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 emitPIXLD(const Instruction
*);
118 void emitFlow(const Instruction
*);
120 inline void defId(const ValueDef
&, const int pos
);
121 inline void srcId(const ValueRef
&, const int pos
);
122 inline void srcId(const ValueRef
*, const int pos
);
123 inline void srcId(const Instruction
*, int s
, const int pos
);
125 inline void srcAddr32(const ValueRef
&, const int pos
); // address / 4
127 inline bool isLIMM(const ValueRef
&, DataType ty
, bool mod
= false);
130 #define GK110_GPR_ZERO 255
133 if (i->src(s).mod.neg()) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
135 if (i->src(s).mod.abs()) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
137 #define NOT_(b, s) if (i->src(s).mod & Modifier(NV50_IR_MOD_NOT)) \
138 code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
140 #define FTZ_(b) if (i->ftz) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
142 #define SAT_(b) if (i->saturate) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
144 #define RND_(b, t) emitRoundMode##t(i->rnd, 0x##b)
146 #define SDATA(a) ((a).rep()->reg.data)
147 #define DDATA(a) ((a).rep()->reg.data)
149 void CodeEmitterGK110::srcId(const ValueRef
& src
, const int pos
)
151 code
[pos
/ 32] |= (src
.get() ? SDATA(src
).id
: GK110_GPR_ZERO
) << (pos
% 32);
154 void CodeEmitterGK110::srcId(const ValueRef
*src
, const int pos
)
156 code
[pos
/ 32] |= (src
? SDATA(*src
).id
: GK110_GPR_ZERO
) << (pos
% 32);
159 void CodeEmitterGK110::srcId(const Instruction
*insn
, int s
, int pos
)
161 int r
= insn
->srcExists(s
) ? SDATA(insn
->src(s
)).id
: GK110_GPR_ZERO
;
162 code
[pos
/ 32] |= r
<< (pos
% 32);
165 void CodeEmitterGK110::srcAddr32(const ValueRef
& src
, const int pos
)
167 code
[pos
/ 32] |= (SDATA(src
).offset
>> 2) << (pos
% 32);
170 void CodeEmitterGK110::defId(const ValueDef
& def
, const int pos
)
172 code
[pos
/ 32] |= (def
.get() ? DDATA(def
).id
: GK110_GPR_ZERO
) << (pos
% 32);
175 bool CodeEmitterGK110::isLIMM(const ValueRef
& ref
, DataType ty
, bool mod
)
177 const ImmediateValue
*imm
= ref
.get()->asImm();
179 return imm
&& (imm
->reg
.data
.u32
& ((ty
== TYPE_F32
) ? 0xfff : 0xfff00000));
183 CodeEmitterGK110::emitRoundMode(RoundMode rnd
, const int pos
, const int rintPos
)
189 case ROUND_MI
: rint
= true; /* fall through */ case ROUND_M
: n
= 1; break;
190 case ROUND_PI
: rint
= true; /* fall through */ case ROUND_P
: n
= 2; break;
191 case ROUND_ZI
: rint
= true; /* fall through */ case ROUND_Z
: n
= 3; break;
193 rint
= rnd
== ROUND_NI
;
195 assert(rnd
== ROUND_N
|| rnd
== ROUND_NI
);
198 code
[pos
/ 32] |= n
<< (pos
% 32);
199 if (rint
&& rintPos
>= 0)
200 code
[rintPos
/ 32] |= 1 << (rintPos
% 32);
204 CodeEmitterGK110::emitRoundModeF(RoundMode rnd
, const int pos
)
209 case ROUND_M
: n
= 1; break;
210 case ROUND_P
: n
= 2; break;
211 case ROUND_Z
: n
= 3; break;
214 assert(rnd
== ROUND_N
);
217 code
[pos
/ 32] |= n
<< (pos
% 32);
221 CodeEmitterGK110::emitRoundModeI(RoundMode rnd
, const int pos
)
226 case ROUND_MI
: n
= 1; break;
227 case ROUND_PI
: n
= 2; break;
228 case ROUND_ZI
: n
= 3; break;
231 assert(rnd
== ROUND_NI
);
234 code
[pos
/ 32] |= n
<< (pos
% 32);
237 void CodeEmitterGK110::emitCondCode(CondCode cc
, int pos
, uint8_t mask
)
242 case CC_FL
: n
= 0x00; break;
243 case CC_LT
: n
= 0x01; break;
244 case CC_EQ
: n
= 0x02; break;
245 case CC_LE
: n
= 0x03; break;
246 case CC_GT
: n
= 0x04; break;
247 case CC_NE
: n
= 0x05; break;
248 case CC_GE
: n
= 0x06; break;
249 case CC_LTU
: n
= 0x09; break;
250 case CC_EQU
: n
= 0x0a; break;
251 case CC_LEU
: n
= 0x0b; break;
252 case CC_GTU
: n
= 0x0c; break;
253 case CC_NEU
: n
= 0x0d; break;
254 case CC_GEU
: n
= 0x0e; break;
255 case CC_TR
: n
= 0x0f; break;
256 case CC_NO
: n
= 0x10; break;
257 case CC_NC
: n
= 0x11; break;
258 case CC_NS
: n
= 0x12; break;
259 case CC_NA
: n
= 0x13; break;
260 case CC_A
: n
= 0x14; break;
261 case CC_S
: n
= 0x15; break;
262 case CC_C
: n
= 0x16; break;
263 case CC_O
: n
= 0x17; break;
266 assert(!"invalid condition code");
269 code
[pos
/ 32] |= (n
& mask
) << (pos
% 32);
273 CodeEmitterGK110::emitPredicate(const Instruction
*i
)
275 if (i
->predSrc
>= 0) {
276 srcId(i
->src(i
->predSrc
), 18);
277 if (i
->cc
== CC_NOT_P
)
278 code
[0] |= 8 << 18; // negate
279 assert(i
->getPredicate()->reg
.file
== FILE_PREDICATE
);
286 CodeEmitterGK110::setCAddress14(const ValueRef
& src
)
288 const int32_t addr
= src
.get()->asSym()->reg
.data
.offset
/ 4;
290 code
[0] |= (addr
& 0x01ff) << 23;
291 code
[1] |= (addr
& 0x3e00) >> 9;
295 CodeEmitterGK110::setShortImmediate(const Instruction
*i
, const int s
)
297 const uint32_t u32
= i
->getSrc(s
)->asImm()->reg
.data
.u32
;
298 const uint64_t u64
= i
->getSrc(s
)->asImm()->reg
.data
.u64
;
300 if (i
->sType
== TYPE_F32
) {
301 assert(!(u32
& 0x00000fff));
302 code
[0] |= ((u32
& 0x001ff000) >> 12) << 23;
303 code
[1] |= ((u32
& 0x7fe00000) >> 21);
304 code
[1] |= ((u32
& 0x80000000) >> 4);
306 if (i
->sType
== TYPE_F64
) {
307 assert(!(u64
& 0x00000fffffffffffULL
));
308 code
[0] |= ((u64
& 0x001ff00000000000ULL
) >> 44) << 23;
309 code
[1] |= ((u64
& 0x7fe0000000000000ULL
) >> 53);
310 code
[1] |= ((u64
& 0x8000000000000000ULL
) >> 36);
312 assert((u32
& 0xfff00000) == 0 || (u32
& 0xfff00000) == 0xfff00000);
313 code
[0] |= (u32
& 0x001ff) << 23;
314 code
[1] |= (u32
& 0x7fe00) >> 9;
315 code
[1] |= (u32
& 0x80000) << 8;
320 CodeEmitterGK110::setImmediate32(const Instruction
*i
, const int s
,
323 uint32_t u32
= i
->getSrc(s
)->asImm()->reg
.data
.u32
;
326 ImmediateValue
imm(i
->getSrc(s
)->asImm(), i
->sType
);
328 u32
= imm
.reg
.data
.u32
;
331 code
[0] |= u32
<< 23;
336 CodeEmitterGK110::emitForm_L(const Instruction
*i
, uint32_t opc
, uint8_t ctg
,
346 for (int s
= 0; s
< 3 && i
->srcExists(s
); ++s
) {
347 switch (i
->src(s
).getFile()) {
349 srcId(i
->src(s
), s
? 42 : 10);
352 setImmediate32(i
, s
, mod
);
362 CodeEmitterGK110::emitForm_C(const Instruction
*i
, uint32_t opc
, uint8_t ctg
)
371 switch (i
->src(0).getFile()) {
372 case FILE_MEMORY_CONST
:
373 code
[1] |= 0x4 << 28;
374 setCAddress14(i
->src(0));
377 code
[1] |= 0xc << 28;
378 srcId(i
->src(0), 23);
386 // 0x2 for GPR, c[] and 0x1 for short immediate
388 CodeEmitterGK110::emitForm_21(const Instruction
*i
, uint32_t opc2
,
391 const bool imm
= i
->srcExists(1) && i
->src(1).getFile() == FILE_IMMEDIATE
;
394 if (i
->srcExists(2) && i
->src(2).getFile() == FILE_MEMORY_CONST
)
399 code
[1] = opc1
<< 20;
402 code
[1] = (0xc << 28) | (opc2
<< 20);
409 for (int s
= 0; s
< 3 && i
->srcExists(s
); ++s
) {
410 switch (i
->src(s
).getFile()) {
411 case FILE_MEMORY_CONST
:
412 code
[1] &= (s
== 2) ? ~(0x4 << 28) : ~(0x8 << 28);
413 setCAddress14(i
->src(s
));
414 code
[1] |= i
->getSrc(s
)->reg
.fileIndex
<< 5;
417 setShortImmediate(i
, s
);
420 srcId(i
->src(s
), s
? ((s
== 2) ? 42 : s1
) : 10);
423 // ignore here, can be predicate or flags, but must not be address
431 assert(imm
|| (code
[1] & (0xc << 28)));
435 CodeEmitterGK110::modNegAbsF32_3b(const Instruction
*i
, const int s
)
437 if (i
->src(s
).mod
.abs()) code
[1] &= ~(1 << 27);
438 if (i
->src(s
).mod
.neg()) code
[1] ^= (1 << 27);
442 CodeEmitterGK110::emitNOP(const Instruction
*i
)
444 code
[0] = 0x00003c02;
445 code
[1] = 0x85800000;
450 code
[0] = 0x001c3c02;
454 CodeEmitterGK110::emitFMAD(const Instruction
*i
)
456 assert(!isLIMM(i
->src(1), TYPE_F32
));
458 emitForm_21(i
, 0x0c0, 0x940);
465 bool neg1
= (i
->src(0).mod
^ i
->src(1).mod
).neg();
477 CodeEmitterGK110::emitFMUL(const Instruction
*i
)
479 bool neg
= (i
->src(0).mod
^ i
->src(1).mod
).neg();
481 assert(i
->postFactor
>= -3 && i
->postFactor
<= 3);
483 if (isLIMM(i
->src(1), TYPE_F32
)) {
484 emitForm_L(i
, 0x200, 0x2, Modifier(0));
491 assert(i
->postFactor
== 0);
493 emitForm_21(i
, 0x234, 0xc34);
494 code
[1] |= ((i
->postFactor
> 0) ?
495 (7 - i
->postFactor
) : (0 - i
->postFactor
)) << 12;
512 CodeEmitterGK110::emitIMUL(const Instruction
*i
)
514 assert(!i
->src(0).mod
.neg() && !i
->src(1).mod
.neg());
515 assert(!i
->src(0).mod
.abs() && !i
->src(1).mod
.abs());
517 if (isLIMM(i
->src(1), TYPE_S32
)) {
518 emitForm_L(i
, 0x280, 2, Modifier(0));
520 assert(i
->subOp
!= NV50_IR_SUBOP_MUL_HIGH
);
522 if (i
->sType
== TYPE_S32
)
525 emitForm_21(i
, 0x21c, 0xc1c);
527 if (i
->subOp
== NV50_IR_SUBOP_MUL_HIGH
)
529 if (i
->sType
== TYPE_S32
)
535 CodeEmitterGK110::emitFADD(const Instruction
*i
)
537 if (isLIMM(i
->src(1), TYPE_F32
)) {
538 assert(i
->rnd
== ROUND_N
);
539 assert(!i
->saturate
);
541 Modifier mod
= i
->src(1).mod
^
542 Modifier(i
->op
== OP_SUB
? NV50_IR_MOD_NEG
: 0);
544 emitForm_L(i
, 0x400, 0, mod
);
550 emitForm_21(i
, 0x22c, 0xc2c);
558 modNegAbsF32_3b(i
, 1);
559 if (i
->op
== OP_SUB
) code
[1] ^= 1 << 27;
563 if (i
->op
== OP_SUB
) code
[1] ^= 1 << 16;
569 CodeEmitterGK110::emitUADD(const Instruction
*i
)
571 uint8_t addOp
= (i
->src(0).mod
.neg() << 1) | i
->src(1).mod
.neg();
576 assert(!i
->src(0).mod
.abs() && !i
->src(1).mod
.abs());
578 if (isLIMM(i
->src(1), TYPE_S32
)) {
579 emitForm_L(i
, 0x400, 1, Modifier((addOp
& 1) ? NV50_IR_MOD_NEG
: 0));
584 assert(!i
->defExists(1));
585 assert(i
->flagsSrc
< 0);
589 emitForm_21(i
, 0x208, 0xc08);
591 assert(addOp
!= 3); // would be add-plus-one
593 code
[1] |= addOp
<< 19;
596 code
[1] |= 1 << 18; // write carry
597 if (i
->flagsSrc
>= 0)
598 code
[1] |= 1 << 14; // add carry
606 CodeEmitterGK110::emitIMAD(const Instruction
*i
)
609 (i
->src(2).mod
.neg() << 1) | (i
->src(0).mod
.neg() ^ i
->src(1).mod
.neg());
611 emitForm_21(i
, 0x100, 0xa00);
614 code
[1] |= addOp
<< 26;
616 if (i
->sType
== TYPE_S32
)
617 code
[1] |= (1 << 19) | (1 << 24);
623 if (i
->subOp
== NV50_IR_SUBOP_MUL_HIGH
)
630 CodeEmitterGK110::emitISAD(const Instruction
*i
)
632 assert(i
->dType
== TYPE_S32
|| i
->dType
== TYPE_U32
);
634 emitForm_21(i
, 0x1fc, 0xb74);
636 if (i
->dType
== TYPE_S32
)
641 CodeEmitterGK110::emitNOT(const Instruction
*i
)
643 code
[0] = 0x0003fc02; // logop(mov2) dst, 0, not src
644 code
[1] = 0x22003800;
650 switch (i
->src(0).getFile()) {
652 code
[1] |= 0xc << 28;
653 srcId(i
->src(0), 23);
655 case FILE_MEMORY_CONST
:
656 code
[1] |= 0x4 << 28;
657 setCAddress14(i
->src(1));
666 CodeEmitterGK110::emitLogicOp(const Instruction
*i
, uint8_t subOp
)
668 if (isLIMM(i
->src(1), TYPE_S32
)) {
669 emitForm_L(i
, 0x200, 0, i
->src(1).mod
);
670 code
[1] |= subOp
<< 24;
673 emitForm_21(i
, 0x220, 0xc20);
674 code
[1] |= subOp
<< 12;
681 CodeEmitterGK110::emitPOPC(const Instruction
*i
)
683 assert(!isLIMM(i
->src(1), TYPE_S32
, true));
685 emitForm_21(i
, 0x204, 0xc04);
688 if (!(code
[0] & 0x1))
693 CodeEmitterGK110::emitINSBF(const Instruction
*i
)
695 emitForm_21(i
, 0x1f8, 0xb78);
699 CodeEmitterGK110::emitShift(const Instruction
*i
)
701 if (i
->op
== OP_SHR
) {
702 emitForm_21(i
, 0x214, 0xc14);
703 if (isSignedType(i
->dType
))
706 emitForm_21(i
, 0x224, 0xc24);
709 if (i
->subOp
== NV50_IR_SUBOP_SHIFT_WRAP
)
714 CodeEmitterGK110::emitPreOp(const Instruction
*i
)
716 emitForm_C(i
, 0x248, 0x2);
718 if (i
->op
== OP_PREEX2
)
726 CodeEmitterGK110::emitSFnOp(const Instruction
*i
, uint8_t subOp
)
728 code
[0] = 0x00000002 | (subOp
<< 23);
729 code
[1] = 0x84000000;
734 srcId(i
->src(0), 10);
742 CodeEmitterGK110::emitMINMAX(const Instruction
*i
)
766 emitForm_21(i
, op2
, op1
);
768 if (i
->dType
== TYPE_S32
)
770 code
[1] |= (i
->op
== OP_MIN
) ? 0x1c00 : 0x3c00; // [!]pt
776 modNegAbsF32_3b(i
, 1);
784 CodeEmitterGK110::emitCVT(const Instruction
*i
)
786 const bool f2f
= isFloatType(i
->dType
) && isFloatType(i
->sType
);
787 const bool f2i
= !isFloatType(i
->dType
) && isFloatType(i
->sType
);
788 const bool i2f
= isFloatType(i
->dType
) && !isFloatType(i
->sType
);
790 bool sat
= i
->saturate
;
791 bool abs
= i
->src(0).mod
.abs();
792 bool neg
= i
->src(0).mod
.neg();
794 RoundMode rnd
= i
->rnd
;
797 case OP_CEIL
: rnd
= f2f
? ROUND_PI
: ROUND_P
; break;
798 case OP_FLOOR
: rnd
= f2f
? ROUND_MI
: ROUND_M
; break;
799 case OP_TRUNC
: rnd
= f2f
? ROUND_ZI
: ROUND_Z
; break;
800 case OP_SAT
: sat
= true; break;
801 case OP_NEG
: neg
= !neg
; break;
802 case OP_ABS
: abs
= true; neg
= false; break;
809 if (i
->op
== OP_NEG
&& i
->dType
== TYPE_U32
)
818 else if (f2i
) op
= 0x258;
819 else if (i2f
) op
= 0x25c;
822 emitForm_C(i
, op
, 0x2);
825 if (neg
) code
[1] |= 1 << 16;
826 if (abs
) code
[1] |= 1 << 20;
827 if (sat
) code
[1] |= 1 << 21;
829 emitRoundMode(rnd
, 32 + 10, f2f
? (32 + 13) : -1);
831 code
[0] |= typeSizeofLog2(dType
) << 10;
832 code
[0] |= typeSizeofLog2(i
->sType
) << 12;
834 if (isSignedIntType(dType
))
836 if (isSignedIntType(i
->sType
))
841 CodeEmitterGK110::emitSET(const CmpInstruction
*i
)
845 if (i
->def(0).getFile() == FILE_PREDICATE
) {
847 case TYPE_F32
: op2
= 0x1d8; op1
= 0xb58; break;
848 case TYPE_F64
: op2
= 0x1c0; op1
= 0xb40; break;
854 emitForm_21(i
, op2
, op1
);
858 if (!(code
[0] & 0x1)) {
862 modNegAbsF32_3b(i
, 1);
866 // normal DST field is negated predicate result
867 code
[0] = (code
[0] & ~0xfc) | ((code
[0] << 3) & 0xe0);
874 case TYPE_F32
: op2
= 0x000; op1
= 0x820; break;
875 case TYPE_F64
: op2
= 0x080; op1
= 0x900; break;
881 emitForm_21(i
, op2
, op1
);
885 if (!(code
[0] & 0x1)) {
889 modNegAbsF32_3b(i
, 1);
893 if (i
->sType
== TYPE_S32
)
896 if (i
->op
!= OP_SET
) {
898 case OP_SET_AND
: code
[1] |= 0x0 << 16; break;
899 case OP_SET_OR
: code
[1] |= 0x1 << 16; break;
900 case OP_SET_XOR
: code
[1] |= 0x2 << 16; break;
905 srcId(i
->src(2), 0x2a);
907 code
[1] |= 0x7 << 10;
909 emitCondCode(i
->setCond
,
910 isFloatType(i
->sType
) ? 0x33 : 0x34,
911 isFloatType(i
->sType
) ? 0xf : 0x7);
915 CodeEmitterGK110::emitSLCT(const CmpInstruction
*i
)
917 CondCode cc
= i
->setCond
;
918 if (i
->src(2).mod
.neg())
919 cc
= reverseCondCode(cc
);
921 if (i
->dType
== TYPE_F32
) {
922 emitForm_21(i
, 0x1d0, 0xb50);
924 emitCondCode(cc
, 0x33, 0xf);
926 emitForm_21(i
, 0x1a4, 0xb20);
927 emitCondCode(cc
, 0x34, 0x7);
931 void CodeEmitterGK110::emitSELP(const Instruction
*i
)
933 emitForm_21(i
, 0x250, 0x050);
935 if ((i
->cc
== CC_NOT_P
) ^ (bool)(i
->src(2).mod
& Modifier(NV50_IR_MOD_NOT
)))
939 void CodeEmitterGK110::emitTEXBAR(const Instruction
*i
)
941 code
[0] = 0x00000002 | (i
->subOp
<< 23);
942 code
[1] = 0x77000000;
947 void CodeEmitterGK110::emitTEXCSAA(const TexInstruction
*i
)
949 code
[0] = 0x00000002;
950 code
[1] = 0x76c00000;
952 code
[1] |= i
->tex
.r
<< 9;
953 // code[1] |= i->tex.s << (9 + 8);
956 code
[0] |= 0x80000000;
959 srcId(i
->src(0), 10);
963 isNextIndependentTex(const TexInstruction
*i
)
965 if (!i
->next
|| !isTextureOp(i
->next
->op
))
967 if (i
->getDef(0)->interfers(i
->next
->getSrc(0)))
969 return !i
->next
->srcExists(1) || !i
->getDef(0)->interfers(i
->next
->getSrc(1));
973 CodeEmitterGK110::emitTEX(const TexInstruction
*i
)
975 const bool ind
= i
->tex
.rIndirectSrc
>= 0;
978 code
[0] = 0x00000002;
981 code
[1] = 0x7e000000;
984 code
[1] = 0x7e800000;
987 code
[1] = 0x78000000;
990 code
[1] = 0x7dc00000;
993 code
[1] = 0x7d800000;
999 code
[0] = 0x00000002;
1000 code
[1] = 0x76000000;
1001 code
[1] |= i
->tex
.r
<< 9;
1004 code
[0] = 0x00000002;
1005 code
[1] = 0x76800000;
1006 code
[1] |= i
->tex
.r
<< 9;
1009 code
[0] = 0x00000002;
1010 code
[1] = 0x70000000;
1011 code
[1] |= i
->tex
.r
<< 13;
1014 code
[0] = 0x00000001;
1015 code
[1] = 0x70000000;
1016 code
[1] |= i
->tex
.r
<< 15;
1019 code
[0] = 0x00000001;
1020 code
[1] = 0x60000000;
1021 code
[1] |= i
->tex
.r
<< 15;
1026 code
[1] |= isNextIndependentTex(i
) ? 0x1 : 0x2; // t : p mode
1028 if (i
->tex
.liveOnly
)
1029 code
[0] |= 0x80000000;
1033 case OP_TXB
: code
[1] |= 0x2000; break;
1034 case OP_TXL
: code
[1] |= 0x3000; break;
1038 case OP_TXLQ
: break;
1040 assert(!"invalid texture op");
1044 if (i
->op
== OP_TXF
) {
1045 if (!i
->tex
.levelZero
)
1048 if (i
->tex
.levelZero
) {
1052 if (i
->op
!= OP_TXD
&& i
->tex
.derivAll
)
1057 code
[1] |= i
->tex
.mask
<< 2;
1059 const int src1
= (i
->predSrc
== 1) ? 2 : 1; // if predSrc == 1, !srcExists(2)
1061 defId(i
->def(0), 2);
1062 srcId(i
->src(0), 10);
1065 if (i
->op
== OP_TXG
) code
[1] |= i
->tex
.gatherComp
<< 13;
1068 code
[1] |= (i
->tex
.target
.isCube() ? 3 : (i
->tex
.target
.getDim() - 1)) << 7;
1069 if (i
->tex
.target
.isArray())
1071 if (i
->tex
.target
.isShadow())
1073 if (i
->tex
.target
== TEX_TARGET_2D_MS
||
1074 i
->tex
.target
== TEX_TARGET_2D_MS_ARRAY
)
1077 if (i
->srcExists(src1
) && i
->src(src1
).getFile() == FILE_IMMEDIATE
) {
1081 if (i
->tex
.useOffsets
) {
1083 case OP_TXF
: code
[1] |= 0x200; break;
1084 default: code
[1] |= 0x800; break;
1090 CodeEmitterGK110::emitTXQ(const TexInstruction
*i
)
1092 code
[0] = 0x00000002;
1093 code
[1] = 0x75400001;
1095 switch (i
->tex
.query
) {
1096 case TXQ_DIMS
: code
[0] |= 0x01 << 25; break;
1097 case TXQ_TYPE
: code
[0] |= 0x02 << 25; break;
1098 case TXQ_SAMPLE_POSITION
: code
[0] |= 0x05 << 25; break;
1099 case TXQ_FILTER
: code
[0] |= 0x10 << 25; break;
1100 case TXQ_LOD
: code
[0] |= 0x12 << 25; break;
1101 case TXQ_BORDER_COLOUR
: code
[0] |= 0x16 << 25; break;
1103 assert(!"invalid texture query");
1107 code
[1] |= i
->tex
.mask
<< 2;
1108 code
[1] |= i
->tex
.r
<< 9;
1109 if (/*i->tex.sIndirectSrc >= 0 || */i
->tex
.rIndirectSrc
>= 0)
1110 code
[1] |= 0x08000000;
1112 defId(i
->def(0), 2);
1113 srcId(i
->src(0), 10);
1119 CodeEmitterGK110::emitQUADOP(const Instruction
*i
, uint8_t qOp
, uint8_t laneMask
)
1121 code
[0] = 0x00000002 | ((qOp
& 1) << 31);
1122 code
[1] = 0x7fc00000 | (qOp
>> 1) | (laneMask
<< 12);
1124 defId(i
->def(0), 2);
1125 srcId(i
->src(0), 10);
1126 srcId(i
->srcExists(1) ? i
->src(1) : i
->src(0), 23);
1128 if (i
->op
== OP_QUADOP
&& progType
!= Program::TYPE_FRAGMENT
)
1129 code
[1] |= 1 << 9; // dall
1135 CodeEmitterGK110::emitPIXLD(const Instruction
*i
)
1137 emitForm_L(i
, 0x7f4, 2, Modifier(0));
1138 code
[1] |= i
->subOp
<< 2;
1139 code
[1] |= 0x00070000;
1143 CodeEmitterGK110::emitFlow(const Instruction
*i
)
1145 const FlowInstruction
*f
= i
->asFlow();
1147 unsigned mask
; // bit 0: predicate, bit 1: target
1149 code
[0] = 0x00000000;
1153 code
[1] = f
->absolute
? 0x10800000 : 0x12000000;
1154 if (i
->srcExists(0) && i
->src(0).getFile() == FILE_MEMORY_CONST
)
1159 code
[1] = f
->absolute
? 0x11000000 : 0x13000000;
1160 if (i
->srcExists(0) && i
->src(0).getFile() == FILE_MEMORY_CONST
)
1165 case OP_EXIT
: code
[1] = 0x18000000; mask
= 1; break;
1166 case OP_RET
: code
[1] = 0x19000000; mask
= 1; break;
1167 case OP_DISCARD
: code
[1] = 0x19800000; mask
= 1; break;
1168 case OP_BREAK
: code
[1] = 0x1a000000; mask
= 1; break;
1169 case OP_CONT
: code
[1] = 0x1a800000; mask
= 1; break;
1171 case OP_JOINAT
: code
[1] = 0x14800000; mask
= 2; break;
1172 case OP_PREBREAK
: code
[1] = 0x15000000; mask
= 2; break;
1173 case OP_PRECONT
: code
[1] = 0x15800000; mask
= 2; break;
1174 case OP_PRERET
: code
[1] = 0x13800000; mask
= 2; break;
1176 case OP_QUADON
: code
[1] = 0x1b000000; mask
= 0; break;
1177 case OP_QUADPOP
: code
[1] = 0x1c000000; mask
= 0; break;
1178 case OP_BRKPT
: code
[1] = 0x00000000; mask
= 0; break;
1180 assert(!"invalid flow operation");
1186 if (i
->flagsSrc
< 0)
1198 if (f
->op
== OP_CALL
) {
1200 assert(f
->absolute
);
1201 uint32_t pcAbs
= targNVC0
->getBuiltinOffset(f
->target
.builtin
);
1202 addReloc(RelocEntry::TYPE_BUILTIN
, 0, pcAbs
, 0xff800000, 23);
1203 addReloc(RelocEntry::TYPE_BUILTIN
, 1, pcAbs
, 0x007fffff, -9);
1205 assert(!f
->absolute
);
1206 int32_t pcRel
= f
->target
.fn
->binPos
- (codeSize
+ 8);
1207 code
[0] |= (pcRel
& 0x1ff) << 23;
1208 code
[1] |= (pcRel
>> 9) & 0x7fff;
1212 int32_t pcRel
= f
->target
.bb
->binPos
- (codeSize
+ 8);
1213 // currently we don't want absolute branches
1214 assert(!f
->absolute
);
1215 code
[0] |= (pcRel
& 0x1ff) << 23;
1216 code
[1] |= (pcRel
>> 9) & 0x7fff;
1221 CodeEmitterGK110::emitPFETCH(const Instruction
*i
)
1223 uint32_t prim
= i
->src(0).get()->reg
.data
.u32
;
1225 code
[0] = 0x00000002 | ((prim
& 0xff) << 23);
1226 code
[1] = 0x7f800000;
1230 defId(i
->def(0), 2);
1231 srcId(i
->src(1), 10);
1235 CodeEmitterGK110::emitVFETCH(const Instruction
*i
)
1237 unsigned int size
= typeSizeof(i
->dType
);
1238 uint32_t offset
= i
->src(0).get()->reg
.data
.offset
;
1240 code
[0] = 0x00000002 | (offset
<< 23);
1241 code
[1] = 0x7ec00000 | (offset
>> 9);
1242 code
[1] |= (size
/ 4 - 1) << 18;
1247 if (i
->getSrc(0)->reg
.file
== FILE_SHADER_OUTPUT
)
1248 code
[0] |= 0x200; // yes, TCPs can read from *outputs* of other threads
1253 defId(i
->def(0), 2);
1254 srcId(i
->src(0).getIndirect(0), 10);
1255 srcId(i
->src(0).getIndirect(1), 32 + 10); // vertex address
1259 CodeEmitterGK110::emitEXPORT(const Instruction
*i
)
1261 unsigned int size
= typeSizeof(i
->dType
);
1262 uint32_t offset
= i
->src(0).get()->reg
.data
.offset
;
1264 code
[0] = 0x00000002 | (offset
<< 23);
1265 code
[1] = 0x7f000000 | (offset
>> 9);
1266 code
[1] |= (size
/ 4 - 1) << 18;
1275 assert(i
->src(1).getFile() == FILE_GPR
);
1277 srcId(i
->src(0).getIndirect(0), 10);
1278 srcId(i
->src(0).getIndirect(1), 32 + 10); // vertex base address
1279 srcId(i
->src(1), 2);
1283 CodeEmitterGK110::emitOUT(const Instruction
*i
)
1285 assert(i
->src(0).getFile() == FILE_GPR
);
1287 emitForm_21(i
, 0x1f0, 0xb70);
1289 if (i
->op
== OP_EMIT
)
1291 if (i
->op
== OP_RESTART
|| i
->subOp
== NV50_IR_SUBOP_EMIT_RESTART
)
1296 CodeEmitterGK110::emitInterpMode(const Instruction
*i
)
1298 code
[1] |= i
->ipa
<< 21; // TODO: INTERP_SAMPLEID
1302 CodeEmitterGK110::emitINTERP(const Instruction
*i
)
1304 const uint32_t base
= i
->getSrc(0)->reg
.data
.offset
;
1306 code
[0] = 0x00000002 | (base
<< 31);
1307 code
[1] = 0x74800000 | (base
>> 1);
1312 if (i
->op
== OP_PINTERP
)
1313 srcId(i
->src(1), 23);
1315 code
[0] |= 0xff << 23;
1317 srcId(i
->src(0).getIndirect(0), 10);
1321 defId(i
->def(0), 2);
1323 if (i
->getSampleMode() == NV50_IR_INTERP_OFFSET
)
1324 srcId(i
->src(i
->op
== OP_PINTERP
? 2 : 1), 32 + 10);
1326 code
[1] |= 0xff << 10;
1330 CodeEmitterGK110::emitLoadStoreType(DataType ty
, const int pos
)
1362 assert(!"invalid ld/st type");
1365 code
[pos
/ 32] |= n
<< (pos
% 32);
1369 CodeEmitterGK110::emitCachingMode(CacheMode c
, const int pos
)
1390 assert(!"invalid caching mode");
1393 code
[pos
/ 32] |= n
<< (pos
% 32);
1397 CodeEmitterGK110::emitSTORE(const Instruction
*i
)
1399 int32_t offset
= SDATA(i
->src(0)).offset
;
1401 switch (i
->src(0).getFile()) {
1402 case FILE_MEMORY_GLOBAL
: code
[1] = 0xe0000000; code
[0] = 0x00000000; break;
1403 case FILE_MEMORY_LOCAL
: code
[1] = 0x7a800000; code
[0] = 0x00000002; break;
1404 case FILE_MEMORY_SHARED
: code
[1] = 0x7ac00000; code
[0] = 0x00000002; break;
1406 assert(!"invalid memory file");
1410 if (i
->src(0).getFile() != FILE_MEMORY_GLOBAL
)
1413 if (code
[0] & 0x2) {
1414 emitLoadStoreType(i
->dType
, 0x33);
1415 if (i
->src(0).getFile() == FILE_MEMORY_LOCAL
)
1416 emitCachingMode(i
->cache
, 0x2f);
1418 emitLoadStoreType(i
->dType
, 0x38);
1419 emitCachingMode(i
->cache
, 0x3b);
1421 code
[0] |= offset
<< 23;
1422 code
[1] |= offset
>> 9;
1426 srcId(i
->src(1), 2);
1427 srcId(i
->src(0).getIndirect(0), 10);
1431 CodeEmitterGK110::emitLOAD(const Instruction
*i
)
1433 int32_t offset
= SDATA(i
->src(0)).offset
;
1435 switch (i
->src(0).getFile()) {
1436 case FILE_MEMORY_GLOBAL
: code
[1] = 0xc0000000; code
[0] = 0x00000000; break;
1437 case FILE_MEMORY_LOCAL
: code
[1] = 0x7a000000; code
[0] = 0x00000002; break;
1438 case FILE_MEMORY_SHARED
: code
[1] = 0x7ac00000; code
[0] = 0x00000002; break;
1439 case FILE_MEMORY_CONST
:
1440 if (!i
->src(0).isIndirect(0) && typeSizeof(i
->dType
) == 4) {
1445 code
[0] = 0x00000002;
1446 code
[1] = 0x7c800000 | (i
->src(0).get()->reg
.fileIndex
<< 7);
1449 assert(!"invalid memory file");
1453 if (code
[0] & 0x2) {
1455 emitLoadStoreType(i
->dType
, 0x33);
1456 if (i
->src(0).getFile() == FILE_MEMORY_LOCAL
)
1457 emitCachingMode(i
->cache
, 0x2f);
1459 emitLoadStoreType(i
->dType
, 0x38);
1460 emitCachingMode(i
->cache
, 0x3b);
1462 code
[0] |= offset
<< 23;
1463 code
[1] |= offset
>> 9;
1467 defId(i
->def(0), 2);
1468 srcId(i
->src(0).getIndirect(0), 10);
1472 CodeEmitterGK110::getSRegEncoding(const ValueRef
& ref
)
1474 switch (SDATA(ref
).sv
.sv
) {
1475 case SV_LANEID
: return 0x00;
1476 case SV_PHYSID
: return 0x03;
1477 case SV_VERTEX_COUNT
: return 0x10;
1478 case SV_INVOCATION_ID
: return 0x11;
1479 case SV_YDIR
: return 0x12;
1480 case SV_TID
: return 0x21 + SDATA(ref
).sv
.index
;
1481 case SV_CTAID
: return 0x25 + SDATA(ref
).sv
.index
;
1482 case SV_NTID
: return 0x29 + SDATA(ref
).sv
.index
;
1483 case SV_GRIDID
: return 0x2c;
1484 case SV_NCTAID
: return 0x2d + SDATA(ref
).sv
.index
;
1485 case SV_LBASE
: return 0x34;
1486 case SV_SBASE
: return 0x30;
1487 case SV_CLOCK
: return 0x50 + SDATA(ref
).sv
.index
;
1489 assert(!"no sreg for system value");
1495 CodeEmitterGK110::emitMOV(const Instruction
*i
)
1497 if (i
->src(0).getFile() == FILE_SYSTEM_VALUE
) {
1498 code
[0] = 0x00000002 | (getSRegEncoding(i
->src(0)) << 23);
1499 code
[1] = 0x86400000;
1501 defId(i
->def(0), 2);
1503 if (i
->src(0).getFile() == FILE_IMMEDIATE
) {
1504 code
[0] = 0x00000002 | (i
->lanes
<< 14);
1505 code
[1] = 0x74000000;
1507 defId(i
->def(0), 2);
1508 setImmediate32(i
, 0, Modifier(0));
1510 if (i
->src(0).getFile() == FILE_PREDICATE
) {
1511 code
[0] = 0x00000002;
1512 code
[1] = 0x84401c07;
1514 defId(i
->def(0), 2);
1515 srcId(i
->src(0), 14);
1517 emitForm_C(i
, 0x24c, 2);
1518 code
[1] |= i
->lanes
<< 10;
1523 CodeEmitterGK110::emitInstruction(Instruction
*insn
)
1525 const unsigned int size
= (writeIssueDelays
&& !(codeSize
& 0x3f)) ? 16 : 8;
1527 if (insn
->encSize
!= 8) {
1528 ERROR("skipping unencodable instruction: ");
1532 if (codeSize
+ size
> codeSizeLimit
) {
1533 ERROR("code emitter output buffer too small\n");
1537 if (writeIssueDelays
) {
1538 int id
= (codeSize
& 0x3f) / 8 - 1;
1541 code
[0] = 0x00000000; // cf issue delay "instruction"
1542 code
[1] = 0x08000000;
1546 uint32_t *data
= code
- (id
* 2 + 2);
1549 case 0: data
[0] |= insn
->sched
<< 2; break;
1550 case 1: data
[0] |= insn
->sched
<< 10; break;
1551 case 2: data
[0] |= insn
->sched
<< 18; break;
1552 case 3: data
[0] |= insn
->sched
<< 26; data
[1] |= insn
->sched
>> 6; break;
1553 case 4: data
[1] |= insn
->sched
<< 2; break;
1554 case 5: data
[1] |= insn
->sched
<< 10; break;
1555 case 6: data
[1] |= insn
->sched
<< 18; break;
1562 // assert that instructions with multiple defs don't corrupt registers
1563 for (int d
= 0; insn
->defExists(d
); ++d
)
1564 assert(insn
->asTex() || insn
->def(d
).rep()->reg
.data
.id
>= 0);
1598 if (isFloatType(insn
->dType
))
1604 if (isFloatType(insn
->dType
))
1611 if (isFloatType(insn
->dType
))
1623 emitLogicOp(insn
, 0);
1626 emitLogicOp(insn
, 1);
1629 emitLogicOp(insn
, 2);
1639 emitSET(insn
->asCmp());
1645 emitSLCT(insn
->asCmp());
1689 emitTEX(insn
->asTex());
1692 emitTXQ(insn
->asTex());
1717 emitQUADOP(insn
, insn
->subOp
, insn
->lanes
);
1720 emitQUADOP(insn
, insn
->src(0).mod
.neg() ? 0x66 : 0x99, 0x4);
1723 emitQUADOP(insn
, insn
->src(0).mod
.neg() ? 0x5a : 0xa5, 0x5);
1735 ERROR("operation should have been eliminated");
1741 ERROR("operation should have been lowered\n");
1744 ERROR("unknow op\n");
1757 CodeEmitterGK110::getMinEncodingSize(const Instruction
*i
) const
1759 // No more short instruction encodings.
1764 CodeEmitterGK110::prepareEmission(Function
*func
)
1766 const Target
*targ
= func
->getProgram()->getTarget();
1768 CodeEmitter::prepareEmission(func
);
1770 if (targ
->hasSWSched
)
1771 calculateSchedDataNVC0(targ
, func
);
1774 CodeEmitterGK110::CodeEmitterGK110(const TargetNVC0
*target
)
1775 : CodeEmitter(target
),
1777 writeIssueDelays(target
->hasSWSched
)
1780 codeSize
= codeSizeLimit
= 0;
1785 TargetNVC0::createCodeEmitterGK110(Program::Type type
)
1787 CodeEmitterGK110
*emit
= new CodeEmitterGK110(this);
1788 emit
->setProgramType(type
);
1792 } // namespace nv50_ir