 Chromium Code Reviews
 Chromium Code Reviews Issue 11191029:
  Use VLDR instead of VMOVs from GPR when a 64-bit double can't be encoded as a VMOV immediate.  (Closed) 
  Base URL: https://v8.googlecode.com/svn/branches/bleeding_edge
    
  
    Issue 11191029:
  Use VLDR instead of VMOVs from GPR when a 64-bit double can't be encoded as a VMOV immediate.  (Closed) 
  Base URL: https://v8.googlecode.com/svn/branches/bleeding_edge| Index: src/arm/assembler-arm.cc | 
| diff --git a/src/arm/assembler-arm.cc b/src/arm/assembler-arm.cc | 
| index ccf7208a8f6aa6c56cb73483d66a2a3ca1cca79d..a27fb572479f76ef9e58b1a2a5cda6aaf166c184 100644 | 
| --- a/src/arm/assembler-arm.cc | 
| +++ b/src/arm/assembler-arm.cc | 
| @@ -50,7 +50,6 @@ bool CpuFeatures::initialized_ = false; | 
| unsigned CpuFeatures::supported_ = 0; | 
| unsigned CpuFeatures::found_by_runtime_probing_ = 0; | 
| - | 
| 
ulan
2012/10/18 13:36:53
This line seems to be accidentally deleted.
 | 
| // Get the CPU features enabled by the build. For cross compilation the | 
| // preprocessor symbols CAN_USE_ARMV7_INSTRUCTIONS and CAN_USE_VFP3_INSTRUCTIONS | 
| // can be defined to enable ARMv7 and VFPv3 instructions when building the | 
| @@ -198,7 +197,7 @@ Operand::Operand(Handle<Object> handle) { | 
| } else { | 
| // no relocation needed | 
| imm32_ = reinterpret_cast<intptr_t>(obj); | 
| - rmode_ = RelocInfo::NONE; | 
| + rmode_ = RelocInfo::NONE32; | 
| } | 
| } | 
| @@ -273,8 +272,11 @@ const Instr kPopRegPattern = | 
| // mov lr, pc | 
| const Instr kMovLrPc = al | MOV | kRegister_pc_Code | kRegister_lr_Code * B12; | 
| // ldr rd, [pc, #offset] | 
| -const Instr kLdrPCMask = kCondMask | 15 * B24 | 7 * B20 | 15 * B16; | 
| -const Instr kLdrPCPattern = al | 5 * B24 | L | kRegister_pc_Code * B16; | 
| +const Instr kLdrPCMask = 15 * B24 | 7 * B20 | 15 * B16; | 
| +const Instr kLdrPCPattern = 5 * B24 | L | kRegister_pc_Code * B16; | 
| +// vldr dd, [pc, #offset] | 
| +const Instr kVldrDPCMask = 15 * B24 | 3 * B20 | 15 * B16 | 15 * B8; | 
| +const Instr kVldrDPCPattern = 13 * B24 | L | kRegister_pc_Code * B16 | 11 * B8; | 
| // blxcc rm | 
| const Instr kBlxRegMask = | 
| 15 * B24 | 15 * B20 | 15 * B16 | 15 * B12 | 15 * B8 | 15 * B4; | 
| @@ -350,6 +352,7 @@ Assembler::Assembler(Isolate* arg_isolate, void* buffer, int buffer_size) | 
| pc_ = buffer_; | 
| reloc_info_writer.Reposition(buffer_ + buffer_size, pc_); | 
| num_pending_reloc_info_ = 0; | 
| + num_pending_64_bit_reloc_info_ = 0; | 
| next_buffer_check_ = 0; | 
| const_pool_blocked_nesting_ = 0; | 
| no_const_pool_before_ = 0; | 
| @@ -376,6 +379,7 @@ void Assembler::GetCode(CodeDesc* desc) { | 
| // Emit constant pool if necessary. | 
| CheckConstPool(true, false); | 
| ASSERT(num_pending_reloc_info_ == 0); | 
| + ASSERT(num_pending_64_bit_reloc_info_ == 0); | 
| // Set up code descriptor. | 
| desc->buffer = buffer_; | 
| @@ -422,6 +426,11 @@ bool Assembler::IsLdrRegisterImmediate(Instr instr) { | 
| } | 
| +bool Assembler::IsVldrDRegisterImmediate(Instr instr) { | 
| + return (instr & (15 * B24 | 3 * B20 | 15 * B8)) == (13 * B24 | B20 | 11 * B8); | 
| +} | 
| + | 
| + | 
| int Assembler::GetLdrRegisterImmediateOffset(Instr instr) { | 
| ASSERT(IsLdrRegisterImmediate(instr)); | 
| bool positive = (instr & B23) == B23; | 
| @@ -430,6 +439,15 @@ int Assembler::GetLdrRegisterImmediateOffset(Instr instr) { | 
| } | 
| +int Assembler::GetVldrDRegisterImmediateOffset(Instr instr) { | 
| + ASSERT(IsVldrDRegisterImmediate(instr)); | 
| + bool positive = (instr & B23) == B23; | 
| + int offset = instr & kOff8Mask; // Zero extended offset. | 
| + offset <<= 2; | 
| + return positive ? offset : -offset; | 
| +} | 
| + | 
| + | 
| Instr Assembler::SetLdrRegisterImmediateOffset(Instr instr, int offset) { | 
| ASSERT(IsLdrRegisterImmediate(instr)); | 
| bool positive = offset >= 0; | 
| @@ -441,6 +459,18 @@ Instr Assembler::SetLdrRegisterImmediateOffset(Instr instr, int offset) { | 
| return (instr & ~kOff12Mask) | offset; | 
| } | 
| +Instr Assembler::SetVldrDRegisterImmediateOffset(Instr instr, int offset) { | 
| + ASSERT(IsVldrDRegisterImmediate(instr)); | 
| + ASSERT((offset & ~3) == offset); // Must be 64-bit aligned. | 
| + bool positive = offset >= 0; | 
| + if (!positive) offset = -offset; | 
| + ASSERT(is_uint10(offset)); | 
| + // Set bit indicating whether the offset should be added. | 
| + instr = (instr & ~B23) | (positive ? B23 : 0); | 
| + // Set the actual offset. Its bottom 2 bits are zero. | 
| + return (instr & ~kOff8Mask) | (offset >> 2); | 
| +} | 
| + | 
| bool Assembler::IsStrRegisterImmediate(Instr instr) { | 
| return (instr & (B27 | B26 | B25 | B22 | B20)) == B26; | 
| @@ -527,7 +557,14 @@ bool Assembler::IsLdrRegFpNegOffset(Instr instr) { | 
| bool Assembler::IsLdrPcImmediateOffset(Instr instr) { | 
| // Check the instruction is indeed a | 
| // ldr<cond> <Rd>, [pc +/- offset_12]. | 
| - return (instr & (kLdrPCMask & ~kCondMask)) == 0x051f0000; | 
| + return (instr & kLdrPCMask) == kLdrPCPattern; | 
| +} | 
| + | 
| + | 
| +bool Assembler::IsVldrDPcImmediateOffset(Instr instr) { | 
| + // Check the instruction is indeed a | 
| + // vldr<cond> <Dd>, [pc +/- offset_12]. | 
| + return (instr & kVldrDPCMask) == kVldrDPCPattern; | 
| } | 
| @@ -809,7 +846,7 @@ bool Operand::must_use_constant_pool(const Assembler* assembler) const { | 
| #endif // def DEBUG | 
| if (assembler != NULL && assembler->predictable_code_size()) return true; | 
| return Serializer::enabled(); | 
| - } else if (rmode_ == RelocInfo::NONE) { | 
| + } else if (RelocInfo::IsNone(rmode_)) { | 
| return false; | 
| } | 
| return true; | 
| @@ -2017,27 +2054,9 @@ void Assembler::vmov(const DwVfpRegister dst, | 
| // The double can be encoded in the instruction. | 
| emit(cond | 0xE*B24 | 0xB*B20 | dst.code()*B12 | 0xB*B8 | enc); | 
| } else { | 
| - // Synthesise the double from ARM immediates. This could be implemented | 
| - // using vldr from a constant pool. | 
| - uint32_t lo, hi; | 
| - DoubleAsTwoUInt32(imm, &lo, &hi); | 
| - mov(ip, Operand(lo)); | 
| - | 
| - if (scratch.is(no_reg)) { | 
| - // Move the low part of the double into the lower of the corresponsing S | 
| - // registers of D register dst. | 
| - vmov(dst.low(), ip, cond); | 
| - | 
| - // Move the high part of the double into the higher of the corresponsing S | 
| - // registers of D register dst. | 
| - mov(ip, Operand(hi)); | 
| - vmov(dst.high(), ip, cond); | 
| - } else { | 
| - // Move the low and high parts of the double to a D register in one | 
| - // instruction. | 
| - mov(scratch, Operand(hi)); | 
| - vmov(dst, ip, scratch, cond); | 
| - } | 
| + RecordRelocInfo(imm); | 
| + vldr(dst, MemOperand(pc, 0), cond); | 
| + // TODO(jfb) Constant blinding, denorm to zero, no NaN. | 
| } | 
| } | 
| @@ -2551,6 +2570,7 @@ void Assembler::db(uint8_t data) { | 
| // to write pure data with no pointers and the constant pool should | 
| // be emitted before using db. | 
| ASSERT(num_pending_reloc_info_ == 0); | 
| + ASSERT(num_pending_64_bit_reloc_info_ == 0); | 
| CheckBuffer(); | 
| *reinterpret_cast<uint8_t*>(pc_) = data; | 
| pc_ += sizeof(uint8_t); | 
| @@ -2562,6 +2582,7 @@ void Assembler::dd(uint32_t data) { | 
| // to write pure data with no pointers and the constant pool should | 
| // be emitted before using dd. | 
| ASSERT(num_pending_reloc_info_ == 0); | 
| + ASSERT(num_pending_64_bit_reloc_info_ == 0); | 
| CheckBuffer(); | 
| *reinterpret_cast<uint32_t*>(pc_) = data; | 
| pc_ += sizeof(uint32_t); | 
| @@ -2582,16 +2603,9 @@ void Assembler::RecordRelocInfo(RelocInfo::Mode rmode, intptr_t data) { | 
| || RelocInfo::IsConstPool(rmode)); | 
| // These modes do not need an entry in the constant pool. | 
| } else { | 
| - ASSERT(num_pending_reloc_info_ < kMaxNumPendingRelocInfo); | 
| - if (num_pending_reloc_info_ == 0) { | 
| - first_const_pool_use_ = pc_offset(); | 
| - } | 
| - pending_reloc_info_[num_pending_reloc_info_++] = rinfo; | 
| - // Make sure the constant pool is not emitted in place of the next | 
| - // instruction for which we just recorded relocation info. | 
| - BlockConstPoolFor(1); | 
| + RecordRelocInfoConstantPoolEntryHelper(rinfo); | 
| } | 
| - if (rinfo.rmode() != RelocInfo::NONE) { | 
| + if (!RelocInfo::IsNone(rinfo.rmode())) { | 
| // Don't record external references unless the heap will be serialized. | 
| if (rmode == RelocInfo::EXTERNAL_REFERENCE) { | 
| #ifdef DEBUG | 
| @@ -2617,14 +2631,36 @@ void Assembler::RecordRelocInfo(RelocInfo::Mode rmode, intptr_t data) { | 
| } | 
| } | 
| +void Assembler::RecordRelocInfo(double data) { | 
| + // We do not try to reuse pool constants. | 
| + RelocInfo rinfo(pc_, data); | 
| + RecordRelocInfoConstantPoolEntryHelper(rinfo); | 
| +} | 
| + | 
| + | 
| +void Assembler::RecordRelocInfoConstantPoolEntryHelper(const RelocInfo& rinfo) { | 
| + ASSERT(num_pending_reloc_info_ < kMaxNumPendingRelocInfo); | 
| + if (num_pending_reloc_info_ == 0) { | 
| + first_const_pool_use_ = pc_offset(); | 
| + } | 
| + pending_reloc_info_[num_pending_reloc_info_++] = rinfo; | 
| + if (rinfo.rmode() == RelocInfo::NONE64) { | 
| + ++num_pending_64_bit_reloc_info_; | 
| + } | 
| + ASSERT(num_pending_64_bit_reloc_info_ <= num_pending_reloc_info_); | 
| + // Make sure the constant pool is not emitted in place of the next | 
| + // instruction for which we just recorded relocation info. | 
| + BlockConstPoolFor(1); | 
| +} | 
| + | 
| void Assembler::BlockConstPoolFor(int instructions) { | 
| int pc_limit = pc_offset() + instructions * kInstrSize; | 
| if (no_const_pool_before_ < pc_limit) { | 
| // If there are some pending entries, the constant pool cannot be blocked | 
| - // further than first_const_pool_use_ + kMaxDistToPool | 
| + // further than constant pool instruction's reach. | 
| ASSERT((num_pending_reloc_info_ == 0) || | 
| - (pc_limit < (first_const_pool_use_ + kMaxDistToPool))); | 
| + (pc_limit - first_const_pool_use_ < kMaxDistToIntPool)); | 
| no_const_pool_before_ = pc_limit; | 
| } | 
| @@ -2646,29 +2682,52 @@ void Assembler::CheckConstPool(bool force_emit, bool require_jump) { | 
| // There is nothing to do if there are no pending constant pool entries. | 
| if (num_pending_reloc_info_ == 0) { | 
| + ASSERT(num_pending_64_bit_reloc_info_ == 0); | 
| // Calculate the offset of the next check. | 
| next_buffer_check_ = pc_offset() + kCheckPoolInterval; | 
| return; | 
| } | 
| - // We emit a constant pool when: | 
| - // * requested to do so by parameter force_emit (e.g. after each function). | 
| - // * the distance to the first instruction accessing the constant pool is | 
| - // kAvgDistToPool or more. | 
| - // * no jump is required and the distance to the first instruction accessing | 
| - // the constant pool is at least kMaxDistToPool / 2. | 
| - ASSERT(first_const_pool_use_ >= 0); | 
| - int dist = pc_offset() - first_const_pool_use_; | 
| - if (!force_emit && dist < kAvgDistToPool && | 
| - (require_jump || (dist < (kMaxDistToPool / 2)))) { | 
| - return; | 
| - } | 
| - | 
| // Check that the code buffer is large enough before emitting the constant | 
| // pool (include the jump over the pool and the constant pool marker and | 
| // the gap to the relocation information). | 
| + // Note 64-bit values are wider, and the first one needs to be 64-bit aligned. | 
| int jump_instr = require_jump ? kInstrSize : 0; | 
| - int size = jump_instr + kInstrSize + num_pending_reloc_info_ * kPointerSize; | 
| + int size = kInstrSize + jump_instr + num_pending_reloc_info_ * kPointerSize; | 
| + bool has_fp_values = (num_pending_64_bit_reloc_info_ > 0); | 
| + // 64-bit values must be 64-bit aligned. | 
| + bool require_64_bit_align = has_fp_values && (((uintptr_t)pc_ + size) & 0x3); | 
| + if (require_64_bit_align) { | 
| + size += kInstrSize; | 
| + } | 
| + STATIC_ASSERT(kPointerSize == kDoubleSize / 2); | 
| + size += num_pending_64_bit_reloc_info_ * (kDoubleSize / 2); | 
| + int marker_num = (size - kInstrSize - jump_instr) / 4; | 
| + | 
| + // We emit a constant pool when: | 
| + // * requested to do so by parameter force_emit (e.g. after each function). | 
| + // * the distance from the first instruction accessing the constant pool to | 
| + // any of the constant pool entries will exceed its limit the next | 
| + // time the pool is checked. This is overly restrictive, but we don't emit | 
| + // constant pool entries in-order so it's conservatively correct. | 
| + // * the instruction doesn't require a jump after itself to jump over the | 
| + // constant pool, and we're getting close to running out of range. | 
| + if (!force_emit) { | 
| + ASSERT((first_const_pool_use_ >= 0) && (num_pending_reloc_info_ > 0)); | 
| + int dist = pc_offset() + size - first_const_pool_use_; | 
| + if (has_fp_values) { | 
| + if ((dist < kMaxDistToFPPool - kCheckPoolInterval) && | 
| + (require_jump || (dist < kMaxDistToFPPool / 2))) { | 
| + return; | 
| + } | 
| + } else { | 
| + if ((dist < kMaxDistToIntPool - kCheckPoolInterval) && | 
| + (require_jump || (dist < kMaxDistToIntPool / 2))) { | 
| + return; | 
| + } | 
| + } | 
| + } | 
| + | 
| int needed_space = size + kGap; | 
| while (buffer_space() <= needed_space) GrowBuffer(); | 
| @@ -2686,9 +2745,41 @@ void Assembler::CheckConstPool(bool force_emit, bool require_jump) { | 
| // Put down constant pool marker "Undefined instruction" as specified by | 
| // A5.6 (ARMv7) Instruction set encoding. | 
| - emit(kConstantPoolMarker | num_pending_reloc_info_); | 
| + emit(kConstantPoolMarker | marker_num); | 
| - // Emit constant pool entries. | 
| + if (require_64_bit_align) { | 
| + emit(kConstantPoolMarker); | 
| + } | 
| + | 
| + // Emit 64-bit constant pool entries first: their range is smaller than | 
| + // 32-bit entries. | 
| + for (int i = 0; i < num_pending_reloc_info_; i++) { | 
| + ASSERT(!((uintptr_t)pc_ & 0x3)); // Check 64-bit alignment. | 
| + RelocInfo& rinfo = pending_reloc_info_[i]; | 
| + | 
| + if (rinfo.rmode() != RelocInfo::NONE64) { | 
| + // 32-bit values emitted later. | 
| + continue; | 
| + } | 
| + | 
| + Instr instr = instr_at(rinfo.pc()); | 
| + // Instruction to patch must be 'vldr rd, [pc, #offset]' with offset == 0. | 
| + ASSERT((IsVldrDPcImmediateOffset(instr) && | 
| + GetVldrDRegisterImmediateOffset(instr) == 0)); | 
| + | 
| + int delta = pc_ - rinfo.pc() - kPcLoadDelta; | 
| + ASSERT(is_uint10(delta)); | 
| + | 
| + instr_at_put(rinfo.pc(), SetVldrDRegisterImmediateOffset(instr, delta)); | 
| + | 
| + const double double_data = rinfo.data64(); | 
| + uint64_t uint_data = 0; | 
| + memcpy(&uint_data, &double_data, sizeof(double_data)); | 
| + emit(uint_data & 0xFFFFFFFF); | 
| + emit(uint_data >> 32); | 
| + } | 
| + | 
| + // Emit 32-bit constant pool entries. | 
| for (int i = 0; i < num_pending_reloc_info_; i++) { | 
| RelocInfo& rinfo = pending_reloc_info_[i]; | 
| ASSERT(rinfo.rmode() != RelocInfo::COMMENT && | 
| @@ -2696,10 +2787,15 @@ void Assembler::CheckConstPool(bool force_emit, bool require_jump) { | 
| rinfo.rmode() != RelocInfo::STATEMENT_POSITION && | 
| rinfo.rmode() != RelocInfo::CONST_POOL); | 
| + if (rinfo.rmode() == RelocInfo::NONE64) { | 
| + // 64-bit values emitted earlier. | 
| + continue; | 
| + } | 
| + | 
| Instr instr = instr_at(rinfo.pc()); | 
| // Instruction to patch must be 'ldr rd, [pc, #offset]' with offset == 0. | 
| - ASSERT(IsLdrPcImmediateOffset(instr) && | 
| - GetLdrRegisterImmediateOffset(instr) == 0); | 
| + ASSERT((IsLdrPcImmediateOffset(instr) && | 
| + GetLdrRegisterImmediateOffset(instr) == 0)); | 
| int delta = pc_ - rinfo.pc() - kPcLoadDelta; | 
| // 0 is the smallest delta: | 
| @@ -2713,6 +2809,7 @@ void Assembler::CheckConstPool(bool force_emit, bool require_jump) { | 
| } | 
| num_pending_reloc_info_ = 0; | 
| + num_pending_64_bit_reloc_info_ = 0; | 
| first_const_pool_use_ = -1; | 
| RecordComment("]"); |