2 * Copyright 2011 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.
33 #include "codegen/unordered_set.h"
34 #include "codegen/nv50_ir_util.h"
35 #include "codegen/nv50_ir_graph.h"
37 #include "codegen/nv50_ir_driver.h"
45 OP_UNION
, // unify a new definition and several source values
46 OP_SPLIT
, // $r0d -> { $r0, $r1 } ($r0d and $r0/$r1 will be coalesced)
47 OP_MERGE
, // opposite of split, e.g. combine 2 32 bit into a 64 bit value
48 OP_CONSTRAINT
, // copy values into consecutive registers
49 OP_MOV
, // simple copy, no modifiers allowed
52 OP_ADD
, // NOTE: add u64 + u32 is legal for targets w/o 64-bit integer adds
59 OP_SAD
, // abs(src0 - src1) + src2
71 OP_SAT
, // CLAMP(f32, 0.0, 1.0)
76 OP_SET_AND
, // dst = (src0 CMP src1) & src2
80 OP_SELP
, // dst = src2 ? src0 : src1
81 OP_SLCT
, // dst = (src2 CMP 0) ? src0 : src1
88 OP_EXP
, // exponential (base M_E)
89 OP_LOG
, // natural logarithm
102 OP_BRKPT
, // breakpoint (not related to loops)
103 OP_JOINAT
, // push control flow convergence point
107 OP_MEMBAR
, // memory barrier (mfence, lfence, sfence)
108 OP_VFETCH
, // indirection 0 in attribute space, indirection 1 is vertex base
109 OP_PFETCH
, // fetch base address of vertex src0 (immediate) [+ src1]
110 OP_AFETCH
, // fetch base address of shader input (a[%r1+0x10])
114 OP_EMIT
, // emit vertex
115 OP_RESTART
, // restart primitive
117 OP_TXB
, // texture bias
118 OP_TXL
, // texure lod
119 OP_TXF
, // texel fetch
120 OP_TXQ
, // texture size query
121 OP_TXD
, // texture derivatives
122 OP_TXG
, // texture gather
123 OP_TXLQ
, // texture query lod
124 OP_TEXCSAA
, // texture op for coverage sampling
125 OP_TEXPREP
, // turn cube map array into 2d array coordinates
126 OP_SULDB
, // surface load (raw)
127 OP_SULDP
, // surface load (formatted)
128 OP_SUSTB
, // surface store (raw)
129 OP_SUSTP
, // surface store (formatted)
131 OP_SUREDP
, // surface reduction (atomic op)
132 OP_SULEA
, // surface load effective address
133 OP_SUBFM
, // surface bitfield manipulation
134 OP_SUCLAMP
, // clamp surface coordinates
135 OP_SUEAU
, // surface effective address
136 OP_SUQ
, // surface query
137 OP_MADSP
, // special integer multiply-add
138 OP_TEXBAR
, // texture dependency barrier
141 OP_RDSV
, // read system value
142 OP_WRSV
, // write system value
143 OP_PIXLD
, // get info about raster object or surfaces
147 OP_POPCNT
, // bitcount(src0 & src1)
148 OP_INSBF
, // insert first src1[8:15] bits of src0 into src2 at src1[0:7]
149 OP_EXTBF
, // place bits [K,K+N) of src0 into dst, src1 = 0xNNKK
150 OP_BFIND
, // find highest/lowest set bit
151 OP_PERMT
, // dst = bytes from src2,src0 selected by src1 (nvc0's src order)
153 OP_BAR
, // execution barrier, sources = { id, thread count, predicate }
154 OP_VADD
, // byte/word vector operations
163 OP_CCTL
, // cache control
164 OP_SHFL
, // warp shuffle
166 OP_BUFQ
, // buffer query
170 // various instruction-specific modifier definitions Instruction::subOp
171 // MOV_FINAL marks a MOV originating from an EXPORT (used for placing TEXBARs)
172 #define NV50_IR_SUBOP_MUL_HIGH 1
173 #define NV50_IR_SUBOP_EMIT_RESTART 1
174 #define NV50_IR_SUBOP_LDC_IL 1
175 #define NV50_IR_SUBOP_LDC_IS 2
176 #define NV50_IR_SUBOP_LDC_ISL 3
177 #define NV50_IR_SUBOP_SHIFT_WRAP 1
178 #define NV50_IR_SUBOP_EMU_PRERET 1
179 #define NV50_IR_SUBOP_TEXBAR(n) n
180 #define NV50_IR_SUBOP_MOV_FINAL 1
181 #define NV50_IR_SUBOP_EXTBF_REV 1
182 #define NV50_IR_SUBOP_BFIND_SAMT 1
183 #define NV50_IR_SUBOP_RCPRSQ_64H 1
184 #define NV50_IR_SUBOP_PERMT_F4E 1
185 #define NV50_IR_SUBOP_PERMT_B4E 2
186 #define NV50_IR_SUBOP_PERMT_RC8 3
187 #define NV50_IR_SUBOP_PERMT_ECL 4
188 #define NV50_IR_SUBOP_PERMT_ECR 5
189 #define NV50_IR_SUBOP_PERMT_RC16 6
190 #define NV50_IR_SUBOP_BAR_SYNC 0
191 #define NV50_IR_SUBOP_BAR_ARRIVE 1
192 #define NV50_IR_SUBOP_BAR_RED_AND 2
193 #define NV50_IR_SUBOP_BAR_RED_OR 3
194 #define NV50_IR_SUBOP_BAR_RED_POPC 4
195 #define NV50_IR_SUBOP_MEMBAR_L 1
196 #define NV50_IR_SUBOP_MEMBAR_S 2
197 #define NV50_IR_SUBOP_MEMBAR_M 3
198 #define NV50_IR_SUBOP_MEMBAR_CTA (0 << 2)
199 #define NV50_IR_SUBOP_MEMBAR_GL (1 << 2)
200 #define NV50_IR_SUBOP_MEMBAR_SYS (2 << 2)
201 #define NV50_IR_SUBOP_MEMBAR_DIR(m) ((m) & 0x3)
202 #define NV50_IR_SUBOP_MEMBAR_SCOPE(m) ((m) & ~0x3)
203 #define NV50_IR_SUBOP_MEMBAR(d,s) \
204 (NV50_IR_SUBOP_MEMBAR_##d | NV50_IR_SUBOP_MEMBAR_##s)
205 #define NV50_IR_SUBOP_ATOM_ADD 0
206 #define NV50_IR_SUBOP_ATOM_MIN 1
207 #define NV50_IR_SUBOP_ATOM_MAX 2
208 #define NV50_IR_SUBOP_ATOM_INC 3
209 #define NV50_IR_SUBOP_ATOM_DEC 4
210 #define NV50_IR_SUBOP_ATOM_AND 5
211 #define NV50_IR_SUBOP_ATOM_OR 6
212 #define NV50_IR_SUBOP_ATOM_XOR 7
213 #define NV50_IR_SUBOP_ATOM_CAS 8
214 #define NV50_IR_SUBOP_ATOM_EXCH 9
215 #define NV50_IR_SUBOP_CCTL_IV 5
216 #define NV50_IR_SUBOP_CCTL_IVALL 6
217 #define NV50_IR_SUBOP_SUST_IGN 0
218 #define NV50_IR_SUBOP_SUST_TRAP 1
219 #define NV50_IR_SUBOP_SUST_SDCL 3
220 #define NV50_IR_SUBOP_SULD_ZERO 0
221 #define NV50_IR_SUBOP_SULD_TRAP 1
222 #define NV50_IR_SUBOP_SULD_SDCL 3
223 #define NV50_IR_SUBOP_SUBFM_3D 1
224 #define NV50_IR_SUBOP_SUCLAMP_2D 0x10
225 #define NV50_IR_SUBOP_SUCLAMP_SD(r, d) (( 0 + (r)) | ((d == 2) ? 0x10 : 0))
226 #define NV50_IR_SUBOP_SUCLAMP_PL(r, d) (( 5 + (r)) | ((d == 2) ? 0x10 : 0))
227 #define NV50_IR_SUBOP_SUCLAMP_BL(r, d) ((10 + (r)) | ((d == 2) ? 0x10 : 0))
228 #define NV50_IR_SUBOP_PIXLD_COUNT 0
229 #define NV50_IR_SUBOP_PIXLD_COVMASK 1
230 #define NV50_IR_SUBOP_PIXLD_COVERED 2
231 #define NV50_IR_SUBOP_PIXLD_OFFSET 3
232 #define NV50_IR_SUBOP_PIXLD_CENT_OFFSET 4
233 #define NV50_IR_SUBOP_PIXLD_SAMPLEID 5
234 #define NV50_IR_SUBOP_SHFL_IDX 0
235 #define NV50_IR_SUBOP_SHFL_UP 1
236 #define NV50_IR_SUBOP_SHFL_DOWN 2
237 #define NV50_IR_SUBOP_SHFL_BFLY 3
238 #define NV50_IR_SUBOP_LOAD_LOCKED 1
239 #define NV50_IR_SUBOP_STORE_UNLOCKED 2
240 #define NV50_IR_SUBOP_MADSP_SD 0xffff
241 // Yes, we could represent those with DataType.
242 // Or put the type into operation and have a couple 1000 values in that enum.
243 // This will have to do for now.
244 // The bitfields are supposed to correspond to nve4 ISA.
245 #define NV50_IR_SUBOP_MADSP(a,b,c) (((c) << 8) | ((b) << 4) | (a))
246 #define NV50_IR_SUBOP_V1(d,a,b) (((d) << 10) | ((b) << 5) | (a) | 0x0000)
247 #define NV50_IR_SUBOP_V2(d,a,b) (((d) << 10) | ((b) << 5) | (a) | 0x4000)
248 #define NV50_IR_SUBOP_V4(d,a,b) (((d) << 10) | ((b) << 5) | (a) | 0x8000)
249 #define NV50_IR_SUBOP_Vn(n) ((n) >> 14)
250 #define NV50_IR_SUBOP_VOTE_ALL 0
251 #define NV50_IR_SUBOP_VOTE_ANY 1
252 #define NV50_IR_SUBOP_VOTE_UNI 2
254 #define NV50_IR_SUBOP_MINMAX_LOW 1
255 #define NV50_IR_SUBOP_MINMAX_MED 2
256 #define NV50_IR_SUBOP_MINMAX_HIGH 3
267 TYPE_U64
, // 64 bit operations are only lowered after register allocation
279 CC_NEVER
= CC_FL
, // when used with FILE_FLAGS
282 CC_NOT_P
= CC_EQ
, // when used with FILE_PREDICATE
310 ROUND_M
, // towards -inf
311 ROUND_Z
, // towards 0
312 ROUND_P
, // towards +inf
313 ROUND_NI
, // nearest integer
314 ROUND_MI
, // to integer towards -inf
315 ROUND_ZI
, // to integer towards 0
316 ROUND_PI
, // to integer towards +inf
321 CACHE_CA
, // cache at all levels
322 CACHE_WB
= CACHE_CA
, // cache write back
323 CACHE_CG
, // cache at global level
324 CACHE_CS
, // cache streaming
325 CACHE_CV
, // cache as volatile
326 CACHE_WT
= CACHE_CV
// cache write-through
333 FILE_PREDICATE
, // boolean predicate
334 FILE_FLAGS
, // zero/sign/carry/overflow bits
336 LAST_REGISTER_FILE
= FILE_ADDRESS
,
356 TEX_TARGET_1D_SHADOW
,
357 TEX_TARGET_2D_SHADOW
,
358 TEX_TARGET_CUBE_SHADOW
,
361 TEX_TARGET_2D_MS_ARRAY
,
362 TEX_TARGET_CUBE_ARRAY
,
363 TEX_TARGET_1D_ARRAY_SHADOW
,
364 TEX_TARGET_2D_ARRAY_SHADOW
,
366 TEX_TARGET_RECT_SHADOW
,
367 TEX_TARGET_CUBE_ARRAY_SHADOW
,
440 SV_VERTEX_COUNT
, // gl_PatchVerticesIn
483 class CmpInstruction
;
484 class TexInstruction
;
485 class FlowInstruction
;
490 class ImmediateValue
;
495 int8_t fileIndex
; // signed, may be indirect for CONST[]
496 uint8_t size
; // this should match the Instruction type's size
497 DataType type
; // mainly for pretty printing
499 uint64_t u64
; // immediate values
509 int32_t offset
; // offset from 0 (base of address space)
510 int32_t id
; // register id (< 0 if virtual/unassigned, in units <= 4)
518 // precedence: NOT after SAT after NEG after ABS
519 #define NV50_IR_MOD_ABS (1 << 0)
520 #define NV50_IR_MOD_NEG (1 << 1)
521 #define NV50_IR_MOD_SAT (1 << 2)
522 #define NV50_IR_MOD_NOT (1 << 3)
523 #define NV50_IR_MOD_NEG_ABS (NV50_IR_MOD_NEG | NV50_IR_MOD_ABS)
525 #define NV50_IR_INTERP_MODE_MASK 0x3
526 #define NV50_IR_INTERP_LINEAR (0 << 0)
527 #define NV50_IR_INTERP_PERSPECTIVE (1 << 0)
528 #define NV50_IR_INTERP_FLAT (2 << 0)
529 #define NV50_IR_INTERP_SC (3 << 0) // what exactly is that ?
530 #define NV50_IR_INTERP_SAMPLE_MASK 0xc
531 #define NV50_IR_INTERP_DEFAULT (0 << 2)
532 #define NV50_IR_INTERP_CENTROID (1 << 2)
533 #define NV50_IR_INTERP_OFFSET (2 << 2)
534 #define NV50_IR_INTERP_SAMPLEID (3 << 2)
536 // do we really want this to be a class ?
540 Modifier() : bits(0) { }
541 Modifier(unsigned int m
) : bits(m
) { }
542 Modifier(operation op
);
544 // @return new Modifier applying a after b (asserts if unrepresentable)
545 Modifier
operator*(const Modifier
) const;
546 Modifier
operator*=(const Modifier m
) { *this = *this * m
; return *this; }
547 Modifier
operator==(const Modifier m
) const { return m
.bits
== bits
; }
548 Modifier
operator!=(const Modifier m
) const { return m
.bits
!= bits
; }
550 inline Modifier
operator&(const Modifier m
) const { return bits
& m
.bits
; }
551 inline Modifier
operator|(const Modifier m
) const { return bits
| m
.bits
; }
552 inline Modifier
operator^(const Modifier m
) const { return bits
^ m
.bits
; }
554 operation
getOp() const;
556 inline int neg() const { return (bits
& NV50_IR_MOD_NEG
) ? 1 : 0; }
557 inline int abs() const { return (bits
& NV50_IR_MOD_ABS
) ? 1 : 0; }
559 inline operator bool() const { return bits
? true : false; }
561 void applyTo(ImmediateValue
&imm
) const;
563 int print(char *buf
, size_t size
) const;
572 ValueRef(Value
* = NULL
);
573 ValueRef(const ValueRef
&);
576 inline bool exists() const { return value
!= NULL
; }
579 void set(const ValueRef
&);
580 inline Value
*get() const { return value
; }
581 inline Value
*rep() const;
583 inline Instruction
*getInsn() const { return insn
; }
584 inline void setInsn(Instruction
*inst
) { insn
= inst
; }
586 inline bool isIndirect(int dim
) const { return indirect
[dim
] >= 0; }
587 inline const ValueRef
*getIndirect(int dim
) const;
589 inline DataFile
getFile() const;
590 inline unsigned getSize() const;
592 // SSA: return eventual (traverse MOVs) literal value, if it exists
593 bool getImmediate(ImmediateValue
&) const;
597 int8_t indirect
[2]; // >= 0 if relative to lvalue in insn->src(indirect[i])
600 bool usedAsPtr
; // for printing
610 ValueDef(Value
* = NULL
);
611 ValueDef(const ValueDef
&);
614 inline bool exists() const { return value
!= NULL
; }
616 inline Value
*get() const { return value
; }
617 inline Value
*rep() const;
619 bool mayReplace(const ValueRef
&);
620 void replace(const ValueRef
&, bool doSet
); // replace all uses of the old value
622 inline Instruction
*getInsn() const { return insn
; }
623 inline void setInsn(Instruction
*inst
) { insn
= inst
; }
625 inline DataFile
getFile() const;
626 inline unsigned getSize() const;
628 inline void setSSA(LValue
*);
629 inline const LValue
*preSSA() const;
632 Value
*value
; // should make this LValue * ...
633 LValue
*origin
; // pre SSA value
643 virtual Value
*clone(ClonePolicy
<Function
>&) const = 0;
645 virtual int print(char *, size_t, DataType ty
= TYPE_NONE
) const = 0;
647 virtual bool equals(const Value
*, bool strict
= false) const;
648 virtual bool interfers(const Value
*) const;
649 virtual bool isUniform() const { return true; }
651 inline Value
*rep() const { return join
; }
653 inline Instruction
*getUniqueInsn() const;
654 inline Instruction
*getInsn() const; // use when uniqueness is certain
656 inline int refCount() { return uses
.size(); }
658 inline LValue
*asLValue();
659 inline Symbol
*asSym();
660 inline ImmediateValue
*asImm();
661 inline const Symbol
*asSym() const;
662 inline const ImmediateValue
*asImm() const;
664 inline bool inFile(DataFile f
) { return reg
.file
== f
; }
666 static inline Value
*get(Iterator
&);
668 unordered_set
<ValueRef
*> uses
;
669 std::list
<ValueDef
*> defs
;
670 typedef unordered_set
<ValueRef
*>::iterator UseIterator
;
671 typedef unordered_set
<ValueRef
*>::const_iterator UseCIterator
;
672 typedef std::list
<ValueDef
*>::iterator DefIterator
;
673 typedef std::list
<ValueDef
*>::const_iterator DefCIterator
;
678 // TODO: these should be in LValue:
683 class LValue
: public Value
686 LValue(Function
*, DataFile file
);
687 LValue(Function
*, LValue
*);
690 virtual bool isUniform() const;
692 virtual LValue
*clone(ClonePolicy
<Function
>&) const;
694 virtual int print(char *, size_t, DataType ty
= TYPE_NONE
) const;
697 unsigned compMask
: 8; // compound/component mask
698 unsigned compound
: 1; // used by RA, value involved in split/merge
700 unsigned fixedReg
: 1; // set & used by RA, earlier just use (id < 0)
701 unsigned noSpill
: 1; // do not spill (e.g. if spill temporary already)
704 class Symbol
: public Value
707 Symbol(Program
*, DataFile file
= FILE_MEMORY_CONST
, ubyte fileIdx
= 0);
710 virtual Symbol
*clone(ClonePolicy
<Function
>&) const;
712 virtual bool equals(const Value
*that
, bool strict
) const;
714 virtual bool isUniform() const;
716 virtual int print(char *, size_t, DataType ty
= TYPE_NONE
) const;
718 // print with indirect values
719 int print(char *, size_t, Value
*, Value
*, DataType ty
= TYPE_NONE
) const;
721 inline void setFile(DataFile file
, ubyte fileIndex
= 0)
724 reg
.fileIndex
= fileIndex
;
727 inline void setOffset(int32_t offset
);
728 inline void setAddress(Symbol
*base
, int32_t offset
);
729 inline void setSV(SVSemantic sv
, uint32_t idx
= 0);
731 inline const Symbol
*getBase() const { return baseSym
; }
734 Symbol
*baseSym
; // array base for Symbols representing array elements
737 class ImmediateValue
: public Value
741 ImmediateValue(Program
*, uint32_t);
742 ImmediateValue(Program
*, float);
743 ImmediateValue(Program
*, double);
744 // NOTE: not added to program with
745 ImmediateValue(const ImmediateValue
*, DataType ty
);
746 ~ImmediateValue() { };
748 virtual ImmediateValue
*clone(ClonePolicy
<Function
>&) const;
750 virtual bool equals(const Value
*that
, bool strict
) const;
752 // these only work if 'type' is valid (we mostly use untyped literals):
753 bool isInteger(const int ival
) const; // ival is cast to this' type
754 bool isNegative() const;
759 // for constant folding:
760 ImmediateValue
operator+(const ImmediateValue
&) const;
761 ImmediateValue
operator-(const ImmediateValue
&) const;
762 ImmediateValue
operator*(const ImmediateValue
&) const;
763 ImmediateValue
operator/(const ImmediateValue
&) const;
765 ImmediateValue
& operator=(const ImmediateValue
&); // only sets value !
767 bool compare(CondCode cc
, float fval
) const;
769 virtual int print(char *, size_t, DataType ty
= TYPE_NONE
) const;
776 Instruction(Function
*, operation
, DataType
);
777 virtual ~Instruction();
779 virtual Instruction
*clone(ClonePolicy
<Function
>&,
780 Instruction
* = NULL
) const;
782 void setDef(int i
, Value
*);
783 void setSrc(int s
, Value
*);
784 void setSrc(int s
, const ValueRef
&);
785 void swapSources(int a
, int b
);
786 void moveSources(int s
, int delta
);
787 bool setIndirect(int s
, int dim
, Value
*);
789 inline ValueRef
& src(int s
) { return srcs
[s
]; }
790 inline ValueDef
& def(int s
) { return defs
[s
]; }
791 inline const ValueRef
& src(int s
) const { return srcs
[s
]; }
792 inline const ValueDef
& def(int s
) const { return defs
[s
]; }
794 inline Value
*getDef(int d
) const { return defs
[d
].get(); }
795 inline Value
*getSrc(int s
) const { return srcs
[s
].get(); }
796 inline Value
*getIndirect(int s
, int dim
) const;
798 inline bool defExists(unsigned d
) const
800 return d
< defs
.size() && defs
[d
].exists();
802 inline bool srcExists(unsigned s
) const
804 return s
< srcs
.size() && srcs
[s
].exists();
807 inline bool constrainedDefs() const;
809 bool setPredicate(CondCode ccode
, Value
*);
810 inline Value
*getPredicate() const;
811 bool writesPredicate() const;
812 inline bool isPredicated() const { return predSrc
>= 0; }
814 inline void setFlagsSrc(int s
, Value
*);
815 inline void setFlagsDef(int d
, Value
*);
816 inline bool usesFlags() const { return flagsSrc
>= 0; }
818 unsigned int defCount() const { return defs
.size(); };
819 unsigned int defCount(unsigned int mask
, bool singleFile
= false) const;
820 unsigned int srcCount() const { return srcs
.size(); };
821 unsigned int srcCount(unsigned int mask
, bool singleFile
= false) const;
823 // save & remove / set indirect[0,1] and predicate source
824 void takeExtraSources(int s
, Value
*[3]);
825 void putExtraSources(int s
, Value
*[3]);
827 inline void setType(DataType type
) { dType
= sType
= type
; }
829 inline void setType(DataType dtype
, DataType stype
)
835 inline bool isPseudo() const { return op
< OP_MOV
; }
838 bool isCommutationLegal(const Instruction
*) const; // must be adjacent !
839 bool isActionEqual(const Instruction
*) const;
840 bool isResultEqual(const Instruction
*) const;
842 // check whether the defs interfere with srcs and defs of another instruction
843 bool canCommuteDefDef(const Instruction
*) const;
844 bool canCommuteDefSrc(const Instruction
*) const;
848 inline CmpInstruction
*asCmp();
849 inline TexInstruction
*asTex();
850 inline FlowInstruction
*asFlow();
851 inline const TexInstruction
*asTex() const;
852 inline const CmpInstruction
*asCmp() const;
853 inline const FlowInstruction
*asFlow() const;
859 int serial
; // CFG order
862 DataType dType
; // destination or defining type
863 DataType sType
; // source or secondary type
868 uint16_t subOp
; // quadop, 1 for mul-high, etc.
870 unsigned encSize
: 4; // encoding size in bytes
871 unsigned saturate
: 1; // to [0.0f, 1.0f]
872 unsigned join
: 1; // converge control flow (use OP_JOIN until end)
873 unsigned fixed
: 1; // prevent dead code elimination
874 unsigned terminator
: 1; // end of basic block
875 unsigned ftz
: 1; // flush denormal to zero
876 unsigned dnz
: 1; // denormals, NaN are zero
877 unsigned ipa
: 4; // interpolation mode
879 unsigned perPatch
: 1;
880 unsigned exit
: 1; // terminate program after insn
881 unsigned mask
: 4; // for vector ops
883 int8_t postFactor
; // MUL/DIV(if < 0) by 1 << postFactor
889 uint32_t sched
; // scheduling data (NOTE: maybe move to separate storage)
894 std::deque
<ValueDef
> defs
; // no gaps !
895 std::deque
<ValueRef
> srcs
; // no gaps !
897 // instruction specific methods:
898 // (don't want to subclass, would need more constructors and memory pools)
900 inline void setInterpolate(unsigned int mode
) { ipa
= mode
; }
902 unsigned int getInterpMode() const { return ipa
& 0x3; }
903 unsigned int getSampleMode() const { return ipa
& 0xc; }
911 TXQ_DIMS
, /* x, y, z, levels */
912 TXQ_TYPE
, /* ?, ?, samples, ? */
920 class TexInstruction
: public Instruction
926 Target(TexTarget targ
= TEX_TARGET_2D
) : target(targ
) { }
928 const char *getName() const { return descTable
[target
].name
; }
929 unsigned int getArgCount() const { return descTable
[target
].argc
; }
930 unsigned int getDim() const { return descTable
[target
].dim
; }
931 int isArray() const { return descTable
[target
].array
? 1 : 0; }
932 int isCube() const { return descTable
[target
].cube
? 1 : 0; }
933 int isShadow() const { return descTable
[target
].shadow
? 1 : 0; }
935 return target
== TEX_TARGET_2D_MS
|| target
== TEX_TARGET_2D_MS_ARRAY
; }
939 target
= TEX_TARGET_2D_ARRAY
;
941 target
= TEX_TARGET_2D
;
945 Target
& operator=(TexTarget targ
)
947 assert(targ
< TEX_TARGET_COUNT
);
952 inline bool operator==(TexTarget targ
) const { return target
== targ
; }
953 inline bool operator!=(TexTarget targ
) const { return target
!= targ
; }
955 enum TexTarget
getEnum() const { return target
; }
968 static const struct Desc descTable
[TEX_TARGET_COUNT
];
971 enum TexTarget target
;
984 static const struct ImgFormatDesc formatTable
[IMG_FORMAT_COUNT
];
987 TexInstruction(Function
*, operation
);
988 virtual ~TexInstruction();
990 virtual TexInstruction
*clone(ClonePolicy
<Function
>&,
991 Instruction
* = NULL
) const;
993 inline void setTexture(Target targ
, uint8_t r
, uint8_t s
)
1000 void setIndirectR(Value
*);
1001 void setIndirectS(Value
*);
1002 inline Value
*getIndirectR() const;
1003 inline Value
*getIndirectS() const;
1011 int8_t rIndirectSrc
;
1012 int8_t sIndirectSrc
;
1017 bool liveOnly
; // only execute on live pixels of a quad (optimization)
1021 int8_t useOffsets
; // 0, 1, or 4 for textureGatherOffsets
1022 int8_t offset
[3]; // only used on nv50
1024 enum TexQuery query
;
1025 const struct ImgFormatDesc
*format
;
1030 ValueRef offset
[4][3];
1033 class CmpInstruction
: public Instruction
1036 CmpInstruction(Function
*, operation
);
1038 virtual CmpInstruction
*clone(ClonePolicy
<Function
>&,
1039 Instruction
* = NULL
) const;
1041 void setCondition(CondCode cond
) { setCond
= cond
; }
1042 CondCode
getCondition() const { return setCond
; }
1048 class FlowInstruction
: public Instruction
1051 FlowInstruction(Function
*, operation
, void *target
);
1053 virtual FlowInstruction
*clone(ClonePolicy
<Function
>&,
1054 Instruction
* = NULL
) const;
1057 unsigned allWarp
: 1;
1058 unsigned absolute
: 1;
1060 unsigned builtin
: 1; // true for calls to emulation code
1061 unsigned indirect
: 1; // target in src(0)
1073 BasicBlock(Function
*);
1076 BasicBlock
*clone(ClonePolicy
<Function
>&) const;
1078 inline int getId() const { return id
; }
1079 inline unsigned int getInsnCount() const { return numInsns
; }
1080 inline bool isTerminated() const { return exit
&& exit
->terminator
; }
1082 bool dominatedBy(BasicBlock
*bb
);
1083 inline bool reachableBy(const BasicBlock
*by
, const BasicBlock
*term
);
1085 // returns mask of conditional out blocks
1086 // e.g. 3 for IF { .. } ELSE { .. } ENDIF, 1 for IF { .. } ENDIF
1087 unsigned int initiatesSimpleConditional() const;
1090 Function
*getFunction() const { return func
; }
1091 Program
*getProgram() const { return program
; }
1093 Instruction
*getEntry() const { return entry
; } // first non-phi instruction
1094 Instruction
*getPhi() const { return phi
; }
1095 Instruction
*getFirst() const { return phi
? phi
: entry
; }
1096 Instruction
*getExit() const { return exit
; }
1098 void insertHead(Instruction
*);
1099 void insertTail(Instruction
*);
1100 void insertBefore(Instruction
*, Instruction
*);
1101 void insertAfter(Instruction
*, Instruction
*);
1102 void remove(Instruction
*);
1103 void permuteAdjacent(Instruction
*, Instruction
*);
1105 BasicBlock
*idom() const;
1107 // NOTE: currently does not rebuild the dominator tree
1108 BasicBlock
*splitBefore(Instruction
*, bool attach
= true);
1109 BasicBlock
*splitAfter(Instruction
*, bool attach
= true);
1111 DLList
& getDF() { return df
; }
1112 DLList::Iterator
iterDF() { return df
.iterator(); }
1114 static inline BasicBlock
*get(Iterator
&);
1115 static inline BasicBlock
*get(Graph::Node
*);
1118 Graph::Node cfg
; // first edge is branch *taken* (the ELSE branch)
1127 Instruction
*joinAt
; // for quick reference
1129 bool explicitCont
; // loop headers: true if loop contains continue stmts
1139 unsigned int numInsns
;
1145 void splitCommon(Instruction
*, BasicBlock
*, bool attach
);
1151 Function(Program
*, const char *name
, uint32_t label
);
1154 static inline Function
*get(Graph::Node
*node
);
1156 inline Program
*getProgram() const { return prog
; }
1157 inline const char *getName() const { return name
; }
1158 inline int getId() const { return id
; }
1159 inline uint32_t getLabel() const { return label
; }
1162 void printLiveIntervals() const;
1163 void printCFGraph(const char *filePath
);
1165 bool setEntry(BasicBlock
*);
1166 bool setExit(BasicBlock
*);
1168 unsigned int orderInstructions(ArrayList
&);
1170 inline void add(BasicBlock
*bb
, int& id
) { allBBlocks
.insert(bb
, id
); }
1171 inline void add(Instruction
*insn
, int& id
) { allInsns
.insert(insn
, id
); }
1172 inline void add(LValue
*lval
, int& id
) { allLValues
.insert(lval
, id
); }
1174 inline LValue
*getLValue(int id
);
1176 void buildLiveSets();
1177 void buildDefSets();
1178 bool convertToSSA();
1181 std::deque
<ValueDef
> ins
;
1182 std::deque
<ValueRef
> outs
;
1183 std::deque
<Value
*> clobbers
;
1186 Graph::Node
*cfgExit
;
1188 Graph::Node call
; // node in the call graph
1190 BasicBlock
**bbArray
; // BBs in emission order
1193 unsigned int loopNestingBound
;
1201 uint32_t tlsBase
; // base address for l[] space (if no stack pointer is used)
1204 ArrayList allBBlocks
;
1206 ArrayList allLValues
;
1209 void buildLiveSetsPreSSA(BasicBlock
*, const int sequence
);
1210 void buildDefSetsPreSSA(BasicBlock
*bb
, const int seq
);
1215 const char *const name
;
1222 CG_STAGE_SSA
, // expected directly before register allocation
1232 TYPE_TESSELLATION_CONTROL
,
1233 TYPE_TESSELLATION_EVAL
,
1239 Program(Type type
, Target
*targ
);
1244 Type
getType() const { return progType
; }
1246 inline void add(Function
*fn
, int& id
) { allFuncs
.insert(fn
, id
); }
1247 inline void del(Function
*fn
, int& id
) { allFuncs
.remove(id
); }
1248 inline void add(Value
*rval
, int& id
) { allRValues
.insert(rval
, id
); }
1250 bool makeFromTGSI(struct nv50_ir_prog_info
*);
1251 bool makeFromSM4(struct nv50_ir_prog_info
*);
1252 bool convertToSSA();
1253 bool optimizeSSA(int level
);
1254 bool optimizePostRA(int level
);
1255 bool registerAllocation();
1256 bool emitBinary(struct nv50_ir_prog_info
*);
1258 const Target
*getTarget() const { return target
; }
1261 void emitSymbolTable(struct nv50_ir_prog_info
*);
1271 ArrayList allRValues
;
1275 uint32_t tlsSize
; // size required for FILE_MEMORY_LOCAL
1279 MemoryPool mem_Instruction
;
1280 MemoryPool mem_CmpInstruction
;
1281 MemoryPool mem_TexInstruction
;
1282 MemoryPool mem_FlowInstruction
;
1283 MemoryPool mem_LValue
;
1284 MemoryPool mem_Symbol
;
1285 MemoryPool mem_ImmediateValue
;
1290 void *targetPriv
; // e.g. to carry information between passes
1292 const struct nv50_ir_prog_info
*driver
; // for driver configuration
1294 void releaseInstruction(Instruction
*);
1295 void releaseValue(Value
*);
1298 // TODO: add const version
1302 bool run(Program
*, bool ordered
= false, bool skipPhi
= false);
1303 bool run(Function
*, bool ordered
= false, bool skipPhi
= false);
1306 // return false to continue with next entity on next higher level
1307 virtual bool visit(Function
*) { return true; }
1308 virtual bool visit(BasicBlock
*) { return true; }
1309 virtual bool visit(Instruction
*) { return false; }
1311 bool doRun(Program
*, bool ordered
, bool skipPhi
);
1312 bool doRun(Function
*, bool ordered
, bool skipPhi
);
1320 // =============================================================================
1322 #include "codegen/nv50_ir_inlines.h"
1324 } // namespace nv50_ir
1326 #endif // __NV50_IR_H__