i
->dType
== TYPE_U32
);
642 emitForm_21(i
, 0x1f4, 0xb74);
644 if (i
->dType
== TYPE_S32
)
649 CodeEmitterGK110::emitNOT(const Instruction
*i
)
651 code
[0] = 0x0003fc02; // logop(mov2) dst, 0, not src
652 code
[1] = 0x22003800;
658 switch (i
->src(0).getFile()) {
660 code
[1] |= 0xc << 28;
661 srcId(i
->src(0), 23);
663 case FILE_MEMORY_CONST
:
664 code
[1] |= 0x4 << 28;
665 setCAddress14(i
->src(1));
674 CodeEmitterGK110::emitLogicOp(const Instruction
*i
, uint8_t subOp
)
676 if (isLIMM(i
->src(1), TYPE_S32
)) {
677 emitForm_L(i
, 0x200, 0, i
->src(1).mod
);
678 code
[1] |= subOp
<< 24;
681 emitForm_21(i
, 0x220, 0xc20);
682 code
[1] |= subOp
<< 12;
689 CodeEmitterGK110::emitPOPC(const Instruction
*i
)
691 assert(!isLIMM(i
->src(1), TYPE_S32
, true));
693 emitForm_21(i
, 0x204, 0xc04);
696 if (!(code
[0] & 0x1))
701 CodeEmitterGK110::emitINSBF(const Instruction
*i
)
703 emitForm_21(i
, 0x1f8, 0xb78);
707 CodeEmitterGK110::emitEXTBF(const Instruction
*i
)
709 emitForm_21(i
, 0x600, 0xc00);
711 if (i
->dType
== TYPE_S32
)
713 if (i
->subOp
== NV50_IR_SUBOP_EXTBF_REV
)
718 CodeEmitterGK110::emitBFIND(const Instruction
*i
)
720 emitForm_C(i
, 0x218, 0x2);
722 if (i
->dType
== TYPE_S32
)
724 if (i
->src(0).mod
== Modifier(NV50_IR_MOD_NOT
))
726 if (i
->subOp
== NV50_IR_SUBOP_BFIND_SAMT
)
731 CodeEmitterGK110::emitShift(const Instruction
*i
)
733 if (i
->op
== OP_SHR
) {
734 emitForm_21(i
, 0x214, 0xc14);
735 if (isSignedType(i
->dType
))
738 emitForm_21(i
, 0x224, 0xc24);
741 if (i
->subOp
== NV50_IR_SUBOP_SHIFT_WRAP
)
746 CodeEmitterGK110::emitPreOp(const Instruction
*i
)
748 emitForm_C(i
, 0x248, 0x2);
750 if (i
->op
== OP_PREEX2
)
758 CodeEmitterGK110::emitSFnOp(const Instruction
*i
, uint8_t subOp
)
760 code
[0] = 0x00000002 | (subOp
<< 23);
761 code
[1] = 0x84000000;
766 srcId(i
->src(0), 10);
774 CodeEmitterGK110::emitMINMAX(const Instruction
*i
)
798 emitForm_21(i
, op2
, op1
);
800 if (i
->dType
== TYPE_S32
)
802 code
[1] |= (i
->op
== OP_MIN
) ? 0x1c00 : 0x3c00; // [!]pt
808 modNegAbsF32_3b(i
, 1);
816 CodeEmitterGK110::emitCVT(const Instruction
*i
)
818 const bool f2f
= isFloatType(i
->dType
) && isFloatType(i
->sType
);
819 const bool f2i
= !isFloatType(i
->dType
) && isFloatType(i
->sType
);
820 const bool i2f
= isFloatType(i
->dType
) && !isFloatType(i
->sType
);
822 bool sat
= i
->saturate
;
823 bool abs
= i
->src(0).mod
.abs();
824 bool neg
= i
->src(0).mod
.neg();
826 RoundMode rnd
= i
->rnd
;
829 case OP_CEIL
: rnd
= f2f
? ROUND_PI
: ROUND_P
; break;
830 case OP_FLOOR
: rnd
= f2f
? ROUND_MI
: ROUND_M
; break;
831 case OP_TRUNC
: rnd
= f2f
? ROUND_ZI
: ROUND_Z
; break;
832 case OP_SAT
: sat
= true; break;
833 case OP_NEG
: neg
= !neg
; break;
834 case OP_ABS
: abs
= true; neg
= false; break;
841 if (i
->op
== OP_NEG
&& i
->dType
== TYPE_U32
)
850 else if (f2i
) op
= 0x258;
851 else if (i2f
) op
= 0x25c;
854 emitForm_C(i
, op
, 0x2);
857 if (neg
) code
[1] |= 1 << 16;
858 if (abs
) code
[1] |= 1 << 20;
859 if (sat
) code
[1] |= 1 << 21;
861 emitRoundMode(rnd
, 32 + 10, f2f
? (32 + 13) : -1);
863 code
[0] |= typeSizeofLog2(dType
) << 10;
864 code
[0] |= typeSizeofLog2(i
->sType
) << 12;
866 if (isSignedIntType(dType
))
868 if (isSignedIntType(i
->sType
))
873 CodeEmitterGK110::emitSET(const CmpInstruction
*i
)
877 if (i
->def(0).getFile() == FILE_PREDICATE
) {
879 case TYPE_F32
: op2
= 0x1d8; op1
= 0xb58; break;
880 case TYPE_F64
: op2
= 0x1c0; op1
= 0xb40; break;
886 emitForm_21(i
, op2
, op1
);
890 if (!(code
[0] & 0x1)) {
894 modNegAbsF32_3b(i
, 1);
898 // normal DST field is negated predicate result
899 code
[0] = (code
[0] & ~0xfc) | ((code
[0] << 3) & 0xe0);
906 case TYPE_F32
: op2
= 0x000; op1
= 0x820; break;
907 case TYPE_F64
: op2
= 0x080; op1
= 0x900; break;
913 emitForm_21(i
, op2
, op1
);
917 if (!(code
[0] & 0x1)) {
921 modNegAbsF32_3b(i
, 1);
925 if (i
->dType
== TYPE_F32
)
928 if (i
->sType
== TYPE_S32
)
931 if (i
->op
!= OP_SET
) {
933 case OP_SET_AND
: code
[1] |= 0x0 << 16; break;
934 case OP_SET_OR
: code
[1] |= 0x1 << 16; break;
935 case OP_SET_XOR
: code
[1] |= 0x2 << 16; break;
940 srcId(i
->src(2), 0x2a);
942 code
[1] |= 0x7 << 10;
944 emitCondCode(i
->setCond
,
945 isFloatType(i
->sType
) ? 0x33 : 0x34,
946 isFloatType(i
->sType
) ? 0xf : 0x7);
950 CodeEmitterGK110::emitSLCT(const CmpInstruction
*i
)
952 CondCode cc
= i
->setCond
;
953 if (i
->src(2).mod
.neg())
954 cc
= reverseCondCode(cc
);
956 if (i
->dType
== TYPE_F32
) {
957 emitForm_21(i
, 0x1d0, 0xb50);
959 emitCondCode(cc
, 0x33, 0xf);
961 emitForm_21(i
, 0x1a0, 0xb20);
962 emitCondCode(cc
, 0x34, 0x7);
966 void CodeEmitterGK110::emitSELP(const Instruction
*i
)
968 emitForm_21(i
, 0x250, 0x050);
970 if ((i
->cc
== CC_NOT_P
) ^ (bool)(i
->src(2).mod
& Modifier(NV50_IR_MOD_NOT
)))
974 void CodeEmitterGK110::emitTEXBAR(const Instruction
*i
)
976 code
[0] = 0x0000003e | (i
->subOp
<< 23);
977 code
[1] = 0x77000000;
982 void CodeEmitterGK110::emitTEXCSAA(const TexInstruction
*i
)
984 code
[0] = 0x00000002;
985 code
[1] = 0x76c00000;
987 code
[1] |= i
->tex
.r
<< 9;
988 // code[1] |= i->tex.s << (9 + 8);
991 code
[0] |= 0x80000000;
994 srcId(i
->src(0), 10);
998 isNextIndependentTex(const TexInstruction
*i
)
1000 if (!i
->next
|| !isTextureOp(i
->next
->op
))
1002 if (i
->getDef(0)->interfers(i
->next
->getSrc(0)))
1004 return !i
->next
->srcExists(1) || !i
->getDef(0)->interfers(i
->next
->getSrc(1));
1008 CodeEmitterGK110::emitTEX(const TexInstruction
*i
)
1010 const bool ind
= i
->tex
.rIndirectSrc
>= 0;
1013 code
[0] = 0x00000002;
1016 code
[1] = 0x7e000000;
1019 code
[1] = 0x7e800000;
1022 code
[1] = 0x78000000;
1025 code
[1] = 0x7dc00000;
1028 code
[1] = 0x7d800000;
1034 code
[0] = 0x00000002;
1035 code
[1] = 0x76000000;
1036 code
[1] |= i
->tex
.r
<< 9;
1039 code
[0] = 0x00000002;
1040 code
[1] = 0x76800000;
1041 code
[1] |= i
->tex
.r
<< 9;
1044 code
[0] = 0x00000002;
1045 code
[1] = 0x70000000;
1046 code
[1] |= i
->tex
.r
<< 13;
1049 code
[0] = 0x00000001;
1050 code
[1] = 0x70000000;
1051 code
[1] |= i
->tex
.r
<< 15;
1054 code
[0] = 0x00000001;
1055 code
[1] = 0x60000000;
1056 code
[1] |= i
->tex
.r
<< 15;
1061 code
[1] |= isNextIndependentTex(i
) ? 0x1 : 0x2; // t : p mode
1063 if (i
->tex
.liveOnly
)
1064 code
[0] |= 0x80000000;
1068 case OP_TXB
: code
[1] |= 0x2000; break;
1069 case OP_TXL
: code
[1] |= 0x3000; break;
1073 case OP_TXLQ
: break;
1075 assert(!"invalid texture op");
1079 if (i
->op
== OP_TXF
) {
1080 if (!i
->tex
.levelZero
)
1083 if (i
->tex
.levelZero
) {
1087 if (i
->op
!= OP_TXD
&& i
->tex
.derivAll
)
1092 code
[1] |= i
->tex
.mask
<< 2;
1094 const int src1
= (i
->predSrc
== 1) ? 2 : 1; // if predSrc == 1, !srcExists(2)
1096 defId(i
->def(0), 2);
1097 srcId(i
->src(0), 10);
1100 if (i
->op
== OP_TXG
) code
[1] |= i
->tex
.gatherComp
<< 13;
1103 code
[1] |= (i
->tex
.target
.isCube() ? 3 : (i
->tex
.target
.getDim() - 1)) << 7;
1104 if (i
->tex
.target
.isArray())
1106 if (i
->tex
.target
.isShadow())
1108 if (i
->tex
.target
== TEX_TARGET_2D_MS
||
1109 i
->tex
.target
== TEX_TARGET_2D_MS_ARRAY
)
1112 if (i
->srcExists(src1
) && i
->src(src1
).getFile() == FILE_IMMEDIATE
) {
1116 if (i
->tex
.useOffsets
== 1) {
1118 case OP_TXF
: code
[1] |= 0x200; break;
1119 default: code
[1] |= 0x800; break;
1122 if (i
->tex
.useOffsets
== 4)
1127 CodeEmitterGK110::emitTXQ(const TexInstruction
*i
)
1129 code
[0] = 0x00000002;
1130 code
[1] = 0x75400001;
1132 switch (i
->tex
.query
) {
1133 case TXQ_DIMS
: code
[0] |= 0x01 << 25; break;
1134 case TXQ_TYPE
: code
[0] |= 0x02 << 25; break;
1135 case TXQ_SAMPLE_POSITION
: code
[0] |= 0x05 << 25; break;
1136 case TXQ_FILTER
: code
[0] |= 0x10 << 25; break;
1137 case TXQ_LOD
: code
[0] |= 0x12 << 25; break;
1138 case TXQ_BORDER_COLOUR
: code
[0] |= 0x16 << 25; break;
1140 assert(!"invalid texture query");
1144 code
[1] |= i
->tex
.mask
<< 2;
1145 code
[1] |= i
->tex
.r
<< 9;
1146 if (/*i->tex.sIndirectSrc >= 0 || */i
->tex
.rIndirectSrc
>= 0)
1147 code
[1] |= 0x08000000;
1149 defId(i
->def(0), 2);
1150 srcId(i
->src(0), 10);
1156 CodeEmitterGK110::emitQUADOP(const Instruction
*i
, uint8_t qOp
, uint8_t laneMask
)
1158 code
[0] = 0x00000002 | ((qOp
& 1) << 31);
1159 code
[1] = 0x7fc00000 | (qOp
>> 1) | (laneMask
<< 12);
1161 defId(i
->def(0), 2);
1162 srcId(i
->src(0), 10);
1163 srcId(i
->srcExists(1) ? i
->src(1) : i
->src(0), 23);
1165 if (i
->op
== OP_QUADOP
&& progType
!= Program::TYPE_FRAGMENT
)
1166 code
[1] |= 1 << 9; // dall
1172 CodeEmitterGK110::emitPIXLD(const Instruction
*i
)
1174 emitForm_L(i
, 0x7f4, 2, Modifier(0));
1175 code
[1] |= i
->subOp
<< 2;
1176 code
[1] |= 0x00070000;
1180 CodeEmitterGK110::emitFlow(const Instruction
*i
)
1182 const FlowInstruction
*f
= i
->asFlow();
1184 unsigned mask
; // bit 0: predicate, bit 1: target
1186 code
[0] = 0x00000000;
1190 code
[1] = f
->absolute
? 0x10800000 : 0x12000000;
1191 if (i
->srcExists(0) && i
->src(0).getFile() == FILE_MEMORY_CONST
)
1196 code
[1] = f
->absolute
? 0x11000000 : 0x13000000;
1197 if (i
->srcExists(0) && i
->src(0).getFile() == FILE_MEMORY_CONST
)
1202 case OP_EXIT
: code
[1] = 0x18000000; mask
= 1; break;
1203 case OP_RET
: code
[1] = 0x19000000; mask
= 1; break;
1204 case OP_DISCARD
: code
[1] = 0x19800000; mask
= 1; break;
1205 case OP_BREAK
: code
[1] = 0x1a000000; mask
= 1; break;
1206 case OP_CONT
: code
[1] = 0x1a800000; mask
= 1; break;
1208 case OP_JOINAT
: code
[1] = 0x14800000; mask
= 2; break;
1209 case OP_PREBREAK
: code
[1] = 0x15000000; mask
= 2; break;
1210 case OP_PRECONT
: code
[1] = 0x15800000; mask
= 2; break;
1211 case OP_PRERET
: code
[1] = 0x13800000; mask
= 2; break;
1213 case OP_QUADON
: code
[1] = 0x1b800000; mask
= 0; break;
1214 case OP_QUADPOP
: code
[1] = 0x1c000000; mask
= 0; break;
1215 case OP_BRKPT
: code
[1] = 0x00000000; mask
= 0; break;
1217 assert(!"invalid flow operation");
1223 if (i
->flagsSrc
< 0)
1235 if (f
->op
== OP_CALL
) {
1237 assert(f
->absolute
);
1238 uint32_t pcAbs
= targNVC0
->getBuiltinOffset(f
->target
.builtin
);
1239 addReloc(RelocEntry::TYPE_BUILTIN
, 0, pcAbs
, 0xff800000, 23);
1240 addReloc(RelocEntry::TYPE_BUILTIN
, 1, pcAbs
, 0x007fffff, -9);
1242 assert(!f
->absolute
);
1243 int32_t pcRel
= f
->target
.fn
->binPos
- (codeSize
+ 8);
1244 code
[0] |= (pcRel
& 0x1ff) << 23;
1245 code
[1] |= (pcRel
>> 9) & 0x7fff;
1249 int32_t pcRel
= f
->target
.bb
->binPos
- (codeSize
+ 8);
1250 // currently we don't want absolute branches
1251 assert(!f
->absolute
);
1252 code
[0] |= (pcRel
& 0x1ff) << 23;
1253 code
[1] |= (pcRel
>> 9) & 0x7fff;
1258 CodeEmitterGK110::emitPFETCH(const Instruction
*i
)
1260 uint32_t prim
= i
->src(0).get()->reg
.data
.u32
;
1262 code
[0] = 0x00000002 | ((prim
& 0xff) << 23);
1263 code
[1] = 0x7f800000;
1267 defId(i
->def(0), 2);
1268 srcId(i
->src(1), 10);
1272 CodeEmitterGK110::emitVFETCH(const Instruction
*i
)
1274 unsigned int size
= typeSizeof(i
->dType
);
1275 uint32_t offset
= i
->src(0).get()->reg
.data
.offset
;
1277 code
[0] = 0x00000002 | (offset
<< 23);
1278 code
[1] = 0x7ec00000 | (offset
>> 9);
1279 code
[1] |= (size
/ 4 - 1) << 18;
1284 if (i
->getSrc(0)->reg
.file
== FILE_SHADER_OUTPUT
)
1285 code
[0] |= 0x200; // yes, TCPs can read from *outputs* of other threads
1290 defId(i
->def(0), 2);
1291 srcId(i
->src(0).getIndirect(0), 10);
1292 srcId(i
->src(0).getIndirect(1), 32 + 10); // vertex address
1296 CodeEmitterGK110::emitEXPORT(const Instruction
*i
)
1298 unsigned int size
= typeSizeof(i
->dType
);
1299 uint32_t offset
= i
->src(0).get()->reg
.data
.offset
;
1301 code
[0] = 0x00000002 | (offset
<< 23);
1302 code
[1] = 0x7f000000 | (offset
>> 9);
1303 code
[1] |= (size
/ 4 - 1) << 18;
1312 assert(i
->src(1).getFile() == FILE_GPR
);
1314 srcId(i
->src(0).getIndirect(0), 10);
1315 srcId(i
->src(0).getIndirect(1), 32 + 10); // vertex base address
1316 srcId(i
->src(1), 2);
1320 CodeEmitterGK110::emitOUT(const Instruction
*i
)
1322 assert(i
->src(0).getFile() == FILE_GPR
);
1324 emitForm_21(i
, 0x1f0, 0xb70);
1326 if (i
->op
== OP_EMIT
)
1328 if (i
->op
== OP_RESTART
|| i
->subOp
== NV50_IR_SUBOP_EMIT_RESTART
)
1333 CodeEmitterGK110::emitInterpMode(const Instruction
*i
)
1335 code
[1] |= (i
->ipa
& 0x3) << 21; // TODO: INTERP_SAMPLEID
1336 code
[1] |= (i
->ipa
& 0xc) << (19 - 2);
1340 CodeEmitterGK110::emitINTERP(const Instruction
*i
)
1342 const uint32_t base
= i
->getSrc(0)->reg
.data
.offset
;
1344 code
[0] = 0x00000002 | (base
<< 31);
1345 code
[1] = 0x74800000 | (base
>> 1);
1350 if (i
->op
== OP_PINTERP
)
1351 srcId(i
->src(1), 23);
1353 code
[0] |= 0xff << 23;
1355 srcId(i
->src(0).getIndirect(0), 10);
1359 defId(i
->def(0), 2);
1361 if (i
->getSampleMode() == NV50_IR_INTERP_OFFSET
)
1362 srcId(i
->src(i
->op
== OP_PINTERP
? 2 : 1), 32 + 10);
1364 code
[1] |= 0xff << 10;
1368 CodeEmitterGK110::emitLoadStoreType(DataType ty
, const int pos
)
1400 assert(!"invalid ld/st type");
1403 code
[pos
/ 32] |= n
<< (pos
% 32);
1407 CodeEmitterGK110::emitCachingMode(CacheMode c
, const int pos
)
1428 assert(!"invalid caching mode");
1431 code
[pos
/ 32] |= n
<< (pos
% 32);
1435 CodeEmitterGK110::emitSTORE(const Instruction
*i
)
1437 int32_t offset
= SDATA(i
->src(0)).offset
;
1439 switch (i
->src(0).getFile()) {
1440 case FILE_MEMORY_GLOBAL
: code
[1] = 0xe0000000; code
[0] = 0x00000000; break;
1441 case FILE_MEMORY_LOCAL
: code
[1] = 0x7a800000; code
[0] = 0x00000002; break;
1442 case FILE_MEMORY_SHARED
: code
[1] = 0x7ac00000; code
[0] = 0x00000002; break;
1444 assert(!"invalid memory file");
1448 if (i
->src(0).getFile() != FILE_MEMORY_GLOBAL
)
1451 if (code
[0] & 0x2) {
1452 emitLoadStoreType(i
->dType
, 0x33);
1453 if (i
->src(0).getFile() == FILE_MEMORY_LOCAL
)
1454 emitCachingMode(i
->cache
, 0x2f);
1456 emitLoadStoreType(i
->dType
, 0x38);
1457 emitCachingMode(i
->cache
, 0x3b);
1459 code
[0] |= offset
<< 23;
1460 code
[1] |= offset
>> 9;
1464 srcId(i
->src(1), 2);
1465 srcId(i
->src(0).getIndirect(0), 10);
1469 CodeEmitterGK110::emitLOAD(const Instruction
*i
)
1471 int32_t offset
= SDATA(i
->src(0)).offset
;
1473 switch (i
->src(0).getFile()) {
1474 case FILE_MEMORY_GLOBAL
: code
[1] = 0xc0000000; code
[0] = 0x00000000; break;
1475 case FILE_MEMORY_LOCAL
: code
[1] = 0x7a000000; code
[0] = 0x00000002; break;
1476 case FILE_MEMORY_SHARED
: code
[1] = 0x7ac00000; code
[0] = 0x00000002; break;
1477 case FILE_MEMORY_CONST
:
1478 if (!i
->src(0).isIndirect(0) && typeSizeof(i
->dType
) == 4) {
1483 code
[0] = 0x00000002;
1484 code
[1] = 0x7c800000 | (i
->src(0).get()->reg
.fileIndex
<< 7);
1485 code
[1] |= i
->subOp
<< 15;
1488 assert(!"invalid memory file");
1492 if (code
[0] & 0x2) {
1494 emitLoadStoreType(i
->dType
, 0x33);
1495 if (i
->src(0).getFile() == FILE_MEMORY_LOCAL
)
1496 emitCachingMode(i
->cache
, 0x2f);
1498 emitLoadStoreType(i
->dType
, 0x38);
1499 emitCachingMode(i
->cache
, 0x3b);
1501 code
[0] |= offset
<< 23;
1502 code
[1] |= offset
>> 9;
1506 defId(i
->def(0), 2);
1507 srcId(i
->src(0).getIndirect(0), 10);
1511 CodeEmitterGK110::getSRegEncoding(const ValueRef
& ref
)
1513 switch (SDATA(ref
).sv
.sv
) {
1514 case SV_LANEID
: return 0x00;
1515 case SV_PHYSID
: return 0x03;
1516 case SV_VERTEX_COUNT
: return 0x10;
1517 case SV_INVOCATION_ID
: return 0x11;
1518 case SV_YDIR
: return 0x12;
1519 case SV_TID
: return 0x21 + SDATA(ref
).sv
.index
;
1520 case SV_CTAID
: return 0x25 + SDATA(ref
).sv
.index
;
1521 case SV_NTID
: return 0x29 + SDATA(ref
).sv
.index
;
1522 case SV_GRIDID
: return 0x2c;
1523 case SV_NCTAID
: return 0x2d + SDATA(ref
).sv
.index
;
1524 case SV_LBASE
: return 0x34;
1525 case SV_SBASE
: return 0x30;
1526 case SV_CLOCK
: return 0x50 + SDATA(ref
).sv
.index
;
1528 assert(!"no sreg for system value");
1534 CodeEmitterGK110::emitMOV(const Instruction
*i
)
1536 if (i
->src(0).getFile() == FILE_SYSTEM_VALUE
) {
1537 code
[0] = 0x00000002 | (getSRegEncoding(i
->src(0)) << 23);
1538 code
[1] = 0x86400000;
1540 defId(i
->def(0), 2);
1542 if (i
->src(0).getFile() == FILE_IMMEDIATE
) {
1543 code
[0] = 0x00000002 | (i
->lanes
<< 14);
1544 code
[1] = 0x74000000;
1546 defId(i
->def(0), 2);
1547 setImmediate32(i
, 0, Modifier(0));
1549 if (i
->src(0).getFile() == FILE_PREDICATE
) {
1550 code
[0] = 0x00000002;
1551 code
[1] = 0x84401c07;
1553 defId(i
->def(0), 2);
1554 srcId(i
->src(0), 14);
1556 emitForm_C(i
, 0x24c, 2);
1557 code
[1] |= i
->lanes
<< 10;
1562 CodeEmitterGK110::emitInstruction(Instruction
*insn
)
1564 const unsigned int size
= (writeIssueDelays
&& !(codeSize
& 0x3f)) ? 16 : 8;
1566 if (insn
->encSize
!= 8) {
1567 ERROR("skipping unencodable instruction: ");
1571 if (codeSize
+ size
> codeSizeLimit
) {
1572 ERROR("code emitter output buffer too small\n");
1576 if (writeIssueDelays
) {
1577 int id
= (codeSize
& 0x3f) / 8 - 1;
1580 code
[0] = 0x00000000; // cf issue delay "instruction"
1581 code
[1] = 0x08000000;
1585 uint32_t *data
= code
- (id
* 2 + 2);
1588 case 0: data
[0] |= insn
->sched
<< 2; break;
1589 case 1: data
[0] |= insn
->sched
<< 10; break;
1590 case 2: data
[0] |= insn
->sched
<< 18; break;
1591 case 3: data
[0] |= insn
->sched
<< 26; data
[1] |= insn
->sched
>> 6; break;
1592 case 4: data
[1] |= insn
->sched
<< 2; break;
1593 case 5: data
[1] |= insn
->sched
<< 10; break;
1594 case 6: data
[1] |= insn
->sched
<< 18; break;
1601 // assert that instructions with multiple defs don't corrupt registers
1602 for (int d
= 0; insn
->defExists(d
); ++d
)
1603 assert(insn
->asTex() || insn
->def(d
).rep()->reg
.data
.id
>= 0);
1637 if (isFloatType(insn
->dType
))
1643 if (isFloatType(insn
->dType
))
1650 if (isFloatType(insn
->dType
))
1662 emitLogicOp(insn
, 0);
1665 emitLogicOp(insn
, 1);
1668 emitLogicOp(insn
, 2);
1678 emitSET(insn
->asCmp());
1684 emitSLCT(insn
->asCmp());
1728 emitTEX(insn
->asTex());
1731 emitTXQ(insn
->asTex());
1756 emitQUADOP(insn
, insn
->subOp
, insn
->lanes
);
1759 emitQUADOP(insn
, insn
->src(0).mod
.neg() ? 0x66 : 0x99, 0x4);
1762 emitQUADOP(insn
, insn
->src(0).mod
.neg() ? 0x5a : 0xa5, 0x5);
1783 ERROR("operation should have been eliminated");
1789 ERROR("operation should have been lowered\n");
1792 ERROR("unknow op\n");
1805 CodeEmitterGK110::getMinEncodingSize(const Instruction
*i
) const
1807 // No more short instruction encodings.
1812 CodeEmitterGK110::prepareEmission(Function
*func
)
1814 const Target
*targ
= func
->getProgram()->getTarget();
1816 CodeEmitter::prepareEmission(func
);
1818 if (targ
->hasSWSched
)
1819 calculateSchedDataNVC0(targ
, func
);
1822 CodeEmitterGK110::CodeEmitterGK110(const TargetNVC0
*target
)
1823 : CodeEmitter(target
),
1825 writeIssueDelays(target
->hasSWSched
)
1828 codeSize
= codeSizeLimit
= 0;
1833 TargetNVC0::createCodeEmitterGK110(Program::Type type
)
1835 CodeEmitterGK110
*emit
= new CodeEmitterGK110(this);
1836 emit
->setProgramType(type
);
1840 } // namespace nv50_ir