src/amd/llvm/ac_llvm_helper.cpp

   1 /*
   2  * Copyright 2014 Advanced Micro Devices, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the
   6  * "Software"), to deal in the Software without restriction, including
   7  * without limitation the rights to use, copy, modify, merge, publish,
   8  * distribute, sub license, and/or sell copies of the Software, and to
   9  * permit persons to whom the Software is furnished to do so, subject to
  10  * the following conditions:
  11  *
  12  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  13  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  14  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  15  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
  16  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  17  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  18  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  19  *
  20  * The above copyright notice and this permission notice (including the
  21  * next paragraph) shall be included in all copies or substantial portions
  22  * of the Software.
  23  *
  24  */
  25
  26 #include <cstring>
  27
  28 #include "ac_binary.h"
  29 #include "ac_llvm_util.h"
  30 #include "ac_llvm_build.h"
  31
  32 #include "util/macros.h"
  33
  34 #include <llvm-c/Core.h>
  35 #include <llvm/Target/TargetMachine.h>
  36 #include <llvm/IR/IRBuilder.h>
  37 #include <llvm/Analysis/TargetLibraryInfo.h>
  38 #include <llvm/Transforms/IPO.h>
  39
  40 #include <llvm/IR/LegacyPassManager.h>
  41
  42 void ac_add_attr_dereferenceable(LLVMValueRef val, uint64_t bytes)
  43 {
  44    llvm::Argument *A = llvm::unwrap<llvm::Argument>(val);
  45    A->addAttr(llvm::Attribute::getWithDereferenceableBytes(A->getContext(), bytes));
  46 }
  47
  48 bool ac_is_sgpr_param(LLVMValueRef arg)
  49 {
  50         llvm::Argument *A = llvm::unwrap<llvm::Argument>(arg);
  51         llvm::AttributeList AS = A->getParent()->getAttributes();
  52         unsigned ArgNo = A->getArgNo();
  53         return AS.hasAttribute(ArgNo + 1, llvm::Attribute::InReg);
  54 }
  55
  56 LLVMValueRef ac_llvm_get_called_value(LLVMValueRef call)
  57 {
  58         return LLVMGetCalledValue(call);
  59 }
  60
  61 bool ac_llvm_is_function(LLVMValueRef v)
  62 {
  63         return LLVMGetValueKind(v) == LLVMFunctionValueKind;
  64 }
  65
  66 LLVMModuleRef ac_create_module(LLVMTargetMachineRef tm, LLVMContextRef ctx)
  67 {
  68    llvm::TargetMachine *TM = reinterpret_cast<llvm::TargetMachine*>(tm);
  69    LLVMModuleRef module = LLVMModuleCreateWithNameInContext("mesa-shader", ctx);
  70
  71    llvm::unwrap(module)->setTargetTriple(TM->getTargetTriple().getTriple());
  72    llvm::unwrap(module)->setDataLayout(TM->createDataLayout());
  73    return module;
  74 }
  75
  76 LLVMBuilderRef ac_create_builder(LLVMContextRef ctx,
  77                                  enum ac_float_mode float_mode)
  78 {
  79         LLVMBuilderRef builder = LLVMCreateBuilderInContext(ctx);
  80
  81         llvm::FastMathFlags flags;
  82
  83         switch (float_mode) {
  84         case AC_FLOAT_MODE_DEFAULT:
  85         case AC_FLOAT_MODE_DENORM_FLUSH_TO_ZERO:
  86                 break;
  87
  88         case AC_FLOAT_MODE_DEFAULT_OPENGL:
  89                 /* Allow optimizations to treat the sign of a zero argument or
  90                  * result as insignificant.
  91                  */
  92                 flags.setNoSignedZeros(); /* nsz */
  93
  94                 /* Allow optimizations to use the reciprocal of an argument
  95                  * rather than perform division.
  96                  */
  97                 flags.setAllowReciprocal(); /* arcp */
  98
  99                 /* Allow floating-point contraction (e.g. fusing a multiply
 100                  * followed by an addition into a fused multiply-and-add).
 101                  */
 102                 flags.setAllowContract(); /* contract */
 103
 104                 llvm::unwrap(builder)->setFastMathFlags(flags);
 105                 break;
 106         }
 107
 108         return builder;
 109 }
 110
 111 /* Return the original state of inexact math. */
 112 bool ac_disable_inexact_math(LLVMBuilderRef builder)
 113 {
 114         auto *b = llvm::unwrap(builder);
 115         llvm::FastMathFlags flags = b->getFastMathFlags();
 116
 117         if (!flags.allowContract())
 118                 return false;
 119
 120         flags.setAllowContract(false);
 121         b->setFastMathFlags(flags);
 122         return true;
 123 }
 124
 125 void ac_restore_inexact_math(LLVMBuilderRef builder, bool value)
 126 {
 127         auto *b = llvm::unwrap(builder);
 128         llvm::FastMathFlags flags = b->getFastMathFlags();
 129
 130         if (flags.allowContract() == value)
 131                 return;
 132
 133         flags.setAllowContract(value);
 134         b->setFastMathFlags(flags);
 135 }
 136
 137 LLVMTargetLibraryInfoRef
 138 ac_create_target_library_info(const char *triple)
 139 {
 140         return reinterpret_cast<LLVMTargetLibraryInfoRef>(new llvm::TargetLibraryInfoImpl(llvm::Triple(triple)));
 141 }
 142
 143 void
 144 ac_dispose_target_library_info(LLVMTargetLibraryInfoRef library_info)
 145 {
 146         delete reinterpret_cast<llvm::TargetLibraryInfoImpl *>(library_info);
 147 }
 148
 149 /* Implementation of raw_pwrite_stream that works on malloc()ed memory for
 150  * better compatibility with C code. */
 151 struct raw_memory_ostream : public llvm::raw_pwrite_stream {
 152         char *buffer;
 153         size_t written;
 154         size_t bufsize;
 155
 156         raw_memory_ostream()
 157         {
 158                 buffer = NULL;
 159                 written = 0;
 160                 bufsize = 0;
 161                 SetUnbuffered();
 162         }
 163
 164         ~raw_memory_ostream()
 165         {
 166                 free(buffer);
 167         }
 168
 169         void clear()
 170         {
 171                 written = 0;
 172         }
 173
 174         void take(char *&out_buffer, size_t &out_size)
 175         {
 176                 out_buffer = buffer;
 177                 out_size = written;
 178                 buffer = NULL;
 179                 written = 0;
 180                 bufsize = 0;
 181         }
 182
 183         void flush() = delete;
 184
 185         void write_impl(const char *ptr, size_t size) override
 186         {
 187                 if (unlikely(written + size < written))
 188                         abort();
 189                 if (written + size > bufsize) {
 190                         bufsize = MAX3(1024, written + size, bufsize / 3 * 4);
 191                         buffer = (char *)realloc(buffer, bufsize);
 192                         if (!buffer) {
 193                                 fprintf(stderr, "amd: out of memory allocating ELF buffer\n");
 194                                 abort();
 195                         }
 196                 }
 197                 memcpy(buffer + written, ptr, size);
 198                 written += size;
 199         }
 200
 201         void pwrite_impl(const char *ptr, size_t size, uint64_t offset) override
 202         {
 203                 assert(offset == (size_t)offset &&
 204                        offset + size >= offset && offset + size <= written);
 205                 memcpy(buffer + offset, ptr, size);
 206         }
 207
 208         uint64_t current_pos() const override
 209         {
 210                 return written;
 211         }
 212 };
 213
 214 /* The LLVM compiler is represented as a pass manager containing passes for
 215  * optimizations, instruction selection, and code generation.
 216  */
 217 struct ac_compiler_passes {
 218         raw_memory_ostream ostream; /* ELF shader binary stream */
 219         llvm::legacy::PassManager passmgr; /* list of passes */
 220 };
 221
 222 struct ac_compiler_passes *ac_create_llvm_passes(LLVMTargetMachineRef tm)
 223 {
 224         struct ac_compiler_passes *p = new ac_compiler_passes();
 225         if (!p)
 226                 return NULL;
 227
 228         llvm::TargetMachine *TM = reinterpret_cast<llvm::TargetMachine*>(tm);
 229
 230         if (TM->addPassesToEmitFile(p->passmgr, p->ostream,
 231                                     nullptr,
 232 #if LLVM_VERSION_MAJOR >= 10
 233                                     llvm::CGFT_ObjectFile)) {
 234 #else
 235                                     llvm::TargetMachine::CGFT_ObjectFile)) {
 236 #endif
 237                 fprintf(stderr, "amd: TargetMachine can't emit a file of this type!\n");
 238                 delete p;
 239                 return NULL;
 240         }
 241         return p;
 242 }
 243
 244 void ac_destroy_llvm_passes(struct ac_compiler_passes *p)
 245 {
 246         delete p;
 247 }
 248
 249 /* This returns false on failure. */
 250 bool ac_compile_module_to_elf(struct ac_compiler_passes *p, LLVMModuleRef module,
 251                               char **pelf_buffer, size_t *pelf_size)
 252 {
 253         p->passmgr.run(*llvm::unwrap(module));
 254         p->ostream.take(*pelf_buffer, *pelf_size);
 255         return true;
 256 }
 257
 258 void ac_llvm_add_barrier_noop_pass(LLVMPassManagerRef passmgr)
 259 {
 260         llvm::unwrap(passmgr)->add(llvm::createBarrierNoopPass());
 261 }
 262
 263 void ac_enable_global_isel(LLVMTargetMachineRef tm)
 264 {
 265   reinterpret_cast<llvm::TargetMachine*>(tm)->setGlobalISel(true);
 266 }
 267
 268 LLVMValueRef ac_build_atomic_rmw(struct ac_llvm_context *ctx, LLVMAtomicRMWBinOp op,
 269                                  LLVMValueRef ptr, LLVMValueRef val,
 270                                  const char *sync_scope) {
 271         llvm::AtomicRMWInst::BinOp binop;
 272         switch (op) {
 273         case LLVMAtomicRMWBinOpXchg:
 274                 binop = llvm::AtomicRMWInst::Xchg;
 275                 break;
 276         case LLVMAtomicRMWBinOpAdd:
 277                 binop = llvm::AtomicRMWInst::Add;
 278                 break;
 279         case LLVMAtomicRMWBinOpSub:
 280                 binop = llvm::AtomicRMWInst::Sub;
 281                 break;
 282         case LLVMAtomicRMWBinOpAnd:
 283                 binop = llvm::AtomicRMWInst::And;
 284                 break;
 285         case LLVMAtomicRMWBinOpNand:
 286                 binop = llvm::AtomicRMWInst::Nand;
 287                 break;
 288         case LLVMAtomicRMWBinOpOr:
 289                 binop = llvm::AtomicRMWInst::Or;
 290                 break;
 291         case LLVMAtomicRMWBinOpXor:
 292                 binop = llvm::AtomicRMWInst::Xor;
 293                 break;
 294         case LLVMAtomicRMWBinOpMax:
 295                 binop = llvm::AtomicRMWInst::Max;
 296                 break;
 297         case LLVMAtomicRMWBinOpMin:
 298                 binop = llvm::AtomicRMWInst::Min;
 299                 break;
 300         case LLVMAtomicRMWBinOpUMax:
 301                 binop = llvm::AtomicRMWInst::UMax;
 302                 break;
 303         case LLVMAtomicRMWBinOpUMin:
 304                 binop = llvm::AtomicRMWInst::UMin;
 305                 break;
 306         default:
 307                 unreachable(!"invalid LLVMAtomicRMWBinOp");
 308            break;
 309         }
 310         unsigned SSID = llvm::unwrap(ctx->context)->getOrInsertSyncScopeID(sync_scope);
 311         return llvm::wrap(llvm::unwrap(ctx->builder)->CreateAtomicRMW(
 312                 binop, llvm::unwrap(ptr), llvm::unwrap(val),
 313                 llvm::AtomicOrdering::SequentiallyConsistent, SSID));
 314 }
 315
 316 LLVMValueRef ac_build_atomic_cmp_xchg(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 317                                       LLVMValueRef cmp, LLVMValueRef val,
 318                                       const char *sync_scope) {
 319         unsigned SSID = llvm::unwrap(ctx->context)->getOrInsertSyncScopeID(sync_scope);
 320         return llvm::wrap(llvm::unwrap(ctx->builder)->CreateAtomicCmpXchg(
 321                           llvm::unwrap(ptr), llvm::unwrap(cmp), llvm::unwrap(val),
 322                           llvm::AtomicOrdering::SequentiallyConsistent,
 323                           llvm::AtomicOrdering::SequentiallyConsistent, SSID));
 324 }