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 #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                 break;
  86         case AC_FLOAT_MODE_NO_SIGNED_ZEROS_FP_MATH:
  87                 flags.setNoSignedZeros();
  88                 llvm::unwrap(builder)->setFastMathFlags(flags);
  89                 break;
  90         case AC_FLOAT_MODE_UNSAFE_FP_MATH:
  91                 flags.setFast();
  92                 llvm::unwrap(builder)->setFastMathFlags(flags);
  93                 break;
  94         }
  95
  96         return builder;
  97 }
  98
  99 LLVMTargetLibraryInfoRef
 100 ac_create_target_library_info(const char *triple)
 101 {
 102         return reinterpret_cast<LLVMTargetLibraryInfoRef>(new llvm::TargetLibraryInfoImpl(llvm::Triple(triple)));
 103 }
 104
 105 void
 106 ac_dispose_target_library_info(LLVMTargetLibraryInfoRef library_info)
 107 {
 108         delete reinterpret_cast<llvm::TargetLibraryInfoImpl *>(library_info);
 109 }
 110
 111 /* Implementation of raw_pwrite_stream that works on malloc()ed memory for
 112  * better compatibility with C code. */
 113 struct raw_memory_ostream : public llvm::raw_pwrite_stream {
 114         char *buffer;
 115         size_t written;
 116         size_t bufsize;
 117
 118         raw_memory_ostream()
 119         {
 120                 buffer = NULL;
 121                 written = 0;
 122                 bufsize = 0;
 123                 SetUnbuffered();
 124         }
 125
 126         ~raw_memory_ostream()
 127         {
 128                 free(buffer);
 129         }
 130
 131         void clear()
 132         {
 133                 written = 0;
 134         }
 135
 136         void take(char *&out_buffer, size_t &out_size)
 137         {
 138                 out_buffer = buffer;
 139                 out_size = written;
 140                 buffer = NULL;
 141                 written = 0;
 142                 bufsize = 0;
 143         }
 144
 145         void flush() = delete;
 146
 147         void write_impl(const char *ptr, size_t size) override
 148         {
 149                 if (unlikely(written + size < written))
 150                         abort();
 151                 if (written + size > bufsize) {
 152                         bufsize = MAX3(1024, written + size, bufsize / 3 * 4);
 153                         buffer = (char *)realloc(buffer, bufsize);
 154                         if (!buffer) {
 155                                 fprintf(stderr, "amd: out of memory allocating ELF buffer\n");
 156                                 abort();
 157                         }
 158                 }
 159                 memcpy(buffer + written, ptr, size);
 160                 written += size;
 161         }
 162
 163         void pwrite_impl(const char *ptr, size_t size, uint64_t offset) override
 164         {
 165                 assert(offset == (size_t)offset &&
 166                        offset + size >= offset && offset + size <= written);
 167                 memcpy(buffer + offset, ptr, size);
 168         }
 169
 170         uint64_t current_pos() const override
 171         {
 172                 return written;
 173         }
 174 };
 175
 176 /* The LLVM compiler is represented as a pass manager containing passes for
 177  * optimizations, instruction selection, and code generation.
 178  */
 179 struct ac_compiler_passes {
 180         raw_memory_ostream ostream; /* ELF shader binary stream */
 181         llvm::legacy::PassManager passmgr; /* list of passes */
 182 };
 183
 184 struct ac_compiler_passes *ac_create_llvm_passes(LLVMTargetMachineRef tm)
 185 {
 186         struct ac_compiler_passes *p = new ac_compiler_passes();
 187         if (!p)
 188                 return NULL;
 189
 190         llvm::TargetMachine *TM = reinterpret_cast<llvm::TargetMachine*>(tm);
 191
 192         if (TM->addPassesToEmitFile(p->passmgr, p->ostream,
 193                                     nullptr,
 194                                     llvm::TargetMachine::CGFT_ObjectFile)) {
 195                 fprintf(stderr, "amd: TargetMachine can't emit a file of this type!\n");
 196                 delete p;
 197                 return NULL;
 198         }
 199         return p;
 200 }
 201
 202 void ac_destroy_llvm_passes(struct ac_compiler_passes *p)
 203 {
 204         delete p;
 205 }
 206
 207 /* This returns false on failure. */
 208 bool ac_compile_module_to_elf(struct ac_compiler_passes *p, LLVMModuleRef module,
 209                               char **pelf_buffer, size_t *pelf_size)
 210 {
 211         p->passmgr.run(*llvm::unwrap(module));
 212         p->ostream.take(*pelf_buffer, *pelf_size);
 213         return true;
 214 }
 215
 216 void ac_llvm_add_barrier_noop_pass(LLVMPassManagerRef passmgr)
 217 {
 218         llvm::unwrap(passmgr)->add(llvm::createBarrierNoopPass());
 219 }
 220
 221 void ac_enable_global_isel(LLVMTargetMachineRef tm)
 222 {
 223   reinterpret_cast<llvm::TargetMachine*>(tm)->setGlobalISel(true);
 224 }
 225
 226 LLVMValueRef ac_build_atomic_rmw(struct ac_llvm_context *ctx, LLVMAtomicRMWBinOp op,
 227                                  LLVMValueRef ptr, LLVMValueRef val,
 228                                  const char *sync_scope) {
 229         llvm::AtomicRMWInst::BinOp binop;
 230         switch (op) {
 231         case LLVMAtomicRMWBinOpXchg:
 232                 binop = llvm::AtomicRMWInst::Xchg;
 233                 break;
 234         case LLVMAtomicRMWBinOpAdd:
 235                 binop = llvm::AtomicRMWInst::Add;
 236                 break;
 237         case LLVMAtomicRMWBinOpSub:
 238                 binop = llvm::AtomicRMWInst::Sub;
 239                 break;
 240         case LLVMAtomicRMWBinOpAnd:
 241                 binop = llvm::AtomicRMWInst::And;
 242                 break;
 243         case LLVMAtomicRMWBinOpNand:
 244                 binop = llvm::AtomicRMWInst::Nand;
 245                 break;
 246         case LLVMAtomicRMWBinOpOr:
 247                 binop = llvm::AtomicRMWInst::Or;
 248                 break;
 249         case LLVMAtomicRMWBinOpXor:
 250                 binop = llvm::AtomicRMWInst::Xor;
 251                 break;
 252         case LLVMAtomicRMWBinOpMax:
 253                 binop = llvm::AtomicRMWInst::Max;
 254                 break;
 255         case LLVMAtomicRMWBinOpMin:
 256                 binop = llvm::AtomicRMWInst::Min;
 257                 break;
 258         case LLVMAtomicRMWBinOpUMax:
 259                 binop = llvm::AtomicRMWInst::UMax;
 260                 break;
 261         case LLVMAtomicRMWBinOpUMin:
 262                 binop = llvm::AtomicRMWInst::UMin;
 263                 break;
 264         default:
 265                 unreachable(!"invalid LLVMAtomicRMWBinOp");
 266            break;
 267         }
 268         unsigned SSID = llvm::unwrap(ctx->context)->getOrInsertSyncScopeID(sync_scope);
 269         return llvm::wrap(llvm::unwrap(ctx->builder)->CreateAtomicRMW(
 270                 binop, llvm::unwrap(ptr), llvm::unwrap(val),
 271                 llvm::AtomicOrdering::SequentiallyConsistent, SSID));
 272 }
 273
 274 LLVMValueRef ac_build_atomic_cmp_xchg(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 275                                       LLVMValueRef cmp, LLVMValueRef val,
 276                                       const char *sync_scope) {
 277         unsigned SSID = llvm::unwrap(ctx->context)->getOrInsertSyncScopeID(sync_scope);
 278         return llvm::wrap(llvm::unwrap(ctx->builder)->CreateAtomicCmpXchg(
 279                           llvm::unwrap(ptr), llvm::unwrap(cmp), llvm::unwrap(val),
 280                           llvm::AtomicOrdering::SequentiallyConsistent,
 281                           llvm::AtomicOrdering::SequentiallyConsistent, SSID));
 282 }