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 BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
18 * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF
19 * OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33 #include "nv50_ir_util.h"
34 #include "nv50_ir_graph.h"
36 #include "nv50_ir_driver.h"
44 OP_UNION
, // unify a new definition and several source values
45 OP_SPLIT
, // $r0d -> { $r0, $r1 } ($r0d and $r0/$r1 will be coalesced)
46 OP_MERGE
, // opposite of split, e.g. combine 2 32 bit into a 64 bit value
47 OP_CONSTRAINT
, // copy values into consecutive registers
58 OP_SAD
, // abs(src0 - src1) + src2
69 OP_SAT
, // CLAMP(f32, 0.0, 1.0)
74 OP_SET_AND
, // dst = (src0 CMP src1) & src2
78 OP_SELP
, // dst = src2 ? src0 : src1
79 OP_SLCT
, // dst = (src2 CMP 0) ? src0 : src1
86 OP_EXP
, // exponential (base M_E)
87 OP_LOG
, // natural logarithm
100 OP_BRKPT
, // breakpoint (not related to loops)
101 OP_JOINAT
, // push control flow convergence point
106 OP_VFETCH
, // indirection 0 in attribute space, indirection 1 is vertex base
107 OP_PFETCH
, // fetch base address of vertex src0 (immediate) [+ src1]
111 OP_EMIT
, // emit vertex
112 OP_RESTART
, // restart primitive
114 OP_TXB
, // texture bias
115 OP_TXL
, // texure lod
116 OP_TXF
, // texel fetch
117 OP_TXQ
, // texture size query
118 OP_TXD
, // texture derivatives
119 OP_TXG
, // texture gather
121 OP_SULD
, // surface load
122 OP_SUST
, // surface store
125 OP_RDSV
, // read system value
126 OP_WRSV
, // write system value
131 OP_POPCNT
, // bitcount(src0 & src1)
132 OP_INSBF
, // insert first src1[8:15] bits of src0 into src2 at src1[0:7]
137 #define NV50_IR_SUBOP_MUL_HIGH 1
138 #define NV50_IR_SUBOP_EMIT_RESTART 1
139 #define NV50_IR_SUBOP_LDC_IL 1
140 #define NV50_IR_SUBOP_LDC_IS 2
141 #define NV50_IR_SUBOP_LDC_ISL 3
142 #define NV50_IR_SUBOP_SHIFT_WRAP 1
153 TYPE_U64
, // 64 bit operations are only lowered after register allocation
165 CC_NEVER
= CC_FL
, // when used with FILE_FLAGS
168 CC_NOT_P
= CC_EQ
, // when used with FILE_PREDICATE
196 ROUND_M
, // towards -inf
197 ROUND_Z
, // towards 0
198 ROUND_P
, // towards +inf
199 ROUND_NI
, // nearest integer
200 ROUND_MI
, // to integer towards -inf
201 ROUND_ZI
, // to integer towards 0
202 ROUND_PI
, // to integer towards +inf
207 CACHE_CA
, // cache at all levels
208 CACHE_WB
= CACHE_CA
, // cache write back
209 CACHE_CG
, // cache at global level
210 CACHE_CS
, // cache streaming
211 CACHE_CV
, // cache as volatile
212 CACHE_WT
= CACHE_CV
// cache write-through
219 FILE_PREDICATE
, // boolean predicate
220 FILE_FLAGS
, // zero/sign/carry/overflow bits
240 TEX_TARGET_1D_SHADOW
,
241 TEX_TARGET_2D_SHADOW
,
242 TEX_TARGET_CUBE_SHADOW
,
245 TEX_TARGET_2D_MS_ARRAY
,
246 TEX_TARGET_CUBE_ARRAY
,
247 TEX_TARGET_1D_ARRAY_SHADOW
,
248 TEX_TARGET_2D_ARRAY_SHADOW
,
250 TEX_TARGET_RECT_SHADOW
,
251 TEX_TARGET_CUBE_ARRAY_SHADOW
,
263 SV_VERTEX_COUNT
, // gl_PatchVerticesIn
296 class CmpInstruction
;
297 class TexInstruction
;
298 class FlowInstruction
;
303 class ImmediateValue
;
308 int8_t fileIndex
; // signed, may be indirect for CONST[]
309 uint8_t size
; // this should match the Instruction type's size
310 DataType type
; // mainly for pretty printing
312 uint64_t u64
; // immediate values
322 int32_t offset
; // offset from 0 (base of address space)
323 int32_t id
; // register id (< 0 if virtual/unassigned)
331 // precedence: NOT after SAT after NEG after ABS
332 #define NV50_IR_MOD_ABS (1 << 0)
333 #define NV50_IR_MOD_NEG (1 << 1)
334 #define NV50_IR_MOD_SAT (1 << 2)
335 #define NV50_IR_MOD_NOT (1 << 3)
336 #define NV50_IR_MOD_NEG_ABS (NV50_IR_MOD_NEG | NV50_IR_MOD_ABS)
338 #define NV50_IR_INTERP_MODE_MASK 0x3
339 #define NV50_IR_INTERP_LINEAR (0 << 0)
340 #define NV50_IR_INTERP_PERSPECTIVE (1 << 0)
341 #define NV50_IR_INTERP_FLAT (2 << 0)
342 #define NV50_IR_INTERP_SC (3 << 0) // what exactly is that ?
343 #define NV50_IR_INTERP_SAMPLE_MASK 0xc
344 #define NV50_IR_INTERP_DEFAULT (0 << 2)
345 #define NV50_IR_INTERP_CENTROID (1 << 2)
346 #define NV50_IR_INTERP_OFFSET (2 << 2)
347 #define NV50_IR_INTERP_SAMPLEID (3 << 2)
349 // do we really want this to be a class ?
353 Modifier() : bits(0) { }
354 Modifier(unsigned int m
) : bits(m
) { }
355 Modifier(operation op
);
357 // @return new Modifier applying a after b (asserts if unrepresentable)
358 Modifier
operator*(const Modifier
) const;
359 Modifier
operator==(const Modifier m
) const { return m
.bits
== bits
; }
360 Modifier
operator!=(const Modifier m
) const { return m
.bits
!= bits
; }
362 inline Modifier
operator&(const Modifier m
) const { return bits
& m
.bits
; }
363 inline Modifier
operator|(const Modifier m
) const { return bits
| m
.bits
; }
364 inline Modifier
operator^(const Modifier m
) const { return bits
^ m
.bits
; }
366 operation
getOp() const;
368 inline int neg() const { return (bits
& NV50_IR_MOD_NEG
) ? 1 : 0; }
369 inline int abs() const { return (bits
& NV50_IR_MOD_ABS
) ? 1 : 0; }
371 inline operator bool() { return bits
? true : false; }
373 void applyTo(ImmediateValue
&imm
) const;
375 int print(char *buf
, size_t size
) const;
385 ValueRef(const ValueRef
&);
388 inline bool exists() const { return value
!= NULL
; }
391 void set(const ValueRef
&);
392 inline Value
*get() const { return value
; }
393 inline Value
*rep() const;
395 inline Instruction
*getInsn() const { return insn
; }
396 inline void setInsn(Instruction
*inst
) { insn
= inst
; }
398 inline bool isIndirect(int dim
) const { return indirect
[dim
] >= 0; }
399 inline const ValueRef
*getIndirect(int dim
) const;
401 inline DataFile
getFile() const;
402 inline unsigned getSize() const;
404 // SSA: return eventual (traverse MOVs) literal value, if it exists
405 ImmediateValue
*getImmediate() const;
409 int8_t indirect
[2]; // >= 0 if relative to lvalue in insn->src(indirect[i])
412 bool usedAsPtr
; // for printing
423 ValueDef(const ValueDef
&);
426 inline bool exists() const { return value
!= NULL
; }
428 inline Value
*get() const { return value
; }
429 inline Value
*rep() const;
431 void replace(Value
*, bool doSet
); // replace all uses of the old value
433 inline Instruction
*getInsn() const { return insn
; }
434 inline void setInsn(Instruction
*inst
) { insn
= inst
; }
436 inline DataFile
getFile() const;
437 inline unsigned getSize() const;
439 inline void setSSA(LValue
*);
440 inline const LValue
*preSSA() const;
443 Value
*value
; // should make this LValue * ...
444 LValue
*origin
; // pre SSA value
453 virtual Value
*clone(Function
*) const { return NULL
; }
455 virtual int print(char *, size_t, DataType ty
= TYPE_NONE
) const = 0;
457 virtual bool equals(const Value
*, bool strict
= false) const;
458 virtual bool interfers(const Value
*) const;
460 inline Value
*rep() const { return join
; }
462 inline Instruction
*getUniqueInsn() const;
463 inline Instruction
*getInsn() const; // use when uniqueness is certain
465 inline int refCount() { return uses
.size(); }
467 inline LValue
*asLValue();
468 inline Symbol
*asSym();
469 inline ImmediateValue
*asImm();
470 inline const Symbol
*asSym() const;
471 inline const ImmediateValue
*asImm() const;
473 bool coalesce(Value
*, bool force
= false);
475 inline bool inFile(DataFile f
) { return reg
.file
== f
; }
477 static inline Value
*get(Iterator
&);
479 std::list
<ValueRef
*> uses
;
480 std::list
<ValueDef
*> defs
;
481 typedef std::list
<ValueRef
*>::iterator UseIterator
;
482 typedef std::list
<ValueRef
*>::const_iterator UseCIterator
;
483 typedef std::list
<ValueDef
*>::iterator DefIterator
;
484 typedef std::list
<ValueDef
*>::const_iterator DefCIterator
;
489 // TODO: these should be in LValue:
494 class LValue
: public Value
497 LValue(Function
*, DataFile file
);
498 LValue(Function
*, LValue
*);
500 virtual Value
*clone(Function
*) const;
502 virtual int print(char *, size_t, DataType ty
= TYPE_NONE
) const;
510 class Symbol
: public Value
513 Symbol(Program
*, DataFile file
= FILE_MEMORY_CONST
, ubyte fileIdx
= 0);
515 virtual Value
*clone(Function
*) const;
517 virtual bool equals(const Value
*that
, bool strict
) const;
519 virtual int print(char *, size_t, DataType ty
= TYPE_NONE
) const;
521 // print with indirect values
522 int print(char *, size_t, Value
*, Value
*, DataType ty
= TYPE_NONE
) const;
524 inline void setFile(DataFile file
, ubyte fileIndex
= 0)
527 reg
.fileIndex
= fileIndex
;
530 inline void setOffset(int32_t offset
);
531 inline void setAddress(Symbol
*base
, int32_t offset
);
532 inline void setSV(SVSemantic sv
, uint32_t idx
= 0);
534 inline const Symbol
*getBase() const { return baseSym
; }
537 Symbol
*baseSym
; // array base for Symbols representing array elements
540 class ImmediateValue
: public Value
543 ImmediateValue(Program
*, uint32_t);
544 ImmediateValue(Program
*, float);
545 ImmediateValue(Program
*, double);
547 // NOTE: not added to program with
548 ImmediateValue(const ImmediateValue
*, DataType ty
);
550 virtual bool equals(const Value
*that
, bool strict
) const;
552 // these only work if 'type' is valid (we mostly use untyped literals):
553 bool isInteger(const int ival
) const; // ival is cast to this' type
554 bool isNegative() const;
559 // for constant folding:
560 ImmediateValue
operator+(const ImmediateValue
&) const;
561 ImmediateValue
operator-(const ImmediateValue
&) const;
562 ImmediateValue
operator*(const ImmediateValue
&) const;
563 ImmediateValue
operator/(const ImmediateValue
&) const;
565 bool compare(CondCode cc
, float fval
) const;
567 virtual int print(char *, size_t, DataType ty
= TYPE_NONE
) const;
574 Instruction(Function
*, operation
, DataType
);
575 virtual ~Instruction();
577 virtual Instruction
*clone(bool deep
) const;
579 void setDef(int i
, Value
*);
580 void setSrc(int s
, Value
*);
581 void setSrc(int s
, const ValueRef
&);
582 void swapSources(int a
, int b
);
583 bool setIndirect(int s
, int dim
, Value
*);
585 inline ValueRef
& src(int s
) { return srcs
[s
]; }
586 inline ValueDef
& def(int s
) { return defs
[s
]; }
587 inline const ValueRef
& src(int s
) const { return srcs
[s
]; }
588 inline const ValueDef
& def(int s
) const { return defs
[s
]; }
590 inline Value
*getDef(int d
) const { return defs
[d
].get(); }
591 inline Value
*getSrc(int s
) const { return srcs
[s
].get(); }
592 inline Value
*getIndirect(int s
, int dim
) const;
594 inline bool defExists(unsigned d
) const
596 return d
< defs
.size() && defs
[d
].exists();
598 inline bool srcExists(unsigned s
) const
600 return s
< srcs
.size() && srcs
[s
].exists();
603 inline bool constrainedDefs() const { return defExists(1); }
605 bool setPredicate(CondCode ccode
, Value
*);
606 inline Value
*getPredicate() const;
607 bool writesPredicate() const;
609 inline void setFlagsSrc(int s
, Value
*);
610 inline void setFlagsDef(int d
, Value
*);
612 unsigned int defCount() const { return defs
.size(); };
613 unsigned int defCount(unsigned int mask
) const;
614 unsigned int srcCount() const { return srcs
.size(); };
615 unsigned int srcCount(unsigned int mask
) const;
617 // save & remove / set indirect[0,1] and predicate source
618 void takeExtraSources(int s
, Value
*[3]);
619 void putExtraSources(int s
, Value
*[3]);
621 inline void setType(DataType type
) { dType
= sType
= type
; }
623 inline void setType(DataType dtype
, DataType stype
)
629 inline bool isPseudo() const { return op
< OP_MOV
; }
632 bool isCommutationLegal(const Instruction
*) const; // must be adjacent !
633 bool isActionEqual(const Instruction
*) const;
634 bool isResultEqual(const Instruction
*) const;
638 inline CmpInstruction
*asCmp();
639 inline TexInstruction
*asTex();
640 inline FlowInstruction
*asFlow();
641 inline const TexInstruction
*asTex() const;
642 inline const CmpInstruction
*asCmp() const;
643 inline const FlowInstruction
*asFlow() const;
649 int serial
; // CFG order
652 DataType dType
; // destination or defining type
653 DataType sType
; // source or secondary type
658 uint8_t subOp
; // quadop, 1 for mul-high, etc.
660 unsigned encSize
: 4; // encoding size in bytes
661 unsigned saturate
: 1; // to [0.0f, 1.0f]
662 unsigned join
: 1; // converge control flow (use OP_JOIN until end)
663 unsigned fixed
: 1; // prevent dead code elimination
664 unsigned terminator
: 1; // end of basic block
666 unsigned ftz
: 1; // flush denormal to zero
667 unsigned dnz
: 1; // denormals, NaN are zero
668 unsigned ipa
: 4; // interpolation mode
670 unsigned perPatch
: 1;
671 unsigned exit
: 1; // terminate program after insn
673 int8_t postFactor
; // MUL/DIV(if < 0) by 1 << postFactor
682 std::deque
<ValueDef
> defs
; // no gaps !
683 std::deque
<ValueRef
> srcs
; // no gaps !
685 // instruction specific methods:
686 // (don't want to subclass, would need more constructors and memory pools)
688 inline void setInterpolate(unsigned int mode
) { ipa
= mode
; }
690 unsigned int getInterpMode() const { return ipa
& 0x3; }
691 unsigned int getSampleMode() const { return ipa
& 0xc; }
696 void cloneBase(Instruction
*clone
, bool deep
) const;
710 class TexInstruction
: public Instruction
716 Target(TexTarget targ
= TEX_TARGET_2D
) : target(targ
) { }
718 const char *getName() const { return descTable
[target
].name
; }
719 unsigned int getArgCount() const { return descTable
[target
].argc
; }
720 unsigned int getDim() const { return descTable
[target
].dim
; }
721 int isArray() const { return descTable
[target
].array
? 1 : 0; }
722 int isCube() const { return descTable
[target
].cube
? 1 : 0; }
723 int isShadow() const { return descTable
[target
].shadow
? 1 : 0; }
725 Target
& operator=(TexTarget targ
)
727 assert(targ
< TEX_TARGET_COUNT
);
731 inline bool operator==(TexTarget targ
) const { return target
== targ
; }
744 static const struct Desc descTable
[TEX_TARGET_COUNT
];
747 enum TexTarget target
;
751 TexInstruction(Function
*, operation
);
752 virtual ~TexInstruction();
754 virtual Instruction
*clone(bool deep
) const;
756 inline void setTexture(Target targ
, uint8_t r
, uint8_t s
)
763 inline Value
*getIndirectR() const;
764 inline Value
*getIndirectS() const;
778 bool liveOnly
; // only execute on live pixels of a quad (optimization)
782 int8_t useOffsets
; // 0, 1, or 4 for textureGatherOffsets
792 class CmpInstruction
: public Instruction
795 CmpInstruction(Function
*, operation
);
797 virtual Instruction
*clone(bool deep
) const;
799 void setCondition(CondCode cond
) { setCond
= cond
; }
800 CondCode
getCondition() const { return setCond
; }
806 class FlowInstruction
: public Instruction
809 FlowInstruction(Function
*, operation
, BasicBlock
*target
);
812 unsigned allWarp
: 1;
813 unsigned absolute
: 1;
815 unsigned builtin
: 1; // true for calls to emulation code
827 BasicBlock(Function
*);
830 inline int getId() const { return id
; }
831 inline unsigned int getInsnCount() const { return numInsns
; }
832 inline bool isTerminated() const { return exit
&& exit
->terminator
; }
834 bool dominatedBy(BasicBlock
*bb
);
835 inline bool reachableBy(BasicBlock
*by
, BasicBlock
*term
);
837 // returns mask of conditional out blocks
838 // e.g. 3 for IF { .. } ELSE { .. } ENDIF, 1 for IF { .. } ENDIF
839 unsigned int initiatesSimpleConditional() const;
842 Function
*getFunction() const { return func
; }
843 Program
*getProgram() const { return program
; }
845 Instruction
*getEntry() const { return entry
; } // first non-phi instruction
846 Instruction
*getPhi() const { return phi
; }
847 Instruction
*getFirst() const { return phi
? phi
: entry
; }
848 Instruction
*getExit() const { return exit
; }
850 void insertHead(Instruction
*);
851 void insertTail(Instruction
*);
852 void insertBefore(Instruction
*, Instruction
*);
853 void insertAfter(Instruction
*, Instruction
*);
854 void remove(Instruction
*);
855 void permuteAdjacent(Instruction
*, Instruction
*);
857 BasicBlock
*idom() const;
859 // NOTE: currently does not rebuild the dominator tree
860 BasicBlock
*splitBefore(Instruction
*, bool attach
= true);
861 BasicBlock
*splitAfter(Instruction
*, bool attach
= true);
863 DLList
& getDF() { return df
; }
864 DLList::Iterator
iterDF() { return df
.iterator(); }
866 static inline BasicBlock
*get(Iterator
&);
867 static inline BasicBlock
*get(Graph::Node
*);
870 Graph::Node cfg
; // first edge is branch *taken* (the ELSE branch)
878 Instruction
*joinAt
; // for quick reference
880 bool explicitCont
; // loop headers: true if loop contains continue stmts
890 unsigned int numInsns
;
896 void splitCommon(Instruction
*, BasicBlock
*, bool attach
);
902 Function(Program
*, const char *name
);
905 inline Program
*getProgram() const { return prog
; }
906 inline const char *getName() const { return name
; }
907 inline int getId() const { return id
; }
910 void printLiveIntervals() const;
911 void printCFGraph(const char *filePath
);
913 bool setEntry(BasicBlock
*);
914 bool setExit(BasicBlock
*);
916 unsigned int orderInstructions(ArrayList
&);
918 inline void add(BasicBlock
*bb
, int& id
) { allBBlocks
.insert(bb
, id
); }
919 inline void add(Instruction
*insn
, int& id
) { allInsns
.insert(insn
, id
); }
920 inline void add(LValue
*lval
, int& id
) { allLValues
.insert(lval
, id
); }
922 inline LValue
*getLValue(int id
);
928 Graph::Node
*cfgExit
;
930 Graph::Node call
; // node in the call graph
932 BasicBlock
**bbArray
; // BBs in emission order
935 unsigned int loopNestingBound
;
941 ArrayList allBBlocks
;
943 ArrayList allLValues
;
946 void buildLiveSetsPreSSA(BasicBlock
*, const int sequence
);
950 const char *const name
;
957 CG_STAGE_SSA
, // expected directly before register allocation
967 TYPE_TESSELLATION_CONTROL
,
968 TYPE_TESSELLATION_EVAL
,
974 Program(Type type
, Target
*targ
);
979 Type
getType() const { return progType
; }
981 inline void add(Function
*fn
, int& id
) { allFuncs
.insert(fn
, id
); }
982 inline void add(Value
*rval
, int& id
) { allRValues
.insert(rval
, id
); }
984 bool makeFromTGSI(struct nv50_ir_prog_info
*);
985 bool makeFromSM4(struct nv50_ir_prog_info
*);
987 bool optimizeSSA(int level
);
988 bool optimizePostRA(int level
);
989 bool registerAllocation();
990 bool emitBinary(struct nv50_ir_prog_info
*);
992 const Target
*getTarget() const { return target
; }
1003 ArrayList allRValues
;
1010 MemoryPool mem_Instruction
;
1011 MemoryPool mem_CmpInstruction
;
1012 MemoryPool mem_TexInstruction
;
1013 MemoryPool mem_FlowInstruction
;
1014 MemoryPool mem_LValue
;
1015 MemoryPool mem_Symbol
;
1016 MemoryPool mem_ImmediateValue
;
1020 void releaseInstruction(Instruction
*);
1021 void releaseValue(Value
*);
1024 // TODO: add const version
1028 bool run(Program
*, bool ordered
= false, bool skipPhi
= false);
1029 bool run(Function
*, bool ordered
= false, bool skipPhi
= false);
1032 // return false to continue with next entity on next higher level
1033 virtual bool visit(Function
*) { return true; }
1034 virtual bool visit(BasicBlock
*) { return true; }
1035 virtual bool visit(Instruction
*) { return false; }
1037 bool doRun(Program
*, bool ordered
, bool skipPhi
);
1038 bool doRun(Function
*, bool ordered
, bool skipPhi
);
1046 // =============================================================================
1048 #include "nv50_ir_inlines.h"
1050 } // namespace nv50_ir
1052 #endif // __NV50_IR_H__