| 1 |
/* |
| 2 |
* Copyright 2011 Christoph Bumiller |
| 3 |
* |
| 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: |
| 10 |
* |
| 11 |
* The above copyright notice and this permission notice shall be included in |
| 12 |
* all copies or substantial portions of the Software. |
| 13 |
* |
| 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. |
| 21 |
*/ |
| 22 |
|
| 23 |
#ifndef __NV50_IR_TARGET_H__ |
| 24 |
#define __NV50_IR_TARGET_H__ |
| 25 |
|
| 26 |
#include "nv50_ir.h" |
| 27 |
|
| 28 |
namespace nv50_ir { |
| 29 |
|
| 30 |
struct RelocInfo; |
| 31 |
|
| 32 |
struct RelocEntry |
| 33 |
{ |
| 34 |
enum Type |
| 35 |
{ |
| 36 |
TYPE_CODE, |
| 37 |
TYPE_BUILTIN, |
| 38 |
TYPE_DATA |
| 39 |
}; |
| 40 |
|
| 41 |
uint32_t data; |
| 42 |
uint32_t mask; |
| 43 |
uint32_t offset; |
| 44 |
int8_t bitPos; |
| 45 |
Type type; |
| 46 |
|
| 47 |
inline void apply(uint32_t *binary, const RelocInfo *info) const; |
| 48 |
}; |
| 49 |
|
| 50 |
struct RelocInfo |
| 51 |
{ |
| 52 |
uint32_t codePos; |
| 53 |
uint32_t libPos; |
| 54 |
uint32_t dataPos; |
| 55 |
|
| 56 |
uint32_t count; |
| 57 |
|
| 58 |
RelocEntry entry[0]; |
| 59 |
}; |
| 60 |
|
| 61 |
struct FixupData { |
| 62 |
FixupData(bool force, bool flat, uint8_t alphatest, bool msaa) : |
| 63 |
force_persample_interp(force), flatshade(flat), alphatest(alphatest), msaa(msaa) {} |
| 64 |
bool force_persample_interp; |
| 65 |
bool flatshade; |
| 66 |
uint8_t alphatest; |
| 67 |
bool msaa; |
| 68 |
}; |
| 69 |
|
| 70 |
struct FixupEntry; |
| 71 |
typedef void (*FixupApply)(const FixupEntry*, uint32_t*, const FixupData&); |
| 72 |
|
| 73 |
struct FixupEntry |
| 74 |
{ |
| 75 |
FixupEntry(FixupApply apply, int ipa, int reg, int loc) : |
| 76 |
apply(apply), ipa(ipa), reg(reg), loc(loc) {} |
| 77 |
|
| 78 |
FixupApply apply; |
| 79 |
union { |
| 80 |
struct { |
| 81 |
uint32_t ipa:4; // SC mode used to identify colors |
| 82 |
uint32_t reg:8; // The reg used for perspective division |
| 83 |
uint32_t loc:20; // Let's hope we don't have more than 1M-sized shaders |
| 84 |
}; |
| 85 |
uint32_t val; |
| 86 |
}; |
| 87 |
}; |
| 88 |
|
| 89 |
struct FixupInfo |
| 90 |
{ |
| 91 |
uint32_t count; |
| 92 |
FixupEntry entry[0]; |
| 93 |
}; |
| 94 |
|
| 95 |
class CodeEmitter |
| 96 |
{ |
| 97 |
public: |
| 98 |
CodeEmitter(const Target *); |
| 99 |
virtual ~CodeEmitter() { } |
| 100 |
|
| 101 |
// returns whether the instruction was encodable and written |
| 102 |
virtual bool emitInstruction(Instruction *) = 0; |
| 103 |
|
| 104 |
virtual uint32_t getMinEncodingSize(const Instruction *) const = 0; |
| 105 |
|
| 106 |
void setCodeLocation(void *, uint32_t size); |
| 107 |
inline void *getCodeLocation() const { return code; } |
| 108 |
inline uint32_t getCodeSize() const { return codeSize; } |
| 109 |
|
| 110 |
bool addReloc(RelocEntry::Type, int w, uint32_t data, uint32_t m, |
| 111 |
int s); |
| 112 |
|
| 113 |
inline void *getRelocInfo() const { return relocInfo; } |
| 114 |
|
| 115 |
bool addInterp(int ipa, int reg, FixupApply apply); |
| 116 |
inline void *getFixupInfo() const { return fixupInfo; } |
| 117 |
|
| 118 |
virtual void prepareEmission(Program *); |
| 119 |
virtual void prepareEmission(Function *); |
| 120 |
virtual void prepareEmission(BasicBlock *); |
| 121 |
|
| 122 |
void printBinary() const; |
| 123 |
|
| 124 |
protected: |
| 125 |
const Target *targ; |
| 126 |
|
| 127 |
uint32_t *code; |
| 128 |
uint32_t codeSize; |
| 129 |
uint32_t codeSizeLimit; |
| 130 |
|
| 131 |
RelocInfo *relocInfo; |
| 132 |
FixupInfo *fixupInfo; |
| 133 |
}; |
| 134 |
|
| 135 |
|
| 136 |
enum OpClass |
| 137 |
{ |
| 138 |
OPCLASS_MOVE = 0, |
| 139 |
OPCLASS_LOAD = 1, |
| 140 |
OPCLASS_STORE = 2, |
| 141 |
OPCLASS_ARITH = 3, |
| 142 |
OPCLASS_SHIFT = 4, |
| 143 |
OPCLASS_SFU = 5, |
| 144 |
OPCLASS_LOGIC = 6, |
| 145 |
OPCLASS_COMPARE = 7, |
| 146 |
OPCLASS_CONVERT = 8, |
| 147 |
OPCLASS_ATOMIC = 9, |
| 148 |
OPCLASS_TEXTURE = 10, |
| 149 |
OPCLASS_SURFACE = 11, |
| 150 |
OPCLASS_FLOW = 12, |
| 151 |
OPCLASS_PSEUDO = 14, |
| 152 |
OPCLASS_VECTOR = 15, |
| 153 |
OPCLASS_BITFIELD = 16, |
| 154 |
OPCLASS_CONTROL = 17, |
| 155 |
OPCLASS_OTHER = 18 |
| 156 |
}; |
| 157 |
|
| 158 |
class Target |
| 159 |
{ |
| 160 |
public: |
| 161 |
Target(bool m, bool j, bool s) : hasJoin(m), joinAnterior(j), hasSWSched(s) { } |
| 162 |
virtual ~Target() { } |
| 163 |
|
| 164 |
static Target *create(uint32_t chipset); |
| 165 |
static void destroy(Target *); |
| 166 |
|
| 167 |
// 0x50 and 0x84 to 0xaf for nv50 |
| 168 |
// 0xc0 to 0xdf for nvc0 |
| 169 |
inline uint32_t getChipset() const { return chipset; } |
| 170 |
|
| 171 |
virtual CodeEmitter *getCodeEmitter(Program::Type) = 0; |
| 172 |
|
| 173 |
// Drivers should upload this so we can use it from all programs. |
| 174 |
// The address chosen is supplied to the relocation routine. |
| 175 |
virtual void getBuiltinCode(const uint32_t **code, uint32_t *size) const = 0; |
| 176 |
|
| 177 |
virtual void parseDriverInfo(const struct nv50_ir_prog_info *info, |
| 178 |
const struct nv50_ir_prog_info_out *info_out) { |
| 179 |
if (info_out->type == MESA_SHADER_COMPUTE) { |
| 180 |
threads = info->prop.cp.numThreads[0] * |
| 181 |
info->prop.cp.numThreads[1] * |
| 182 |
info->prop.cp.numThreads[2]; |
| 183 |
if (threads == 0) |
| 184 |
threads = info->target >= NVISA_GK104_CHIPSET ? 1024 : 512; |
| 185 |
} else { |
| 186 |
threads = 32; // doesn't matter, just not too big. |
| 187 |
} |
| 188 |
} |
| 189 |
|
| 190 |
virtual bool runLegalizePass(Program *, CGStage stage) const = 0; |
| 191 |
|
| 192 |
public: |
| 193 |
struct OpInfo |
| 194 |
{ |
| 195 |
OpInfo *variants; |
| 196 |
operation op; |
| 197 |
uint16_t srcTypes; |
| 198 |
uint16_t dstTypes; |
| 199 |
uint32_t immdBits; |
| 200 |
uint8_t srcNr; |
| 201 |
uint8_t srcMods[3]; |
| 202 |
uint8_t dstMods; |
| 203 |
uint16_t srcFiles[3]; |
| 204 |
uint16_t dstFiles; |
| 205 |
unsigned int minEncSize : 5; |
| 206 |
unsigned int vector : 1; |
| 207 |
unsigned int predicate : 1; |
| 208 |
unsigned int commutative : 1; |
| 209 |
unsigned int pseudo : 1; |
| 210 |
unsigned int flow : 1; |
| 211 |
unsigned int hasDest : 1; |
| 212 |
unsigned int terminator : 1; |
| 213 |
}; |
| 214 |
|
| 215 |
inline const OpInfo& getOpInfo(const Instruction *) const; |
| 216 |
inline const OpInfo& getOpInfo(const operation) const; |
| 217 |
|
| 218 |
inline DataFile nativeFile(DataFile f) const; |
| 219 |
|
| 220 |
virtual bool insnCanLoad(const Instruction *insn, int s, |
| 221 |
const Instruction *ld) const = 0; |
| 222 |
virtual bool insnCanLoadOffset(const Instruction *insn, int s, |
| 223 |
int offset) const = 0; |
| 224 |
virtual bool isOpSupported(operation, DataType) const = 0; |
| 225 |
virtual bool isAccessSupported(DataFile, DataType) const = 0; |
| 226 |
virtual bool isModSupported(const Instruction *, |
| 227 |
int s, Modifier) const = 0; |
| 228 |
virtual bool isSatSupported(const Instruction *) const = 0; |
| 229 |
virtual bool isPostMultiplySupported(operation op, float f, |
| 230 |
int& e) const { return false; } |
| 231 |
virtual bool mayPredicate(const Instruction *, |
| 232 |
const Value *) const = 0; |
| 233 |
|
| 234 |
// whether @insn can be issued together with @next (order matters) |
| 235 |
virtual bool canDualIssue(const Instruction *insn, |
| 236 |
const Instruction *next) const { return false; } |
| 237 |
virtual int getLatency(const Instruction *) const { return 1; } |
| 238 |
virtual int getThroughput(const Instruction *) const { return 1; } |
| 239 |
|
| 240 |
virtual unsigned int getFileSize(DataFile) const = 0; |
| 241 |
virtual unsigned int getFileUnit(DataFile) const = 0; |
| 242 |
|
| 243 |
virtual uint32_t getSVAddress(DataFile, const Symbol *) const = 0; |
| 244 |
|
| 245 |
public: |
| 246 |
const bool hasJoin; // true if instructions have a join modifier |
| 247 |
const bool joinAnterior; // true if join is executed before the op |
| 248 |
const bool hasSWSched; // true if code should provide scheduling data |
| 249 |
|
| 250 |
static const uint8_t operationSrcNr[]; |
| 251 |
static const OpClass operationClass[]; |
| 252 |
|
| 253 |
static inline uint8_t getOpSrcNr(operation op) |
| 254 |
{ |
| 255 |
return operationSrcNr[op]; |
| 256 |
} |
| 257 |
static inline OpClass getOpClass(operation op) |
| 258 |
{ |
| 259 |
return operationClass[op]; |
| 260 |
} |
| 261 |
|
| 262 |
protected: |
| 263 |
uint32_t chipset; |
| 264 |
uint32_t threads; |
| 265 |
|
| 266 |
DataFile nativeFileMap[DATA_FILE_COUNT]; |
| 267 |
|
| 268 |
OpInfo opInfo[OP_LAST + 1]; |
| 269 |
}; |
| 270 |
|
| 271 |
const Target::OpInfo& Target::getOpInfo(const Instruction *insn) const |
| 272 |
{ |
| 273 |
return opInfo[MIN2(insn->op, OP_LAST)]; |
| 274 |
} |
| 275 |
|
| 276 |
const Target::OpInfo& Target::getOpInfo(const operation op) const |
| 277 |
{ |
| 278 |
return opInfo[op]; |
| 279 |
} |
| 280 |
|
| 281 |
inline DataFile Target::nativeFile(DataFile f) const |
| 282 |
{ |
| 283 |
return nativeFileMap[f]; |
| 284 |
} |
| 285 |
|
| 286 |
} // namespace nv50_ir |
| 287 |
|
| 288 |
#endif // __NV50_IR_TARGET_H__ |