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 emitShift(const Instruction
*);
99 void emitSFnOp(const Instruction
*, uint8_t subOp
);
101 void emitCVT(const Instruction
*);
102 void emitMINMAX(const Instruction
*);
103 void emitPreOp(const Instruction
*);
105 void emitSET(const CmpInstruction
*);
106 void emitSLCT(const CmpInstruction
*);
107 void emitSELP(const Instruction
*);
109 void emitTEXBAR(const Instruction
*);
110 void emitTEX(const TexInstruction
*);
111 void emitTEXCSAA(const TexInstruction
*);
112 void emitTXQ(const TexInstruction
*);
114 void emitQUADOP(const Instruction
*, uint8_t qOp
, uint8_t laneMask
);
116 void emitFlow(const Instruction
*);
118 inline void defId(const ValueDef
&, const int pos
);
119 inline void srcId(const ValueRef
&, const int pos
);
120 inline void srcId(const ValueRef
*, const int pos
);
121 inline void srcId(const Instruction
*, int s
, const int pos
);
123 inline void srcAddr32(const ValueRef
&, const int pos
); // address / 4
125 inline bool isLIMM(const ValueRef
&, DataType ty
, bool mod
= false);
128 #define GK110_GPR_ZERO 255
131 if (i->src(s).mod.neg()) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
133 if (i->src(s).mod.abs()) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
135 #define NOT_(b, s) if (i->src(s).mod & Modifier(NV50_IR_MOD_NOT)) \
136 code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
138 #define FTZ_(b) if (i->ftz) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
140 #define SAT_(b) if (i->saturate) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
142 #define RND_(b, t) emitRoundMode##t(i->rnd, 0x##b)
144 #define SDATA(a) ((a).rep()->reg.data)
145 #define DDATA(a) ((a).rep()->reg.data)
147 void CodeEmitterGK110::srcId(const ValueRef
& src
, const int pos
)
149 code
[pos
/ 32] |= (src
.get() ? SDATA(src
).id
: GK110_GPR_ZERO
) << (pos
% 32);
152 void CodeEmitterGK110::srcId(const ValueRef
*src
, const int pos
)
154 code
[pos
/ 32] |= (src
? SDATA(*src
).id
: GK110_GPR_ZERO
) << (pos
% 32);
157 void CodeEmitterGK110::srcId(const Instruction
*insn
, int s
, int pos
)
159 int r
= insn
->srcExists(s
) ? SDATA(insn
->src(s
)).id
: GK110_GPR_ZERO
;
160 code
[pos
/ 32] |= r
<< (pos
% 32);
163 void CodeEmitterGK110::srcAddr32(const ValueRef
& src
, const int pos
)
165 code
[pos
/ 32] |= (SDATA(src
).offset
>> 2) << (pos
% 32);
168 void CodeEmitterGK110::defId(const ValueDef
& def
, const int pos
)
170 code
[pos
/ 32] |= (def
.get() ? DDATA(def
).id
: GK110_GPR_ZERO
) << (pos
% 32);
173 bool CodeEmitterGK110::isLIMM(const ValueRef
& ref
, DataType ty
, bool mod
)
175 const ImmediateValue
*imm
= ref
.get()->asImm();
177 return imm
&& (imm
->reg
.data
.u32
& ((ty
== TYPE_F32
) ? 0xfff : 0xfff00000));
181 CodeEmitterGK110::emitRoundMode(RoundMode rnd
, const int pos
, const int rintPos
)
187 case ROUND_MI
: rint
= true; /* fall through */ case ROUND_M
: n
= 1; break;
188 case ROUND_PI
: rint
= true; /* fall through */ case ROUND_P
: n
= 2; break;
189 case ROUND_ZI
: rint
= true; /* fall through */ case ROUND_Z
: n
= 3; break;
191 rint
= rnd
== ROUND_NI
;
193 assert(rnd
== ROUND_N
|| rnd
== ROUND_NI
);
196 code
[pos
/ 32] |= n
<< (pos
% 32);
197 if (rint
&& rintPos
>= 0)
198 code
[rintPos
/ 32] |= 1 << (rintPos
% 32);
202 CodeEmitterGK110::emitRoundModeF(RoundMode rnd
, const int pos
)
207 case ROUND_M
: n
= 1; break;
208 case ROUND_P
: n
= 2; break;
209 case ROUND_Z
: n
= 3; break;
212 assert(rnd
== ROUND_N
);
215 code
[pos
/ 32] |= n
<< (pos
% 32);
219 CodeEmitterGK110::emitRoundModeI(RoundMode rnd
, const int pos
)
224 case ROUND_MI
: n
= 1; break;
225 case ROUND_PI
: n
= 2; break;
226 case ROUND_ZI
: n
= 3; break;
229 assert(rnd
== ROUND_NI
);
232 code
[pos
/ 32] |= n
<< (pos
% 32);
235 void CodeEmitterGK110::emitCondCode(CondCode cc
, int pos
, uint8_t mask
)
240 case CC_FL
: n
= 0x00; break;
241 case CC_LT
: n
= 0x01; break;
242 case CC_EQ
: n
= 0x02; break;
243 case CC_LE
: n
= 0x03; break;
244 case CC_GT
: n
= 0x04; break;
245 case CC_NE
: n
= 0x05; break;
246 case CC_GE
: n
= 0x06; break;
247 case CC_LTU
: n
= 0x09; break;
248 case CC_EQU
: n
= 0x0a; break;
249 case CC_LEU
: n
= 0x0b; break;
250 case CC_GTU
: n
= 0x0c; break;
251 case CC_NEU
: n
= 0x0d; break;
252 case CC_GEU
: n
= 0x0e; break;
253 case CC_TR
: n
= 0x0f; break;
254 case CC_NO
: n
= 0x10; break;
255 case CC_NC
: n
= 0x11; break;
256 case CC_NS
: n
= 0x12; break;
257 case CC_NA
: n
= 0x13; break;
258 case CC_A
: n
= 0x14; break;
259 case CC_S
: n
= 0x15; break;
260 case CC_C
: n
= 0x16; break;
261 case CC_O
: n
= 0x17; break;
264 assert(!"invalid condition code");
267 code
[pos
/ 32] |= (n
& mask
) << (pos
% 32);
271 CodeEmitterGK110::emitPredicate(const Instruction
*i
)
273 if (i
->predSrc
>= 0) {
274 srcId(i
->src(i
->predSrc
), 18);
275 if (i
->cc
== CC_NOT_P
)
276 code
[0] |= 8 << 18; // negate
277 assert(i
->getPredicate()->reg
.file
== FILE_PREDICATE
);
284 CodeEmitterGK110::setCAddress14(const ValueRef
& src
)
286 const int32_t addr
= src
.get()->asSym()->reg
.data
.offset
/ 4;
288 code
[0] |= (addr
& 0x01ff) << 23;
289 code
[1] |= (addr
& 0x3e00) >> 9;
293 CodeEmitterGK110::setShortImmediate(const Instruction
*i
, const int s
)
295 const uint32_t u32
= i
->getSrc(s
)->asImm()->reg
.data
.u32
;
296 const uint64_t u64
= i
->getSrc(s
)->asImm()->reg
.data
.u64
;
298 if (i
->sType
== TYPE_F32
) {
299 assert(!(u32
& 0x00000fff));
300 code
[0] |= ((u32
& 0x001ff000) >> 12) << 23;
301 code
[1] |= ((u32
& 0x7fe00000) >> 21);
302 code
[1] |= ((u32
& 0x80000000) >> 4);
304 if (i
->sType
== TYPE_F64
) {
305 assert(!(u64
& 0x00000fffffffffffULL
));
306 code
[0] |= ((u64
& 0x001ff00000000000ULL
) >> 44) << 23;
307 code
[1] |= ((u64
& 0x7fe0000000000000ULL
) >> 53);
308 code
[1] |= ((u64
& 0x8000000000000000ULL
) >> 36);
310 assert((u32
& 0xfff00000) == 0 || (u32
& 0xfff00000) == 0xfff00000);
311 code
[0] |= (u32
& 0x001ff) << 23;
312 code
[1] |= (u32
& 0x7fe00) >> 9;
313 code
[1] |= (u32
& 0x80000) << 8;
318 CodeEmitterGK110::setImmediate32(const Instruction
*i
, const int s
,
321 uint32_t u32
= i
->getSrc(s
)->asImm()->reg
.data
.u32
;
324 ImmediateValue
imm(i
->getSrc(s
)->asImm(), i
->sType
);
326 u32
= imm
.reg
.data
.u32
;
329 code
[0] |= u32
<< 23;
334 CodeEmitterGK110::emitForm_L(const Instruction
*i
, uint32_t opc
, uint8_t ctg
,
344 for (int s
= 0; s
< 3 && i
->srcExists(s
); ++s
) {
345 switch (i
->src(s
).getFile()) {
347 srcId(i
->src(s
), s
? 42 : 10);
350 setImmediate32(i
, s
, mod
);
360 CodeEmitterGK110::emitForm_C(const Instruction
*i
, uint32_t opc
, uint8_t ctg
)
369 switch (i
->src(0).getFile()) {
370 case FILE_MEMORY_CONST
:
371 code
[1] |= 0x4 << 28;
372 setCAddress14(i
->src(0));
375 code
[1] |= 0xc << 28;
376 srcId(i
->src(0), 23);
384 // 0x2 for GPR, c[] and 0x1 for short immediate
386 CodeEmitterGK110::emitForm_21(const Instruction
*i
, uint32_t opc2
,
389 const bool imm
= i
->srcExists(1) && i
->src(1).getFile() == FILE_IMMEDIATE
;
392 if (i
->srcExists(2) && i
->src(2).getFile() == FILE_MEMORY_CONST
)
397 code
[1] = opc1
<< 20;
400 code
[1] = (0xc << 28) | (opc2
<< 20);
407 for (int s
= 0; s
< 3 && i
->srcExists(s
); ++s
) {
408 switch (i
->src(s
).getFile()) {
409 case FILE_MEMORY_CONST
:
410 code
[1] &= (s
== 2) ? ~(0x4 << 28) : ~(0x8 << 28);
411 setCAddress14(i
->src(s
));
412 code
[1] |= i
->getSrc(s
)->reg
.fileIndex
<< 5;
415 setShortImmediate(i
, s
);
418 srcId(i
->src(s
), s
? ((s
== 2) ? 42 : s1
) : 10);
421 // ignore here, can be predicate or flags, but must not be address
429 assert(imm
|| (code
[1] & (0xc << 28)));
433 CodeEmitterGK110::modNegAbsF32_3b(const Instruction
*i
, const int s
)
435 if (i
->src(s
).mod
.abs()) code
[1] &= ~(1 << 27);
436 if (i
->src(s
).mod
.neg()) code
[1] ^= (1 << 27);
440 CodeEmitterGK110::emitNOP(const Instruction
*i
)
442 code
[0] = 0x00003c02;
443 code
[1] = 0x85800000;
448 code
[0] = 0x001c3c02;
452 CodeEmitterGK110::emitFMAD(const Instruction
*i
)
454 assert(!isLIMM(i
->src(1), TYPE_F32
));
456 emitForm_21(i
, 0x0c0, 0x940);
463 bool neg1
= (i
->src(0).mod
^ i
->src(1).mod
).neg();
475 CodeEmitterGK110::emitFMUL(const Instruction
*i
)
477 bool neg
= (i
->src(0).mod
^ i
->src(1).mod
).neg();
479 assert(i
->postFactor
>= -3 && i
->postFactor
<= 3);
481 if (isLIMM(i
->src(1), TYPE_F32
)) {
482 emitForm_L(i
, 0x200, 0x2, Modifier(0));
489 assert(i
->postFactor
== 0);
491 emitForm_21(i
, 0x234, 0xc34);
508 CodeEmitterGK110::emitIMUL(const Instruction
*i
)
510 assert(!i
->src(0).mod
.neg() && !i
->src(1).mod
.neg());
511 assert(!i
->src(0).mod
.abs() && !i
->src(1).mod
.abs());
513 if (isLIMM(i
->src(1), TYPE_S32
)) {
514 emitForm_L(i
, 0x280, 2, Modifier(0));
516 assert(i
->subOp
!= NV50_IR_SUBOP_MUL_HIGH
);
518 if (i
->sType
== TYPE_S32
)
521 emitForm_21(i
, 0x21c, 0xc1c);
523 if (i
->subOp
== NV50_IR_SUBOP_MUL_HIGH
)
525 if (i
->sType
== TYPE_S32
)
531 CodeEmitterGK110::emitFADD(const Instruction
*i
)
533 if (isLIMM(i
->src(1), TYPE_F32
)) {
534 assert(i
->rnd
== ROUND_N
);
535 assert(!i
->saturate
);
537 emitForm_L(i
, 0x400, 0, i
->src(1).mod
);
543 emitForm_21(i
, 0x22c, 0xc2c);
551 modNegAbsF32_3b(i
, 1);
560 CodeEmitterGK110::emitUADD(const Instruction
*i
)
562 uint8_t addOp
= (i
->src(0).mod
.neg() << 1) | i
->src(1).mod
.neg();
567 assert(!i
->src(0).mod
.abs() && !i
->src(1).mod
.abs());
569 if (isLIMM(i
->src(1), TYPE_S32
)) {
570 emitForm_L(i
, 0x400, 1, Modifier((addOp
& 1) ? NV50_IR_MOD_NEG
: 0));
575 assert(!i
->defExists(1));
576 assert(i
->flagsSrc
< 0);
580 emitForm_21(i
, 0x208, 0xc08);
582 assert(addOp
!= 3); // would be add-plus-one
584 code
[1] |= addOp
<< 19;
587 code
[1] |= 1 << 18; // write carry
588 if (i
->flagsSrc
>= 0)
589 code
[1] |= 1 << 14; // add carry
597 CodeEmitterGK110::emitIMAD(const Instruction
*i
)
600 (i
->src(2).mod
.neg() << 1) | (i
->src(0).mod
.neg() ^ i
->src(1).mod
.neg());
602 emitForm_21(i
, 0x100, 0xa00);
605 code
[1] |= addOp
<< 26;
607 if (i
->sType
== TYPE_S32
)
608 code
[1] |= (1 << 19) | (1 << 24);
614 if (i
->subOp
== NV50_IR_SUBOP_MUL_HIGH
)
621 CodeEmitterGK110::emitISAD(const Instruction
*i
)
623 assert(i
->dType
== TYPE_S32
|| i
->dType
== TYPE_U32
);
625 emitForm_21(i
, 0x1fc, 0xb74);
627 if (i
->dType
== TYPE_S32
)
632 CodeEmitterGK110::emitNOT(const Instruction
*i
)
634 code
[0] = 0x0003fc02; // logop(mov2) dst, 0, not src
635 code
[1] = 0x22003800;
641 switch (i
->src(0).getFile()) {
643 code
[1] |= 0xc << 28;
644 srcId(i
->src(0), 23);
646 case FILE_MEMORY_CONST
:
647 code
[1] |= 0x4 << 28;
648 setCAddress14(i
->src(1));
657 CodeEmitterGK110::emitLogicOp(const Instruction
*i
, uint8_t subOp
)
659 assert(!(i
->src(0).mod
& Modifier(NV50_IR_MOD_NOT
))); // XXX: find me
661 if (isLIMM(i
->src(1), TYPE_S32
)) {
662 emitForm_L(i
, 0x200, 0, i
->src(1).mod
);
663 code
[1] |= subOp
<< 24;
665 emitForm_21(i
, 0x220, 0xc20);
666 code
[1] |= subOp
<< 12;
669 assert(!(code
[0] & 0x1) || !(i
->src(1).mod
& Modifier(NV50_IR_MOD_NOT
)));
673 CodeEmitterGK110::emitPOPC(const Instruction
*i
)
675 assert(!isLIMM(i
->src(1), TYPE_S32
, true));
677 emitForm_21(i
, 0x204, 0xc04);
680 if (!(code
[0] & 0x1))
685 CodeEmitterGK110::emitINSBF(const Instruction
*i
)
687 emitForm_21(i
, 0x1f8, 0xb78);
691 CodeEmitterGK110::emitShift(const Instruction
*i
)
693 const bool sar
= i
->op
== OP_SHR
&& isSignedType(i
->sType
);
696 emitForm_21(i
, 0x214, 0x014);
699 if (i
->op
== OP_SHR
) {
700 // this is actually RSHF
701 emitForm_21(i
, 0x27c, 0x87c);
702 code
[1] |= GK110_GPR_ZERO
<< 10;
704 // this is actually LSHF
705 emitForm_21(i
, 0x1fc, 0xb7c);
706 code
[1] |= GK110_GPR_ZERO
<< 10;
709 if (i
->subOp
== NV50_IR_SUBOP_SHIFT_WRAP
) {
712 // XXX: find wrap modifier for SHR S32
717 CodeEmitterGK110::emitPreOp(const Instruction
*i
)
719 emitForm_21(i
, 0x248, -1);
721 if (i
->op
== OP_PREEX2
)
729 CodeEmitterGK110::emitSFnOp(const Instruction
*i
, uint8_t subOp
)
731 code
[0] = 0x00000002 | (subOp
<< 23);
732 code
[1] = 0x84000000;
737 srcId(i
->src(0), 10);
745 CodeEmitterGK110::emitMINMAX(const Instruction
*i
)
769 emitForm_21(i
, op2
, op1
);
771 if (i
->dType
== TYPE_S32
)
773 code
[1] |= (i
->op
== OP_MIN
) ? 0x1c00 : 0x3c00; // [!]pt
779 modNegAbsF32_3b(i
, 1);
787 CodeEmitterGK110::emitCVT(const Instruction
*i
)
789 const bool f2f
= isFloatType(i
->dType
) && isFloatType(i
->sType
);
790 const bool f2i
= !isFloatType(i
->dType
) && isFloatType(i
->sType
);
791 const bool i2f
= isFloatType(i
->dType
) && !isFloatType(i
->sType
);
793 bool sat
= i
->saturate
;
794 bool abs
= i
->src(0).mod
.abs();
795 bool neg
= i
->src(0).mod
.neg();
797 RoundMode rnd
= i
->rnd
;
800 case OP_CEIL
: rnd
= f2f
? ROUND_PI
: ROUND_P
; break;
801 case OP_FLOOR
: rnd
= f2f
? ROUND_MI
: ROUND_M
; break;
802 case OP_TRUNC
: rnd
= f2f
? ROUND_ZI
: ROUND_Z
; break;
803 case OP_SAT
: sat
= true; break;
804 case OP_NEG
: neg
= !neg
; break;
805 case OP_ABS
: abs
= true; neg
= false; break;
812 if (i
->op
== OP_NEG
&& i
->dType
== TYPE_U32
)
821 else if (f2i
) op
= 0x258;
822 else if (i2f
) op
= 0x25c;
825 emitForm_C(i
, op
, 0x2);
828 if (neg
) code
[1] |= 1 << 16;
829 if (abs
) code
[1] |= 1 << 20;
830 if (sat
) code
[1] |= 1 << 21;
832 emitRoundMode(rnd
, 32 + 10, f2f
? (32 + 13) : -1);
834 code
[0] |= typeSizeofLog2(dType
) << 10;
835 code
[0] |= typeSizeofLog2(i
->sType
) << 12;
837 if (isSignedIntType(dType
))
839 if (isSignedIntType(i
->sType
))
844 CodeEmitterGK110::emitSET(const CmpInstruction
*i
)
848 if (i
->def(0).getFile() == FILE_PREDICATE
) {
850 case TYPE_F32
: op2
= 0x1d8; op1
= 0xb58; break;
851 case TYPE_F64
: op2
= 0x1c0; op1
= 0xb40; break;
857 emitForm_21(i
, op2
, op1
);
861 if (!(code
[0] & 0x1)) {
865 modNegAbsF32_3b(i
, 1);
869 // normal DST field is negated predicate result
870 code
[0] = (code
[0] & ~0xfc) | ((code
[0] << 3) & 0xe0);
877 case TYPE_F32
: op2
= 0x000; op1
= 0x820; break;
878 case TYPE_F64
: op2
= 0x080; op1
= 0x900; break;
884 emitForm_21(i
, op2
, op1
);
888 if (!(code
[0] & 0x1)) {
892 modNegAbsF32_3b(i
, 1);
896 if (i
->sType
== TYPE_S32
)
899 if (i
->op
!= OP_SET
) {
901 case OP_SET_AND
: code
[1] |= 0x0 << 16; break;
902 case OP_SET_OR
: code
[1] |= 0x1 << 16; break;
903 case OP_SET_XOR
: code
[1] |= 0x2 << 16; break;
908 srcId(i
->src(2), 0x2a);
910 code
[1] |= 0x7 << 10;
912 emitCondCode(i
->setCond
,
913 isFloatType(i
->sType
) ? 0x33 : 0x34,
914 isFloatType(i
->sType
) ? 0xf : 0x7);
918 CodeEmitterGK110::emitSLCT(const CmpInstruction
*i
)
920 CondCode cc
= i
->setCond
;
921 if (i
->src(2).mod
.neg())
922 cc
= reverseCondCode(cc
);
924 if (i
->dType
== TYPE_F32
) {
925 emitForm_21(i
, 0x1d0, 0xb50);
927 emitCondCode(cc
, 0x33, 0xf);
929 emitForm_21(i
, 0x1a4, 0xb20);
930 emitCondCode(cc
, 0x34, 0x7);
934 void CodeEmitterGK110::emitSELP(const Instruction
*i
)
936 emitForm_21(i
, 0x250, 0x050);
938 if ((i
->cc
== CC_NOT_P
) ^ (bool)(i
->src(2).mod
& Modifier(NV50_IR_MOD_NOT
)))
942 void CodeEmitterGK110::emitTEXBAR(const Instruction
*i
)
944 code
[0] = 0x00000002 | (i
->subOp
<< 23);
945 code
[1] = 0x77000000;
950 void CodeEmitterGK110::emitTEXCSAA(const TexInstruction
*i
)
952 code
[0] = 0x00000002;
953 code
[1] = 0x76c00000;
955 code
[1] |= i
->tex
.r
<< 9;
956 // code[1] |= i->tex.s << (9 + 8);
959 code
[0] |= 0x80000000;
962 srcId(i
->src(0), 10);
966 isNextIndependentTex(const TexInstruction
*i
)
968 if (!i
->next
|| !isTextureOp(i
->next
->op
))
970 if (i
->getDef(0)->interfers(i
->next
->getSrc(0)))
972 return !i
->next
->srcExists(1) || !i
->getDef(0)->interfers(i
->next
->getSrc(1));
976 CodeEmitterGK110::emitTEX(const TexInstruction
*i
)
978 const bool ind
= i
->tex
.rIndirectSrc
>= 0;
981 code
[0] = 0x00000002;
984 code
[1] = 0x7e000000;
987 code
[1] = 0x78000000;
990 code
[1] = 0x7d800000;
996 code
[0] = 0x00000002;
997 code
[1] = 0x76000000;
998 code
[1] |= i
->tex
.r
<< 9;
1001 code
[0] = 0x00000002;
1002 code
[1] = 0x70000000;
1003 code
[1] |= i
->tex
.r
<< 13;
1006 code
[0] = 0x00000001;
1007 code
[1] = 0x60000000;
1008 code
[1] |= i
->tex
.r
<< 15;
1013 code
[1] |= isNextIndependentTex(i
) ? 0x1 : 0x2; // t : p mode
1015 if (i
->tex
.liveOnly
)
1016 code
[0] |= 0x80000000;
1020 case OP_TXB
: code
[1] |= 0x2000; break;
1021 case OP_TXL
: code
[1] |= 0x3000; break;
1023 case OP_TXG
: break; // XXX
1026 assert(!"invalid texture op");
1030 if (i
->op
== OP_TXF
) {
1031 if (!i
->tex
.levelZero
)
1034 if (i
->tex
.levelZero
) {
1038 if (i
->op
!= OP_TXD
&& i
->tex
.derivAll
)
1043 code
[1] |= i
->tex
.mask
<< 2;
1045 const int src1
= (i
->predSrc
== 1) ? 2 : 1; // if predSrc == 1, !srcExists(2)
1047 defId(i
->def(0), 2);
1048 srcId(i
->src(0), 10);
1051 // if (i->op == OP_TXG) code[0] |= i->tex.gatherComp << 5;
1054 code
[1] |= (i
->tex
.target
.isCube() ? 3 : (i
->tex
.target
.getDim() - 1)) << 7;
1055 if (i
->tex
.target
.isArray())
1057 if (i
->tex
.target
.isShadow())
1059 if (i
->tex
.target
== TEX_TARGET_2D_MS
||
1060 i
->tex
.target
== TEX_TARGET_2D_MS_ARRAY
)
1063 if (i
->srcExists(src1
) && i
->src(src1
).getFile() == FILE_IMMEDIATE
) {
1067 if (i
->tex
.useOffsets
) {
1069 case OP_TXF
: code
[1] |= 0x200; break;
1070 default: code
[1] |= 0x800; break;
1076 CodeEmitterGK110::emitTXQ(const TexInstruction
*i
)
1078 code
[0] = 0x00000002;
1079 code
[1] = 0x75400001;
1081 switch (i
->tex
.query
) {
1082 case TXQ_DIMS
: code
[0] |= 0x01 << 25; break;
1083 case TXQ_TYPE
: code
[0] |= 0x02 << 25; break;
1084 case TXQ_SAMPLE_POSITION
: code
[0] |= 0x05 << 25; break;
1085 case TXQ_FILTER
: code
[0] |= 0x10 << 25; break;
1086 case TXQ_LOD
: code
[0] |= 0x12 << 25; break;
1087 case TXQ_BORDER_COLOUR
: code
[0] |= 0x16 << 25; break;
1089 assert(!"invalid texture query");
1093 code
[1] |= i
->tex
.mask
<< 2;
1094 code
[1] |= i
->tex
.r
<< 9;
1095 if (/*i->tex.sIndirectSrc >= 0 || */i
->tex
.rIndirectSrc
>= 0)
1096 code
[1] |= 0x08000000;
1098 defId(i
->def(0), 2);
1099 srcId(i
->src(0), 10);
1105 CodeEmitterGK110::emitQUADOP(const Instruction
*i
, uint8_t qOp
, uint8_t laneMask
)
1111 CodeEmitterGK110::emitFlow(const Instruction
*i
)
1113 const FlowInstruction
*f
= i
->asFlow();
1115 unsigned mask
; // bit 0: predicate, bit 1: target
1117 code
[0] = 0x00000000;
1121 code
[1] = f
->absolute
? 0x10800000 : 0x12000000;
1122 if (i
->srcExists(0) && i
->src(0).getFile() == FILE_MEMORY_CONST
)
1127 code
[1] = f
->absolute
? 0x11000000 : 0x13000000;
1128 if (i
->srcExists(0) && i
->src(0).getFile() == FILE_MEMORY_CONST
)
1133 case OP_EXIT
: code
[1] = 0x18000000; mask
= 1; break;
1134 case OP_RET
: code
[1] = 0x19000000; mask
= 1; break;
1135 case OP_DISCARD
: code
[1] = 0x19800000; mask
= 1; break;
1136 case OP_BREAK
: code
[1] = 0x1a000000; mask
= 1; break;
1137 case OP_CONT
: code
[1] = 0x1a800000; mask
= 1; break;
1139 case OP_JOINAT
: code
[1] = 0x14800000; mask
= 2; break;
1140 case OP_PREBREAK
: code
[1] = 0x15000000; mask
= 2; break;
1141 case OP_PRECONT
: code
[1] = 0x15800000; mask
= 2; break;
1142 case OP_PRERET
: code
[1] = 0x13800000; mask
= 2; break;
1144 case OP_QUADON
: code
[1] = 0x1b000000; mask
= 0; break;
1145 case OP_QUADPOP
: code
[1] = 0x1c000000; mask
= 0; break;
1146 case OP_BRKPT
: code
[1] = 0x00000000; mask
= 0; break;
1148 assert(!"invalid flow operation");
1154 if (i
->flagsSrc
< 0)
1166 if (f
->op
== OP_CALL
) {
1168 assert(f
->absolute
);
1169 uint32_t pcAbs
= targNVC0
->getBuiltinOffset(f
->target
.builtin
);
1170 addReloc(RelocEntry::TYPE_BUILTIN
, 0, pcAbs
, 0xff800000, 23);
1171 addReloc(RelocEntry::TYPE_BUILTIN
, 1, pcAbs
, 0x007fffff, -9);
1173 assert(!f
->absolute
);
1174 int32_t pcRel
= f
->target
.fn
->binPos
- (codeSize
+ 8);
1175 code
[0] |= (pcRel
& 0x1ff) << 23;
1176 code
[1] |= (pcRel
>> 9) & 0x7fff;
1180 int32_t pcRel
= f
->target
.bb
->binPos
- (codeSize
+ 8);
1181 // currently we don't want absolute branches
1182 assert(!f
->absolute
);
1183 code
[0] |= (pcRel
& 0x1ff) << 23;
1184 code
[1] |= (pcRel
>> 9) & 0x7fff;
1189 CodeEmitterGK110::emitPFETCH(const Instruction
*i
)
1191 uint32_t prim
= i
->src(0).get()->reg
.data
.u32
;
1193 code
[0] = 0x00000002 | ((prim
& 0xff) << 23);
1194 code
[1] = 0x7f800000;
1198 defId(i
->def(0), 2);
1199 srcId(i
->src(1), 10);
1203 CodeEmitterGK110::emitVFETCH(const Instruction
*i
)
1205 uint32_t offset
= i
->src(0).get()->reg
.data
.offset
;
1207 code
[0] = 0x00000002 | (offset
<< 23);
1208 code
[1] = 0x7ec00000 | (offset
>> 9);
1213 if (i
->getSrc(0)->reg
.file
== FILE_SHADER_OUTPUT
)
1214 code
[0] |= 0x200; // yes, TCPs can read from *outputs* of other threads
1219 defId(i
->def(0), 2);
1220 srcId(i
->src(0).getIndirect(0), 10);
1221 srcId(i
->src(0).getIndirect(1), 32 + 10); // vertex address
1225 CodeEmitterGK110::emitEXPORT(const Instruction
*i
)
1227 uint32_t offset
= i
->src(0).get()->reg
.data
.offset
;
1229 code
[0] = 0x00000002 | (offset
<< 23);
1230 code
[1] = 0x7f000000 | (offset
>> 9);
1239 assert(i
->src(1).getFile() == FILE_GPR
);
1241 srcId(i
->src(0).getIndirect(0), 10);
1242 srcId(i
->src(0).getIndirect(1), 32 + 10); // vertex base address
1243 srcId(i
->src(1), 2);
1247 CodeEmitterGK110::emitOUT(const Instruction
*i
)
1249 assert(i
->src(0).getFile() == FILE_GPR
);
1251 emitForm_21(i
, 0x1f0, 0xb70);
1253 if (i
->op
== OP_EMIT
)
1255 if (i
->op
== OP_RESTART
|| i
->subOp
== NV50_IR_SUBOP_EMIT_RESTART
)
1260 CodeEmitterGK110::emitInterpMode(const Instruction
*i
)
1262 code
[1] |= i
->ipa
<< 21; // TODO: INTERP_SAMPLEID
1266 CodeEmitterGK110::emitINTERP(const Instruction
*i
)
1268 const uint32_t base
= i
->getSrc(0)->reg
.data
.offset
;
1270 code
[0] = 0x00000002 | (base
<< 31);
1271 code
[1] = 0x74800000 | (base
>> 1);
1276 if (i
->op
== OP_PINTERP
)
1277 srcId(i
->src(1), 23);
1279 code
[0] |= 0xff << 23;
1281 srcId(i
->src(0).getIndirect(0), 10);
1285 defId(i
->def(0), 2);
1287 if (i
->getSampleMode() == NV50_IR_INTERP_OFFSET
)
1288 srcId(i
->src(i
->op
== OP_PINTERP
? 2 : 1), 32 + 10);
1290 code
[1] |= 0xff << 10;
1294 CodeEmitterGK110::emitLoadStoreType(DataType ty
, const int pos
)
1326 assert(!"invalid ld/st type");
1329 code
[pos
/ 32] |= n
<< (pos
% 32);
1333 CodeEmitterGK110::emitCachingMode(CacheMode c
, const int pos
)
1354 assert(!"invalid caching mode");
1357 code
[pos
/ 32] |= n
<< (pos
% 32);
1361 CodeEmitterGK110::emitSTORE(const Instruction
*i
)
1363 int32_t offset
= SDATA(i
->src(0)).offset
;
1365 switch (i
->src(0).getFile()) {
1366 case FILE_MEMORY_GLOBAL
: code
[1] = 0xe0000000; code
[0] = 0x00000000; break;
1367 case FILE_MEMORY_LOCAL
: code
[1] = 0x7a800000; code
[0] = 0x00000002; break;
1368 case FILE_MEMORY_SHARED
: code
[1] = 0x7ac00000; code
[0] = 0x00000002; break;
1370 assert(!"invalid memory file");
1374 if (i
->src(0).getFile() != FILE_MEMORY_GLOBAL
)
1377 if (code
[0] & 0x2) {
1378 emitLoadStoreType(i
->dType
, 0x33);
1379 if (i
->src(0).getFile() == FILE_MEMORY_LOCAL
)
1380 emitCachingMode(i
->cache
, 0x2f);
1382 emitLoadStoreType(i
->dType
, 0x38);
1383 emitCachingMode(i
->cache
, 0x3b);
1385 code
[0] |= offset
<< 23;
1386 code
[1] |= offset
>> 9;
1390 srcId(i
->src(1), 2);
1391 srcId(i
->src(0).getIndirect(0), 10);
1395 CodeEmitterGK110::emitLOAD(const Instruction
*i
)
1397 int32_t offset
= SDATA(i
->src(0)).offset
;
1399 switch (i
->src(0).getFile()) {
1400 case FILE_MEMORY_GLOBAL
: code
[1] = 0xc0000000; code
[0] = 0x00000000; break;
1401 case FILE_MEMORY_LOCAL
: code
[1] = 0x7a000000; code
[0] = 0x00000002; break;
1402 case FILE_MEMORY_SHARED
: code
[1] = 0x7ac00000; code
[0] = 0x00000002; break;
1403 case FILE_MEMORY_CONST
:
1404 if (!i
->src(0).isIndirect(0) && typeSizeof(i
->dType
) == 4) {
1409 code
[0] = 0x00000002;
1410 code
[1] = 0x7c800000 | (i
->src(0).get()->reg
.fileIndex
<< 7);
1413 assert(!"invalid memory file");
1417 if (code
[0] & 0x2) {
1419 emitLoadStoreType(i
->dType
, 0x33);
1420 if (i
->src(0).getFile() == FILE_MEMORY_LOCAL
)
1421 emitCachingMode(i
->cache
, 0x2f);
1423 emitLoadStoreType(i
->dType
, 0x38);
1424 emitCachingMode(i
->cache
, 0x3b);
1426 code
[0] |= offset
<< 23;
1427 code
[1] |= offset
>> 9;
1431 defId(i
->def(0), 2);
1432 srcId(i
->src(0).getIndirect(0), 10);
1436 CodeEmitterGK110::getSRegEncoding(const ValueRef
& ref
)
1438 switch (SDATA(ref
).sv
.sv
) {
1439 case SV_LANEID
: return 0x00;
1440 case SV_PHYSID
: return 0x03;
1441 case SV_VERTEX_COUNT
: return 0x10;
1442 case SV_INVOCATION_ID
: return 0x11;
1443 case SV_YDIR
: return 0x12;
1444 case SV_TID
: return 0x21 + SDATA(ref
).sv
.index
;
1445 case SV_CTAID
: return 0x25 + SDATA(ref
).sv
.index
;
1446 case SV_NTID
: return 0x29 + SDATA(ref
).sv
.index
;
1447 case SV_GRIDID
: return 0x2c;
1448 case SV_NCTAID
: return 0x2d + SDATA(ref
).sv
.index
;
1449 case SV_LBASE
: return 0x34;
1450 case SV_SBASE
: return 0x30;
1451 case SV_CLOCK
: return 0x50 + SDATA(ref
).sv
.index
;
1453 assert(!"no sreg for system value");
1459 CodeEmitterGK110::emitMOV(const Instruction
*i
)
1461 if (i
->src(0).getFile() == FILE_SYSTEM_VALUE
) {
1462 code
[0] = 0x00000002 | (getSRegEncoding(i
->src(0)) << 23);
1463 code
[1] = 0x86400000;
1465 defId(i
->def(0), 2);
1467 if (i
->src(0).getFile() == FILE_IMMEDIATE
) {
1468 code
[0] = 0x00000002 | (i
->lanes
<< 14);
1469 code
[1] = 0x74000000;
1471 defId(i
->def(0), 2);
1472 setImmediate32(i
, 0, Modifier(0));
1474 if (i
->src(0).getFile() == FILE_PREDICATE
) {
1475 code
[0] = 0x00000002;
1476 code
[1] = 0x84401c07;
1478 defId(i
->def(0), 2);
1479 srcId(i
->src(0), 14);
1481 emitForm_C(i
, 0x24c, 2);
1482 code
[1] |= i
->lanes
<< 10;
1487 CodeEmitterGK110::emitInstruction(Instruction
*insn
)
1489 const unsigned int size
= (writeIssueDelays
&& !(codeSize
& 0x3f)) ? 16 : 8;
1491 if (insn
->encSize
!= 8) {
1492 ERROR("skipping unencodable instruction: ");
1496 if (codeSize
+ size
> codeSizeLimit
) {
1497 ERROR("code emitter output buffer too small\n");
1501 if (writeIssueDelays
) {
1502 int id
= (codeSize
& 0x3f) / 8 - 1;
1505 code
[0] = 0x00000000; // cf issue delay "instruction"
1506 code
[1] = 0x08000000;
1510 uint32_t *data
= code
- (id
* 2 + 2);
1513 case 0: data
[0] |= insn
->sched
<< 2; break;
1514 case 1: data
[0] |= insn
->sched
<< 10; break;
1515 case 2: data
[0] |= insn
->sched
<< 18; break;
1516 case 3: data
[0] |= insn
->sched
<< 26; data
[1] |= insn
->sched
>> 6; break;
1517 case 4: data
[1] |= insn
->sched
<< 2; break;
1518 case 5: data
[1] |= insn
->sched
<< 10; break;
1519 case 6: data
[1] |= insn
->sched
<< 18; break;
1526 // assert that instructions with multiple defs don't corrupt registers
1527 for (int d
= 0; insn
->defExists(d
); ++d
)
1528 assert(insn
->asTex() || insn
->def(d
).rep()->reg
.data
.id
>= 0);
1562 if (isFloatType(insn
->dType
))
1568 if (isFloatType(insn
->dType
))
1575 if (isFloatType(insn
->dType
))
1587 emitLogicOp(insn
, 0);
1590 emitLogicOp(insn
, 1);
1593 emitLogicOp(insn
, 2);
1603 emitSET(insn
->asCmp());
1609 emitSLCT(insn
->asCmp());
1651 emitTEX(insn
->asTex());
1654 emitTXQ(insn
->asTex());
1676 emitQUADOP(insn
, insn
->subOp
, insn
->lanes
);
1679 emitQUADOP(insn
, insn
->src(0).mod
.neg() ? 0x66 : 0x99, 0x4);
1682 emitQUADOP(insn
, insn
->src(0).mod
.neg() ? 0x5a : 0xa5, 0x5);
1694 ERROR("operation should have been eliminated");
1700 ERROR("operation should have been lowered\n");
1703 ERROR("unknow op\n");
1716 CodeEmitterGK110::getMinEncodingSize(const Instruction
*i
) const
1718 // No more short instruction encodings.
1723 CodeEmitterGK110::prepareEmission(Function
*func
)
1725 const Target
*targ
= func
->getProgram()->getTarget();
1727 CodeEmitter::prepareEmission(func
);
1729 if (targ
->hasSWSched
)
1730 calculateSchedDataNVC0(targ
, func
);
1733 CodeEmitterGK110::CodeEmitterGK110(const TargetNVC0
*target
)
1734 : CodeEmitter(target
),
1736 writeIssueDelays(target
->hasSWSched
)
1739 codeSize
= codeSizeLimit
= 0;
1744 TargetNVC0::createCodeEmitterGK110(Program::Type type
)
1746 CodeEmitterGK110
*emit
= new CodeEmitterGK110(this);
1747 emit
->setProgramType(type
);
1751 } // namespace nv50_ir