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
*);
78 void emitATOM(const Instruction
*);
79 void emitCCTL(const Instruction
*);
81 void emitINTERP(const Instruction
*);
82 void emitAFETCH(const Instruction
*);
83 void emitPFETCH(const Instruction
*);
84 void emitVFETCH(const Instruction
*);
85 void emitEXPORT(const Instruction
*);
86 void emitOUT(const Instruction
*);
88 void emitUADD(const Instruction
*);
89 void emitFADD(const Instruction
*);
90 void emitDADD(const Instruction
*);
91 void emitIMUL(const Instruction
*);
92 void emitFMUL(const Instruction
*);
93 void emitDMUL(const Instruction
*);
94 void emitIMAD(const Instruction
*);
95 void emitISAD(const Instruction
*);
96 void emitFMAD(const Instruction
*);
97 void emitDMAD(const Instruction
*);
99 void emitNOT(const Instruction
*);
100 void emitLogicOp(const Instruction
*, uint8_t subOp
);
101 void emitPOPC(const Instruction
*);
102 void emitINSBF(const Instruction
*);
103 void emitEXTBF(const Instruction
*);
104 void emitBFIND(const Instruction
*);
105 void emitShift(const Instruction
*);
107 void emitSFnOp(const Instruction
*, uint8_t subOp
);
109 void emitCVT(const Instruction
*);
110 void emitMINMAX(const Instruction
*);
111 void emitPreOp(const Instruction
*);
113 void emitSET(const CmpInstruction
*);
114 void emitSLCT(const CmpInstruction
*);
115 void emitSELP(const Instruction
*);
117 void emitTEXBAR(const Instruction
*);
118 void emitTEX(const TexInstruction
*);
119 void emitTEXCSAA(const TexInstruction
*);
120 void emitTXQ(const TexInstruction
*);
122 void emitQUADOP(const Instruction
*, uint8_t qOp
, uint8_t laneMask
);
124 void emitPIXLD(const Instruction
*);
126 void emitBAR(const Instruction
*);
127 void emitMEMBAR(const Instruction
*);
129 void emitFlow(const Instruction
*);
131 inline void defId(const ValueDef
&, const int pos
);
132 inline void srcId(const ValueRef
&, const int pos
);
133 inline void srcId(const ValueRef
*, const int pos
);
134 inline void srcId(const Instruction
*, int s
, const int pos
);
136 inline void srcAddr32(const ValueRef
&, const int pos
); // address / 4
138 inline bool isLIMM(const ValueRef
&, DataType ty
, bool mod
= false);
141 #define GK110_GPR_ZERO 255
144 if (i->src(s).mod.neg()) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
146 if (i->src(s).mod.abs()) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
148 #define NOT_(b, s) if (i->src(s).mod & Modifier(NV50_IR_MOD_NOT)) \
149 code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
151 #define FTZ_(b) if (i->ftz) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
152 #define DNZ_(b) if (i->dnz) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
154 #define SAT_(b) if (i->saturate) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
156 #define RND_(b, t) emitRoundMode##t(i->rnd, 0x##b)
158 #define SDATA(a) ((a).rep()->reg.data)
159 #define DDATA(a) ((a).rep()->reg.data)
161 void CodeEmitterGK110::srcId(const ValueRef
& src
, const int pos
)
163 code
[pos
/ 32] |= (src
.get() ? SDATA(src
).id
: GK110_GPR_ZERO
) << (pos
% 32);
166 void CodeEmitterGK110::srcId(const ValueRef
*src
, const int pos
)
168 code
[pos
/ 32] |= (src
? SDATA(*src
).id
: GK110_GPR_ZERO
) << (pos
% 32);
171 void CodeEmitterGK110::srcId(const Instruction
*insn
, int s
, int pos
)
173 int r
= insn
->srcExists(s
) ? SDATA(insn
->src(s
)).id
: GK110_GPR_ZERO
;
174 code
[pos
/ 32] |= r
<< (pos
% 32);
177 void CodeEmitterGK110::srcAddr32(const ValueRef
& src
, const int pos
)
179 code
[pos
/ 32] |= (SDATA(src
).offset
>> 2) << (pos
% 32);
182 void CodeEmitterGK110::defId(const ValueDef
& def
, const int pos
)
184 code
[pos
/ 32] |= (def
.get() ? DDATA(def
).id
: GK110_GPR_ZERO
) << (pos
% 32);
187 bool CodeEmitterGK110::isLIMM(const ValueRef
& ref
, DataType ty
, bool mod
)
189 const ImmediateValue
*imm
= ref
.get()->asImm();
191 return imm
&& (imm
->reg
.data
.u32
& ((ty
== TYPE_F32
) ? 0xfff : 0xfff00000));
195 CodeEmitterGK110::emitRoundMode(RoundMode rnd
, const int pos
, const int rintPos
)
201 case ROUND_MI
: rint
= true; /* fall through */ case ROUND_M
: n
= 1; break;
202 case ROUND_PI
: rint
= true; /* fall through */ case ROUND_P
: n
= 2; break;
203 case ROUND_ZI
: rint
= true; /* fall through */ case ROUND_Z
: n
= 3; break;
205 rint
= rnd
== ROUND_NI
;
207 assert(rnd
== ROUND_N
|| rnd
== ROUND_NI
);
210 code
[pos
/ 32] |= n
<< (pos
% 32);
211 if (rint
&& rintPos
>= 0)
212 code
[rintPos
/ 32] |= 1 << (rintPos
% 32);
216 CodeEmitterGK110::emitRoundModeF(RoundMode rnd
, const int pos
)
221 case ROUND_M
: n
= 1; break;
222 case ROUND_P
: n
= 2; break;
223 case ROUND_Z
: n
= 3; break;
226 assert(rnd
== ROUND_N
);
229 code
[pos
/ 32] |= n
<< (pos
% 32);
233 CodeEmitterGK110::emitRoundModeI(RoundMode rnd
, const int pos
)
238 case ROUND_MI
: n
= 1; break;
239 case ROUND_PI
: n
= 2; break;
240 case ROUND_ZI
: n
= 3; break;
243 assert(rnd
== ROUND_NI
);
246 code
[pos
/ 32] |= n
<< (pos
% 32);
249 void CodeEmitterGK110::emitCondCode(CondCode cc
, int pos
, uint8_t mask
)
254 case CC_FL
: n
= 0x00; break;
255 case CC_LT
: n
= 0x01; break;
256 case CC_EQ
: n
= 0x02; break;
257 case CC_LE
: n
= 0x03; break;
258 case CC_GT
: n
= 0x04; break;
259 case CC_NE
: n
= 0x05; break;
260 case CC_GE
: n
= 0x06; break;
261 case CC_LTU
: n
= 0x09; break;
262 case CC_EQU
: n
= 0x0a; break;
263 case CC_LEU
: n
= 0x0b; break;
264 case CC_GTU
: n
= 0x0c; break;
265 case CC_NEU
: n
= 0x0d; break;
266 case CC_GEU
: n
= 0x0e; break;
267 case CC_TR
: n
= 0x0f; break;
268 case CC_NO
: n
= 0x10; break;
269 case CC_NC
: n
= 0x11; break;
270 case CC_NS
: n
= 0x12; break;
271 case CC_NA
: n
= 0x13; break;
272 case CC_A
: n
= 0x14; break;
273 case CC_S
: n
= 0x15; break;
274 case CC_C
: n
= 0x16; break;
275 case CC_O
: n
= 0x17; break;
278 assert(!"invalid condition code");
281 code
[pos
/ 32] |= (n
& mask
) << (pos
% 32);
285 CodeEmitterGK110::emitPredicate(const Instruction
*i
)
287 if (i
->predSrc
>= 0) {
288 srcId(i
->src(i
->predSrc
), 18);
289 if (i
->cc
== CC_NOT_P
)
290 code
[0] |= 8 << 18; // negate
291 assert(i
->getPredicate()->reg
.file
== FILE_PREDICATE
);
298 CodeEmitterGK110::setCAddress14(const ValueRef
& src
)
300 const Storage
& res
= src
.get()->asSym()->reg
;
301 const int32_t addr
= res
.data
.offset
/ 4;
303 code
[0] |= (addr
& 0x01ff) << 23;
304 code
[1] |= (addr
& 0x3e00) >> 9;
305 code
[1] |= res
.fileIndex
<< 5;
309 CodeEmitterGK110::setShortImmediate(const Instruction
*i
, const int s
)
311 const uint32_t u32
= i
->getSrc(s
)->asImm()->reg
.data
.u32
;
312 const uint64_t u64
= i
->getSrc(s
)->asImm()->reg
.data
.u64
;
314 if (i
->sType
== TYPE_F32
) {
315 assert(!(u32
& 0x00000fff));
316 code
[0] |= ((u32
& 0x001ff000) >> 12) << 23;
317 code
[1] |= ((u32
& 0x7fe00000) >> 21);
318 code
[1] |= ((u32
& 0x80000000) >> 4);
320 if (i
->sType
== TYPE_F64
) {
321 assert(!(u64
& 0x00000fffffffffffULL
));
322 code
[0] |= ((u64
& 0x001ff00000000000ULL
) >> 44) << 23;
323 code
[1] |= ((u64
& 0x7fe0000000000000ULL
) >> 53);
324 code
[1] |= ((u64
& 0x8000000000000000ULL
) >> 36);
326 assert((u32
& 0xfff00000) == 0 || (u32
& 0xfff00000) == 0xfff00000);
327 code
[0] |= (u32
& 0x001ff) << 23;
328 code
[1] |= (u32
& 0x7fe00) >> 9;
329 code
[1] |= (u32
& 0x80000) << 8;
334 CodeEmitterGK110::setImmediate32(const Instruction
*i
, const int s
,
337 uint32_t u32
= i
->getSrc(s
)->asImm()->reg
.data
.u32
;
340 ImmediateValue
imm(i
->getSrc(s
)->asImm(), i
->sType
);
342 u32
= imm
.reg
.data
.u32
;
345 code
[0] |= u32
<< 23;
350 CodeEmitterGK110::emitForm_L(const Instruction
*i
, uint32_t opc
, uint8_t ctg
,
360 for (int s
= 0; s
< 3 && i
->srcExists(s
); ++s
) {
361 switch (i
->src(s
).getFile()) {
363 srcId(i
->src(s
), s
? 42 : 10);
366 setImmediate32(i
, s
, mod
);
376 CodeEmitterGK110::emitForm_C(const Instruction
*i
, uint32_t opc
, uint8_t ctg
)
385 switch (i
->src(0).getFile()) {
386 case FILE_MEMORY_CONST
:
387 code
[1] |= 0x4 << 28;
388 setCAddress14(i
->src(0));
391 code
[1] |= 0xc << 28;
392 srcId(i
->src(0), 23);
400 // 0x2 for GPR, c[] and 0x1 for short immediate
402 CodeEmitterGK110::emitForm_21(const Instruction
*i
, uint32_t opc2
,
405 const bool imm
= i
->srcExists(1) && i
->src(1).getFile() == FILE_IMMEDIATE
;
408 if (i
->srcExists(2) && i
->src(2).getFile() == FILE_MEMORY_CONST
)
413 code
[1] = opc1
<< 20;
416 code
[1] = (0xc << 28) | (opc2
<< 20);
423 for (int s
= 0; s
< 3 && i
->srcExists(s
); ++s
) {
424 switch (i
->src(s
).getFile()) {
425 case FILE_MEMORY_CONST
:
426 code
[1] &= (s
== 2) ? ~(0x4 << 28) : ~(0x8 << 28);
427 setCAddress14(i
->src(s
));
430 setShortImmediate(i
, s
);
433 srcId(i
->src(s
), s
? ((s
== 2) ? 42 : s1
) : 10);
436 if (i
->op
== OP_SELP
) {
437 assert(s
== 2 && i
->src(s
).getFile() == FILE_PREDICATE
);
438 srcId(i
->src(s
), 42);
440 // ignore here, can be predicate or flags, but must not be address
448 assert(imm
|| (code
[1] & (0xc << 28)));
452 CodeEmitterGK110::modNegAbsF32_3b(const Instruction
*i
, const int s
)
454 if (i
->src(s
).mod
.abs()) code
[1] &= ~(1 << 27);
455 if (i
->src(s
).mod
.neg()) code
[1] ^= (1 << 27);
459 CodeEmitterGK110::emitNOP(const Instruction
*i
)
461 code
[0] = 0x00003c02;
462 code
[1] = 0x85800000;
467 code
[0] = 0x001c3c02;
471 CodeEmitterGK110::emitFMAD(const Instruction
*i
)
473 assert(!isLIMM(i
->src(1), TYPE_F32
));
475 emitForm_21(i
, 0x0c0, 0x940);
483 bool neg1
= (i
->src(0).mod
^ i
->src(1).mod
).neg();
495 CodeEmitterGK110::emitDMAD(const Instruction
*i
)
497 assert(!i
->saturate
);
500 emitForm_21(i
, 0x1b8, 0xb38);
505 bool neg1
= (i
->src(0).mod
^ i
->src(1).mod
).neg();
517 CodeEmitterGK110::emitFMUL(const Instruction
*i
)
519 bool neg
= (i
->src(0).mod
^ i
->src(1).mod
).neg();
521 assert(i
->postFactor
>= -3 && i
->postFactor
<= 3);
523 if (isLIMM(i
->src(1), TYPE_F32
)) {
524 emitForm_L(i
, 0x200, 0x2, Modifier(0));
532 assert(i
->postFactor
== 0);
534 emitForm_21(i
, 0x234, 0xc34);
535 code
[1] |= ((i
->postFactor
> 0) ?
536 (7 - i
->postFactor
) : (0 - i
->postFactor
)) << 12;
554 CodeEmitterGK110::emitDMUL(const Instruction
*i
)
556 bool neg
= (i
->src(0).mod
^ i
->src(1).mod
).neg();
558 assert(!i
->postFactor
);
559 assert(!i
->saturate
);
563 emitForm_21(i
, 0x240, 0xc40);
577 CodeEmitterGK110::emitIMUL(const Instruction
*i
)
579 assert(!i
->src(0).mod
.neg() && !i
->src(1).mod
.neg());
580 assert(!i
->src(0).mod
.abs() && !i
->src(1).mod
.abs());
582 if (isLIMM(i
->src(1), TYPE_S32
)) {
583 emitForm_L(i
, 0x280, 2, Modifier(0));
585 if (i
->subOp
== NV50_IR_SUBOP_MUL_HIGH
)
587 if (i
->sType
== TYPE_S32
)
590 emitForm_21(i
, 0x21c, 0xc1c);
592 if (i
->subOp
== NV50_IR_SUBOP_MUL_HIGH
)
594 if (i
->sType
== TYPE_S32
)
600 CodeEmitterGK110::emitFADD(const Instruction
*i
)
602 if (isLIMM(i
->src(1), TYPE_F32
)) {
603 assert(i
->rnd
== ROUND_N
);
604 assert(!i
->saturate
);
606 Modifier mod
= i
->src(1).mod
^
607 Modifier(i
->op
== OP_SUB
? NV50_IR_MOD_NEG
: 0);
609 emitForm_L(i
, 0x400, 0, mod
);
615 emitForm_21(i
, 0x22c, 0xc2c);
624 modNegAbsF32_3b(i
, 1);
625 if (i
->op
== OP_SUB
) code
[1] ^= 1 << 27;
629 if (i
->op
== OP_SUB
) code
[1] ^= 1 << 16;
635 CodeEmitterGK110::emitDADD(const Instruction
*i
)
637 assert(!i
->saturate
);
640 emitForm_21(i
, 0x238, 0xc38);
645 modNegAbsF32_3b(i
, 1);
646 if (i
->op
== OP_SUB
) code
[1] ^= 1 << 27;
650 if (i
->op
== OP_SUB
) code
[1] ^= 1 << 16;
655 CodeEmitterGK110::emitUADD(const Instruction
*i
)
657 uint8_t addOp
= (i
->src(0).mod
.neg() << 1) | i
->src(1).mod
.neg();
662 assert(!i
->src(0).mod
.abs() && !i
->src(1).mod
.abs());
664 if (isLIMM(i
->src(1), TYPE_S32
)) {
665 emitForm_L(i
, 0x400, 1, Modifier((addOp
& 1) ? NV50_IR_MOD_NEG
: 0));
670 assert(!i
->defExists(1));
671 assert(i
->flagsSrc
< 0);
675 emitForm_21(i
, 0x208, 0xc08);
677 assert(addOp
!= 3); // would be add-plus-one
679 code
[1] |= addOp
<< 19;
682 code
[1] |= 1 << 18; // write carry
683 if (i
->flagsSrc
>= 0)
684 code
[1] |= 1 << 14; // add carry
692 CodeEmitterGK110::emitIMAD(const Instruction
*i
)
695 (i
->src(2).mod
.neg() << 1) | (i
->src(0).mod
.neg() ^ i
->src(1).mod
.neg());
697 emitForm_21(i
, 0x100, 0xa00);
700 code
[1] |= addOp
<< 26;
702 if (i
->sType
== TYPE_S32
)
703 code
[1] |= (1 << 19) | (1 << 24);
705 if (i
->subOp
== NV50_IR_SUBOP_MUL_HIGH
)
708 if (i
->flagsDef
>= 0) code
[1] |= 1 << 18;
709 if (i
->flagsSrc
>= 0) code
[1] |= 1 << 20;
715 CodeEmitterGK110::emitISAD(const Instruction
*i
)
717 assert(i
->dType
== TYPE_S32
|| i
->dType
== TYPE_U32
);
719 emitForm_21(i
, 0x1f4, 0xb74);
721 if (i
->dType
== TYPE_S32
)
726 CodeEmitterGK110::emitNOT(const Instruction
*i
)
728 code
[0] = 0x0003fc02; // logop(mov2) dst, 0, not src
729 code
[1] = 0x22003800;
735 switch (i
->src(0).getFile()) {
737 code
[1] |= 0xc << 28;
738 srcId(i
->src(0), 23);
740 case FILE_MEMORY_CONST
:
741 code
[1] |= 0x4 << 28;
742 setCAddress14(i
->src(1));
751 CodeEmitterGK110::emitLogicOp(const Instruction
*i
, uint8_t subOp
)
753 if (isLIMM(i
->src(1), TYPE_S32
)) {
754 emitForm_L(i
, 0x200, 0, i
->src(1).mod
);
755 code
[1] |= subOp
<< 24;
758 emitForm_21(i
, 0x220, 0xc20);
759 code
[1] |= subOp
<< 12;
766 CodeEmitterGK110::emitPOPC(const Instruction
*i
)
768 assert(!isLIMM(i
->src(1), TYPE_S32
, true));
770 emitForm_21(i
, 0x204, 0xc04);
773 if (!(code
[0] & 0x1))
778 CodeEmitterGK110::emitINSBF(const Instruction
*i
)
780 emitForm_21(i
, 0x1f8, 0xb78);
784 CodeEmitterGK110::emitEXTBF(const Instruction
*i
)
786 emitForm_21(i
, 0x600, 0xc00);
788 if (i
->dType
== TYPE_S32
)
790 if (i
->subOp
== NV50_IR_SUBOP_EXTBF_REV
)
795 CodeEmitterGK110::emitBFIND(const Instruction
*i
)
797 emitForm_C(i
, 0x218, 0x2);
799 if (i
->dType
== TYPE_S32
)
801 if (i
->src(0).mod
== Modifier(NV50_IR_MOD_NOT
))
803 if (i
->subOp
== NV50_IR_SUBOP_BFIND_SAMT
)
808 CodeEmitterGK110::emitShift(const Instruction
*i
)
810 if (i
->op
== OP_SHR
) {
811 emitForm_21(i
, 0x214, 0xc14);
812 if (isSignedType(i
->dType
))
815 emitForm_21(i
, 0x224, 0xc24);
818 if (i
->subOp
== NV50_IR_SUBOP_SHIFT_WRAP
)
823 CodeEmitterGK110::emitPreOp(const Instruction
*i
)
825 emitForm_C(i
, 0x248, 0x2);
827 if (i
->op
== OP_PREEX2
)
835 CodeEmitterGK110::emitSFnOp(const Instruction
*i
, uint8_t subOp
)
837 code
[0] = 0x00000002 | (subOp
<< 23);
838 code
[1] = 0x84000000;
843 srcId(i
->src(0), 10);
851 CodeEmitterGK110::emitMINMAX(const Instruction
*i
)
875 emitForm_21(i
, op2
, op1
);
877 if (i
->dType
== TYPE_S32
)
879 code
[1] |= (i
->op
== OP_MIN
) ? 0x1c00 : 0x3c00; // [!]pt
885 modNegAbsF32_3b(i
, 1);
893 CodeEmitterGK110::emitCVT(const Instruction
*i
)
895 const bool f2f
= isFloatType(i
->dType
) && isFloatType(i
->sType
);
896 const bool f2i
= !isFloatType(i
->dType
) && isFloatType(i
->sType
);
897 const bool i2f
= isFloatType(i
->dType
) && !isFloatType(i
->sType
);
899 bool sat
= i
->saturate
;
900 bool abs
= i
->src(0).mod
.abs();
901 bool neg
= i
->src(0).mod
.neg();
903 RoundMode rnd
= i
->rnd
;
906 case OP_CEIL
: rnd
= f2f
? ROUND_PI
: ROUND_P
; break;
907 case OP_FLOOR
: rnd
= f2f
? ROUND_MI
: ROUND_M
; break;
908 case OP_TRUNC
: rnd
= f2f
? ROUND_ZI
: ROUND_Z
; break;
909 case OP_SAT
: sat
= true; break;
910 case OP_NEG
: neg
= !neg
; break;
911 case OP_ABS
: abs
= true; neg
= false; break;
918 if (i
->op
== OP_NEG
&& i
->dType
== TYPE_U32
)
927 else if (f2i
) op
= 0x258;
928 else if (i2f
) op
= 0x25c;
931 emitForm_C(i
, op
, 0x2);
934 if (neg
) code
[1] |= 1 << 16;
935 if (abs
) code
[1] |= 1 << 20;
936 if (sat
) code
[1] |= 1 << 21;
938 emitRoundMode(rnd
, 32 + 10, f2f
? (32 + 13) : -1);
940 code
[0] |= typeSizeofLog2(dType
) << 10;
941 code
[0] |= typeSizeofLog2(i
->sType
) << 12;
942 code
[1] |= i
->subOp
<< 12;
944 if (isSignedIntType(dType
))
946 if (isSignedIntType(i
->sType
))
951 CodeEmitterGK110::emitSET(const CmpInstruction
*i
)
955 if (i
->def(0).getFile() == FILE_PREDICATE
) {
957 case TYPE_F32
: op2
= 0x1d8; op1
= 0xb58; break;
958 case TYPE_F64
: op2
= 0x1c0; op1
= 0xb40; break;
964 emitForm_21(i
, op2
, op1
);
968 if (!(code
[0] & 0x1)) {
972 modNegAbsF32_3b(i
, 1);
976 // normal DST field is negated predicate result
977 code
[0] = (code
[0] & ~0xfc) | ((code
[0] << 3) & 0xe0);
984 case TYPE_F32
: op2
= 0x000; op1
= 0x800; break;
985 case TYPE_F64
: op2
= 0x080; op1
= 0x900; break;
991 emitForm_21(i
, op2
, op1
);
995 if (!(code
[0] & 0x1)) {
999 modNegAbsF32_3b(i
, 1);
1003 if (i
->dType
== TYPE_F32
) {
1004 if (isFloatType(i
->sType
))
1010 if (i
->sType
== TYPE_S32
)
1013 if (i
->op
!= OP_SET
) {
1015 case OP_SET_AND
: code
[1] |= 0x0 << 16; break;
1016 case OP_SET_OR
: code
[1] |= 0x1 << 16; break;
1017 case OP_SET_XOR
: code
[1] |= 0x2 << 16; break;
1022 srcId(i
->src(2), 0x2a);
1024 code
[1] |= 0x7 << 10;
1026 emitCondCode(i
->setCond
,
1027 isFloatType(i
->sType
) ? 0x33 : 0x34,
1028 isFloatType(i
->sType
) ? 0xf : 0x7);
1032 CodeEmitterGK110::emitSLCT(const CmpInstruction
*i
)
1034 CondCode cc
= i
->setCond
;
1035 if (i
->src(2).mod
.neg())
1036 cc
= reverseCondCode(cc
);
1038 if (i
->dType
== TYPE_F32
) {
1039 emitForm_21(i
, 0x1d0, 0xb50);
1041 emitCondCode(cc
, 0x33, 0xf);
1043 emitForm_21(i
, 0x1a0, 0xb20);
1044 emitCondCode(cc
, 0x34, 0x7);
1048 void CodeEmitterGK110::emitSELP(const Instruction
*i
)
1050 emitForm_21(i
, 0x250, 0x050);
1052 if (i
->src(2).mod
& Modifier(NV50_IR_MOD_NOT
))
1056 void CodeEmitterGK110::emitTEXBAR(const Instruction
*i
)
1058 code
[0] = 0x0000003e | (i
->subOp
<< 23);
1059 code
[1] = 0x77000000;
1064 void CodeEmitterGK110::emitTEXCSAA(const TexInstruction
*i
)
1066 code
[0] = 0x00000002;
1067 code
[1] = 0x76c00000;
1069 code
[1] |= i
->tex
.r
<< 9;
1070 // code[1] |= i->tex.s << (9 + 8);
1072 if (i
->tex
.liveOnly
)
1073 code
[0] |= 0x80000000;
1075 defId(i
->def(0), 2);
1076 srcId(i
->src(0), 10);
1080 isNextIndependentTex(const TexInstruction
*i
)
1082 if (!i
->next
|| !isTextureOp(i
->next
->op
))
1084 if (i
->getDef(0)->interfers(i
->next
->getSrc(0)))
1086 return !i
->next
->srcExists(1) || !i
->getDef(0)->interfers(i
->next
->getSrc(1));
1090 CodeEmitterGK110::emitTEX(const TexInstruction
*i
)
1092 const bool ind
= i
->tex
.rIndirectSrc
>= 0;
1095 code
[0] = 0x00000002;
1098 code
[1] = 0x7e000000;
1101 code
[1] = 0x7e800000;
1104 code
[1] = 0x78000000;
1107 code
[1] = 0x7dc00000;
1110 code
[1] = 0x7d800000;
1116 code
[0] = 0x00000002;
1117 code
[1] = 0x76000000;
1118 code
[1] |= i
->tex
.r
<< 9;
1121 code
[0] = 0x00000002;
1122 code
[1] = 0x76800000;
1123 code
[1] |= i
->tex
.r
<< 9;
1126 code
[0] = 0x00000002;
1127 code
[1] = 0x70000000;
1128 code
[1] |= i
->tex
.r
<< 13;
1131 code
[0] = 0x00000001;
1132 code
[1] = 0x70000000;
1133 code
[1] |= i
->tex
.r
<< 15;
1136 code
[0] = 0x00000001;
1137 code
[1] = 0x60000000;
1138 code
[1] |= i
->tex
.r
<< 15;
1143 code
[1] |= isNextIndependentTex(i
) ? 0x1 : 0x2; // t : p mode
1145 if (i
->tex
.liveOnly
)
1146 code
[0] |= 0x80000000;
1150 case OP_TXB
: code
[1] |= 0x2000; break;
1151 case OP_TXL
: code
[1] |= 0x3000; break;
1155 case OP_TXLQ
: break;
1157 assert(!"invalid texture op");
1161 if (i
->op
== OP_TXF
) {
1162 if (!i
->tex
.levelZero
)
1165 if (i
->tex
.levelZero
) {
1169 if (i
->op
!= OP_TXD
&& i
->tex
.derivAll
)
1174 code
[1] |= i
->tex
.mask
<< 2;
1176 const int src1
= (i
->predSrc
== 1) ? 2 : 1; // if predSrc == 1, !srcExists(2)
1178 defId(i
->def(0), 2);
1179 srcId(i
->src(0), 10);
1182 if (i
->op
== OP_TXG
) code
[1] |= i
->tex
.gatherComp
<< 13;
1185 code
[1] |= (i
->tex
.target
.isCube() ? 3 : (i
->tex
.target
.getDim() - 1)) << 7;
1186 if (i
->tex
.target
.isArray())
1188 if (i
->tex
.target
.isShadow())
1190 if (i
->tex
.target
== TEX_TARGET_2D_MS
||
1191 i
->tex
.target
== TEX_TARGET_2D_MS_ARRAY
)
1194 if (i
->srcExists(src1
) && i
->src(src1
).getFile() == FILE_IMMEDIATE
) {
1198 if (i
->tex
.useOffsets
== 1) {
1200 case OP_TXF
: code
[1] |= 0x200; break;
1201 case OP_TXD
: code
[1] |= 0x00400000; break;
1202 default: code
[1] |= 0x800; break;
1205 if (i
->tex
.useOffsets
== 4)
1210 CodeEmitterGK110::emitTXQ(const TexInstruction
*i
)
1212 code
[0] = 0x00000002;
1213 code
[1] = 0x75400001;
1215 switch (i
->tex
.query
) {
1216 case TXQ_DIMS
: code
[0] |= 0x01 << 25; break;
1217 case TXQ_TYPE
: code
[0] |= 0x02 << 25; break;
1218 case TXQ_SAMPLE_POSITION
: code
[0] |= 0x05 << 25; break;
1219 case TXQ_FILTER
: code
[0] |= 0x10 << 25; break;
1220 case TXQ_LOD
: code
[0] |= 0x12 << 25; break;
1221 case TXQ_BORDER_COLOUR
: code
[0] |= 0x16 << 25; break;
1223 assert(!"invalid texture query");
1227 code
[1] |= i
->tex
.mask
<< 2;
1228 code
[1] |= i
->tex
.r
<< 9;
1229 if (/*i->tex.sIndirectSrc >= 0 || */i
->tex
.rIndirectSrc
>= 0)
1230 code
[1] |= 0x08000000;
1232 defId(i
->def(0), 2);
1233 srcId(i
->src(0), 10);
1239 CodeEmitterGK110::emitQUADOP(const Instruction
*i
, uint8_t qOp
, uint8_t laneMask
)
1241 code
[0] = 0x00000002 | ((qOp
& 1) << 31);
1242 code
[1] = 0x7fc00000 | (qOp
>> 1) | (laneMask
<< 12);
1244 defId(i
->def(0), 2);
1245 srcId(i
->src(0), 10);
1246 srcId((i
->srcExists(1) && i
->predSrc
!= 1) ? i
->src(1) : i
->src(0), 23);
1248 if (i
->op
== OP_QUADOP
&& progType
!= Program::TYPE_FRAGMENT
)
1249 code
[1] |= 1 << 9; // dall
1255 CodeEmitterGK110::emitPIXLD(const Instruction
*i
)
1257 emitForm_L(i
, 0x7f4, 2, Modifier(0));
1258 code
[1] |= i
->subOp
<< 2;
1259 code
[1] |= 0x00070000;
1263 CodeEmitterGK110::emitBAR(const Instruction
*i
)
1265 code
[0] = 0x00000002;
1266 code
[1] = 0x85400000;
1269 case NV50_IR_SUBOP_BAR_ARRIVE
: code
[1] |= 0x08; break;
1270 case NV50_IR_SUBOP_BAR_RED_AND
: code
[1] |= 0x50; break;
1271 case NV50_IR_SUBOP_BAR_RED_OR
: code
[1] |= 0x90; break;
1272 case NV50_IR_SUBOP_BAR_RED_POPC
: code
[1] |= 0x10; break;
1275 assert(i
->subOp
== NV50_IR_SUBOP_BAR_SYNC
);
1281 srcId(i
->src(0), 10);
1282 srcId(i
->src(1), 23);
1285 void CodeEmitterGK110::emitMEMBAR(const Instruction
*i
)
1287 code
[0] = 0x00000002 | NV50_IR_SUBOP_MEMBAR_SCOPE(i
->subOp
) << 8;
1288 code
[1] = 0x7cc00000;
1294 CodeEmitterGK110::emitFlow(const Instruction
*i
)
1296 const FlowInstruction
*f
= i
->asFlow();
1298 unsigned mask
; // bit 0: predicate, bit 1: target
1300 code
[0] = 0x00000000;
1304 code
[1] = f
->absolute
? 0x10800000 : 0x12000000;
1305 if (i
->srcExists(0) && i
->src(0).getFile() == FILE_MEMORY_CONST
)
1310 code
[1] = f
->absolute
? 0x11000000 : 0x13000000;
1311 if (i
->srcExists(0) && i
->src(0).getFile() == FILE_MEMORY_CONST
)
1316 case OP_EXIT
: code
[1] = 0x18000000; mask
= 1; break;
1317 case OP_RET
: code
[1] = 0x19000000; mask
= 1; break;
1318 case OP_DISCARD
: code
[1] = 0x19800000; mask
= 1; break;
1319 case OP_BREAK
: code
[1] = 0x1a000000; mask
= 1; break;
1320 case OP_CONT
: code
[1] = 0x1a800000; mask
= 1; break;
1322 case OP_JOINAT
: code
[1] = 0x14800000; mask
= 2; break;
1323 case OP_PREBREAK
: code
[1] = 0x15000000; mask
= 2; break;
1324 case OP_PRECONT
: code
[1] = 0x15800000; mask
= 2; break;
1325 case OP_PRERET
: code
[1] = 0x13800000; mask
= 2; break;
1327 case OP_QUADON
: code
[1] = 0x1b800000; mask
= 0; break;
1328 case OP_QUADPOP
: code
[1] = 0x1c000000; mask
= 0; break;
1329 case OP_BRKPT
: code
[1] = 0x00000000; mask
= 0; break;
1331 assert(!"invalid flow operation");
1337 if (i
->flagsSrc
< 0)
1349 if (f
->op
== OP_CALL
) {
1351 assert(f
->absolute
);
1352 uint32_t pcAbs
= targNVC0
->getBuiltinOffset(f
->target
.builtin
);
1353 addReloc(RelocEntry::TYPE_BUILTIN
, 0, pcAbs
, 0xff800000, 23);
1354 addReloc(RelocEntry::TYPE_BUILTIN
, 1, pcAbs
, 0x007fffff, -9);
1356 assert(!f
->absolute
);
1357 int32_t pcRel
= f
->target
.fn
->binPos
- (codeSize
+ 8);
1358 code
[0] |= (pcRel
& 0x1ff) << 23;
1359 code
[1] |= (pcRel
>> 9) & 0x7fff;
1363 int32_t pcRel
= f
->target
.bb
->binPos
- (codeSize
+ 8);
1364 if (writeIssueDelays
&& !(f
->target
.bb
->binPos
& 0x3f))
1366 // currently we don't want absolute branches
1367 assert(!f
->absolute
);
1368 code
[0] |= (pcRel
& 0x1ff) << 23;
1369 code
[1] |= (pcRel
>> 9) & 0x7fff;
1374 CodeEmitterGK110::emitAFETCH(const Instruction
*i
)
1376 uint32_t offset
= i
->src(0).get()->reg
.data
.offset
& 0x7ff;
1378 code
[0] = 0x00000002 | (offset
<< 23);
1379 code
[1] = 0x7d000000 | (offset
>> 9);
1381 if (i
->getSrc(0)->reg
.file
== FILE_SHADER_OUTPUT
)
1386 defId(i
->def(0), 2);
1387 srcId(i
->src(0).getIndirect(0), 10);
1391 CodeEmitterGK110::emitPFETCH(const Instruction
*i
)
1393 uint32_t prim
= i
->src(0).get()->reg
.data
.u32
;
1395 code
[0] = 0x00000002 | ((prim
& 0xff) << 23);
1396 code
[1] = 0x7f800000;
1400 const int src1
= (i
->predSrc
== 1) ? 2 : 1; // if predSrc == 1, !srcExists(2)
1402 defId(i
->def(0), 2);
1407 CodeEmitterGK110::emitVFETCH(const Instruction
*i
)
1409 unsigned int size
= typeSizeof(i
->dType
);
1410 uint32_t offset
= i
->src(0).get()->reg
.data
.offset
;
1412 code
[0] = 0x00000002 | (offset
<< 23);
1413 code
[1] = 0x7ec00000 | (offset
>> 9);
1414 code
[1] |= (size
/ 4 - 1) << 18;
1418 if (i
->getSrc(0)->reg
.file
== FILE_SHADER_OUTPUT
)
1419 code
[1] |= 0x8; // yes, TCPs can read from *outputs* of other threads
1423 defId(i
->def(0), 2);
1424 srcId(i
->src(0).getIndirect(0), 10);
1425 srcId(i
->src(0).getIndirect(1), 32 + 10); // vertex address
1429 CodeEmitterGK110::emitEXPORT(const Instruction
*i
)
1431 unsigned int size
= typeSizeof(i
->dType
);
1432 uint32_t offset
= i
->src(0).get()->reg
.data
.offset
;
1434 code
[0] = 0x00000002 | (offset
<< 23);
1435 code
[1] = 0x7f000000 | (offset
>> 9);
1436 code
[1] |= (size
/ 4 - 1) << 18;
1443 assert(i
->src(1).getFile() == FILE_GPR
);
1445 srcId(i
->src(0).getIndirect(0), 10);
1446 srcId(i
->src(0).getIndirect(1), 32 + 10); // vertex base address
1447 srcId(i
->src(1), 2);
1451 CodeEmitterGK110::emitOUT(const Instruction
*i
)
1453 assert(i
->src(0).getFile() == FILE_GPR
);
1455 emitForm_21(i
, 0x1f0, 0xb70);
1457 if (i
->op
== OP_EMIT
)
1459 if (i
->op
== OP_RESTART
|| i
->subOp
== NV50_IR_SUBOP_EMIT_RESTART
)
1464 CodeEmitterGK110::emitInterpMode(const Instruction
*i
)
1466 code
[1] |= (i
->ipa
& 0x3) << 21; // TODO: INTERP_SAMPLEID
1467 code
[1] |= (i
->ipa
& 0xc) << (19 - 2);
1471 interpApply(const InterpEntry
*entry
, uint32_t *code
,
1472 bool force_persample_interp
, bool flatshade
)
1474 int ipa
= entry
->ipa
;
1475 int reg
= entry
->reg
;
1476 int loc
= entry
->loc
;
1479 (ipa
& NV50_IR_INTERP_MODE_MASK
) == NV50_IR_INTERP_SC
) {
1480 ipa
= NV50_IR_INTERP_FLAT
;
1482 } else if (force_persample_interp
&&
1483 (ipa
& NV50_IR_INTERP_SAMPLE_MASK
) == NV50_IR_INTERP_DEFAULT
&&
1484 (ipa
& NV50_IR_INTERP_MODE_MASK
) != NV50_IR_INTERP_FLAT
) {
1485 ipa
|= NV50_IR_INTERP_CENTROID
;
1487 code
[loc
+ 1] &= ~(0xf << 19);
1488 code
[loc
+ 1] |= (ipa
& 0x3) << 21;
1489 code
[loc
+ 1] |= (ipa
& 0xc) << (19 - 2);
1490 code
[loc
+ 0] &= ~(0xff << 23);
1491 code
[loc
+ 0] |= reg
<< 23;
1495 CodeEmitterGK110::emitINTERP(const Instruction
*i
)
1497 const uint32_t base
= i
->getSrc(0)->reg
.data
.offset
;
1499 code
[0] = 0x00000002 | (base
<< 31);
1500 code
[1] = 0x74800000 | (base
>> 1);
1505 if (i
->op
== OP_PINTERP
) {
1506 srcId(i
->src(1), 23);
1507 addInterp(i
->ipa
, SDATA(i
->src(1)).id
, interpApply
);
1509 code
[0] |= 0xff << 23;
1510 addInterp(i
->ipa
, 0xff, interpApply
);
1513 srcId(i
->src(0).getIndirect(0), 10);
1517 defId(i
->def(0), 2);
1519 if (i
->getSampleMode() == NV50_IR_INTERP_OFFSET
)
1520 srcId(i
->src(i
->op
== OP_PINTERP
? 2 : 1), 32 + 10);
1522 code
[1] |= 0xff << 10;
1526 CodeEmitterGK110::emitLoadStoreType(DataType ty
, const int pos
)
1558 assert(!"invalid ld/st type");
1561 code
[pos
/ 32] |= n
<< (pos
% 32);
1565 CodeEmitterGK110::emitCachingMode(CacheMode c
, const int pos
)
1586 assert(!"invalid caching mode");
1589 code
[pos
/ 32] |= n
<< (pos
% 32);
1593 CodeEmitterGK110::emitSTORE(const Instruction
*i
)
1595 int32_t offset
= SDATA(i
->src(0)).offset
;
1597 switch (i
->src(0).getFile()) {
1598 case FILE_MEMORY_GLOBAL
: code
[1] = 0xe0000000; code
[0] = 0x00000000; break;
1599 case FILE_MEMORY_LOCAL
: code
[1] = 0x7a800000; code
[0] = 0x00000002; break;
1600 case FILE_MEMORY_SHARED
: code
[1] = 0x7ac00000; code
[0] = 0x00000002; break;
1602 assert(!"invalid memory file");
1606 if (i
->src(0).getFile() != FILE_MEMORY_GLOBAL
)
1609 if (code
[0] & 0x2) {
1610 emitLoadStoreType(i
->dType
, 0x33);
1611 if (i
->src(0).getFile() == FILE_MEMORY_LOCAL
)
1612 emitCachingMode(i
->cache
, 0x2f);
1614 emitLoadStoreType(i
->dType
, 0x38);
1615 emitCachingMode(i
->cache
, 0x3b);
1617 code
[0] |= offset
<< 23;
1618 code
[1] |= offset
>> 9;
1622 srcId(i
->src(1), 2);
1623 srcId(i
->src(0).getIndirect(0), 10);
1624 if (i
->src(0).getFile() == FILE_MEMORY_GLOBAL
&&
1625 i
->src(0).isIndirect(0) &&
1626 i
->getIndirect(0, 0)->reg
.size
== 8)
1631 CodeEmitterGK110::emitLOAD(const Instruction
*i
)
1633 int32_t offset
= SDATA(i
->src(0)).offset
;
1635 switch (i
->src(0).getFile()) {
1636 case FILE_MEMORY_GLOBAL
: code
[1] = 0xc0000000; code
[0] = 0x00000000; break;
1637 case FILE_MEMORY_LOCAL
: code
[1] = 0x7a000000; code
[0] = 0x00000002; break;
1638 case FILE_MEMORY_SHARED
: code
[1] = 0x7a400000; code
[0] = 0x00000002; break;
1639 case FILE_MEMORY_CONST
:
1640 if (!i
->src(0).isIndirect(0) && typeSizeof(i
->dType
) == 4) {
1645 code
[0] = 0x00000002;
1646 code
[1] = 0x7c800000 | (i
->src(0).get()->reg
.fileIndex
<< 7);
1647 code
[1] |= i
->subOp
<< 15;
1650 assert(!"invalid memory file");
1654 if (code
[0] & 0x2) {
1656 emitLoadStoreType(i
->dType
, 0x33);
1657 if (i
->src(0).getFile() == FILE_MEMORY_LOCAL
)
1658 emitCachingMode(i
->cache
, 0x2f);
1660 emitLoadStoreType(i
->dType
, 0x38);
1661 emitCachingMode(i
->cache
, 0x3b);
1663 code
[0] |= offset
<< 23;
1664 code
[1] |= offset
>> 9;
1668 defId(i
->def(0), 2);
1669 if (i
->getIndirect(0, 0)) {
1670 srcId(i
->src(0).getIndirect(0), 10);
1671 if (i
->getIndirect(0, 0)->reg
.size
== 8)
1674 code
[0] |= 255 << 10;
1679 CodeEmitterGK110::getSRegEncoding(const ValueRef
& ref
)
1681 switch (SDATA(ref
).sv
.sv
) {
1682 case SV_LANEID
: return 0x00;
1683 case SV_PHYSID
: return 0x03;
1684 case SV_VERTEX_COUNT
: return 0x10;
1685 case SV_INVOCATION_ID
: return 0x11;
1686 case SV_YDIR
: return 0x12;
1687 case SV_THREAD_KILL
: return 0x13;
1688 case SV_TID
: return 0x21 + SDATA(ref
).sv
.index
;
1689 case SV_CTAID
: return 0x25 + SDATA(ref
).sv
.index
;
1690 case SV_NTID
: return 0x29 + SDATA(ref
).sv
.index
;
1691 case SV_GRIDID
: return 0x2c;
1692 case SV_NCTAID
: return 0x2d + SDATA(ref
).sv
.index
;
1693 case SV_LBASE
: return 0x34;
1694 case SV_SBASE
: return 0x30;
1695 case SV_CLOCK
: return 0x50 + SDATA(ref
).sv
.index
;
1697 assert(!"no sreg for system value");
1703 CodeEmitterGK110::emitMOV(const Instruction
*i
)
1705 if (i
->src(0).getFile() == FILE_SYSTEM_VALUE
) {
1706 code
[0] = 0x00000002 | (getSRegEncoding(i
->src(0)) << 23);
1707 code
[1] = 0x86400000;
1709 defId(i
->def(0), 2);
1711 if (i
->src(0).getFile() == FILE_IMMEDIATE
) {
1712 code
[0] = 0x00000002 | (i
->lanes
<< 14);
1713 code
[1] = 0x74000000;
1715 defId(i
->def(0), 2);
1716 setImmediate32(i
, 0, Modifier(0));
1718 if (i
->src(0).getFile() == FILE_PREDICATE
) {
1719 code
[0] = 0x00000002;
1720 code
[1] = 0x84401c07;
1722 defId(i
->def(0), 2);
1723 srcId(i
->src(0), 14);
1725 emitForm_C(i
, 0x24c, 2);
1726 code
[1] |= i
->lanes
<< 10;
1731 uses64bitAddress(const Instruction
*ldst
)
1733 return ldst
->src(0).getFile() == FILE_MEMORY_GLOBAL
&&
1734 ldst
->src(0).isIndirect(0) &&
1735 ldst
->getIndirect(0, 0)->reg
.size
== 8;
1739 CodeEmitterGK110::emitATOM(const Instruction
*i
)
1741 code
[0] = 0x00000002;
1742 if (i
->subOp
== NV50_IR_SUBOP_ATOM_CAS
)
1743 code
[1] = 0x77800000;
1745 code
[1] = 0x68000000;
1748 case NV50_IR_SUBOP_ATOM_CAS
: break;
1749 case NV50_IR_SUBOP_ATOM_EXCH
: code
[1] |= 0x04000000; break;
1750 default: code
[1] |= i
->subOp
<< 23; break;
1754 case TYPE_U32
: break;
1755 case TYPE_S32
: code
[1] |= 0x00100000; break;
1756 case TYPE_U64
: code
[1] |= 0x00200000; break;
1757 case TYPE_F32
: code
[1] |= 0x00300000; break;
1758 case TYPE_B128
: code
[1] |= 0x00400000; break; /* TODO: U128 */
1759 case TYPE_S64
: code
[1] |= 0x00500000; break;
1760 default: assert(!"unsupported type"); break;
1765 /* TODO: cas: check that src regs line up */
1766 /* TODO: cas: flip bits if $r255 is used */
1767 srcId(i
->src(1), 23);
1769 if (i
->defExists(0))
1770 defId(i
->def(0), 2);
1772 code
[0] |= 255 << 2;
1774 const int32_t offset
= SDATA(i
->src(0)).offset
;
1775 assert(offset
< 0x80000 && offset
>= -0x80000);
1776 code
[0] |= (offset
& 1) << 31;
1777 code
[1] |= (offset
& 0xffffe) >> 1;
1779 if (i
->getIndirect(0, 0)) {
1780 srcId(i
->getIndirect(0, 0), 10);
1781 if (i
->getIndirect(0, 0)->reg
.size
== 8)
1784 code
[0] |= 255 << 10;
1789 CodeEmitterGK110::emitCCTL(const Instruction
*i
)
1791 int32_t offset
= SDATA(i
->src(0)).offset
;
1793 code
[0] = 0x00000002 | (i
->subOp
<< 2);
1795 if (i
->src(0).getFile() == FILE_MEMORY_GLOBAL
) {
1796 code
[1] = 0x7b000000;
1798 code
[1] = 0x7c000000;
1801 code
[0] |= offset
<< 23;
1802 code
[1] |= offset
>> 9;
1804 if (uses64bitAddress(i
))
1806 srcId(i
->src(0).getIndirect(0), 10);
1812 CodeEmitterGK110::emitInstruction(Instruction
*insn
)
1814 const unsigned int size
= (writeIssueDelays
&& !(codeSize
& 0x3f)) ? 16 : 8;
1816 if (insn
->encSize
!= 8) {
1817 ERROR("skipping unencodable instruction: ");
1821 if (codeSize
+ size
> codeSizeLimit
) {
1822 ERROR("code emitter output buffer too small\n");
1826 if (writeIssueDelays
) {
1827 int id
= (codeSize
& 0x3f) / 8 - 1;
1830 code
[0] = 0x00000000; // cf issue delay "instruction"
1831 code
[1] = 0x08000000;
1835 uint32_t *data
= code
- (id
* 2 + 2);
1838 case 0: data
[0] |= insn
->sched
<< 2; break;
1839 case 1: data
[0] |= insn
->sched
<< 10; break;
1840 case 2: data
[0] |= insn
->sched
<< 18; break;
1841 case 3: data
[0] |= insn
->sched
<< 26; data
[1] |= insn
->sched
>> 6; break;
1842 case 4: data
[1] |= insn
->sched
<< 2; break;
1843 case 5: data
[1] |= insn
->sched
<< 10; break;
1844 case 6: data
[1] |= insn
->sched
<< 18; break;
1851 // assert that instructions with multiple defs don't corrupt registers
1852 for (int d
= 0; insn
->defExists(d
); ++d
)
1853 assert(insn
->asTex() || insn
->def(d
).rep()->reg
.data
.id
>= 0);
1890 if (insn
->dType
== TYPE_F64
)
1892 else if (isFloatType(insn
->dType
))
1898 if (insn
->dType
== TYPE_F64
)
1900 else if (isFloatType(insn
->dType
))
1907 if (insn
->dType
== TYPE_F64
)
1909 else if (isFloatType(insn
->dType
))
1921 emitLogicOp(insn
, 0);
1924 emitLogicOp(insn
, 1);
1927 emitLogicOp(insn
, 2);
1937 emitSET(insn
->asCmp());
1943 emitSLCT(insn
->asCmp());
1958 if (insn
->def(0).getFile() == FILE_PREDICATE
||
1959 insn
->src(0).getFile() == FILE_PREDICATE
)
1965 emitSFnOp(insn
, 5 + 2 * insn
->subOp
);
1968 emitSFnOp(insn
, 4 + 2 * insn
->subOp
);
1993 emitTEX(insn
->asTex());
1996 emitTXQ(insn
->asTex());
2021 emitQUADOP(insn
, insn
->subOp
, insn
->lanes
);
2024 emitQUADOP(insn
, insn
->src(0).mod
.neg() ? 0x66 : 0x99, 0x4);
2027 emitQUADOP(insn
, insn
->src(0).mod
.neg() ? 0x5a : 0xa5, 0x5);
2060 ERROR("operation should have been eliminated");
2066 ERROR("operation should have been lowered\n");
2069 ERROR("unknown op: %u\n", insn
->op
);
2082 CodeEmitterGK110::getMinEncodingSize(const Instruction
*i
) const
2084 // No more short instruction encodings.
2089 CodeEmitterGK110::prepareEmission(Function
*func
)
2091 const Target
*targ
= func
->getProgram()->getTarget();
2093 CodeEmitter::prepareEmission(func
);
2095 if (targ
->hasSWSched
)
2096 calculateSchedDataNVC0(targ
, func
);
2099 CodeEmitterGK110::CodeEmitterGK110(const TargetNVC0
*target
)
2100 : CodeEmitter(target
),
2102 writeIssueDelays(target
->hasSWSched
)
2105 codeSize
= codeSizeLimit
= 0;
2110 TargetNVC0::createCodeEmitterGK110(Program::Type type
)
2112 CodeEmitterGK110
*emit
= new CodeEmitterGK110(this);
2113 emit
->setProgramType(type
);
2117 } // namespace nv50_ir