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 LLVMTargetLibraryInfoRef
 125 ac_create_target_library_info(const char *triple)
 126 {
 127         return reinterpret_cast<LLVMTargetLibraryInfoRef>(new llvm::TargetLibraryInfoImpl(llvm::Triple(triple)));
 128 }
 129
 130 void
 131 ac_dispose_target_library_info(LLVMTargetLibraryInfoRef library_info)
 132 {
 133         delete reinterpret_cast<llvm::TargetLibraryInfoImpl *>(library_info);
 134 }
 135
 136 /* Implementation of raw_pwrite_stream that works on malloc()ed memory for
 137  * better compatibility with C code. */
 138 struct raw_memory_ostream : public llvm::raw_pwrite_stream {
 139         char *buffer;
 140         size_t written;
 141         size_t bufsize;
 142
 143         raw_memory_ostream()
 144         {
 145                 buffer = NULL;
 146                 written = 0;
 147                 bufsize = 0;
 148                 SetUnbuffered();
 149         }
 150
 151         ~raw_memory_ostream()
 152         {
 153                 free(buffer);
 154         }
 155
 156         void clear()
 157         {
 158                 written = 0;
 159         }
 160
 161         void take(char *&out_buffer, size_t &out_size)
 162         {
 163                 out_buffer = buffer;
 164                 out_size = written;
 165                 buffer = NULL;
 166                 written = 0;
 167                 bufsize = 0;
 168         }
 169
 170         void flush() = delete;
 171
 172         void write_impl(const char *ptr, size_t size) override
 173         {
 174                 if (unlikely(written + size < written))
 175                         abort();
 176                 if (written + size > bufsize) {
 177                         bufsize = MAX3(1024, written + size, bufsize / 3 * 4);
 178                         buffer = (char *)realloc(buffer, bufsize);
 179                         if (!buffer) {
 180                                 fprintf(stderr, "amd: out of memory allocating ELF buffer\n");
 181                                 abort();
 182                         }
 183                 }
 184                 memcpy(buffer + written, ptr, size);
 185                 written += size;
 186         }
 187
 188         void pwrite_impl(const char *ptr, size_t size, uint64_t offset) override
 189         {
 190                 assert(offset == (size_t)offset &&
 191                        offset + size >= offset && offset + size <= written);
 192                 memcpy(buffer + offset, ptr, size);
 193         }
 194
 195         uint64_t current_pos() const override
 196         {
 197                 return written;
 198         }
 199 };
 200
 201 /* The LLVM compiler is represented as a pass manager containing passes for
 202  * optimizations, instruction selection, and code generation.
 203  */
 204 struct ac_compiler_passes {
 205         raw_memory_ostream ostream; /* ELF shader binary stream */
 206         llvm::legacy::PassManager passmgr; /* list of passes */
 207 };
 208
 209 struct ac_compiler_passes *ac_create_llvm_passes(LLVMTargetMachineRef tm)
 210 {
 211         struct ac_compiler_passes *p = new ac_compiler_passes();
 212         if (!p)
 213                 return NULL;
 214
 215         llvm::TargetMachine *TM = reinterpret_cast<llvm::TargetMachine*>(tm);
 216
 217         if (TM->addPassesToEmitFile(p->passmgr, p->ostream,
 218                                     nullptr,
 219 #if LLVM_VERSION_MAJOR >= 10
 220                                     llvm::CGFT_ObjectFile)) {
 221 #else
 222                                     llvm::TargetMachine::CGFT_ObjectFile)) {
 223 #endif
 224                 fprintf(stderr, "amd: TargetMachine can't emit a file of this type!\n");
 225                 delete p;
 226                 return NULL;
 227         }
 228         return p;
 229 }
 230
 231 void ac_destroy_llvm_passes(struct ac_compiler_passes *p)
 232 {
 233         delete p;
 234 }
 235
 236 /* This returns false on failure. */
 237 bool ac_compile_module_to_elf(struct ac_compiler_passes *p, LLVMModuleRef module,
 238                               char **pelf_buffer, size_t *pelf_size)
 239 {
 240         p->passmgr.run(*llvm::unwrap(module));
 241         p->ostream.take(*pelf_buffer, *pelf_size);
 242         return true;
 243 }
 244
 245 void ac_llvm_add_barrier_noop_pass(LLVMPassManagerRef passmgr)
 246 {
 247         llvm::unwrap(passmgr)->add(llvm::createBarrierNoopPass());
 248 }
 249
 250 void ac_enable_global_isel(LLVMTargetMachineRef tm)
 251 {
 252   reinterpret_cast<llvm::TargetMachine*>(tm)->setGlobalISel(true);
 253 }
 254
 255 LLVMValueRef ac_build_atomic_rmw(struct ac_llvm_context *ctx, LLVMAtomicRMWBinOp op,
 256                                  LLVMValueRef ptr, LLVMValueRef val,
 257                                  const char *sync_scope) {
 258         llvm::AtomicRMWInst::BinOp binop;
 259         switch (op) {
 260         case LLVMAtomicRMWBinOpXchg:
 261                 binop = llvm::AtomicRMWInst::Xchg;
 262                 break;
 263         case LLVMAtomicRMWBinOpAdd:
 264                 binop = llvm::AtomicRMWInst::Add;
 265                 break;
 266         case LLVMAtomicRMWBinOpSub:
 267                 binop = llvm::AtomicRMWInst::Sub;
 268                 break;
 269         case LLVMAtomicRMWBinOpAnd:
 270                 binop = llvm::AtomicRMWInst::And;
 271                 break;
 272         case LLVMAtomicRMWBinOpNand:
 273                 binop = llvm::AtomicRMWInst::Nand;
 274                 break;
 275         case LLVMAtomicRMWBinOpOr:
 276                 binop = llvm::AtomicRMWInst::Or;
 277                 break;
 278         case LLVMAtomicRMWBinOpXor:
 279                 binop = llvm::AtomicRMWInst::Xor;
 280                 break;
 281         case LLVMAtomicRMWBinOpMax:
 282                 binop = llvm::AtomicRMWInst::Max;
 283                 break;
 284         case LLVMAtomicRMWBinOpMin:
 285                 binop = llvm::AtomicRMWInst::Min;
 286                 break;
 287         case LLVMAtomicRMWBinOpUMax:
 288                 binop = llvm::AtomicRMWInst::UMax;
 289                 break;
 290         case LLVMAtomicRMWBinOpUMin:
 291                 binop = llvm::AtomicRMWInst::UMin;
 292                 break;
 293         default:
 294                 unreachable(!"invalid LLVMAtomicRMWBinOp");
 295            break;
 296         }
 297         unsigned SSID = llvm::unwrap(ctx->context)->getOrInsertSyncScopeID(sync_scope);
 298         return llvm::wrap(llvm::unwrap(ctx->builder)->CreateAtomicRMW(
 299                 binop, llvm::unwrap(ptr), llvm::unwrap(val),
 300                 llvm::AtomicOrdering::SequentiallyConsistent, SSID));
 301 }
 302
 303 LLVMValueRef ac_build_atomic_cmp_xchg(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 304                                       LLVMValueRef cmp, LLVMValueRef val,
 305                                       const char *sync_scope) {
 306         unsigned SSID = llvm::unwrap(ctx->context)->getOrInsertSyncScopeID(sync_scope);
 307         return llvm::wrap(llvm::unwrap(ctx->builder)->CreateAtomicCmpXchg(
 308                           llvm::unwrap(ptr), llvm::unwrap(cmp), llvm::unwrap(val),
 309                           llvm::AtomicOrdering::SequentiallyConsistent,
 310                           llvm::AtomicOrdering::SequentiallyConsistent, SSID));
 311 }