src/amd/common/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 /* based on Marek's patch to lp_bld_misc.cpp */
  27
  28 // Workaround http://llvm.org/PR23628
  29 #pragma push_macro("DEBUG")
  30 #undef DEBUG
  31
  32 #include <cstring>
  33
  34 #include "ac_binary.h"
  35 #include "ac_llvm_util.h"
  36 #include "ac_llvm_build.h"
  37
  38 #include "util/macros.h"
  39
  40 #include <llvm-c/Core.h>
  41 #include <llvm/Target/TargetMachine.h>
  42 #include <llvm/IR/IRBuilder.h>
  43 #include <llvm/Analysis/TargetLibraryInfo.h>
  44 #include <llvm/Transforms/IPO.h>
  45
  46 #include <llvm/IR/LegacyPassManager.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 bool ac_is_sgpr_param(LLVMValueRef arg)
  55 {
  56         llvm::Argument *A = llvm::unwrap<llvm::Argument>(arg);
  57         llvm::AttributeList AS = A->getParent()->getAttributes();
  58         unsigned ArgNo = A->getArgNo();
  59         return AS.hasAttribute(ArgNo + 1, llvm::Attribute::InReg);
  60 }
  61
  62 LLVMValueRef ac_llvm_get_called_value(LLVMValueRef call)
  63 {
  64         return LLVMGetCalledValue(call);
  65 }
  66
  67 bool ac_llvm_is_function(LLVMValueRef v)
  68 {
  69         return LLVMGetValueKind(v) == LLVMFunctionValueKind;
  70 }
  71
  72 LLVMModuleRef ac_create_module(LLVMTargetMachineRef tm, LLVMContextRef ctx)
  73 {
  74    llvm::TargetMachine *TM = reinterpret_cast<llvm::TargetMachine*>(tm);
  75    LLVMModuleRef module = LLVMModuleCreateWithNameInContext("mesa-shader", ctx);
  76
  77    llvm::unwrap(module)->setTargetTriple(TM->getTargetTriple().getTriple());
  78    llvm::unwrap(module)->setDataLayout(TM->createDataLayout());
  79    return module;
  80 }
  81
  82 LLVMBuilderRef ac_create_builder(LLVMContextRef ctx,
  83                                  enum ac_float_mode float_mode)
  84 {
  85         LLVMBuilderRef builder = LLVMCreateBuilderInContext(ctx);
  86
  87         llvm::FastMathFlags flags;
  88
  89         switch (float_mode) {
  90         case AC_FLOAT_MODE_DEFAULT:
  91                 break;
  92         case AC_FLOAT_MODE_NO_SIGNED_ZEROS_FP_MATH:
  93                 flags.setNoSignedZeros();
  94                 llvm::unwrap(builder)->setFastMathFlags(flags);
  95                 break;
  96         case AC_FLOAT_MODE_UNSAFE_FP_MATH:
  97                 flags.setFast();
  98                 llvm::unwrap(builder)->setFastMathFlags(flags);
  99                 break;
 100         }
 101
 102         return builder;
 103 }
 104
 105 LLVMTargetLibraryInfoRef
 106 ac_create_target_library_info(const char *triple)
 107 {
 108         return reinterpret_cast<LLVMTargetLibraryInfoRef>(new llvm::TargetLibraryInfoImpl(llvm::Triple(triple)));
 109 }
 110
 111 void
 112 ac_dispose_target_library_info(LLVMTargetLibraryInfoRef library_info)
 113 {
 114         delete reinterpret_cast<llvm::TargetLibraryInfoImpl *>(library_info);
 115 }
 116
 117 /* Implementation of raw_pwrite_stream that works on malloc()ed memory for
 118  * better compatibility with C code. */
 119 struct raw_memory_ostream : public llvm::raw_pwrite_stream {
 120         char *buffer;
 121         size_t written;
 122         size_t bufsize;
 123
 124         raw_memory_ostream()
 125         {
 126                 buffer = NULL;
 127                 written = 0;
 128                 bufsize = 0;
 129                 SetUnbuffered();
 130         }
 131
 132         ~raw_memory_ostream()
 133         {
 134                 free(buffer);
 135         }
 136
 137         void clear()
 138         {
 139                 written = 0;
 140         }
 141
 142         void take(char *&out_buffer, size_t &out_size)
 143         {
 144                 out_buffer = buffer;
 145                 out_size = written;
 146                 buffer = NULL;
 147                 written = 0;
 148                 bufsize = 0;
 149         }
 150
 151         void flush() = delete;
 152
 153         void write_impl(const char *ptr, size_t size) override
 154         {
 155                 if (unlikely(written + size < written))
 156                         abort();
 157                 if (written + size > bufsize) {
 158                         bufsize = MAX3(1024, written + size, bufsize / 3 * 4);
 159                         buffer = (char *)realloc(buffer, bufsize);
 160                         if (!buffer) {
 161                                 fprintf(stderr, "amd: out of memory allocating ELF buffer\n");
 162                                 abort();
 163                         }
 164                 }
 165                 memcpy(buffer + written, ptr, size);
 166                 written += size;
 167         }
 168
 169         void pwrite_impl(const char *ptr, size_t size, uint64_t offset) override
 170         {
 171                 assert(offset == (size_t)offset &&
 172                        offset + size >= offset && offset + size <= written);
 173                 memcpy(buffer + offset, ptr, size);
 174         }
 175
 176         uint64_t current_pos() const override
 177         {
 178                 return written;
 179         }
 180 };
 181
 182 /* The LLVM compiler is represented as a pass manager containing passes for
 183  * optimizations, instruction selection, and code generation.
 184  */
 185 struct ac_compiler_passes {
 186         raw_memory_ostream ostream; /* ELF shader binary stream */
 187         llvm::legacy::PassManager passmgr; /* list of passes */
 188 };
 189
 190 struct ac_compiler_passes *ac_create_llvm_passes(LLVMTargetMachineRef tm)
 191 {
 192         struct ac_compiler_passes *p = new ac_compiler_passes();
 193         if (!p)
 194                 return NULL;
 195
 196         llvm::TargetMachine *TM = reinterpret_cast<llvm::TargetMachine*>(tm);
 197
 198         if (TM->addPassesToEmitFile(p->passmgr, p->ostream,
 199                                     nullptr,
 200                                     llvm::TargetMachine::CGFT_ObjectFile)) {
 201                 fprintf(stderr, "amd: TargetMachine can't emit a file of this type!\n");
 202                 delete p;
 203                 return NULL;
 204         }
 205         return p;
 206 }
 207
 208 void ac_destroy_llvm_passes(struct ac_compiler_passes *p)
 209 {
 210         delete p;
 211 }
 212
 213 /* This returns false on failure. */
 214 bool ac_compile_module_to_elf(struct ac_compiler_passes *p, LLVMModuleRef module,
 215                               char **pelf_buffer, size_t *pelf_size)
 216 {
 217         p->passmgr.run(*llvm::unwrap(module));
 218         p->ostream.take(*pelf_buffer, *pelf_size);
 219         return true;
 220 }
 221
 222 void ac_llvm_add_barrier_noop_pass(LLVMPassManagerRef passmgr)
 223 {
 224         llvm::unwrap(passmgr)->add(llvm::createBarrierNoopPass());
 225 }
 226
 227 void ac_enable_global_isel(LLVMTargetMachineRef tm)
 228 {
 229   reinterpret_cast<llvm::TargetMachine*>(tm)->setGlobalISel(true);
 230 }
 231
 232 LLVMValueRef ac_build_atomic_rmw(struct ac_llvm_context *ctx, LLVMAtomicRMWBinOp op,
 233                                  LLVMValueRef ptr, LLVMValueRef val,
 234                                  const char *sync_scope) {
 235         llvm::AtomicRMWInst::BinOp binop;
 236         switch (op) {
 237         case LLVMAtomicRMWBinOpXchg:
 238                 binop = llvm::AtomicRMWInst::Xchg;
 239                 break;
 240         case LLVMAtomicRMWBinOpAdd:
 241                 binop = llvm::AtomicRMWInst::Add;
 242                 break;
 243         case LLVMAtomicRMWBinOpSub:
 244                 binop = llvm::AtomicRMWInst::Sub;
 245                 break;
 246         case LLVMAtomicRMWBinOpAnd:
 247                 binop = llvm::AtomicRMWInst::And;
 248                 break;
 249         case LLVMAtomicRMWBinOpNand:
 250                 binop = llvm::AtomicRMWInst::Nand;
 251                 break;
 252         case LLVMAtomicRMWBinOpOr:
 253                 binop = llvm::AtomicRMWInst::Or;
 254                 break;
 255         case LLVMAtomicRMWBinOpXor:
 256                 binop = llvm::AtomicRMWInst::Xor;
 257                 break;
 258         case LLVMAtomicRMWBinOpMax:
 259                 binop = llvm::AtomicRMWInst::Max;
 260                 break;
 261         case LLVMAtomicRMWBinOpMin:
 262                 binop = llvm::AtomicRMWInst::Min;
 263                 break;
 264         case LLVMAtomicRMWBinOpUMax:
 265                 binop = llvm::AtomicRMWInst::UMax;
 266                 break;
 267         case LLVMAtomicRMWBinOpUMin:
 268                 binop = llvm::AtomicRMWInst::UMin;
 269                 break;
 270         default:
 271                 unreachable(!"invalid LLVMAtomicRMWBinOp");
 272            break;
 273         }
 274         unsigned SSID = llvm::unwrap(ctx->context)->getOrInsertSyncScopeID(sync_scope);
 275         return llvm::wrap(llvm::unwrap(ctx->builder)->CreateAtomicRMW(
 276                 binop, llvm::unwrap(ptr), llvm::unwrap(val),
 277                 llvm::AtomicOrdering::SequentiallyConsistent, SSID));
 278 }
 279
 280 LLVMValueRef ac_build_atomic_cmp_xchg(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 281                                       LLVMValueRef cmp, LLVMValueRef val,
 282                                       const char *sync_scope) {
 283         unsigned SSID = llvm::unwrap(ctx->context)->getOrInsertSyncScopeID(sync_scope);
 284         return llvm::wrap(llvm::unwrap(ctx->builder)->CreateAtomicCmpXchg(
 285                           llvm::unwrap(ptr), llvm::unwrap(cmp), llvm::unwrap(val),
 286                           llvm::AtomicOrdering::SequentiallyConsistent,
 287                           llvm::AtomicOrdering::SequentiallyConsistent, SSID));
 288 }