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         case AC_FLOAT_MODE_NO_SIGNED_ZEROS_FP_MATH:
  88                 flags.setNoSignedZeros();
  89                 llvm::unwrap(builder)->setFastMathFlags(flags);
  90                 break;
  91         }
  92
  93         return builder;
  94 }
  95
  96 LLVMTargetLibraryInfoRef
  97 ac_create_target_library_info(const char *triple)
  98 {
  99         return reinterpret_cast<LLVMTargetLibraryInfoRef>(new llvm::TargetLibraryInfoImpl(llvm::Triple(triple)));
 100 }
 101
 102 void
 103 ac_dispose_target_library_info(LLVMTargetLibraryInfoRef library_info)
 104 {
 105         delete reinterpret_cast<llvm::TargetLibraryInfoImpl *>(library_info);
 106 }
 107
 108 /* Implementation of raw_pwrite_stream that works on malloc()ed memory for
 109  * better compatibility with C code. */
 110 struct raw_memory_ostream : public llvm::raw_pwrite_stream {
 111         char *buffer;
 112         size_t written;
 113         size_t bufsize;
 114
 115         raw_memory_ostream()
 116         {
 117                 buffer = NULL;
 118                 written = 0;
 119                 bufsize = 0;
 120                 SetUnbuffered();
 121         }
 122
 123         ~raw_memory_ostream()
 124         {
 125                 free(buffer);
 126         }
 127
 128         void clear()
 129         {
 130                 written = 0;
 131         }
 132
 133         void take(char *&out_buffer, size_t &out_size)
 134         {
 135                 out_buffer = buffer;
 136                 out_size = written;
 137                 buffer = NULL;
 138                 written = 0;
 139                 bufsize = 0;
 140         }
 141
 142         void flush() = delete;
 143
 144         void write_impl(const char *ptr, size_t size) override
 145         {
 146                 if (unlikely(written + size < written))
 147                         abort();
 148                 if (written + size > bufsize) {
 149                         bufsize = MAX3(1024, written + size, bufsize / 3 * 4);
 150                         buffer = (char *)realloc(buffer, bufsize);
 151                         if (!buffer) {
 152                                 fprintf(stderr, "amd: out of memory allocating ELF buffer\n");
 153                                 abort();
 154                         }
 155                 }
 156                 memcpy(buffer + written, ptr, size);
 157                 written += size;
 158         }
 159
 160         void pwrite_impl(const char *ptr, size_t size, uint64_t offset) override
 161         {
 162                 assert(offset == (size_t)offset &&
 163                        offset + size >= offset && offset + size <= written);
 164                 memcpy(buffer + offset, ptr, size);
 165         }
 166
 167         uint64_t current_pos() const override
 168         {
 169                 return written;
 170         }
 171 };
 172
 173 /* The LLVM compiler is represented as a pass manager containing passes for
 174  * optimizations, instruction selection, and code generation.
 175  */
 176 struct ac_compiler_passes {
 177         raw_memory_ostream ostream; /* ELF shader binary stream */
 178         llvm::legacy::PassManager passmgr; /* list of passes */
 179 };
 180
 181 struct ac_compiler_passes *ac_create_llvm_passes(LLVMTargetMachineRef tm)
 182 {
 183         struct ac_compiler_passes *p = new ac_compiler_passes();
 184         if (!p)
 185                 return NULL;
 186
 187         llvm::TargetMachine *TM = reinterpret_cast<llvm::TargetMachine*>(tm);
 188
 189         if (TM->addPassesToEmitFile(p->passmgr, p->ostream,
 190                                     nullptr,
 191 #if LLVM_VERSION_MAJOR >= 10
 192                                     llvm::CGFT_ObjectFile)) {
 193 #else
 194                                     llvm::TargetMachine::CGFT_ObjectFile)) {
 195 #endif
 196                 fprintf(stderr, "amd: TargetMachine can't emit a file of this type!\n");
 197                 delete p;
 198                 return NULL;
 199         }
 200         return p;
 201 }
 202
 203 void ac_destroy_llvm_passes(struct ac_compiler_passes *p)
 204 {
 205         delete p;
 206 }
 207
 208 /* This returns false on failure. */
 209 bool ac_compile_module_to_elf(struct ac_compiler_passes *p, LLVMModuleRef module,
 210                               char **pelf_buffer, size_t *pelf_size)
 211 {
 212         p->passmgr.run(*llvm::unwrap(module));
 213         p->ostream.take(*pelf_buffer, *pelf_size);
 214         return true;
 215 }
 216
 217 void ac_llvm_add_barrier_noop_pass(LLVMPassManagerRef passmgr)
 218 {
 219         llvm::unwrap(passmgr)->add(llvm::createBarrierNoopPass());
 220 }
 221
 222 void ac_enable_global_isel(LLVMTargetMachineRef tm)
 223 {
 224   reinterpret_cast<llvm::TargetMachine*>(tm)->setGlobalISel(true);
 225 }
 226
 227 LLVMValueRef ac_build_atomic_rmw(struct ac_llvm_context *ctx, LLVMAtomicRMWBinOp op,
 228                                  LLVMValueRef ptr, LLVMValueRef val,
 229                                  const char *sync_scope) {
 230         llvm::AtomicRMWInst::BinOp binop;
 231         switch (op) {
 232         case LLVMAtomicRMWBinOpXchg:
 233                 binop = llvm::AtomicRMWInst::Xchg;
 234                 break;
 235         case LLVMAtomicRMWBinOpAdd:
 236                 binop = llvm::AtomicRMWInst::Add;
 237                 break;
 238         case LLVMAtomicRMWBinOpSub:
 239                 binop = llvm::AtomicRMWInst::Sub;
 240                 break;
 241         case LLVMAtomicRMWBinOpAnd:
 242                 binop = llvm::AtomicRMWInst::And;
 243                 break;
 244         case LLVMAtomicRMWBinOpNand:
 245                 binop = llvm::AtomicRMWInst::Nand;
 246                 break;
 247         case LLVMAtomicRMWBinOpOr:
 248                 binop = llvm::AtomicRMWInst::Or;
 249                 break;
 250         case LLVMAtomicRMWBinOpXor:
 251                 binop = llvm::AtomicRMWInst::Xor;
 252                 break;
 253         case LLVMAtomicRMWBinOpMax:
 254                 binop = llvm::AtomicRMWInst::Max;
 255                 break;
 256         case LLVMAtomicRMWBinOpMin:
 257                 binop = llvm::AtomicRMWInst::Min;
 258                 break;
 259         case LLVMAtomicRMWBinOpUMax:
 260                 binop = llvm::AtomicRMWInst::UMax;
 261                 break;
 262         case LLVMAtomicRMWBinOpUMin:
 263                 binop = llvm::AtomicRMWInst::UMin;
 264                 break;
 265         default:
 266                 unreachable(!"invalid LLVMAtomicRMWBinOp");
 267            break;
 268         }
 269         unsigned SSID = llvm::unwrap(ctx->context)->getOrInsertSyncScopeID(sync_scope);
 270         return llvm::wrap(llvm::unwrap(ctx->builder)->CreateAtomicRMW(
 271                 binop, llvm::unwrap(ptr), llvm::unwrap(val),
 272                 llvm::AtomicOrdering::SequentiallyConsistent, SSID));
 273 }
 274
 275 LLVMValueRef ac_build_atomic_cmp_xchg(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 276                                       LLVMValueRef cmp, LLVMValueRef val,
 277                                       const char *sync_scope) {
 278         unsigned SSID = llvm::unwrap(ctx->context)->getOrInsertSyncScopeID(sync_scope);
 279         return llvm::wrap(llvm::unwrap(ctx->builder)->CreateAtomicCmpXchg(
 280                           llvm::unwrap(ptr), llvm::unwrap(cmp), llvm::unwrap(val),
 281                           llvm::AtomicOrdering::SequentiallyConsistent,
 282                           llvm::AtomicOrdering::SequentiallyConsistent, SSID));
 283 }