Home | History | Annotate | Download | only in Utils
      1 //===-- AMDGPUBaseInfo.cpp - AMDGPU Base encoding information--------------===//
      2 //
      3 //                     The LLVM Compiler Infrastructure
      4 //
      5 // This file is distributed under the University of Illinois Open Source
      6 // License. See LICENSE.TXT for details.
      7 //
      8 //===----------------------------------------------------------------------===//
      9 #include "AMDGPUBaseInfo.h"
     10 #include "AMDGPU.h"
     11 #include "llvm/IR/LLVMContext.h"
     12 #include "llvm/IR/Function.h"
     13 #include "llvm/IR/GlobalValue.h"
     14 #include "llvm/MC/MCContext.h"
     15 #include "llvm/MC/MCSectionELF.h"
     16 #include "llvm/MC/MCSubtargetInfo.h"
     17 #include "llvm/MC/SubtargetFeature.h"
     18 
     19 #define GET_SUBTARGETINFO_ENUM
     20 #include "AMDGPUGenSubtargetInfo.inc"
     21 #undef GET_SUBTARGETINFO_ENUM
     22 
     23 #define GET_REGINFO_ENUM
     24 #include "AMDGPUGenRegisterInfo.inc"
     25 #undef GET_REGINFO_ENUM
     26 
     27 namespace llvm {
     28 namespace AMDGPU {
     29 
     30 IsaVersion getIsaVersion(const FeatureBitset &Features) {
     31 
     32   if (Features.test(FeatureISAVersion7_0_0))
     33     return {7, 0, 0};
     34 
     35   if (Features.test(FeatureISAVersion7_0_1))
     36     return {7, 0, 1};
     37 
     38   if (Features.test(FeatureISAVersion8_0_0))
     39     return {8, 0, 0};
     40 
     41   if (Features.test(FeatureISAVersion8_0_1))
     42     return {8, 0, 1};
     43 
     44   return {0, 0, 0};
     45 }
     46 
     47 void initDefaultAMDKernelCodeT(amd_kernel_code_t &Header,
     48                                const FeatureBitset &Features) {
     49 
     50   IsaVersion ISA = getIsaVersion(Features);
     51 
     52   memset(&Header, 0, sizeof(Header));
     53 
     54   Header.amd_kernel_code_version_major = 1;
     55   Header.amd_kernel_code_version_minor = 0;
     56   Header.amd_machine_kind = 1; // AMD_MACHINE_KIND_AMDGPU
     57   Header.amd_machine_version_major = ISA.Major;
     58   Header.amd_machine_version_minor = ISA.Minor;
     59   Header.amd_machine_version_stepping = ISA.Stepping;
     60   Header.kernel_code_entry_byte_offset = sizeof(Header);
     61   // wavefront_size is specified as a power of 2: 2^6 = 64 threads.
     62   Header.wavefront_size = 6;
     63   // These alignment values are specified in powers of two, so alignment =
     64   // 2^n.  The minimum alignment is 2^4 = 16.
     65   Header.kernarg_segment_alignment = 4;
     66   Header.group_segment_alignment = 4;
     67   Header.private_segment_alignment = 4;
     68 }
     69 
     70 MCSection *getHSATextSection(MCContext &Ctx) {
     71   return Ctx.getELFSection(".hsatext", ELF::SHT_PROGBITS,
     72                            ELF::SHF_ALLOC | ELF::SHF_WRITE |
     73                            ELF::SHF_EXECINSTR |
     74                            ELF::SHF_AMDGPU_HSA_AGENT |
     75                            ELF::SHF_AMDGPU_HSA_CODE);
     76 }
     77 
     78 MCSection *getHSADataGlobalAgentSection(MCContext &Ctx) {
     79   return Ctx.getELFSection(".hsadata_global_agent", ELF::SHT_PROGBITS,
     80                            ELF::SHF_ALLOC | ELF::SHF_WRITE |
     81                            ELF::SHF_AMDGPU_HSA_GLOBAL |
     82                            ELF::SHF_AMDGPU_HSA_AGENT);
     83 }
     84 
     85 MCSection *getHSADataGlobalProgramSection(MCContext &Ctx) {
     86   return  Ctx.getELFSection(".hsadata_global_program", ELF::SHT_PROGBITS,
     87                             ELF::SHF_ALLOC | ELF::SHF_WRITE |
     88                             ELF::SHF_AMDGPU_HSA_GLOBAL);
     89 }
     90 
     91 MCSection *getHSARodataReadonlyAgentSection(MCContext &Ctx) {
     92   return Ctx.getELFSection(".hsarodata_readonly_agent", ELF::SHT_PROGBITS,
     93                            ELF::SHF_ALLOC | ELF::SHF_AMDGPU_HSA_READONLY |
     94                            ELF::SHF_AMDGPU_HSA_AGENT);
     95 }
     96 
     97 bool isGroupSegment(const GlobalValue *GV) {
     98   return GV->getType()->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS;
     99 }
    100 
    101 bool isGlobalSegment(const GlobalValue *GV) {
    102   return GV->getType()->getAddressSpace() == AMDGPUAS::GLOBAL_ADDRESS;
    103 }
    104 
    105 bool isReadOnlySegment(const GlobalValue *GV) {
    106   return GV->getType()->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS;
    107 }
    108 
    109 static const char ShaderTypeAttribute[] = "ShaderType";
    110 
    111 unsigned getShaderType(const Function &F) {
    112   Attribute A = F.getFnAttribute(ShaderTypeAttribute);
    113   unsigned ShaderType = ShaderType::COMPUTE;
    114 
    115   if (A.isStringAttribute()) {
    116     StringRef Str = A.getValueAsString();
    117     if (Str.getAsInteger(0, ShaderType)) {
    118       LLVMContext &Ctx = F.getContext();
    119       Ctx.emitError("can't parse shader type");
    120     }
    121   }
    122   return ShaderType;
    123 }
    124 
    125 bool isSI(const MCSubtargetInfo &STI) {
    126   return STI.getFeatureBits()[AMDGPU::FeatureSouthernIslands];
    127 }
    128 
    129 bool isCI(const MCSubtargetInfo &STI) {
    130   return STI.getFeatureBits()[AMDGPU::FeatureSeaIslands];
    131 }
    132 
    133 bool isVI(const MCSubtargetInfo &STI) {
    134   return STI.getFeatureBits()[AMDGPU::FeatureVolcanicIslands];
    135 }
    136 
    137 unsigned getMCReg(unsigned Reg, const MCSubtargetInfo &STI) {
    138 
    139   switch(Reg) {
    140   default: break;
    141   case AMDGPU::FLAT_SCR:
    142     assert(!isSI(STI));
    143     return isCI(STI) ? AMDGPU::FLAT_SCR_ci : AMDGPU::FLAT_SCR_vi;
    144 
    145   case AMDGPU::FLAT_SCR_LO:
    146     assert(!isSI(STI));
    147     return isCI(STI) ? AMDGPU::FLAT_SCR_LO_ci : AMDGPU::FLAT_SCR_LO_vi;
    148 
    149   case AMDGPU::FLAT_SCR_HI:
    150     assert(!isSI(STI));
    151     return isCI(STI) ? AMDGPU::FLAT_SCR_HI_ci : AMDGPU::FLAT_SCR_HI_vi;
    152   }
    153   return Reg;
    154 }
    155 
    156 } // End namespace AMDGPU
    157 } // End namespace llvm
    158