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 <llvm-c/Core.h>
  29 #include <llvm/Target/TargetMachine.h>
  30 #include <llvm/IR/IRBuilder.h>
  31 #include <llvm/Analysis/TargetLibraryInfo.h>
  32 #include <llvm/Transforms/IPO.h>
  33
  34 #include <llvm/IR/LegacyPassManager.h>
  35
  36 /* DO NOT REORDER THE HEADERS
  37  * The LLVM headers need to all be included before any Mesa header,
  38  * as they use the `restrict` keyword in ways that are incompatible
  39  * with our #define in include/c99_compat.h
  40  */
  41
  42 #include "ac_binary.h"
  43 #include "ac_llvm_util.h"
  44 #include "ac_llvm_build.h"
  45
  46 #include "util/macros.h"
  47
  48 void ac_add_attr_dereferenceable(LLVMValueRef val, uint64_t bytes)
  49 {
  50    llvm::Argument *A = llvm::unwrap<llvm::Argument>(val);
  51    A->addAttr(llvm::Attribute::getWithDereferenceableBytes(A->getContext(), bytes));
  52 }
  53
  54 void ac_add_attr_alignment(LLVMValueRef val, uint64_t bytes)
  55 {
  56 #if LLVM_VERSION_MAJOR >= 10
  57         llvm::Argument *A = llvm::unwrap<llvm::Argument>(val);
  58         A->addAttr(llvm::Attribute::getWithAlignment(A->getContext(), llvm::Align(bytes)));
  59 #else
  60         /* Avoid unused parameter warnings. */
  61         (void)val;
  62         (void)bytes;
  63 #endif
  64 }
  65
  66 bool ac_is_sgpr_param(LLVMValueRef arg)
  67 {
  68         llvm::Argument *A = llvm::unwrap<llvm::Argument>(arg);
  69         llvm::AttributeList AS = A->getParent()->getAttributes();
  70         unsigned ArgNo = A->getArgNo();
  71         return AS.hasAttribute(ArgNo + 1, llvm::Attribute::InReg);
  72 }
  73
  74 LLVMValueRef ac_llvm_get_called_value(LLVMValueRef call)
  75 {
  76         return LLVMGetCalledValue(call);
  77 }
  78
  79 bool ac_llvm_is_function(LLVMValueRef v)
  80 {
  81         return LLVMGetValueKind(v) == LLVMFunctionValueKind;
  82 }
  83
  84 LLVMModuleRef ac_create_module(LLVMTargetMachineRef tm, LLVMContextRef ctx)
  85 {
  86    llvm::TargetMachine *TM = reinterpret_cast<llvm::TargetMachine*>(tm);
  87    LLVMModuleRef module = LLVMModuleCreateWithNameInContext("mesa-shader", ctx);
  88
  89    llvm::unwrap(module)->setTargetTriple(TM->getTargetTriple().getTriple());
  90    llvm::unwrap(module)->setDataLayout(TM->createDataLayout());
  91    return module;
  92 }
  93
  94 LLVMBuilderRef ac_create_builder(LLVMContextRef ctx,
  95                                  enum ac_float_mode float_mode)
  96 {
  97         LLVMBuilderRef builder = LLVMCreateBuilderInContext(ctx);
  98
  99         llvm::FastMathFlags flags;
 100
 101         switch (float_mode) {
 102         case AC_FLOAT_MODE_DEFAULT:
 103         case AC_FLOAT_MODE_DENORM_FLUSH_TO_ZERO:
 104                 break;
 105
 106         case AC_FLOAT_MODE_DEFAULT_OPENGL:
 107                 /* Allow optimizations to treat the sign of a zero argument or
 108                  * result as insignificant.
 109                  */
 110                 flags.setNoSignedZeros(); /* nsz */
 111
 112                 /* Allow optimizations to use the reciprocal of an argument
 113                  * rather than perform division.
 114                  */
 115                 flags.setAllowReciprocal(); /* arcp */
 116
 117                 llvm::unwrap(builder)->setFastMathFlags(flags);
 118                 break;
 119         }
 120
 121         return builder;
 122 }
 123
 124 void ac_enable_signed_zeros(struct ac_llvm_context *ctx)
 125 {
 126         if (ctx->float_mode == AC_FLOAT_MODE_DEFAULT_OPENGL) {
 127                 auto *b = llvm::unwrap(ctx->builder);
 128                 llvm::FastMathFlags flags = b->getFastMathFlags();
 129
 130                 /* This disables the optimization of (x + 0), which is used
 131                  * to convert negative zero to positive zero.
 132                  */
 133                 flags.setNoSignedZeros(false);
 134                 b->setFastMathFlags(flags);
 135         }
 136 }
 137
 138 void ac_disable_signed_zeros(struct ac_llvm_context *ctx)
 139 {
 140         if (ctx->float_mode == AC_FLOAT_MODE_DEFAULT_OPENGL) {
 141                 auto *b = llvm::unwrap(ctx->builder);
 142                 llvm::FastMathFlags flags = b->getFastMathFlags();
 143
 144                 flags.setNoSignedZeros();
 145                 b->setFastMathFlags(flags);
 146         }
 147 }
 148
 149 LLVMTargetLibraryInfoRef
 150 ac_create_target_library_info(const char *triple)
 151 {
 152         return reinterpret_cast<LLVMTargetLibraryInfoRef>(new llvm::TargetLibraryInfoImpl(llvm::Triple(triple)));
 153 }
 154
 155 void
 156 ac_dispose_target_library_info(LLVMTargetLibraryInfoRef library_info)
 157 {
 158         delete reinterpret_cast<llvm::TargetLibraryInfoImpl *>(library_info);
 159 }
 160
 161 /* Implementation of raw_pwrite_stream that works on malloc()ed memory for
 162  * better compatibility with C code. */
 163 struct raw_memory_ostream : public llvm::raw_pwrite_stream {
 164         char *buffer;
 165         size_t written;
 166         size_t bufsize;
 167
 168         raw_memory_ostream()
 169         {
 170                 buffer = NULL;
 171                 written = 0;
 172                 bufsize = 0;
 173                 SetUnbuffered();
 174         }
 175
 176         ~raw_memory_ostream()
 177         {
 178                 free(buffer);
 179         }
 180
 181         void clear()
 182         {
 183                 written = 0;
 184         }
 185
 186         void take(char *&out_buffer, size_t &out_size)
 187         {
 188                 out_buffer = buffer;
 189                 out_size = written;
 190                 buffer = NULL;
 191                 written = 0;
 192                 bufsize = 0;
 193         }
 194
 195         void flush() = delete;
 196
 197         void write_impl(const char *ptr, size_t size) override
 198         {
 199                 if (unlikely(written + size < written))
 200                         abort();
 201                 if (written + size > bufsize) {
 202                         bufsize = MAX3(1024, written + size, bufsize / 3 * 4);
 203                         buffer = (char *)realloc(buffer, bufsize);
 204                         if (!buffer) {
 205                                 fprintf(stderr, "amd: out of memory allocating ELF buffer\n");
 206                                 abort();
 207                         }
 208                 }
 209                 memcpy(buffer + written, ptr, size);
 210                 written += size;
 211         }
 212
 213         void pwrite_impl(const char *ptr, size_t size, uint64_t offset) override
 214         {
 215                 assert(offset == (size_t)offset &&
 216                        offset + size >= offset && offset + size <= written);
 217                 memcpy(buffer + offset, ptr, size);
 218         }
 219
 220         uint64_t current_pos() const override
 221         {
 222                 return written;
 223         }
 224 };
 225
 226 /* The LLVM compiler is represented as a pass manager containing passes for
 227  * optimizations, instruction selection, and code generation.
 228  */
 229 struct ac_compiler_passes {
 230         raw_memory_ostream ostream; /* ELF shader binary stream */
 231         llvm::legacy::PassManager passmgr; /* list of passes */
 232 };
 233
 234 struct ac_compiler_passes *ac_create_llvm_passes(LLVMTargetMachineRef tm)
 235 {
 236         struct ac_compiler_passes *p = new ac_compiler_passes();
 237         if (!p)
 238                 return NULL;
 239
 240         llvm::TargetMachine *TM = reinterpret_cast<llvm::TargetMachine*>(tm);
 241
 242         if (TM->addPassesToEmitFile(p->passmgr, p->ostream,
 243                                     nullptr,
 244 #if LLVM_VERSION_MAJOR >= 10
 245                                     llvm::CGFT_ObjectFile)) {
 246 #else
 247                                     llvm::TargetMachine::CGFT_ObjectFile)) {
 248 #endif
 249                 fprintf(stderr, "amd: TargetMachine can't emit a file of this type!\n");
 250                 delete p;
 251                 return NULL;
 252         }
 253         return p;
 254 }
 255
 256 void ac_destroy_llvm_passes(struct ac_compiler_passes *p)
 257 {
 258         delete p;
 259 }
 260
 261 /* This returns false on failure. */
 262 bool ac_compile_module_to_elf(struct ac_compiler_passes *p, LLVMModuleRef module,
 263                               char **pelf_buffer, size_t *pelf_size)
 264 {
 265         p->passmgr.run(*llvm::unwrap(module));
 266         p->ostream.take(*pelf_buffer, *pelf_size);
 267         return true;
 268 }
 269
 270 void ac_llvm_add_barrier_noop_pass(LLVMPassManagerRef passmgr)
 271 {
 272         llvm::unwrap(passmgr)->add(llvm::createBarrierNoopPass());
 273 }
 274
 275 void ac_enable_global_isel(LLVMTargetMachineRef tm)
 276 {
 277   reinterpret_cast<llvm::TargetMachine*>(tm)->setGlobalISel(true);
 278 }
 279
 280 LLVMValueRef ac_build_atomic_rmw(struct ac_llvm_context *ctx, LLVMAtomicRMWBinOp op,
 281                                  LLVMValueRef ptr, LLVMValueRef val,
 282                                  const char *sync_scope) {
 283         llvm::AtomicRMWInst::BinOp binop;
 284         switch (op) {
 285         case LLVMAtomicRMWBinOpXchg:
 286                 binop = llvm::AtomicRMWInst::Xchg;
 287                 break;
 288         case LLVMAtomicRMWBinOpAdd:
 289                 binop = llvm::AtomicRMWInst::Add;
 290                 break;
 291         case LLVMAtomicRMWBinOpSub:
 292                 binop = llvm::AtomicRMWInst::Sub;
 293                 break;
 294         case LLVMAtomicRMWBinOpAnd:
 295                 binop = llvm::AtomicRMWInst::And;
 296                 break;
 297         case LLVMAtomicRMWBinOpNand:
 298                 binop = llvm::AtomicRMWInst::Nand;
 299                 break;
 300         case LLVMAtomicRMWBinOpOr:
 301                 binop = llvm::AtomicRMWInst::Or;
 302                 break;
 303         case LLVMAtomicRMWBinOpXor:
 304                 binop = llvm::AtomicRMWInst::Xor;
 305                 break;
 306         case LLVMAtomicRMWBinOpMax:
 307                 binop = llvm::AtomicRMWInst::Max;
 308                 break;
 309         case LLVMAtomicRMWBinOpMin:
 310                 binop = llvm::AtomicRMWInst::Min;
 311                 break;
 312         case LLVMAtomicRMWBinOpUMax:
 313                 binop = llvm::AtomicRMWInst::UMax;
 314                 break;
 315         case LLVMAtomicRMWBinOpUMin:
 316                 binop = llvm::AtomicRMWInst::UMin;
 317                 break;
 318         default:
 319                 unreachable(!"invalid LLVMAtomicRMWBinOp");
 320            break;
 321         }
 322         unsigned SSID = llvm::unwrap(ctx->context)->getOrInsertSyncScopeID(sync_scope);
 323         return llvm::wrap(llvm::unwrap(ctx->builder)->CreateAtomicRMW(
 324                 binop, llvm::unwrap(ptr), llvm::unwrap(val),
 325                 llvm::AtomicOrdering::SequentiallyConsistent, SSID));
 326 }
 327
 328 LLVMValueRef ac_build_atomic_cmp_xchg(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 329                                       LLVMValueRef cmp, LLVMValueRef val,
 330                                       const char *sync_scope) {
 331         unsigned SSID = llvm::unwrap(ctx->context)->getOrInsertSyncScopeID(sync_scope);
 332         return llvm::wrap(llvm::unwrap(ctx->builder)->CreateAtomicCmpXchg(
 333                           llvm::unwrap(ptr), llvm::unwrap(cmp), llvm::unwrap(val),
 334                           llvm::AtomicOrdering::SequentiallyConsistent,
 335                           llvm::AtomicOrdering::SequentiallyConsistent, SSID));
 336 }