maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity
llvm-project/llvm/lib/Target/AArch64/AArch64FastISel.cpp

3714 lines
126 KiB
C++
Raw Normal View History

AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
//===-- AArch6464FastISel.cpp - AArch64 FastISel implementation -----------===//
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
// This file defines the AArch64-specific support for the FastISel class. Some
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// of the target-specific code is generated by tablegen in the file
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
// AArch64GenFastISel.inc, which is #included here.
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
//
//===----------------------------------------------------------------------===//
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
#include "AArch64.h"
#include "AArch64Subtarget.h"
Run sort_includes.py on the AArch64 backend. No functionality change. llvm-svn: 213938
2014-07-25 19:42:14 +08:00
#include "AArch64TargetMachine.h"
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
#include "MCTargetDesc/AArch64AddressingModes.h"
[FastISel][AArch64] Add branch weights. Add branch weights to branch instructions, so that the following passes can optimize based on it (i.e. basic block ordering). Fixes <rdar://problem/17887137>. llvm-svn: 214537
2014-08-02 02:39:24 +08:00
#include "llvm/Analysis/BranchProbabilityInfo.h"
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
#include "llvm/CodeGen/CallingConvLower.h"
#include "llvm/CodeGen/FastISel.h"
#include "llvm/CodeGen/FunctionLoweringInfo.h"
#include "llvm/CodeGen/MachineConstantPool.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/IR/CallingConv.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/DerivedTypes.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/GetElementPtrTypeIterator.h"
#include "llvm/IR/GlobalAlias.h"
#include "llvm/IR/GlobalVariable.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/Operator.h"
#include "llvm/Support/CommandLine.h"
using namespace llvm;
namespace {
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
class AArch64FastISel : public FastISel {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
class Address {
public:
typedef enum {
RegBase,
FrameIndexBase
} BaseKind;
private:
BaseKind Kind;
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
AArch64_AM::ShiftExtendType ExtType;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
union {
unsigned Reg;
int FI;
} Base;
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
unsigned OffsetReg;
unsigned Shift;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
int64_t Offset;
[FastISel][AArch64] Add MachO large code model support for function calls. Currently the large code model for MachO uses the GOT to make function calls. Emit the required adrp and ldr instructions to load the address from the GOT. Related to <rdar://problem/17733076>. llvm-svn: 214381
2014-07-31 12:10:40 +08:00
const GlobalValue *GV;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
public:
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
Address() : Kind(RegBase), ExtType(AArch64_AM::InvalidShiftExtend),
OffsetReg(0), Shift(0), Offset(0), GV(nullptr) { Base.Reg = 0; }
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
void setKind(BaseKind K) { Kind = K; }
BaseKind getKind() const { return Kind; }
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
void setExtendType(AArch64_AM::ShiftExtendType E) { ExtType = E; }
AArch64_AM::ShiftExtendType getExtendType() const { return ExtType; }
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
bool isRegBase() const { return Kind == RegBase; }
bool isFIBase() const { return Kind == FrameIndexBase; }
void setReg(unsigned Reg) {
assert(isRegBase() && "Invalid base register access!");
Base.Reg = Reg;
}
unsigned getReg() const {
assert(isRegBase() && "Invalid base register access!");
return Base.Reg;
}
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
void setOffsetReg(unsigned Reg) {
assert(isRegBase() && "Invalid offset register access!");
OffsetReg = Reg;
}
unsigned getOffsetReg() const {
assert(isRegBase() && "Invalid offset register access!");
return OffsetReg;
}
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
void setFI(unsigned FI) {
assert(isFIBase() && "Invalid base frame index access!");
Base.FI = FI;
}
unsigned getFI() const {
assert(isFIBase() && "Invalid base frame index access!");
return Base.FI;
}
void setOffset(int64_t O) { Offset = O; }
int64_t getOffset() { return Offset; }
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
void setShift(unsigned S) { Shift = S; }
unsigned getShift() { return Shift; }
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
[FastISel][AArch64] Add MachO large code model support for function calls. Currently the large code model for MachO uses the GOT to make function calls. Emit the required adrp and ldr instructions to load the address from the GOT. Related to <rdar://problem/17733076>. llvm-svn: 214381
2014-07-31 12:10:40 +08:00
void setGlobalValue(const GlobalValue *G) { GV = G; }
const GlobalValue *getGlobalValue() { return GV; }
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
};
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
/// Subtarget - Keep a pointer to the AArch64Subtarget around so that we can
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
/// make the right decision when generating code for different targets.
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
const AArch64Subtarget *Subtarget;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
LLVMContext *Context;
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
bool fastLowerArguments() override;
bool fastLowerCall(CallLoweringInfo &CLI) override;
bool fastLowerIntrinsicCall(const IntrinsicInst *II) override;
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
private:
// Selection routines.
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
bool selectAddSub(const Instruction *I);
[FastISel][AArch64] Add support for non-native types for logical ops. Extend the logical ops selection to also support non-native types such as i1, i8, and i16. Fixes rdar://problem/18330589. llvm-svn: 217732
2014-09-14 07:46:28 +08:00
bool selectLogicalOp(const Instruction *I, unsigned ISDOpcode);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
bool SelectLoad(const Instruction *I);
bool SelectStore(const Instruction *I);
bool SelectBranch(const Instruction *I);
bool SelectIndirectBr(const Instruction *I);
bool SelectCmp(const Instruction *I);
bool SelectSelect(const Instruction *I);
bool SelectFPExt(const Instruction *I);
bool SelectFPTrunc(const Instruction *I);
bool SelectFPToInt(const Instruction *I, bool Signed);
bool SelectIntToFP(const Instruction *I, bool Signed);
bool SelectRem(const Instruction *I, unsigned ISDOpcode);
bool SelectRet(const Instruction *I);
bool SelectTrunc(const Instruction *I);
bool SelectIntExt(const Instruction *I);
bool SelectMul(const Instruction *I);
[FastISel][AArch64] Add support for variable shift. This adds the missing variable shift support for value type i8, i16, and i32. This fixes <rdar://problem/18095685>. llvm-svn: 216242
2014-08-22 07:06:07 +08:00
bool SelectShift(const Instruction *I);
[FastISel][AArch64] Add basic bitcast support for conversion between float and int. Fixes <rdar://problem/17867078>. llvm-svn: 214389
2014-07-31 14:25:37 +08:00
bool SelectBitCast(const Instruction *I);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Utility helper routines.
bool isTypeLegal(Type *Ty, MVT &VT);
bool isLoadStoreTypeLegal(Type *Ty, MVT &VT);
[FastISel][AArch64] Use a new helper function to determine if a value type is supported. NFCI. FastISel for AArch64 supports more value types than are actually legal. Use a dedicated helper function to reflect this. It is very similar to the isLoadStoreTypeLegal function, with the exception that vector types are not supported yet. llvm-svn: 216984
2014-09-03 06:33:53 +08:00
bool isTypeSupported(Type *Ty, MVT &VT);
[FastISel][AArch64] Don't fold instructions that are not in the same basic block. This fix checks first if the instruction to be folded (e.g. sign-/zero-extend, or shift) is in the same machine basic block as the instruction we are folding into. Not doing so can result in incorrect code, because the value might not be live-out of the basic block, where the value is defined. This fixes rdar://problem/18169495. llvm-svn: 216700
2014-08-29 08:19:21 +08:00
bool isValueAvailable(const Value *V) const;
Revert "[FastISel][AArch64] Don't fold instructions too aggressively into the memory operation." Quentin pointed out that this is not the correct approach and there is a better and easier solution. llvm-svn: 216632
2014-08-28 07:09:40 +08:00
bool ComputeAddress(const Value *Obj, Address &Addr, Type *Ty = nullptr);
[FastISel][AArch64] Add MachO large code model support for function calls. Currently the large code model for MachO uses the GOT to make function calls. Emit the required adrp and ldr instructions to load the address from the GOT. Related to <rdar://problem/17733076>. llvm-svn: 214381
2014-07-31 12:10:40 +08:00
bool ComputeCallAddress(const Value *V, Address &Addr);
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
bool SimplifyAddress(Address &Addr, MVT VT);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
void AddLoadStoreOperands(Address &Addr, const MachineInstrBuilder &MIB,
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
unsigned Flags, unsigned ScaleFactor,
MachineMemOperand *MMO);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
bool IsMemCpySmall(uint64_t Len, unsigned Alignment);
bool TryEmitSmallMemCpy(Address Dest, Address Src, uint64_t Len,
unsigned Alignment);
[FastISel][AArch64] Add branch folding support for the XALU intrinsics. This improves the code generation for the XALU intrinsics when the condition is feeding a branch instruction. This is related to <rdar://problem/17831117>. llvm-svn: 214349
2014-07-31 06:04:34 +08:00
bool foldXALUIntrinsic(AArch64CC::CondCode &CC, const Instruction *I,
const Value *Cond);
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
// Emit helper routines.
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
unsigned emitAddSub(bool UseAdd, MVT RetVT, const Value *LHS,
const Value *RHS, bool SetFlags = false,
bool WantResult = true, bool IsZExt = false);
unsigned emitAddSub_rr(bool UseAdd, MVT RetVT, unsigned LHSReg,
bool LHSIsKill, unsigned RHSReg, bool RHSIsKill,
bool SetFlags = false, bool WantResult = true);
unsigned emitAddSub_ri(bool UseAdd, MVT RetVT, unsigned LHSReg,
bool LHSIsKill, uint64_t Imm, bool SetFlags = false,
bool WantResult = true);
[FastISel][AArch64] Fix address simplification. When a shift with extension or an add with shift and extension cannot be folded into the memory operation, then the address calculation has to be materialized separately. While doing so the code forgot to consider a possible sign-/zero- extension. This fix folds now also the sign-/zero-extension into the add or shift instruction which is used to materialize the address. This fixes rdar://problem/18141718. llvm-svn: 216511
2014-08-27 08:58:30 +08:00
unsigned emitAddSub_rs(bool UseAdd, MVT RetVT, unsigned LHSReg,
bool LHSIsKill, unsigned RHSReg, bool RHSIsKill,
AArch64_AM::ShiftExtendType ShiftType,
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
uint64_t ShiftImm, bool SetFlags = false,
bool WantResult = true);
[FastISel][AArch64] Fix address simplification. When a shift with extension or an add with shift and extension cannot be folded into the memory operation, then the address calculation has to be materialized separately. While doing so the code forgot to consider a possible sign-/zero- extension. This fix folds now also the sign-/zero-extension into the add or shift instruction which is used to materialize the address. This fixes rdar://problem/18141718. llvm-svn: 216511
2014-08-27 08:58:30 +08:00
unsigned emitAddSub_rx(bool UseAdd, MVT RetVT, unsigned LHSReg,
bool LHSIsKill, unsigned RHSReg, bool RHSIsKill,
AArch64_AM::ShiftExtendType ExtType,
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
uint64_t ShiftImm, bool SetFlags = false,
bool WantResult = true);
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Emit functions.
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
bool emitCmp(const Value *LHS, const Value *RHS, bool IsZExt);
bool emitICmp(MVT RetVT, const Value *LHS, const Value *RHS, bool IsZExt);
bool emitICmp_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill, uint64_t Imm);
bool emitFCmp(MVT RetVT, const Value *LHS, const Value *RHS);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
bool EmitLoad(MVT VT, unsigned &ResultReg, Address Addr,
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
MachineMemOperand *MMO = nullptr);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
bool EmitStore(MVT VT, unsigned SrcReg, Address Addr,
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
MachineMemOperand *MMO = nullptr);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
unsigned EmitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, bool isZExt);
unsigned Emiti1Ext(unsigned SrcReg, MVT DestVT, bool isZExt);
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
unsigned emitAdd(MVT RetVT, const Value *LHS, const Value *RHS,
bool SetFlags = false, bool WantResult = true,
bool IsZExt = false);
unsigned emitSub(MVT RetVT, const Value *LHS, const Value *RHS,
bool SetFlags = false, bool WantResult = true,
bool IsZExt = false);
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
unsigned emitSubs_rr(MVT RetVT, unsigned LHSReg, bool LHSIsKill,
unsigned RHSReg, bool RHSIsKill, bool WantResult = true);
unsigned emitSubs_rs(MVT RetVT, unsigned LHSReg, bool LHSIsKill,
unsigned RHSReg, bool RHSIsKill,
AArch64_AM::ShiftExtendType ShiftType, uint64_t ShiftImm,
bool WantResult = true);
[FastISel][AArch64] Add target-specific lowering for logical operations. This change adds support for immediate and shift-left folding into logical operations. This fixes rdar://problem/18223183. llvm-svn: 217118
2014-09-04 09:29:18 +08:00
unsigned emitLogicalOp(unsigned ISDOpc, MVT RetVT, const Value *LHS,
const Value *RHS);
unsigned emitLogicalOp_ri(unsigned ISDOpc, MVT RetVT, unsigned LHSReg,
bool LHSIsKill, uint64_t Imm);
unsigned emitLogicalOp_rs(unsigned ISDOpc, MVT RetVT, unsigned LHSReg,
bool LHSIsKill, unsigned RHSReg, bool RHSIsKill,
uint64_t ShiftImm);
unsigned emitAnd_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill, uint64_t Imm);
[FastISel][AArch64] Create helper functions to create the various multiplies on AArch64. llvm-svn: 214346
2014-07-31 06:04:25 +08:00
unsigned Emit_MUL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
unsigned Op1, bool Op1IsKill);
unsigned Emit_SMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
unsigned Op1, bool Op1IsKill);
unsigned Emit_UMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
unsigned Op1, bool Op1IsKill);
[FastISel][AArch64] Add support for variable shift. This adds the missing variable shift support for value type i8, i16, and i32. This fixes <rdar://problem/18095685>. llvm-svn: 216242
2014-08-22 07:06:07 +08:00
unsigned emitLSL_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
unsigned Op1Reg, bool Op1IsKill);
[FastISel][AArch64] Fold Sign-/Zero-Extend into the shift immediate instruction. llvm-svn: 216510
2014-08-27 08:58:26 +08:00
unsigned emitLSL_ri(MVT RetVT, MVT SrcVT, unsigned Op0Reg, bool Op0IsKill,
uint64_t Imm, bool IsZExt = true);
[FastISel][AArch64] Add support for variable shift. This adds the missing variable shift support for value type i8, i16, and i32. This fixes <rdar://problem/18095685>. llvm-svn: 216242
2014-08-22 07:06:07 +08:00
unsigned emitLSR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
unsigned Op1Reg, bool Op1IsKill);
[FastISel][AArch64] Fold Sign-/Zero-Extend into the shift immediate instruction. llvm-svn: 216510
2014-08-27 08:58:26 +08:00
unsigned emitLSR_ri(MVT RetVT, MVT SrcVT, unsigned Op0Reg, bool Op0IsKill,
uint64_t Imm, bool IsZExt = true);
[FastISel][AArch64] Add support for variable shift. This adds the missing variable shift support for value type i8, i16, and i32. This fixes <rdar://problem/18095685>. llvm-svn: 216242
2014-08-22 07:06:07 +08:00
unsigned emitASR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
unsigned Op1Reg, bool Op1IsKill);
[FastISel][AArch64] Fold Sign-/Zero-Extend into the shift immediate instruction. llvm-svn: 216510
2014-08-27 08:58:26 +08:00
unsigned emitASR_ri(MVT RetVT, MVT SrcVT, unsigned Op0Reg, bool Op0IsKill,
uint64_t Imm, bool IsZExt = false);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
Reapplying [FastISel][AArch64] Cleanup constant materialization code. NFCI. Note: This reapplies r215582 without any modifications. The refactoring wasn't responsible for the buildbot failures. Original commit message: Cleanup and prepare constant materialization code for future commits. llvm-svn: 215752
2014-08-16 02:55:52 +08:00
unsigned AArch64MaterializeInt(const ConstantInt *CI, MVT VT);
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
unsigned AArch64MaterializeFP(const ConstantFP *CFP, MVT VT);
unsigned AArch64MaterializeGV(const GlobalValue *GV);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Call handling routines.
private:
CCAssignFn *CCAssignFnForCall(CallingConv::ID CC) const;
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
bool ProcessCallArgs(CallLoweringInfo &CLI, SmallVectorImpl<MVT> &ArgVTs,
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
unsigned &NumBytes);
[FastISel][AArch64] Fix return type in FastLowerCall. I used the wrong method to obtain the return type inside FinishCall. This fix simply uses the return type from FastLowerCall, which we already determined to be a valid type. Reduced test case from Chad. Thanks. llvm-svn: 213788
2014-07-24 04:03:13 +08:00
bool FinishCall(CallLoweringInfo &CLI, MVT RetVT, unsigned NumBytes);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
public:
// Backend specific FastISel code.
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
unsigned fastMaterializeAlloca(const AllocaInst *AI) override;
unsigned fastMaterializeConstant(const Constant *C) override;
unsigned fastMaterializeFloatZero(const ConstantFP* CF) override;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
[FastISel][AArch64] Move over to target-dependent instruction selection only. This change moves FastISel for AArch64 to target-dependent instruction selection only. This change replicates the existing target-independent behavior, therefore there are no changes to the unit tests or new tests. Future changes will take advantage of this change and update functionality and unit tests. llvm-svn: 216955
2014-09-03 05:32:54 +08:00
explicit AArch64FastISel(FunctionLoweringInfo &FuncInfo,
const TargetLibraryInfo *LibInfo)
: FastISel(FuncInfo, LibInfo, /*SkipTargetIndependentISel=*/true) {
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
Subtarget = &TM.getSubtarget<AArch64Subtarget>();
[FastISel][AArch64] Move over to target-dependent instruction selection only. This change moves FastISel for AArch64 to target-dependent instruction selection only. This change replicates the existing target-independent behavior, therefore there are no changes to the unit tests or new tests. Future changes will take advantage of this change and update functionality and unit tests. llvm-svn: 216955
2014-09-03 05:32:54 +08:00
Context = &FuncInfo.Fn->getContext();
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
bool fastSelectInstruction(const Instruction *I) override;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
#include "AArch64GenFastISel.inc"
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
};
} // end anonymous namespace
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
#include "AArch64GenCallingConv.inc"
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
CCAssignFn *AArch64FastISel::CCAssignFnForCall(CallingConv::ID CC) const {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
if (CC == CallingConv::WebKit_JS)
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
return CC_AArch64_WebKit_JS;
return Subtarget->isTargetDarwin() ? CC_AArch64_DarwinPCS : CC_AArch64_AAPCS;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
unsigned AArch64FastISel::fastMaterializeAlloca(const AllocaInst *AI) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
assert(TLI.getValueType(AI->getType(), true) == MVT::i64 &&
"Alloca should always return a pointer.");
// Don't handle dynamic allocas.
if (!FuncInfo.StaticAllocaMap.count(AI))
return 0;
DenseMap<const AllocaInst *, int>::iterator SI =
FuncInfo.StaticAllocaMap.find(AI);
if (SI != FuncInfo.StaticAllocaMap.end()) {
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
unsigned ResultReg = createResultReg(&AArch64::GPR64spRegClass);
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
ResultReg)
.addFrameIndex(SI->second)
.addImm(0)
.addImm(0);
return ResultReg;
}
return 0;
}
Reapplying [FastISel][AArch64] Cleanup constant materialization code. NFCI. Note: This reapplies r215582 without any modifications. The refactoring wasn't responsible for the buildbot failures. Original commit message: Cleanup and prepare constant materialization code for future commits. llvm-svn: 215752
2014-08-16 02:55:52 +08:00
unsigned AArch64FastISel::AArch64MaterializeInt(const ConstantInt *CI, MVT VT) {
if (VT > MVT::i64)
return 0;
Reapply [FastISel][AArch64] Make use of the zero register when possible (r215591). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: This change materializes now the value "0" from the zero register. The zero register can be folded by several instruction, so no materialization is need at all. Fixes <rdar://problem/17924413>. llvm-svn: 216009
2014-08-20 03:44:02 +08:00
if (!CI->isZero())
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
return fastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue());
Reapply [FastISel][AArch64] Make use of the zero register when possible (r215591). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: This change materializes now the value "0" from the zero register. The zero register can be folded by several instruction, so no materialization is need at all. Fixes <rdar://problem/17924413>. llvm-svn: 216009
2014-08-20 03:44:02 +08:00
// Create a copy from the zero register to materialize a "0" value.
const TargetRegisterClass *RC = (VT == MVT::i64) ? &AArch64::GPR64RegClass
: &AArch64::GPR32RegClass;
unsigned ZeroReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
unsigned ResultReg = createResultReg(RC);
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TargetOpcode::COPY),
ResultReg).addReg(ZeroReg, getKillRegState(true));
Reapply [FastISel][AArch64] Make use of the zero register when possible (r215591). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: This change materializes now the value "0" from the zero register. The zero register can be folded by several instruction, so no materialization is need at all. Fixes <rdar://problem/17924413>. llvm-svn: 216009
2014-08-20 03:44:02 +08:00
return ResultReg;
Reapplying [FastISel][AArch64] Cleanup constant materialization code. NFCI. Note: This reapplies r215582 without any modifications. The refactoring wasn't responsible for the buildbot failures. Original commit message: Cleanup and prepare constant materialization code for future commits. llvm-svn: 215752
2014-08-16 02:55:52 +08:00
}
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
unsigned AArch64FastISel::AArch64MaterializeFP(const ConstantFP *CFP, MVT VT) {
[FastISel][AArch64] Refactor float zero materialization. NFCI. llvm-svn: 216403
2014-08-26 03:58:05 +08:00
// Positive zero (+0.0) has to be materialized with a fmov from the zero
// register, because the immediate version of fmov cannot encode zero.
if (CFP->isNullValue())
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
return fastMaterializeFloatZero(CFP);
[FastISel][AArch64] Refactor float zero materialization. NFCI. llvm-svn: 216403
2014-08-26 03:58:05 +08:00
[ARM64][fast-isel] Fast-isel doesn't know how to handle f128. llvm-svn: 207659
2014-04-30 23:29:57 +08:00
if (VT != MVT::f32 && VT != MVT::f64)
return 0;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
const APFloat Val = CFP->getValueAPF();
Reapplying [FastISel][AArch64] Cleanup constant materialization code. NFCI. Note: This reapplies r215582 without any modifications. The refactoring wasn't responsible for the buildbot failures. Original commit message: Cleanup and prepare constant materialization code for future commits. llvm-svn: 215752
2014-08-16 02:55:52 +08:00
bool Is64Bit = (VT == MVT::f64);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// This checks to see if we can use FMOV instructions to materialize
// a constant, otherwise we have to materialize via the constant pool.
if (TLI.isFPImmLegal(Val, VT)) {
[FastISel][AArch64] Refactor float zero materialization. NFCI. llvm-svn: 216403
2014-08-26 03:58:05 +08:00
int Imm =
Is64Bit ? AArch64_AM::getFP64Imm(Val) : AArch64_AM::getFP32Imm(Val);
assert((Imm != -1) && "Cannot encode floating-point constant.");
Reapplying [FastISel][AArch64] Cleanup constant materialization code. NFCI. Note: This reapplies r215582 without any modifications. The refactoring wasn't responsible for the buildbot failures. Original commit message: Cleanup and prepare constant materialization code for future commits. llvm-svn: 215752
2014-08-16 02:55:52 +08:00
unsigned Opc = Is64Bit ? AArch64::FMOVDi : AArch64::FMOVSi;
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
return fastEmitInst_i(Opc, TLI.getRegClassFor(VT), Imm);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
// Materialize via constant pool. MachineConstantPool wants an explicit
// alignment.
unsigned Align = DL.getPrefTypeAlignment(CFP->getType());
if (Align == 0)
Align = DL.getTypeAllocSize(CFP->getType());
Reapplying [FastISel][AArch64] Cleanup constant materialization code. NFCI. Note: This reapplies r215582 without any modifications. The refactoring wasn't responsible for the buildbot failures. Original commit message: Cleanup and prepare constant materialization code for future commits. llvm-svn: 215752
2014-08-16 02:55:52 +08:00
unsigned CPI = MCP.getConstantPoolIndex(cast<Constant>(CFP), Align);
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
[FastISel][AArch64] Refactor float zero materialization. NFCI. llvm-svn: 216403
2014-08-26 03:58:05 +08:00
ADRPReg).addConstantPoolIndex(CPI, 0, AArch64II::MO_PAGE);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
Reapplying [FastISel][AArch64] Cleanup constant materialization code. NFCI. Note: This reapplies r215582 without any modifications. The refactoring wasn't responsible for the buildbot failures. Original commit message: Cleanup and prepare constant materialization code for future commits. llvm-svn: 215752
2014-08-16 02:55:52 +08:00
unsigned Opc = Is64Bit ? AArch64::LDRDui : AArch64::LDRSui;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
[FastISel][AArch64] Refactor float zero materialization. NFCI. llvm-svn: 216403
2014-08-26 03:58:05 +08:00
.addReg(ADRPReg)
.addConstantPoolIndex(CPI, 0, AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return ResultReg;
}
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
unsigned AArch64FastISel::AArch64MaterializeGV(const GlobalValue *GV) {
[pr19844] Add thread local mode to aliases. This matches gcc's behavior. It also seems natural given that aliases contain other properties that govern how it is accessed (linkage, visibility, dll storage). Clang still has to be updated to expose this feature to C. llvm-svn: 209759
2014-05-29 02:15:43 +08:00
// We can't handle thread-local variables quickly yet.
if (GV->isThreadLocal())
return 0;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
AArch64: disable FastISel for large code model. The code emitted is what would be expected for the small model, so it shouldn't be used when objects can be the full 64-bits away. This fixes MCJIT tests on Linux. llvm-svn: 209585
2014-05-25 03:45:41 +08:00
// MachO still uses GOT for large code-model accesses, but ELF requires
// movz/movk sequences, which FastISel doesn't handle yet.
if (TM.getCodeModel() != CodeModel::Small && !Subtarget->isTargetMachO())
return 0;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
unsigned char OpFlags = Subtarget->ClassifyGlobalReference(GV, TM);
EVT DestEVT = TLI.getValueType(GV->getType(), true);
if (!DestEVT.isSimple())
return 0;
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
ARM64: add constraints to various FastISel operations llvm-svn: 206284
2014-04-15 21:59:53 +08:00
unsigned ResultReg;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
if (OpFlags & AArch64II::MO_GOT) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// ADRP + LDRX
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
ADRPReg)
Reapplying [FastISel][AArch64] Cleanup constant materialization code. NFCI. Note: This reapplies r215582 without any modifications. The refactoring wasn't responsible for the buildbot failures. Original commit message: Cleanup and prepare constant materialization code for future commits. llvm-svn: 215752
2014-08-16 02:55:52 +08:00
.addGlobalAddress(GV, 0, AArch64II::MO_GOT | AArch64II::MO_PAGE);
ARM64: add constraints to various FastISel operations llvm-svn: 206284
2014-04-15 21:59:53 +08:00
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
ResultReg = createResultReg(&AArch64::GPR64RegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::LDRXui),
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
ResultReg)
Reapplying [FastISel][AArch64] Cleanup constant materialization code. NFCI. Note: This reapplies r215582 without any modifications. The refactoring wasn't responsible for the buildbot failures. Original commit message: Cleanup and prepare constant materialization code for future commits. llvm-svn: 215752
2014-08-16 02:55:52 +08:00
.addReg(ADRPReg)
.addGlobalAddress(GV, 0, AArch64II::MO_GOT | AArch64II::MO_PAGEOFF |
AArch64II::MO_NC);
[AArch 64] Use a constant pool load for weak symbol references when using static relocation model and small code model. Summary: currently we generate GOT based relocations for weak symbol references regardless of the underlying relocation model. This should be change so that in static relocation model we use a constant pool load instead. Patch from: Keith Walker Reviewers: Renato Golin, Tim Northover llvm-svn: 217503
2014-09-10 21:54:38 +08:00
} else if (OpFlags & AArch64II::MO_CONSTPOOL) {
// We can't handle addresses loaded from a constant pool quickly yet.
return 0;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
} else {
// ADRP + ADDX
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
Reapplying [FastISel][AArch64] Cleanup constant materialization code. NFCI. Note: This reapplies r215582 without any modifications. The refactoring wasn't responsible for the buildbot failures. Original commit message: Cleanup and prepare constant materialization code for future commits. llvm-svn: 215752
2014-08-16 02:55:52 +08:00
ADRPReg)
.addGlobalAddress(GV, 0, AArch64II::MO_PAGE);
ARM64: add constraints to various FastISel operations llvm-svn: 206284
2014-04-15 21:59:53 +08:00
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
ResultReg = createResultReg(&AArch64::GPR64spRegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
ResultReg)
Reapplying [FastISel][AArch64] Cleanup constant materialization code. NFCI. Note: This reapplies r215582 without any modifications. The refactoring wasn't responsible for the buildbot failures. Original commit message: Cleanup and prepare constant materialization code for future commits. llvm-svn: 215752
2014-08-16 02:55:52 +08:00
.addReg(ADRPReg)
.addGlobalAddress(GV, 0, AArch64II::MO_PAGEOFF | AArch64II::MO_NC)
.addImm(0);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
return ResultReg;
}
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
unsigned AArch64FastISel::fastMaterializeConstant(const Constant *C) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
EVT CEVT = TLI.getValueType(C->getType(), true);
// Only handle simple types.
if (!CEVT.isSimple())
return 0;
MVT VT = CEVT.getSimpleVT();
Reapplying [FastISel][AArch64] Cleanup constant materialization code. NFCI. Note: This reapplies r215582 without any modifications. The refactoring wasn't responsible for the buildbot failures. Original commit message: Cleanup and prepare constant materialization code for future commits. llvm-svn: 215752
2014-08-16 02:55:52 +08:00
if (const auto *CI = dyn_cast<ConstantInt>(C))
return AArch64MaterializeInt(CI, VT);
else if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
return AArch64MaterializeFP(CFP, VT);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
return AArch64MaterializeGV(GV);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return 0;
}
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
unsigned AArch64FastISel::fastMaterializeFloatZero(const ConstantFP* CFP) {
[FastISel][AArch64] Refactor float zero materialization. NFCI. llvm-svn: 216403
2014-08-26 03:58:05 +08:00
assert(CFP->isNullValue() &&
"Floating-point constant is not a positive zero.");
MVT VT;
if (!isTypeLegal(CFP->getType(), VT))
return 0;
if (VT != MVT::f32 && VT != MVT::f64)
return 0;
bool Is64Bit = (VT == MVT::f64);
unsigned ZReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
unsigned Opc = Is64Bit ? AArch64::FMOVXDr : AArch64::FMOVWSr;
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
return fastEmitInst_r(Opc, TLI.getRegClassFor(VT), ZReg, /*IsKill=*/true);
[FastISel][AArch64] Refactor float zero materialization. NFCI. llvm-svn: 216403
2014-08-26 03:58:05 +08:00
}
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Computes the address to get to an object.
Revert "[FastISel][AArch64] Don't fold instructions too aggressively into the memory operation." Quentin pointed out that this is not the correct approach and there is a better and easier solution. llvm-svn: 216632
2014-08-28 07:09:40 +08:00
bool AArch64FastISel::ComputeAddress(const Value *Obj, Address &Addr, Type *Ty)
{
[C++] Use 'nullptr'. Target edition. llvm-svn: 207197
2014-04-25 13:30:21 +08:00
const User *U = nullptr;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
unsigned Opcode = Instruction::UserOp1;
if (const Instruction *I = dyn_cast<Instruction>(Obj)) {
// Don't walk into other basic blocks unless the object is an alloca from
// another block, otherwise it may not have a virtual register assigned.
if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(Obj)) ||
FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
Opcode = I->getOpcode();
U = I;
}
} else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Obj)) {
Opcode = C->getOpcode();
U = C;
}
if (const PointerType *Ty = dyn_cast<PointerType>(Obj->getType()))
if (Ty->getAddressSpace() > 255)
// Fast instruction selection doesn't support the special
// address spaces.
return false;
switch (Opcode) {
default:
break;
case Instruction::BitCast: {
// Look through bitcasts.
Revert "[FastISel][AArch64] Don't fold instructions too aggressively into the memory operation." Quentin pointed out that this is not the correct approach and there is a better and easier solution. llvm-svn: 216632
2014-08-28 07:09:40 +08:00
return ComputeAddress(U->getOperand(0), Addr, Ty);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
case Instruction::IntToPtr: {
// Look past no-op inttoptrs.
if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
Revert "[FastISel][AArch64] Don't fold instructions too aggressively into the memory operation." Quentin pointed out that this is not the correct approach and there is a better and easier solution. llvm-svn: 216632
2014-08-28 07:09:40 +08:00
return ComputeAddress(U->getOperand(0), Addr, Ty);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
break;
}
case Instruction::PtrToInt: {
Revert "[FastISel][AArch64] Don't fold instructions too aggressively into the memory operation." Quentin pointed out that this is not the correct approach and there is a better and easier solution. llvm-svn: 216632
2014-08-28 07:09:40 +08:00
// Look past no-op ptrtoints.
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
Revert "[FastISel][AArch64] Don't fold instructions too aggressively into the memory operation." Quentin pointed out that this is not the correct approach and there is a better and easier solution. llvm-svn: 216632
2014-08-28 07:09:40 +08:00
return ComputeAddress(U->getOperand(0), Addr, Ty);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
break;
}
case Instruction::GetElementPtr: {
Address SavedAddr = Addr;
uint64_t TmpOffset = Addr.getOffset();
// Iterate through the GEP folding the constants into offsets where
// we can.
gep_type_iterator GTI = gep_type_begin(U);
for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e;
++i, ++GTI) {
const Value *Op = *i;
if (StructType *STy = dyn_cast<StructType>(*GTI)) {
const StructLayout *SL = DL.getStructLayout(STy);
unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
TmpOffset += SL->getElementOffset(Idx);
} else {
uint64_t S = DL.getTypeAllocSize(GTI.getIndexedType());
for (;;) {
if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
// Constant-offset addressing.
TmpOffset += CI->getSExtValue() * S;
break;
}
if (canFoldAddIntoGEP(U, Op)) {
// A compatible add with a constant operand. Fold the constant.
ConstantInt *CI =
cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
TmpOffset += CI->getSExtValue() * S;
// Iterate on the other operand.
Op = cast<AddOperator>(Op)->getOperand(0);
continue;
}
// Unsupported
goto unsupported_gep;
}
}
}
// Try to grab the base operand now.
Addr.setOffset(TmpOffset);
Revert "[FastISel][AArch64] Don't fold instructions too aggressively into the memory operation." Quentin pointed out that this is not the correct approach and there is a better and easier solution. llvm-svn: 216632
2014-08-28 07:09:40 +08:00
if (ComputeAddress(U->getOperand(0), Addr, Ty))
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return true;
// We failed, restore everything and try the other options.
Addr = SavedAddr;
unsupported_gep:
break;
}
case Instruction::Alloca: {
const AllocaInst *AI = cast<AllocaInst>(Obj);
DenseMap<const AllocaInst *, int>::iterator SI =
FuncInfo.StaticAllocaMap.find(AI);
if (SI != FuncInfo.StaticAllocaMap.end()) {
Addr.setKind(Address::FrameIndexBase);
Addr.setFI(SI->second);
return true;
}
break;
}
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
case Instruction::Add: {
[FastISel][AArch64] Fold offset into the memory operation. Fold simple offsets into the memory operation: add x0, x0, #8 ldr x0, [x0] --> ldr x0, [x0, #8] Fixes <rdar://problem/17887945>. llvm-svn: 214545
2014-08-02 03:40:16 +08:00
// Adds of constants are common and easy enough.
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
const Value *LHS = U->getOperand(0);
const Value *RHS = U->getOperand(1);
if (isa<ConstantInt>(LHS))
std::swap(LHS, RHS);
if (const ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
[FastISel][AArch64] Fold offset into the memory operation. Fold simple offsets into the memory operation: add x0, x0, #8 ldr x0, [x0] --> ldr x0, [x0, #8] Fixes <rdar://problem/17887945>. llvm-svn: 214545
2014-08-02 03:40:16 +08:00
Addr.setOffset(Addr.getOffset() + (uint64_t)CI->getSExtValue());
Revert "[FastISel][AArch64] Don't fold instructions too aggressively into the memory operation." Quentin pointed out that this is not the correct approach and there is a better and easier solution. llvm-svn: 216632
2014-08-28 07:09:40 +08:00
return ComputeAddress(LHS, Addr, Ty);
[FastISel][AArch64] Fold offset into the memory operation. Fold simple offsets into the memory operation: add x0, x0, #8 ldr x0, [x0] --> ldr x0, [x0, #8] Fixes <rdar://problem/17887945>. llvm-svn: 214545
2014-08-02 03:40:16 +08:00
}
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
Address Backup = Addr;
Revert "[FastISel][AArch64] Don't fold instructions too aggressively into the memory operation." Quentin pointed out that this is not the correct approach and there is a better and easier solution. llvm-svn: 216632
2014-08-28 07:09:40 +08:00
if (ComputeAddress(LHS, Addr, Ty) && ComputeAddress(RHS, Addr, Ty))
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
return true;
Addr = Backup;
[FastISel][AArch64] Fold offset into the memory operation. Fold simple offsets into the memory operation: add x0, x0, #8 ldr x0, [x0] --> ldr x0, [x0, #8] Fixes <rdar://problem/17887945>. llvm-svn: 214545
2014-08-02 03:40:16 +08:00
break;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
case Instruction::Shl:
if (Addr.getOffsetReg())
break;
if (const auto *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
unsigned Val = CI->getZExtValue();
if (Val < 1 || Val > 3)
break;
uint64_t NumBytes = 0;
if (Ty && Ty->isSized()) {
uint64_t NumBits = DL.getTypeSizeInBits(Ty);
NumBytes = NumBits / 8;
if (!isPowerOf2_64(NumBits))
NumBytes = 0;
}
Silencing an MSVC C4334 warning ('<<' : result of 32-bit shift implicitly converted to 64 bits (was 64-bit shift intended?)). NFC. llvm-svn: 216067
2014-08-20 20:14:35 +08:00
if (NumBytes != (1ULL << Val))
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
break;
Addr.setShift(Val);
Addr.setExtendType(AArch64_AM::LSL);
if (const auto *I = dyn_cast<Instruction>(U->getOperand(0)))
Revert "[FastISel][AArch64] Don't fold instructions too aggressively into the memory operation." Quentin pointed out that this is not the correct approach and there is a better and easier solution. llvm-svn: 216632
2014-08-28 07:09:40 +08:00
if (FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB)
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
U = I;
if (const auto *ZE = dyn_cast<ZExtInst>(U))
if (ZE->getOperand(0)->getType()->isIntegerTy(32))
Addr.setExtendType(AArch64_AM::UXTW);
if (const auto *SE = dyn_cast<SExtInst>(U))
if (SE->getOperand(0)->getType()->isIntegerTy(32))
Addr.setExtendType(AArch64_AM::SXTW);
unsigned Reg = getRegForValue(U->getOperand(0));
if (!Reg)
return false;
Addr.setOffsetReg(Reg);
return true;
}
break;
}
if (Addr.getReg()) {
if (!Addr.getOffsetReg()) {
unsigned Reg = getRegForValue(Obj);
if (!Reg)
return false;
Addr.setOffsetReg(Reg);
return true;
}
return false;
}
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
unsigned Reg = getRegForValue(Obj);
if (!Reg)
return false;
Addr.setReg(Reg);
return true;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
[FastISel][AArch64] Add MachO large code model support for function calls. Currently the large code model for MachO uses the GOT to make function calls. Emit the required adrp and ldr instructions to load the address from the GOT. Related to <rdar://problem/17733076>. llvm-svn: 214381
2014-07-31 12:10:40 +08:00
bool AArch64FastISel::ComputeCallAddress(const Value *V, Address &Addr) {
const User *U = nullptr;
unsigned Opcode = Instruction::UserOp1;
bool InMBB = true;
if (const auto *I = dyn_cast<Instruction>(V)) {
Opcode = I->getOpcode();
U = I;
InMBB = I->getParent() == FuncInfo.MBB->getBasicBlock();
} else if (const auto *C = dyn_cast<ConstantExpr>(V)) {
Opcode = C->getOpcode();
U = C;
}
switch (Opcode) {
default: break;
case Instruction::BitCast:
// Look past bitcasts if its operand is in the same BB.
if (InMBB)
return ComputeCallAddress(U->getOperand(0), Addr);
break;
case Instruction::IntToPtr:
// Look past no-op inttoptrs if its operand is in the same BB.
if (InMBB &&
TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
return ComputeCallAddress(U->getOperand(0), Addr);
break;
case Instruction::PtrToInt:
// Look past no-op ptrtoints if its operand is in the same BB.
if (InMBB &&
TLI.getValueType(U->getType()) == TLI.getPointerTy())
return ComputeCallAddress(U->getOperand(0), Addr);
break;
}
if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
Addr.setGlobalValue(GV);
return true;
}
// If all else fails, try to materialize the value in a register.
if (!Addr.getGlobalValue()) {
Addr.setReg(getRegForValue(V));
return Addr.getReg() != 0;
}
return false;
}
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
bool AArch64FastISel::isTypeLegal(Type *Ty, MVT &VT) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
EVT evt = TLI.getValueType(Ty, true);
// Only handle simple types.
if (evt == MVT::Other || !evt.isSimple())
return false;
VT = evt.getSimpleVT();
[ARM64][fast-isel] Fast-isel doesn't know how to handle f128. llvm-svn: 207659
2014-04-30 23:29:57 +08:00
// This is a legal type, but it's not something we handle in fast-isel.
if (VT == MVT::f128)
return false;
// Handle all other legal types, i.e. a register that will directly hold this
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// value.
return TLI.isTypeLegal(VT);
}
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
bool AArch64FastISel::isLoadStoreTypeLegal(Type *Ty, MVT &VT) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
if (isTypeLegal(Ty, VT))
return true;
// If this is a type than can be sign or zero-extended to a basic operation
// go ahead and accept it now. For stores, this reflects truncation.
if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16)
return true;
return false;
}
[FastISel][AArch64] Use a new helper function to determine if a value type is supported. NFCI. FastISel for AArch64 supports more value types than are actually legal. Use a dedicated helper function to reflect this. It is very similar to the isLoadStoreTypeLegal function, with the exception that vector types are not supported yet. llvm-svn: 216984
2014-09-03 06:33:53 +08:00
/// \brief Determine if the value type is supported by FastISel.
///
/// FastISel for AArch64 can handle more value types than are legal. This adds
/// simple value type such as i1, i8, and i16.
/// Vectors on the other side are not supported yet.
bool AArch64FastISel::isTypeSupported(Type *Ty, MVT &VT) {
if (Ty->isVectorTy())
return false;
if (isTypeLegal(Ty, VT))
return true;
// If this is a type than can be sign or zero-extended to a basic operation
// go ahead and accept it now.
if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16)
return true;
return false;
}
[FastISel][AArch64] Don't fold instructions that are not in the same basic block. This fix checks first if the instruction to be folded (e.g. sign-/zero-extend, or shift) is in the same machine basic block as the instruction we are folding into. Not doing so can result in incorrect code, because the value might not be live-out of the basic block, where the value is defined. This fixes rdar://problem/18169495. llvm-svn: 216700
2014-08-29 08:19:21 +08:00
bool AArch64FastISel::isValueAvailable(const Value *V) const {
if (!isa<Instruction>(V))
return true;
const auto *I = cast<Instruction>(V);
if (FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB)
return true;
return false;
}
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
bool AArch64FastISel::SimplifyAddress(Address &Addr, MVT VT) {
unsigned ScaleFactor;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
switch (VT.SimpleTy) {
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
default: return false;
case MVT::i1: // fall-through
case MVT::i8: ScaleFactor = 1; break;
case MVT::i16: ScaleFactor = 2; break;
case MVT::i32: // fall-through
case MVT::f32: ScaleFactor = 4; break;
case MVT::i64: // fall-through
case MVT::f64: ScaleFactor = 8; break;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
bool ImmediateOffsetNeedsLowering = false;
bool RegisterOffsetNeedsLowering = false;
int64_t Offset = Addr.getOffset();
if (((Offset < 0) || (Offset & (ScaleFactor - 1))) && !isInt<9>(Offset))
ImmediateOffsetNeedsLowering = true;
else if (Offset > 0 && !(Offset & (ScaleFactor - 1)) &&
!isUInt<12>(Offset / ScaleFactor))
ImmediateOffsetNeedsLowering = true;
// Cannot encode an offset register and an immediate offset in the same
// instruction. Fold the immediate offset into the load/store instruction and
// emit an additonal add to take care of the offset register.
if (!ImmediateOffsetNeedsLowering && Addr.getOffset() && Addr.isRegBase() &&
Addr.getOffsetReg())
RegisterOffsetNeedsLowering = true;
[FastISel][AArch64] Fix simplify address when the address comes from a shift. When the address comes directly from a shift instruction then the address computation cannot be folded into the memory instruction, because the zero register is not available as a base register. Simplify addess needs to emit the shift instruction and use the result as base register. llvm-svn: 216621
2014-08-28 05:38:33 +08:00
// Cannot encode zero register as base.
if (Addr.isRegBase() && Addr.getOffsetReg() && !Addr.getReg())
RegisterOffsetNeedsLowering = true;
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
// If this is a stack pointer and the offset needs to be simplified then put
AArch64: teach FastISel how to handle offset FrameIndices Previously we were abandonning the attempt, leading to some combination of extra work (when selection of a load/store fails completely) and inferior code (when this leads to a real memcpy call instead of inlining). rdar://problem/17187463 llvm-svn: 210520
2014-06-10 17:52:44 +08:00
// the alloca address into a register, set the base type back to register and
// continue. This should almost never happen.
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
if (ImmediateOffsetNeedsLowering && Addr.isFIBase()) {
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
unsigned ResultReg = createResultReg(&AArch64::GPR64spRegClass);
AArch64: teach FastISel how to handle offset FrameIndices Previously we were abandonning the attempt, leading to some combination of extra work (when selection of a load/store fails completely) and inferior code (when this leads to a real memcpy call instead of inlining). rdar://problem/17187463 llvm-svn: 210520
2014-06-10 17:52:44 +08:00
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
ResultReg)
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
.addFrameIndex(Addr.getFI())
.addImm(0)
.addImm(0);
AArch64: teach FastISel how to handle offset FrameIndices Previously we were abandonning the attempt, leading to some combination of extra work (when selection of a load/store fails completely) and inferior code (when this leads to a real memcpy call instead of inlining). rdar://problem/17187463 llvm-svn: 210520
2014-06-10 17:52:44 +08:00
Addr.setKind(Address::RegBase);
Addr.setReg(ResultReg);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
if (RegisterOffsetNeedsLowering) {
unsigned ResultReg = 0;
[FastISel][AArch64] Fix address simplification. When a shift with extension or an add with shift and extension cannot be folded into the memory operation, then the address calculation has to be materialized separately. While doing so the code forgot to consider a possible sign-/zero- extension. This fix folds now also the sign-/zero-extension into the add or shift instruction which is used to materialize the address. This fixes rdar://problem/18141718. llvm-svn: 216511
2014-08-27 08:58:30 +08:00
if (Addr.getReg()) {
if (Addr.getExtendType() == AArch64_AM::SXTW ||
Addr.getExtendType() == AArch64_AM::UXTW )
ResultReg = emitAddSub_rx(/*UseAdd=*/true, MVT::i64, Addr.getReg(),
/*TODO:IsKill=*/false, Addr.getOffsetReg(),
/*TODO:IsKill=*/false, Addr.getExtendType(),
Addr.getShift());
else
ResultReg = emitAddSub_rs(/*UseAdd=*/true, MVT::i64, Addr.getReg(),
/*TODO:IsKill=*/false, Addr.getOffsetReg(),
/*TODO:IsKill=*/false, AArch64_AM::LSL,
Addr.getShift());
} else {
if (Addr.getExtendType() == AArch64_AM::UXTW)
ResultReg = emitLSL_ri(MVT::i64, MVT::i32, Addr.getOffsetReg(),
/*Op0IsKill=*/false, Addr.getShift(),
/*IsZExt=*/true);
else if (Addr.getExtendType() == AArch64_AM::SXTW)
ResultReg = emitLSL_ri(MVT::i64, MVT::i32, Addr.getOffsetReg(),
/*Op0IsKill=*/false, Addr.getShift(),
/*IsZExt=*/false);
else
ResultReg = emitLSL_ri(MVT::i64, MVT::i64, Addr.getOffsetReg(),
/*Op0IsKill=*/false, Addr.getShift());
}
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
if (!ResultReg)
return false;
Addr.setReg(ResultReg);
Addr.setOffsetReg(0);
Addr.setShift(0);
[FastISel][AArch64] Fix address simplification. When a shift with extension or an add with shift and extension cannot be folded into the memory operation, then the address calculation has to be materialized separately. While doing so the code forgot to consider a possible sign-/zero- extension. This fix folds now also the sign-/zero-extension into the add or shift instruction which is used to materialize the address. This fixes rdar://problem/18141718. llvm-svn: 216511
2014-08-27 08:58:30 +08:00
Addr.setExtendType(AArch64_AM::InvalidShiftExtend);
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
}
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Since the offset is too large for the load/store instruction get the
// reg+offset into a register.
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
if (ImmediateOffsetNeedsLowering) {
unsigned ResultReg = 0;
if (Addr.getReg())
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
ResultReg = fastEmit_ri_(MVT::i64, ISD::ADD, Addr.getReg(),
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
/*IsKill=*/false, Offset, MVT::i64);
else
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
ResultReg = fastEmit_i(MVT::i64, MVT::i64, ISD::Constant, Offset);
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
if (!ResultReg)
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return false;
Addr.setReg(ResultReg);
Addr.setOffset(0);
}
return true;
}
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
void AArch64FastISel::AddLoadStoreOperands(Address &Addr,
const MachineInstrBuilder &MIB,
[FastISel][AArch64] Attach MachineMemOperands to load and store instructions. llvm-svn: 215231
2014-08-09 01:24:10 +08:00
unsigned Flags,
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
unsigned ScaleFactor,
MachineMemOperand *MMO) {
int64_t Offset = Addr.getOffset() / ScaleFactor;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Frame base works a bit differently. Handle it separately.
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
if (Addr.isFIBase()) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
int FI = Addr.getFI();
// FIXME: We shouldn't be using getObjectSize/getObjectAlignment. The size
// and alignment should be based on the VT.
[FastISel][AArch64] Attach MachineMemOperands to load and store instructions. llvm-svn: 215231
2014-08-09 01:24:10 +08:00
MMO = FuncInfo.MF->getMachineMemOperand(
MachinePointerInfo::getFixedStack(FI, Offset), Flags,
MFI.getObjectSize(FI), MFI.getObjectAlignment(FI));
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Now add the rest of the operands.
[FastISel][AArch64] Attach MachineMemOperands to load and store instructions. llvm-svn: 215231
2014-08-09 01:24:10 +08:00
MIB.addFrameIndex(FI).addImm(Offset);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
} else {
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
assert(Addr.isRegBase() && "Unexpected address kind.");
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
const MCInstrDesc &II = MIB->getDesc();
unsigned Idx = (Flags & MachineMemOperand::MOStore) ? 1 : 0;
Addr.setReg(
constrainOperandRegClass(II, Addr.getReg(), II.getNumDefs()+Idx));
Addr.setOffsetReg(
constrainOperandRegClass(II, Addr.getOffsetReg(), II.getNumDefs()+Idx+1));
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
if (Addr.getOffsetReg()) {
assert(Addr.getOffset() == 0 && "Unexpected offset");
bool IsSigned = Addr.getExtendType() == AArch64_AM::SXTW ||
Addr.getExtendType() == AArch64_AM::SXTX;
MIB.addReg(Addr.getReg());
MIB.addReg(Addr.getOffsetReg());
MIB.addImm(IsSigned);
MIB.addImm(Addr.getShift() != 0);
} else {
MIB.addReg(Addr.getReg());
MIB.addImm(Offset);
}
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
[FastISel][AArch64] Attach MachineMemOperands to load and store instructions. llvm-svn: 215231
2014-08-09 01:24:10 +08:00
if (MMO)
MIB.addMemOperand(MMO);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
unsigned AArch64FastISel::emitAddSub(bool UseAdd, MVT RetVT, const Value *LHS,
const Value *RHS, bool SetFlags,
bool WantResult, bool IsZExt) {
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
AArch64_AM::ShiftExtendType ExtendType = AArch64_AM::InvalidShiftExtend;
[FastISel][AArch64] Don't fold the sign-/zero-extend from i1 into the compare. This fixes a bug I introduced in a previous commit (r216033). Sign-/Zero- extension from i1 cannot be folded into the ADDS/SUBS instructions. Instead both operands have to be sign-/zero-extended with separate instructions. Related to <rdar://problem/17913111>. llvm-svn: 216073
2014-08-21 00:34:15 +08:00
bool NeedExtend = false;
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
switch (RetVT.SimpleTy) {
[FastISel][AArch64] Don't fold the sign-/zero-extend from i1 into the compare. This fixes a bug I introduced in a previous commit (r216033). Sign-/Zero- extension from i1 cannot be folded into the ADDS/SUBS instructions. Instead both operands have to be sign-/zero-extended with separate instructions. Related to <rdar://problem/17913111>. llvm-svn: 216073
2014-08-21 00:34:15 +08:00
default:
return 0;
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
case MVT::i1:
[FastISel][AArch64] Don't fold the sign-/zero-extend from i1 into the compare. This fixes a bug I introduced in a previous commit (r216033). Sign-/Zero- extension from i1 cannot be folded into the ADDS/SUBS instructions. Instead both operands have to be sign-/zero-extended with separate instructions. Related to <rdar://problem/17913111>. llvm-svn: 216073
2014-08-21 00:34:15 +08:00
NeedExtend = true;
break;
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
case MVT::i8:
[FastISel][AArch64] Don't fold the sign-/zero-extend from i1 into the compare. This fixes a bug I introduced in a previous commit (r216033). Sign-/Zero- extension from i1 cannot be folded into the ADDS/SUBS instructions. Instead both operands have to be sign-/zero-extended with separate instructions. Related to <rdar://problem/17913111>. llvm-svn: 216073
2014-08-21 00:34:15 +08:00
NeedExtend = true;
ExtendType = IsZExt ? AArch64_AM::UXTB : AArch64_AM::SXTB;
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
break;
case MVT::i16:
[FastISel][AArch64] Don't fold the sign-/zero-extend from i1 into the compare. This fixes a bug I introduced in a previous commit (r216033). Sign-/Zero- extension from i1 cannot be folded into the ADDS/SUBS instructions. Instead both operands have to be sign-/zero-extended with separate instructions. Related to <rdar://problem/17913111>. llvm-svn: 216073
2014-08-21 00:34:15 +08:00
NeedExtend = true;
ExtendType = IsZExt ? AArch64_AM::UXTH : AArch64_AM::SXTH;
break;
case MVT::i32: // fall-through
case MVT::i64:
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
break;
}
[FastISel][AArch64] Don't fold the sign-/zero-extend from i1 into the compare. This fixes a bug I introduced in a previous commit (r216033). Sign-/Zero- extension from i1 cannot be folded into the ADDS/SUBS instructions. Instead both operands have to be sign-/zero-extended with separate instructions. Related to <rdar://problem/17913111>. llvm-svn: 216073
2014-08-21 00:34:15 +08:00
MVT SrcVT = RetVT;
RetVT.SimpleTy = std::max(RetVT.SimpleTy, MVT::i32);
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
// Canonicalize immediates to the RHS first.
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
if (UseAdd && isa<ConstantInt>(LHS) && !isa<ConstantInt>(RHS))
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
std::swap(LHS, RHS);
// Canonicalize shift immediate to the RHS.
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
if (UseAdd && isValueAvailable(LHS))
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
if (const auto *SI = dyn_cast<BinaryOperator>(LHS))
if (isa<ConstantInt>(SI->getOperand(1)))
if (SI->getOpcode() == Instruction::Shl ||
SI->getOpcode() == Instruction::LShr ||
SI->getOpcode() == Instruction::AShr )
std::swap(LHS, RHS);
unsigned LHSReg = getRegForValue(LHS);
if (!LHSReg)
return 0;
bool LHSIsKill = hasTrivialKill(LHS);
[FastISel][AArch64] Don't fold the sign-/zero-extend from i1 into the compare. This fixes a bug I introduced in a previous commit (r216033). Sign-/Zero- extension from i1 cannot be folded into the ADDS/SUBS instructions. Instead both operands have to be sign-/zero-extended with separate instructions. Related to <rdar://problem/17913111>. llvm-svn: 216073
2014-08-21 00:34:15 +08:00
if (NeedExtend)
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
LHSReg = EmitIntExt(SrcVT, LHSReg, RetVT, IsZExt);
unsigned ResultReg = 0;
if (const auto *C = dyn_cast<ConstantInt>(RHS)) {
uint64_t Imm = IsZExt ? C->getZExtValue() : C->getSExtValue();
if (C->isNegative())
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
ResultReg = emitAddSub_ri(!UseAdd, RetVT, LHSReg, LHSIsKill, -Imm,
SetFlags, WantResult);
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
else
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
ResultReg = emitAddSub_ri(UseAdd, RetVT, LHSReg, LHSIsKill, Imm, SetFlags,
WantResult);
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
}
if (ResultReg)
return ResultReg;
[FastISel][AArch64] Don't fold the sign-/zero-extend from i1 into the compare. This fixes a bug I introduced in a previous commit (r216033). Sign-/Zero- extension from i1 cannot be folded into the ADDS/SUBS instructions. Instead both operands have to be sign-/zero-extended with separate instructions. Related to <rdar://problem/17913111>. llvm-svn: 216073
2014-08-21 00:34:15 +08:00
// Only extend the RHS within the instruction if there is a valid extend type.
[FastISel][AArch64] Don't fold instructions that are not in the same basic block. This fix checks first if the instruction to be folded (e.g. sign-/zero-extend, or shift) is in the same machine basic block as the instruction we are folding into. Not doing so can result in incorrect code, because the value might not be live-out of the basic block, where the value is defined. This fixes rdar://problem/18169495. llvm-svn: 216700
2014-08-29 08:19:21 +08:00
if (ExtendType != AArch64_AM::InvalidShiftExtend && isValueAvailable(RHS)) {
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
if (const auto *SI = dyn_cast<BinaryOperator>(RHS))
if (const auto *C = dyn_cast<ConstantInt>(SI->getOperand(1)))
if ((SI->getOpcode() == Instruction::Shl) && (C->getZExtValue() < 4)) {
unsigned RHSReg = getRegForValue(SI->getOperand(0));
if (!RHSReg)
return 0;
bool RHSIsKill = hasTrivialKill(SI->getOperand(0));
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
return emitAddSub_rx(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg,
RHSIsKill, ExtendType, C->getZExtValue(),
SetFlags, WantResult);
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
}
unsigned RHSReg = getRegForValue(RHS);
if (!RHSReg)
return 0;
bool RHSIsKill = hasTrivialKill(RHS);
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
return emitAddSub_rx(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg, RHSIsKill,
ExtendType, 0, SetFlags, WantResult);
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
}
// Check if the shift can be folded into the instruction.
[FastISel][AArch64] Don't fold instructions that are not in the same basic block. This fix checks first if the instruction to be folded (e.g. sign-/zero-extend, or shift) is in the same machine basic block as the instruction we are folding into. Not doing so can result in incorrect code, because the value might not be live-out of the basic block, where the value is defined. This fixes rdar://problem/18169495. llvm-svn: 216700
2014-08-29 08:19:21 +08:00
if (isValueAvailable(RHS))
if (const auto *SI = dyn_cast<BinaryOperator>(RHS)) {
if (const auto *C = dyn_cast<ConstantInt>(SI->getOperand(1))) {
AArch64_AM::ShiftExtendType ShiftType = AArch64_AM::InvalidShiftExtend;
switch (SI->getOpcode()) {
default: break;
case Instruction::Shl: ShiftType = AArch64_AM::LSL; break;
case Instruction::LShr: ShiftType = AArch64_AM::LSR; break;
case Instruction::AShr: ShiftType = AArch64_AM::ASR; break;
}
uint64_t ShiftVal = C->getZExtValue();
if (ShiftType != AArch64_AM::InvalidShiftExtend) {
unsigned RHSReg = getRegForValue(SI->getOperand(0));
if (!RHSReg)
return 0;
bool RHSIsKill = hasTrivialKill(SI->getOperand(0));
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
return emitAddSub_rs(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg,
RHSIsKill, ShiftType, ShiftVal, SetFlags,
WantResult);
[FastISel][AArch64] Don't fold instructions that are not in the same basic block. This fix checks first if the instruction to be folded (e.g. sign-/zero-extend, or shift) is in the same machine basic block as the instruction we are folding into. Not doing so can result in incorrect code, because the value might not be live-out of the basic block, where the value is defined. This fixes rdar://problem/18169495. llvm-svn: 216700
2014-08-29 08:19:21 +08:00
}
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
}
}
unsigned RHSReg = getRegForValue(RHS);
if (!RHSReg)
return 0;
bool RHSIsKill = hasTrivialKill(RHS);
[FastISel][AArch64] Don't fold the sign-/zero-extend from i1 into the compare. This fixes a bug I introduced in a previous commit (r216033). Sign-/Zero- extension from i1 cannot be folded into the ADDS/SUBS instructions. Instead both operands have to be sign-/zero-extended with separate instructions. Related to <rdar://problem/17913111>. llvm-svn: 216073
2014-08-21 00:34:15 +08:00
if (NeedExtend)
RHSReg = EmitIntExt(SrcVT, RHSReg, RetVT, IsZExt);
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
return emitAddSub_rr(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg, RHSIsKill,
SetFlags, WantResult);
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
}
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
unsigned AArch64FastISel::emitAddSub_rr(bool UseAdd, MVT RetVT, unsigned LHSReg,
bool LHSIsKill, unsigned RHSReg,
bool RHSIsKill, bool SetFlags,
bool WantResult) {
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
assert(LHSReg && RHSReg && "Invalid register number.");
if (RetVT != MVT::i32 && RetVT != MVT::i64)
return 0;
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
static const unsigned OpcTable[2][2][2] = {
{ { AArch64::SUBWrr, AArch64::SUBXrr },
{ AArch64::ADDWrr, AArch64::ADDXrr } },
{ { AArch64::SUBSWrr, AArch64::SUBSXrr },
{ AArch64::ADDSWrr, AArch64::ADDSXrr } }
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
};
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
bool Is64Bit = RetVT == MVT::i64;
unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
const TargetRegisterClass *RC =
Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
unsigned ResultReg;
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
if (WantResult)
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
ResultReg = createResultReg(RC);
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
else
ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
const MCInstrDesc &II = TII.get(Opc);
LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
.addReg(LHSReg, getKillRegState(LHSIsKill))
.addReg(RHSReg, getKillRegState(RHSIsKill));
return ResultReg;
}
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
unsigned AArch64FastISel::emitAddSub_ri(bool UseAdd, MVT RetVT, unsigned LHSReg,
bool LHSIsKill, uint64_t Imm,
bool SetFlags, bool WantResult) {
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
assert(LHSReg && "Invalid register number.");
if (RetVT != MVT::i32 && RetVT != MVT::i64)
return 0;
unsigned ShiftImm;
if (isUInt<12>(Imm))
ShiftImm = 0;
else if ((Imm & 0xfff000) == Imm) {
ShiftImm = 12;
Imm >>= 12;
} else
return 0;
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
static const unsigned OpcTable[2][2][2] = {
{ { AArch64::SUBWri, AArch64::SUBXri },
{ AArch64::ADDWri, AArch64::ADDXri } },
{ { AArch64::SUBSWri, AArch64::SUBSXri },
{ AArch64::ADDSWri, AArch64::ADDSXri } }
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
};
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
bool Is64Bit = RetVT == MVT::i64;
unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
const TargetRegisterClass *RC;
if (SetFlags)
RC = Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
else
RC = Is64Bit ? &AArch64::GPR64spRegClass : &AArch64::GPR32spRegClass;
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
unsigned ResultReg;
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
if (WantResult)
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
ResultReg = createResultReg(RC);
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
else
ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
const MCInstrDesc &II = TII.get(Opc);
LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
.addReg(LHSReg, getKillRegState(LHSIsKill))
.addImm(Imm)
.addImm(getShifterImm(AArch64_AM::LSL, ShiftImm));
return ResultReg;
}
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
unsigned AArch64FastISel::emitAddSub_rs(bool UseAdd, MVT RetVT, unsigned LHSReg,
bool LHSIsKill, unsigned RHSReg,
bool RHSIsKill,
[FastISel][AArch64] Fix address simplification. When a shift with extension or an add with shift and extension cannot be folded into the memory operation, then the address calculation has to be materialized separately. While doing so the code forgot to consider a possible sign-/zero- extension. This fix folds now also the sign-/zero-extension into the add or shift instruction which is used to materialize the address. This fixes rdar://problem/18141718. llvm-svn: 216511
2014-08-27 08:58:30 +08:00
AArch64_AM::ShiftExtendType ShiftType,
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
uint64_t ShiftImm, bool SetFlags,
bool WantResult) {
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
assert(LHSReg && RHSReg && "Invalid register number.");
if (RetVT != MVT::i32 && RetVT != MVT::i64)
return 0;
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
static const unsigned OpcTable[2][2][2] = {
{ { AArch64::SUBWrs, AArch64::SUBXrs },
{ AArch64::ADDWrs, AArch64::ADDXrs } },
{ { AArch64::SUBSWrs, AArch64::SUBSXrs },
{ AArch64::ADDSWrs, AArch64::ADDSXrs } }
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
};
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
bool Is64Bit = RetVT == MVT::i64;
unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
const TargetRegisterClass *RC =
Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
unsigned ResultReg;
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
if (WantResult)
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
ResultReg = createResultReg(RC);
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
else
ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
const MCInstrDesc &II = TII.get(Opc);
LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
.addReg(LHSReg, getKillRegState(LHSIsKill))
.addReg(RHSReg, getKillRegState(RHSIsKill))
.addImm(getShifterImm(ShiftType, ShiftImm));
return ResultReg;
}
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
unsigned AArch64FastISel::emitAddSub_rx(bool UseAdd, MVT RetVT, unsigned LHSReg,
bool LHSIsKill, unsigned RHSReg,
bool RHSIsKill,
[FastISel][AArch64] Fix address simplification. When a shift with extension or an add with shift and extension cannot be folded into the memory operation, then the address calculation has to be materialized separately. While doing so the code forgot to consider a possible sign-/zero- extension. This fix folds now also the sign-/zero-extension into the add or shift instruction which is used to materialize the address. This fixes rdar://problem/18141718. llvm-svn: 216511
2014-08-27 08:58:30 +08:00
AArch64_AM::ShiftExtendType ExtType,
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
uint64_t ShiftImm, bool SetFlags,
bool WantResult) {
[FastISel][AArch64] Fix address simplification. When a shift with extension or an add with shift and extension cannot be folded into the memory operation, then the address calculation has to be materialized separately. While doing so the code forgot to consider a possible sign-/zero- extension. This fix folds now also the sign-/zero-extension into the add or shift instruction which is used to materialize the address. This fixes rdar://problem/18141718. llvm-svn: 216511
2014-08-27 08:58:30 +08:00
assert(LHSReg && RHSReg && "Invalid register number.");
if (RetVT != MVT::i32 && RetVT != MVT::i64)
return 0;
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
static const unsigned OpcTable[2][2][2] = {
{ { AArch64::SUBWrx, AArch64::SUBXrx },
{ AArch64::ADDWrx, AArch64::ADDXrx } },
{ { AArch64::SUBSWrx, AArch64::SUBSXrx },
{ AArch64::ADDSWrx, AArch64::ADDSXrx } }
[FastISel][AArch64] Fix address simplification. When a shift with extension or an add with shift and extension cannot be folded into the memory operation, then the address calculation has to be materialized separately. While doing so the code forgot to consider a possible sign-/zero- extension. This fix folds now also the sign-/zero-extension into the add or shift instruction which is used to materialize the address. This fixes rdar://problem/18141718. llvm-svn: 216511
2014-08-27 08:58:30 +08:00
};
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
bool Is64Bit = RetVT == MVT::i64;
unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
const TargetRegisterClass *RC = nullptr;
if (SetFlags)
RC = Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
else
RC = Is64Bit ? &AArch64::GPR64spRegClass : &AArch64::GPR32spRegClass;
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
unsigned ResultReg;
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
if (WantResult)
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
ResultReg = createResultReg(RC);
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
else
ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
const MCInstrDesc &II = TII.get(Opc);
LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
.addReg(LHSReg, getKillRegState(LHSIsKill))
.addReg(RHSReg, getKillRegState(RHSIsKill))
.addImm(getArithExtendImm(ExtType, ShiftImm));
return ResultReg;
}
bool AArch64FastISel::emitCmp(const Value *LHS, const Value *RHS, bool IsZExt) {
Type *Ty = LHS->getType();
EVT EVT = TLI.getValueType(Ty, true);
if (!EVT.isSimple())
return false;
MVT VT = EVT.getSimpleVT();
switch (VT.SimpleTy) {
default:
return false;
case MVT::i1:
case MVT::i8:
case MVT::i16:
case MVT::i32:
case MVT::i64:
return emitICmp(VT, LHS, RHS, IsZExt);
case MVT::f32:
case MVT::f64:
return emitFCmp(VT, LHS, RHS);
}
}
bool AArch64FastISel::emitICmp(MVT RetVT, const Value *LHS, const Value *RHS,
bool IsZExt) {
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
return emitSub(RetVT, LHS, RHS, /*SetFlags=*/true, /*WantResult=*/false,
IsZExt) != 0;
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
}
bool AArch64FastISel::emitICmp_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill,
uint64_t Imm) {
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
return emitAddSub_ri(/*UseAdd=*/false, RetVT, LHSReg, LHSIsKill, Imm,
/*SetFlags=*/true, /*WantResult=*/false) != 0;
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
}
bool AArch64FastISel::emitFCmp(MVT RetVT, const Value *LHS, const Value *RHS) {
if (RetVT != MVT::f32 && RetVT != MVT::f64)
return false;
// Check to see if the 2nd operand is a constant that we can encode directly
// in the compare.
bool UseImm = false;
if (const auto *CFP = dyn_cast<ConstantFP>(RHS))
if (CFP->isZero() && !CFP->isNegative())
UseImm = true;
unsigned LHSReg = getRegForValue(LHS);
if (!LHSReg)
return false;
bool LHSIsKill = hasTrivialKill(LHS);
if (UseImm) {
unsigned Opc = (RetVT == MVT::f64) ? AArch64::FCMPDri : AArch64::FCMPSri;
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc))
.addReg(LHSReg, getKillRegState(LHSIsKill));
return true;
}
unsigned RHSReg = getRegForValue(RHS);
if (!RHSReg)
return false;
bool RHSIsKill = hasTrivialKill(RHS);
unsigned Opc = (RetVT == MVT::f64) ? AArch64::FCMPDrr : AArch64::FCMPSrr;
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc))
.addReg(LHSReg, getKillRegState(LHSIsKill))
.addReg(RHSReg, getKillRegState(RHSIsKill));
return true;
}
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
unsigned AArch64FastISel::emitAdd(MVT RetVT, const Value *LHS, const Value *RHS,
bool SetFlags, bool WantResult, bool IsZExt) {
return emitAddSub(/*UseAdd=*/true, RetVT, LHS, RHS, SetFlags, WantResult,
IsZExt);
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
}
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
unsigned AArch64FastISel::emitSub(MVT RetVT, const Value *LHS, const Value *RHS,
bool SetFlags, bool WantResult, bool IsZExt) {
return emitAddSub(/*UseAdd=*/false, RetVT, LHS, RHS, SetFlags, WantResult,
IsZExt);
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
}
unsigned AArch64FastISel::emitSubs_rr(MVT RetVT, unsigned LHSReg,
bool LHSIsKill, unsigned RHSReg,
bool RHSIsKill, bool WantResult) {
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
return emitAddSub_rr(/*UseAdd=*/false, RetVT, LHSReg, LHSIsKill, RHSReg,
RHSIsKill, /*SetFlags=*/true, WantResult);
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
}
unsigned AArch64FastISel::emitSubs_rs(MVT RetVT, unsigned LHSReg,
bool LHSIsKill, unsigned RHSReg,
bool RHSIsKill,
AArch64_AM::ShiftExtendType ShiftType,
uint64_t ShiftImm, bool WantResult) {
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
return emitAddSub_rs(/*UseAdd=*/false, RetVT, LHSReg, LHSIsKill, RHSReg,
RHSIsKill, ShiftType, ShiftImm, /*SetFlags=*/true,
WantResult);
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
}
[FastISel][AArch64] Add target-specific lowering for logical operations. This change adds support for immediate and shift-left folding into logical operations. This fixes rdar://problem/18223183. llvm-svn: 217118
2014-09-04 09:29:18 +08:00
unsigned AArch64FastISel::emitLogicalOp(unsigned ISDOpc, MVT RetVT,
const Value *LHS, const Value *RHS) {
// Canonicalize immediates to the RHS first.
if (isa<ConstantInt>(LHS) && !isa<ConstantInt>(RHS))
std::swap(LHS, RHS);
// Canonicalize shift immediate to the RHS.
if (isValueAvailable(LHS))
if (const auto *SI = dyn_cast<BinaryOperator>(LHS))
if (isa<ConstantInt>(SI->getOperand(1)))
if (SI->getOpcode() == Instruction::Shl)
std::swap(LHS, RHS);
unsigned LHSReg = getRegForValue(LHS);
if (!LHSReg)
return 0;
bool LHSIsKill = hasTrivialKill(LHS);
unsigned ResultReg = 0;
if (const auto *C = dyn_cast<ConstantInt>(RHS)) {
uint64_t Imm = C->getZExtValue();
ResultReg = emitLogicalOp_ri(ISDOpc, RetVT, LHSReg, LHSIsKill, Imm);
}
if (ResultReg)
return ResultReg;
// Check if the shift can be folded into the instruction.
if (isValueAvailable(RHS))
if (const auto *SI = dyn_cast<BinaryOperator>(RHS))
if (const auto *C = dyn_cast<ConstantInt>(SI->getOperand(1)))
if (SI->getOpcode() == Instruction::Shl) {
uint64_t ShiftVal = C->getZExtValue();
unsigned RHSReg = getRegForValue(SI->getOperand(0));
if (!RHSReg)
return 0;
bool RHSIsKill = hasTrivialKill(SI->getOperand(0));
return emitLogicalOp_rs(ISDOpc, RetVT, LHSReg, LHSIsKill, RHSReg,
RHSIsKill, ShiftVal);
}
unsigned RHSReg = getRegForValue(RHS);
if (!RHSReg)
return 0;
bool RHSIsKill = hasTrivialKill(RHS);
[FastISel][AArch64] Add support for non-native types for logical ops. Extend the logical ops selection to also support non-native types such as i1, i8, and i16. Fixes rdar://problem/18330589. llvm-svn: 217732
2014-09-14 07:46:28 +08:00
MVT VT = std::max(MVT::i32, RetVT.SimpleTy);
ResultReg = fastEmit_rr(VT, VT, ISDOpc, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
if (RetVT >= MVT::i8 && RetVT <= MVT::i16) {
uint64_t Mask = (RetVT == MVT::i8) ? 0xff : 0xffff;
ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
}
return ResultReg;
[FastISel][AArch64] Add target-specific lowering for logical operations. This change adds support for immediate and shift-left folding into logical operations. This fixes rdar://problem/18223183. llvm-svn: 217118
2014-09-04 09:29:18 +08:00
}
unsigned AArch64FastISel::emitLogicalOp_ri(unsigned ISDOpc, MVT RetVT,
unsigned LHSReg, bool LHSIsKill,
uint64_t Imm) {
assert((ISD::AND + 1 == ISD::OR) && (ISD::AND + 2 == ISD::XOR) &&
"ISD nodes are not consecutive!");
static const unsigned OpcTable[3][2] = {
{ AArch64::ANDWri, AArch64::ANDXri },
{ AArch64::ORRWri, AArch64::ORRXri },
{ AArch64::EORWri, AArch64::EORXri }
};
const TargetRegisterClass *RC;
unsigned Opc;
unsigned RegSize;
[FastISel][AArch64] Factor out ANDWri instruction generation into a helper function. NFCI. llvm-svn: 216199
2014-08-22 02:02:25 +08:00
switch (RetVT.SimpleTy) {
default:
return 0;
[FastISel][AArch64] Add support for non-native types for logical ops. Extend the logical ops selection to also support non-native types such as i1, i8, and i16. Fixes rdar://problem/18330589. llvm-svn: 217732
2014-09-14 07:46:28 +08:00
case MVT::i1:
case MVT::i8:
case MVT::i16:
[FastISel][AArch64] Add target-specific lowering for logical operations. This change adds support for immediate and shift-left folding into logical operations. This fixes rdar://problem/18223183. llvm-svn: 217118
2014-09-04 09:29:18 +08:00
case MVT::i32: {
unsigned Idx = ISDOpc - ISD::AND;
Opc = OpcTable[Idx][0];
[FastISel][AArch64] Factor out ANDWri instruction generation into a helper function. NFCI. llvm-svn: 216199
2014-08-22 02:02:25 +08:00
RC = &AArch64::GPR32spRegClass;
RegSize = 32;
break;
[FastISel][AArch64] Add target-specific lowering for logical operations. This change adds support for immediate and shift-left folding into logical operations. This fixes rdar://problem/18223183. llvm-svn: 217118
2014-09-04 09:29:18 +08:00
}
[FastISel][AArch64] Factor out ANDWri instruction generation into a helper function. NFCI. llvm-svn: 216199
2014-08-22 02:02:25 +08:00
case MVT::i64:
[FastISel][AArch64] Add target-specific lowering for logical operations. This change adds support for immediate and shift-left folding into logical operations. This fixes rdar://problem/18223183. llvm-svn: 217118
2014-09-04 09:29:18 +08:00
Opc = OpcTable[ISDOpc - ISD::AND][1];
[FastISel][AArch64] Factor out ANDWri instruction generation into a helper function. NFCI. llvm-svn: 216199
2014-08-22 02:02:25 +08:00
RC = &AArch64::GPR64spRegClass;
RegSize = 64;
break;
}
if (!AArch64_AM::isLogicalImmediate(Imm, RegSize))
return 0;
[FastISel][AArch64] Add support for non-native types for logical ops. Extend the logical ops selection to also support non-native types such as i1, i8, and i16. Fixes rdar://problem/18330589. llvm-svn: 217732
2014-09-14 07:46:28 +08:00
unsigned ResultReg =
fastEmitInst_ri(Opc, RC, LHSReg, LHSIsKill,
AArch64_AM::encodeLogicalImmediate(Imm, RegSize));
if (RetVT >= MVT::i8 && RetVT <= MVT::i16 && ISDOpc != ISD::AND) {
uint64_t Mask = (RetVT == MVT::i8) ? 0xff : 0xffff;
ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
}
return ResultReg;
[FastISel][AArch64] Factor out ANDWri instruction generation into a helper function. NFCI. llvm-svn: 216199
2014-08-22 02:02:25 +08:00
}
[FastISel][AArch64] Add target-specific lowering for logical operations. This change adds support for immediate and shift-left folding into logical operations. This fixes rdar://problem/18223183. llvm-svn: 217118
2014-09-04 09:29:18 +08:00
unsigned AArch64FastISel::emitLogicalOp_rs(unsigned ISDOpc, MVT RetVT,
unsigned LHSReg, bool LHSIsKill,
unsigned RHSReg, bool RHSIsKill,
uint64_t ShiftImm) {
assert((ISD::AND + 1 == ISD::OR) && (ISD::AND + 2 == ISD::XOR) &&
"ISD nodes are not consecutive!");
static const unsigned OpcTable[3][2] = {
{ AArch64::ANDWrs, AArch64::ANDXrs },
{ AArch64::ORRWrs, AArch64::ORRXrs },
{ AArch64::EORWrs, AArch64::EORXrs }
};
const TargetRegisterClass *RC;
unsigned Opc;
switch (RetVT.SimpleTy) {
[FastISel][AArch64] Add support for non-native types for logical ops. Extend the logical ops selection to also support non-native types such as i1, i8, and i16. Fixes rdar://problem/18330589. llvm-svn: 217732
2014-09-14 07:46:28 +08:00
default:
return 0;
case MVT::i1:
case MVT::i8:
case MVT::i16:
case MVT::i32:
Opc = OpcTable[ISDOpc - ISD::AND][0];
RC = &AArch64::GPR32RegClass;
break;
case MVT::i64:
Opc = OpcTable[ISDOpc - ISD::AND][1];
RC = &AArch64::GPR64RegClass;
break;
[FastISel][AArch64] Add target-specific lowering for logical operations. This change adds support for immediate and shift-left folding into logical operations. This fixes rdar://problem/18223183. llvm-svn: 217118
2014-09-04 09:29:18 +08:00
}
[FastISel][AArch64] Add support for non-native types for logical ops. Extend the logical ops selection to also support non-native types such as i1, i8, and i16. Fixes rdar://problem/18330589. llvm-svn: 217732
2014-09-14 07:46:28 +08:00
unsigned ResultReg =
fastEmitInst_rri(Opc, RC, LHSReg, LHSIsKill, RHSReg, RHSIsKill,
AArch64_AM::getShifterImm(AArch64_AM::LSL, ShiftImm));
if (RetVT >= MVT::i8 && RetVT <= MVT::i16) {
uint64_t Mask = (RetVT == MVT::i8) ? 0xff : 0xffff;
ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
}
return ResultReg;
[FastISel][AArch64] Add target-specific lowering for logical operations. This change adds support for immediate and shift-left folding into logical operations. This fixes rdar://problem/18223183. llvm-svn: 217118
2014-09-04 09:29:18 +08:00
}
unsigned AArch64FastISel::emitAnd_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill,
uint64_t Imm) {
return emitLogicalOp_ri(ISD::AND, RetVT, LHSReg, LHSIsKill, Imm);
}
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
bool AArch64FastISel::EmitLoad(MVT VT, unsigned &ResultReg, Address Addr,
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
MachineMemOperand *MMO) {
// Simplify this down to something we can handle.
if (!SimplifyAddress(Addr, VT))
return false;
unsigned ScaleFactor;
switch (VT.SimpleTy) {
default: llvm_unreachable("Unexpected value type.");
case MVT::i1: // fall-through
case MVT::i8: ScaleFactor = 1; break;
case MVT::i16: ScaleFactor = 2; break;
case MVT::i32: // fall-through
case MVT::f32: ScaleFactor = 4; break;
case MVT::i64: // fall-through
case MVT::f64: ScaleFactor = 8; break;
}
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Negative offsets require unscaled, 9-bit, signed immediate offsets.
// Otherwise, we try using scaled, 12-bit, unsigned immediate offsets.
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
bool UseScaled = true;
if ((Addr.getOffset() < 0) || (Addr.getOffset() & (ScaleFactor - 1))) {
UseScaled = false;
ScaleFactor = 1;
}
static const unsigned OpcTable[4][6] = {
{ AArch64::LDURBBi, AArch64::LDURHHi, AArch64::LDURWi, AArch64::LDURXi,
AArch64::LDURSi, AArch64::LDURDi },
{ AArch64::LDRBBui, AArch64::LDRHHui, AArch64::LDRWui, AArch64::LDRXui,
AArch64::LDRSui, AArch64::LDRDui },
{ AArch64::LDRBBroX, AArch64::LDRHHroX, AArch64::LDRWroX, AArch64::LDRXroX,
AArch64::LDRSroX, AArch64::LDRDroX },
{ AArch64::LDRBBroW, AArch64::LDRHHroW, AArch64::LDRWroW, AArch64::LDRXroW,
AArch64::LDRSroW, AArch64::LDRDroW }
};
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
unsigned Opc;
const TargetRegisterClass *RC;
bool VTIsi1 = false;
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
bool UseRegOffset = Addr.isRegBase() && !Addr.getOffset() && Addr.getReg() &&
Addr.getOffsetReg();
unsigned Idx = UseRegOffset ? 2 : UseScaled ? 1 : 0;
if (Addr.getExtendType() == AArch64_AM::UXTW ||
Addr.getExtendType() == AArch64_AM::SXTW)
Idx++;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
switch (VT.SimpleTy) {
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
default: llvm_unreachable("Unexpected value type.");
case MVT::i1: VTIsi1 = true; // Intentional fall-through.
case MVT::i8: Opc = OpcTable[Idx][0]; RC = &AArch64::GPR32RegClass; break;
case MVT::i16: Opc = OpcTable[Idx][1]; RC = &AArch64::GPR32RegClass; break;
case MVT::i32: Opc = OpcTable[Idx][2]; RC = &AArch64::GPR32RegClass; break;
case MVT::i64: Opc = OpcTable[Idx][3]; RC = &AArch64::GPR64RegClass; break;
case MVT::f32: Opc = OpcTable[Idx][4]; RC = &AArch64::FPR32RegClass; break;
case MVT::f64: Opc = OpcTable[Idx][5]; RC = &AArch64::FPR64RegClass; break;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
// Create the base instruction, then add the operands.
ResultReg = createResultReg(RC);
MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
TII.get(Opc), ResultReg);
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
AddLoadStoreOperands(Addr, MIB, MachineMemOperand::MOLoad, ScaleFactor, MMO);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Loading an i1 requires special handling.
if (VTIsi1) {
[FastISel][AArch64] Add target-specific lowering for logical operations. This change adds support for immediate and shift-left folding into logical operations. This fixes rdar://problem/18223183. llvm-svn: 217118
2014-09-04 09:29:18 +08:00
unsigned ANDReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, 1);
[FastISel][AArch64] Factor out ANDWri instruction generation into a helper function. NFCI. llvm-svn: 216199
2014-08-22 02:02:25 +08:00
assert(ANDReg && "Unexpected AND instruction emission failure.");
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
ResultReg = ANDReg;
}
return true;
}
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
bool AArch64FastISel::selectAddSub(const Instruction *I) {
MVT VT;
if (!isTypeSupported(I->getType(), VT))
return false;
unsigned ResultReg;
if (I->getOpcode() == Instruction::Add)
ResultReg = emitAdd(VT, I->getOperand(0), I->getOperand(1));
else if (I->getOpcode() == Instruction::Sub)
ResultReg = emitSub(VT, I->getOperand(0), I->getOperand(1));
else
llvm_unreachable("Unexpected instruction.");
assert(ResultReg && "Couldn't select Add/Sub instruction.");
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
updateValueMap(I, ResultReg);
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
return true;
}
[FastISel][AArch64] Add support for non-native types for logical ops. Extend the logical ops selection to also support non-native types such as i1, i8, and i16. Fixes rdar://problem/18330589. llvm-svn: 217732
2014-09-14 07:46:28 +08:00
bool AArch64FastISel::selectLogicalOp(const Instruction *I, unsigned ISDOpc) {
[FastISel][AArch64] Add target-specific lowering for logical operations. This change adds support for immediate and shift-left folding into logical operations. This fixes rdar://problem/18223183. llvm-svn: 217118
2014-09-04 09:29:18 +08:00
MVT VT;
if (!isTypeSupported(I->getType(), VT))
return false;
unsigned ResultReg =
emitLogicalOp(ISDOpc, VT, I->getOperand(0), I->getOperand(1));
if (!ResultReg)
return false;
updateValueMap(I, ResultReg);
return true;
}
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
bool AArch64FastISel::SelectLoad(const Instruction *I) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
MVT VT;
// Verify we have a legal type before going any further. Currently, we handle
// simple types that will directly fit in a register (i32/f32/i64/f64) or
// those that can be sign or zero-extended to a basic operation (i1/i8/i16).
if (!isLoadStoreTypeLegal(I->getType(), VT) || cast<LoadInst>(I)->isAtomic())
return false;
// See if we can handle this address.
Address Addr;
Revert "[FastISel][AArch64] Don't fold instructions too aggressively into the memory operation." Quentin pointed out that this is not the correct approach and there is a better and easier solution. llvm-svn: 216632
2014-08-28 07:09:40 +08:00
if (!ComputeAddress(I->getOperand(0), Addr, I->getType()))
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return false;
unsigned ResultReg;
[FastISel][AArch64] Attach MachineMemOperands to load and store instructions. llvm-svn: 215231
2014-08-09 01:24:10 +08:00
if (!EmitLoad(VT, ResultReg, Addr, createMachineMemOperandFor(I)))
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return false;
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
updateValueMap(I, ResultReg);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return true;
}
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
bool AArch64FastISel::EmitStore(MVT VT, unsigned SrcReg, Address Addr,
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
MachineMemOperand *MMO) {
// Simplify this down to something we can handle.
if (!SimplifyAddress(Addr, VT))
return false;
unsigned ScaleFactor;
switch (VT.SimpleTy) {
default: llvm_unreachable("Unexpected value type.");
case MVT::i1: // fall-through
case MVT::i8: ScaleFactor = 1; break;
case MVT::i16: ScaleFactor = 2; break;
case MVT::i32: // fall-through
case MVT::f32: ScaleFactor = 4; break;
case MVT::i64: // fall-through
case MVT::f64: ScaleFactor = 8; break;
}
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Negative offsets require unscaled, 9-bit, signed immediate offsets.
// Otherwise, we try using scaled, 12-bit, unsigned immediate offsets.
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
bool UseScaled = true;
if ((Addr.getOffset() < 0) || (Addr.getOffset() & (ScaleFactor - 1))) {
UseScaled = false;
ScaleFactor = 1;
}
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
static const unsigned OpcTable[4][6] = {
{ AArch64::STURBBi, AArch64::STURHHi, AArch64::STURWi, AArch64::STURXi,
AArch64::STURSi, AArch64::STURDi },
{ AArch64::STRBBui, AArch64::STRHHui, AArch64::STRWui, AArch64::STRXui,
AArch64::STRSui, AArch64::STRDui },
{ AArch64::STRBBroX, AArch64::STRHHroX, AArch64::STRWroX, AArch64::STRXroX,
AArch64::STRSroX, AArch64::STRDroX },
{ AArch64::STRBBroW, AArch64::STRHHroW, AArch64::STRWroW, AArch64::STRXroW,
AArch64::STRSroW, AArch64::STRDroW }
};
unsigned Opc;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
bool VTIsi1 = false;
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
bool UseRegOffset = Addr.isRegBase() && !Addr.getOffset() && Addr.getReg() &&
Addr.getOffsetReg();
unsigned Idx = UseRegOffset ? 2 : UseScaled ? 1 : 0;
if (Addr.getExtendType() == AArch64_AM::UXTW ||
Addr.getExtendType() == AArch64_AM::SXTW)
Idx++;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
switch (VT.SimpleTy) {
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
default: llvm_unreachable("Unexpected value type.");
case MVT::i1: VTIsi1 = true;
case MVT::i8: Opc = OpcTable[Idx][0]; break;
case MVT::i16: Opc = OpcTable[Idx][1]; break;
case MVT::i32: Opc = OpcTable[Idx][2]; break;
case MVT::i64: Opc = OpcTable[Idx][3]; break;
case MVT::f32: Opc = OpcTable[Idx][4]; break;
case MVT::f64: Opc = OpcTable[Idx][5]; break;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
// Storing an i1 requires special handling.
[FastISel][AArch64] Use the zero register for stores. Use the zero register directly when possible to avoid an unnecessary register copy and a wasted register at -O0. This also uses integer stores to store a positive floating-point zero. This saves us from materializing the positive zero in a register and then storing it. llvm-svn: 216617
2014-08-28 05:04:52 +08:00
if (VTIsi1 && SrcReg != AArch64::WZR) {
[FastISel][AArch64] Add target-specific lowering for logical operations. This change adds support for immediate and shift-left folding into logical operations. This fixes rdar://problem/18223183. llvm-svn: 217118
2014-09-04 09:29:18 +08:00
unsigned ANDReg = emitAnd_ri(MVT::i32, SrcReg, /*TODO:IsKill=*/false, 1);
[FastISel][AArch64] Factor out ANDWri instruction generation into a helper function. NFCI. llvm-svn: 216199
2014-08-22 02:02:25 +08:00
assert(ANDReg && "Unexpected AND instruction emission failure.");
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
SrcReg = ANDReg;
}
// Create the base instruction, then add the operands.
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
const MCInstrDesc &II = TII.get(Opc);
SrcReg = constrainOperandRegClass(II, SrcReg, II.getNumDefs());
MachineInstrBuilder MIB =
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II).addReg(SrcReg);
Reapply [FastISel][AArch64] Add support for more addressing modes (r215597). Note: This was originally reverted to track down a buildbot error. Reapply without any modifications. Original commit message: FastISel didn't take much advantage of the different addressing modes available to it on AArch64. This commit allows the ComputeAddress method to recognize more addressing modes that allows shifts and sign-/zero-extensions to be folded into the memory operation itself. For Example: lsl x1, x1, #3 --> ldr x0, [x0, x1, lsl #3] ldr x0, [x0, x1] sxtw x1, w1 lsl x1, x1, #3 --> ldr x0, [x0, x1, sxtw #3] ldr x0, [x0, x1] llvm-svn: 216013
2014-08-20 03:44:17 +08:00
AddLoadStoreOperands(Addr, MIB, MachineMemOperand::MOStore, ScaleFactor, MMO);
[FastISel][AArch64] Attach MachineMemOperands to load and store instructions. llvm-svn: 215231
2014-08-09 01:24:10 +08:00
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return true;
}
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
bool AArch64FastISel::SelectStore(const Instruction *I) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
MVT VT;
[FastISel][AArch64] Use the zero register for stores. Use the zero register directly when possible to avoid an unnecessary register copy and a wasted register at -O0. This also uses integer stores to store a positive floating-point zero. This saves us from materializing the positive zero in a register and then storing it. llvm-svn: 216617
2014-08-28 05:04:52 +08:00
const Value *Op0 = I->getOperand(0);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Verify we have a legal type before going any further. Currently, we handle
// simple types that will directly fit in a register (i32/f32/i64/f64) or
// those that can be sign or zero-extended to a basic operation (i1/i8/i16).
if (!isLoadStoreTypeLegal(Op0->getType(), VT) ||
cast<StoreInst>(I)->isAtomic())
return false;
[FastISel][AArch64] Use the zero register for stores. Use the zero register directly when possible to avoid an unnecessary register copy and a wasted register at -O0. This also uses integer stores to store a positive floating-point zero. This saves us from materializing the positive zero in a register and then storing it. llvm-svn: 216617
2014-08-28 05:04:52 +08:00
// Get the value to be stored into a register. Use the zero register directly
[FastISel][AArch64] Fix a comment in my previous commit (r216617). llvm-svn: 216622
2014-08-28 05:40:50 +08:00
// when possible to avoid an unnecessary copy and a wasted register.
[FastISel][AArch64] Use the zero register for stores. Use the zero register directly when possible to avoid an unnecessary register copy and a wasted register at -O0. This also uses integer stores to store a positive floating-point zero. This saves us from materializing the positive zero in a register and then storing it. llvm-svn: 216617
2014-08-28 05:04:52 +08:00
unsigned SrcReg = 0;
if (const auto *CI = dyn_cast<ConstantInt>(Op0)) {
if (CI->isZero())
SrcReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
} else if (const auto *CF = dyn_cast<ConstantFP>(Op0)) {
if (CF->isZero() && !CF->isNegative()) {
VT = MVT::getIntegerVT(VT.getSizeInBits());
SrcReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
}
}
if (!SrcReg)
SrcReg = getRegForValue(Op0);
if (!SrcReg)
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return false;
// See if we can handle this address.
Address Addr;
Revert "[FastISel][AArch64] Don't fold instructions too aggressively into the memory operation." Quentin pointed out that this is not the correct approach and there is a better and easier solution. llvm-svn: 216632
2014-08-28 07:09:40 +08:00
if (!ComputeAddress(I->getOperand(1), Addr, I->getOperand(0)->getType()))
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return false;
[FastISel][AArch64] Attach MachineMemOperands to load and store instructions. llvm-svn: 215231
2014-08-09 01:24:10 +08:00
if (!EmitStore(VT, SrcReg, Addr, createMachineMemOperandFor(I)))
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return false;
return true;
}
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
static AArch64CC::CondCode getCompareCC(CmpInst::Predicate Pred) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
switch (Pred) {
case CmpInst::FCMP_ONE:
case CmpInst::FCMP_UEQ:
default:
// AL is our "false" for now. The other two need more compares.
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
return AArch64CC::AL;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
case CmpInst::ICMP_EQ:
case CmpInst::FCMP_OEQ:
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
return AArch64CC::EQ;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
case CmpInst::ICMP_SGT:
case CmpInst::FCMP_OGT:
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
return AArch64CC::GT;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
case CmpInst::ICMP_SGE:
case CmpInst::FCMP_OGE:
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
return AArch64CC::GE;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
case CmpInst::ICMP_UGT:
case CmpInst::FCMP_UGT:
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
return AArch64CC::HI;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
case CmpInst::FCMP_OLT:
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
return AArch64CC::MI;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
case CmpInst::ICMP_ULE:
case CmpInst::FCMP_OLE:
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
return AArch64CC::LS;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
case CmpInst::FCMP_ORD:
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
return AArch64CC::VC;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
case CmpInst::FCMP_UNO:
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
return AArch64CC::VS;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
case CmpInst::FCMP_UGE:
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
return AArch64CC::PL;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
case CmpInst::ICMP_SLT:
case CmpInst::FCMP_ULT:
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
return AArch64CC::LT;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
case CmpInst::ICMP_SLE:
case CmpInst::FCMP_ULE:
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
return AArch64CC::LE;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
case CmpInst::FCMP_UNE:
case CmpInst::ICMP_NE:
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
return AArch64CC::NE;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
case CmpInst::ICMP_UGE:
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
return AArch64CC::HS;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
case CmpInst::ICMP_ULT:
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
return AArch64CC::LO;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
}
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
bool AArch64FastISel::SelectBranch(const Instruction *I) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
const BranchInst *BI = cast<BranchInst>(I);
[FastISel][AArch64] Move unconditional branch handling into 'SelectBranch'. NFC. llvm-svn: 217054
2014-09-04 01:58:10 +08:00
if (BI->isUnconditional()) {
MachineBasicBlock *MSucc = FuncInfo.MBBMap[BI->getSuccessor(0)];
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
fastEmitBranch(MSucc, BI->getDebugLoc());
[FastISel][AArch64] Move unconditional branch handling into 'SelectBranch'. NFC. llvm-svn: 217054
2014-09-04 01:58:10 +08:00
return true;
}
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
[FastISel][AArch64] Add branch folding support for the XALU intrinsics. This improves the code generation for the XALU intrinsics when the condition is feeding a branch instruction. This is related to <rdar://problem/17831117>. llvm-svn: 214349
2014-07-31 06:04:34 +08:00
AArch64CC::CondCode CC = AArch64CC::NE;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
if (CI->hasOneUse() && (CI->getParent() == I->getParent())) {
// We may not handle every CC for now.
[FastISel][AArch64] Add branch folding support for the XALU intrinsics. This improves the code generation for the XALU intrinsics when the condition is feeding a branch instruction. This is related to <rdar://problem/17831117>. llvm-svn: 214349
2014-07-31 06:04:34 +08:00
CC = getCompareCC(CI->getPredicate());
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
if (CC == AArch64CC::AL)
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return false;
// Emit the cmp.
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
if (!emitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return false;
// Emit the branch.
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
.addImm(CC)
.addMBB(TBB);
[FastISel][AArch64] Add branch weights. Add branch weights to branch instructions, so that the following passes can optimize based on it (i.e. basic block ordering). Fixes <rdar://problem/17887137>. llvm-svn: 214537
2014-08-02 02:39:24 +08:00
// Obtain the branch weight and add the TrueBB to the successor list.
uint32_t BranchWeight = 0;
if (FuncInfo.BPI)
BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
TBB->getBasicBlock());
FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
fastEmitBranch(FBB, DbgLoc);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return true;
}
} else if (TruncInst *TI = dyn_cast<TruncInst>(BI->getCondition())) {
MVT SrcVT;
if (TI->hasOneUse() && TI->getParent() == I->getParent() &&
[FastISel][AArch64] Use a new helper function to determine if a value type is supported. NFCI. FastISel for AArch64 supports more value types than are actually legal. Use a dedicated helper function to reflect this. It is very similar to the isLoadStoreTypeLegal function, with the exception that vector types are not supported yet. llvm-svn: 216984
2014-09-03 06:33:53 +08:00
(isTypeSupported(TI->getOperand(0)->getType(), SrcVT))) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
unsigned CondReg = getRegForValue(TI->getOperand(0));
[FastISel][AArch64] Factor out ANDWri instruction generation into a helper function. NFCI. llvm-svn: 216199
2014-08-22 02:02:25 +08:00
if (!CondReg)
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return false;
[FastISel][AArch64] Factor out ANDWri instruction generation into a helper function. NFCI. llvm-svn: 216199
2014-08-22 02:02:25 +08:00
bool CondIsKill = hasTrivialKill(TI->getOperand(0));
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Issue an extract_subreg to get the lower 32-bits.
[FastISel][AArch64] Factor out ANDWri instruction generation into a helper function. NFCI. llvm-svn: 216199
2014-08-22 02:02:25 +08:00
if (SrcVT == MVT::i64) {
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
CondReg = fastEmitInst_extractsubreg(MVT::i32, CondReg, CondIsKill,
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
AArch64::sub_32);
[FastISel][AArch64] Factor out ANDWri instruction generation into a helper function. NFCI. llvm-svn: 216199
2014-08-22 02:02:25 +08:00
CondIsKill = true;
}
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
[FastISel][AArch64] Add target-specific lowering for logical operations. This change adds support for immediate and shift-left folding into logical operations. This fixes rdar://problem/18223183. llvm-svn: 217118
2014-09-04 09:29:18 +08:00
unsigned ANDReg = emitAnd_ri(MVT::i32, CondReg, CondIsKill, 1);
[FastISel][AArch64] Factor out ANDWri instruction generation into a helper function. NFCI. llvm-svn: 216199
2014-08-22 02:02:25 +08:00
assert(ANDReg && "Unexpected AND instruction emission failure.");
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
emitICmp_ri(MVT::i32, ANDReg, /*IsKill=*/true, 0);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
std::swap(TBB, FBB);
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
CC = AArch64CC::EQ;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
.addImm(CC)
.addMBB(TBB);
[FastISel][AArch64] Add branch weights. Add branch weights to branch instructions, so that the following passes can optimize based on it (i.e. basic block ordering). Fixes <rdar://problem/17887137>. llvm-svn: 214537
2014-08-02 02:39:24 +08:00
// Obtain the branch weight and add the TrueBB to the successor list.
uint32_t BranchWeight = 0;
if (FuncInfo.BPI)
BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
TBB->getBasicBlock());
FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
fastEmitBranch(FBB, DbgLoc);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return true;
}
} else if (const ConstantInt *CI =
dyn_cast<ConstantInt>(BI->getCondition())) {
uint64_t Imm = CI->getZExtValue();
MachineBasicBlock *Target = (Imm == 0) ? FBB : TBB;
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::B))
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
.addMBB(Target);
[FastISel][AArch64] Add branch weights. Add branch weights to branch instructions, so that the following passes can optimize based on it (i.e. basic block ordering). Fixes <rdar://problem/17887137>. llvm-svn: 214537
2014-08-02 02:39:24 +08:00
// Obtain the branch weight and add the target to the successor list.
uint32_t BranchWeight = 0;
if (FuncInfo.BPI)
BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
Target->getBasicBlock());
FuncInfo.MBB->addSuccessor(Target, BranchWeight);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return true;
[FastISel][AArch64] Add branch folding support for the XALU intrinsics. This improves the code generation for the XALU intrinsics when the condition is feeding a branch instruction. This is related to <rdar://problem/17831117>. llvm-svn: 214349
2014-07-31 06:04:34 +08:00
} else if (foldXALUIntrinsic(CC, I, BI->getCondition())) {
// Fake request the condition, otherwise the intrinsic might be completely
// optimized away.
unsigned CondReg = getRegForValue(BI->getCondition());
if (!CondReg)
return false;
// Emit the branch.
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
.addImm(CC)
.addMBB(TBB);
[FastISel][AArch64] Add branch weights. Add branch weights to branch instructions, so that the following passes can optimize based on it (i.e. basic block ordering). Fixes <rdar://problem/17887137>. llvm-svn: 214537
2014-08-02 02:39:24 +08:00
// Obtain the branch weight and add the TrueBB to the successor list.
uint32_t BranchWeight = 0;
if (FuncInfo.BPI)
BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
TBB->getBasicBlock());
FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
[FastISel][AArch64] Add branch folding support for the XALU intrinsics. This improves the code generation for the XALU intrinsics when the condition is feeding a branch instruction. This is related to <rdar://problem/17831117>. llvm-svn: 214349
2014-07-31 06:04:34 +08:00
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
fastEmitBranch(FBB, DbgLoc);
[FastISel][AArch64] Add branch folding support for the XALU intrinsics. This improves the code generation for the XALU intrinsics when the condition is feeding a branch instruction. This is related to <rdar://problem/17831117>. llvm-svn: 214349
2014-07-31 06:04:34 +08:00
return true;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
unsigned CondReg = getRegForValue(BI->getCondition());
if (CondReg == 0)
return false;
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
bool CondRegIsKill = hasTrivialKill(BI->getCondition());
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// We've been divorced from our compare! Our block was split, and
// now our compare lives in a predecessor block. We musn't
// re-compare here, as the children of the compare aren't guaranteed
// live across the block boundary (we *could* check for this).
// Regardless, the compare has been done in the predecessor block,
// and it left a value for us in a virtual register. Ergo, we test
// the one-bit value left in the virtual register.
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
emitICmp_ri(MVT::i32, CondReg, CondRegIsKill, 0);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
std::swap(TBB, FBB);
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
CC = AArch64CC::EQ;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
.addImm(CC)
.addMBB(TBB);
[FastISel][AArch64] Add branch weights. Add branch weights to branch instructions, so that the following passes can optimize based on it (i.e. basic block ordering). Fixes <rdar://problem/17887137>. llvm-svn: 214537
2014-08-02 02:39:24 +08:00
// Obtain the branch weight and add the TrueBB to the successor list.
uint32_t BranchWeight = 0;
if (FuncInfo.BPI)
BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
TBB->getBasicBlock());
FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
fastEmitBranch(FBB, DbgLoc);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return true;
}
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
bool AArch64FastISel::SelectIndirectBr(const Instruction *I) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
const IndirectBrInst *BI = cast<IndirectBrInst>(I);
unsigned AddrReg = getRegForValue(BI->getOperand(0));
if (AddrReg == 0)
return false;
// Emit the indirect branch.
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
const MCInstrDesc &II = TII.get(AArch64::BR);
AddrReg = constrainOperandRegClass(II, AddrReg, II.getNumDefs());
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II).addReg(AddrReg);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Make sure the CFG is up-to-date.
for (unsigned i = 0, e = BI->getNumSuccessors(); i != e; ++i)
FuncInfo.MBB->addSuccessor(FuncInfo.MBBMap[BI->getSuccessor(i)]);
return true;
}
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
bool AArch64FastISel::SelectCmp(const Instruction *I) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
const CmpInst *CI = cast<CmpInst>(I);
[FastISel][AArch64] Improve floating-point compare support. Add support for the last two missing fcmp condition codes: UEQ and ONE. This fixes rdar://problem/18341575. llvm-svn: 217823
2014-09-16 04:47:16 +08:00
// Try to optimize or fold the cmp.
CmpInst::Predicate Predicate = optimizeCmpPredicate(CI);
unsigned ResultReg = 0;
switch (Predicate) {
default:
break;
case CmpInst::FCMP_FALSE:
ResultReg = createResultReg(&AArch64::GPR32RegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
TII.get(TargetOpcode::COPY), ResultReg)
.addReg(AArch64::WZR, getKillRegState(true));
break;
case CmpInst::FCMP_TRUE:
ResultReg = fastEmit_i(MVT::i32, MVT::i32, ISD::Constant, 1);
break;
}
if (ResultReg) {
updateValueMap(I, ResultReg);
return true;
}
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Emit the cmp.
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
if (!emitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return false;
[FastISel][AArch64] Improve floating-point compare support. Add support for the last two missing fcmp condition codes: UEQ and ONE. This fixes rdar://problem/18341575. llvm-svn: 217823
2014-09-16 04:47:16 +08:00
ResultReg = createResultReg(&AArch64::GPR32RegClass);
// FCMP_UEQ and FCMP_ONE cannot be checked with a single instruction. These
// condition codes are inverted, because they are used by CSINC.
static unsigned CondCodeTable[2][2] = {
{ AArch64CC::NE, AArch64CC::VC },
{ AArch64CC::PL, AArch64CC::LE }
};
unsigned *CondCodes = nullptr;
switch (Predicate) {
default:
break;
case CmpInst::FCMP_UEQ:
CondCodes = &CondCodeTable[0][0];
break;
case CmpInst::FCMP_ONE:
CondCodes = &CondCodeTable[1][0];
break;
}
if (CondCodes) {
unsigned TmpReg1 = createResultReg(&AArch64::GPR32RegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::CSINCWr),
TmpReg1)
.addReg(AArch64::WZR, getKillRegState(true))
.addReg(AArch64::WZR, getKillRegState(true))
.addImm(CondCodes[0]);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::CSINCWr),
ResultReg)
.addReg(TmpReg1, getKillRegState(true))
.addReg(AArch64::WZR, getKillRegState(true))
.addImm(CondCodes[1]);
updateValueMap(I, ResultReg);
return true;
}
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Now set a register based on the comparison.
[FastISel][AArch64] Improve floating-point compare support. Add support for the last two missing fcmp condition codes: UEQ and ONE. This fixes rdar://problem/18341575. llvm-svn: 217823
2014-09-16 04:47:16 +08:00
AArch64CC::CondCode CC = getCompareCC(Predicate);
assert((CC != AArch64CC::AL) && "Unexpected condition code.");
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
AArch64CC::CondCode invertedCC = getInvertedCondCode(CC);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::CSINCWr),
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
ResultReg)
[FastISel][AArch64] Improve floating-point compare support. Add support for the last two missing fcmp condition codes: UEQ and ONE. This fixes rdar://problem/18341575. llvm-svn: 217823
2014-09-16 04:47:16 +08:00
.addReg(AArch64::WZR, getKillRegState(true))
.addReg(AArch64::WZR, getKillRegState(true))
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
.addImm(invertedCC);
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
updateValueMap(I, ResultReg);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return true;
}
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
bool AArch64FastISel::SelectSelect(const Instruction *I) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
const SelectInst *SI = cast<SelectInst>(I);
EVT DestEVT = TLI.getValueType(SI->getType(), true);
if (!DestEVT.isSimple())
return false;
MVT DestVT = DestEVT.getSimpleVT();
if (DestVT != MVT::i32 && DestVT != MVT::i64 && DestVT != MVT::f32 &&
DestVT != MVT::f64)
return false;
[FastISel][AArch64] Add select folding support for the XALU intrinsics. This improves the code generation for the XALU intrinsics when the condition is feeding a select instruction. This also updates and enables the XALU unit tests for FastISel. This fixes <rdar://problem/17831117>. llvm-svn: 214350
2014-07-31 06:04:37 +08:00
unsigned SelectOpc;
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
const TargetRegisterClass *RC = nullptr;
[FastISel][AArch64] Add select folding support for the XALU intrinsics. This improves the code generation for the XALU intrinsics when the condition is feeding a select instruction. This also updates and enables the XALU unit tests for FastISel. This fixes <rdar://problem/17831117>. llvm-svn: 214350
2014-07-31 06:04:37 +08:00
switch (DestVT.SimpleTy) {
default: return false;
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
case MVT::i32:
SelectOpc = AArch64::CSELWr; RC = &AArch64::GPR32RegClass; break;
case MVT::i64:
SelectOpc = AArch64::CSELXr; RC = &AArch64::GPR64RegClass; break;
case MVT::f32:
SelectOpc = AArch64::FCSELSrrr; RC = &AArch64::FPR32RegClass; break;
case MVT::f64:
SelectOpc = AArch64::FCSELDrrr; RC = &AArch64::FPR64RegClass; break;
[FastISel][AArch64] Add select folding support for the XALU intrinsics. This improves the code generation for the XALU intrinsics when the condition is feeding a select instruction. This also updates and enables the XALU unit tests for FastISel. This fixes <rdar://problem/17831117>. llvm-svn: 214350
2014-07-31 06:04:37 +08:00
}
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
[FastISel][AArch64] Add select folding support for the XALU intrinsics. This improves the code generation for the XALU intrinsics when the condition is feeding a select instruction. This also updates and enables the XALU unit tests for FastISel. This fixes <rdar://problem/17831117>. llvm-svn: 214350
2014-07-31 06:04:37 +08:00
const Value *Cond = SI->getCondition();
bool NeedTest = true;
AArch64CC::CondCode CC = AArch64CC::NE;
if (foldXALUIntrinsic(CC, I, Cond))
NeedTest = false;
[ARM64] Correctly select ANDWri in FastISel. http://reviews.llvm.org/D3598 llvm-svn: 207917
2014-05-04 01:27:06 +08:00
[FastISel][AArch64] Add select folding support for the XALU intrinsics. This improves the code generation for the XALU intrinsics when the condition is feeding a select instruction. This also updates and enables the XALU unit tests for FastISel. This fixes <rdar://problem/17831117>. llvm-svn: 214350
2014-07-31 06:04:37 +08:00
unsigned CondReg = getRegForValue(Cond);
if (!CondReg)
return false;
bool CondIsKill = hasTrivialKill(Cond);
if (NeedTest) {
[FastISel][AArch64] Add target-specific lowering for logical operations. This change adds support for immediate and shift-left folding into logical operations. This fixes rdar://problem/18223183. llvm-svn: 217118
2014-09-04 09:29:18 +08:00
unsigned ANDReg = emitAnd_ri(MVT::i32, CondReg, CondIsKill, 1);
[FastISel][AArch64] Factor out ANDWri instruction generation into a helper function. NFCI. llvm-svn: 216199
2014-08-22 02:02:25 +08:00
assert(ANDReg && "Unexpected AND instruction emission failure.");
emitICmp_ri(MVT::i32, ANDReg, /*IsKill=*/true, 0);
[FastISel][AArch64] Add select folding support for the XALU intrinsics. This improves the code generation for the XALU intrinsics when the condition is feeding a select instruction. This also updates and enables the XALU unit tests for FastISel. This fixes <rdar://problem/17831117>. llvm-svn: 214350
2014-07-31 06:04:37 +08:00
}
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
[FastISel][AArch64] Add select folding support for the XALU intrinsics. This improves the code generation for the XALU intrinsics when the condition is feeding a select instruction. This also updates and enables the XALU unit tests for FastISel. This fixes <rdar://problem/17831117>. llvm-svn: 214350
2014-07-31 06:04:37 +08:00
unsigned TrueReg = getRegForValue(SI->getTrueValue());
bool TrueIsKill = hasTrivialKill(SI->getTrueValue());
unsigned FalseReg = getRegForValue(SI->getFalseValue());
bool FalseIsKill = hasTrivialKill(SI->getFalseValue());
if (!TrueReg || !FalseReg)
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return false;
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
unsigned ResultReg = fastEmitInst_rri(SelectOpc, RC, TrueReg, TrueIsKill,
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
FalseReg, FalseIsKill, CC);
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
updateValueMap(I, ResultReg);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return true;
}
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
bool AArch64FastISel::SelectFPExt(const Instruction *I) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
Value *V = I->getOperand(0);
if (!I->getType()->isDoubleTy() || !V->getType()->isFloatTy())
return false;
unsigned Op = getRegForValue(V);
if (Op == 0)
return false;
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
unsigned ResultReg = createResultReg(&AArch64::FPR64RegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::FCVTDSr),
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
ResultReg).addReg(Op);
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
updateValueMap(I, ResultReg);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return true;
}
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
bool AArch64FastISel::SelectFPTrunc(const Instruction *I) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
Value *V = I->getOperand(0);
if (!I->getType()->isFloatTy() || !V->getType()->isDoubleTy())
return false;
unsigned Op = getRegForValue(V);
if (Op == 0)
return false;
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
unsigned ResultReg = createResultReg(&AArch64::FPR32RegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::FCVTSDr),
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
ResultReg).addReg(Op);
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
updateValueMap(I, ResultReg);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return true;
}
// FPToUI and FPToSI
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
bool AArch64FastISel::SelectFPToInt(const Instruction *I, bool Signed) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
MVT DestVT;
if (!isTypeLegal(I->getType(), DestVT) || DestVT.isVector())
return false;
unsigned SrcReg = getRegForValue(I->getOperand(0));
if (SrcReg == 0)
return false;
EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType(), true);
[ARM64][fast-isel] Fast-isel doesn't know how to handle f128. llvm-svn: 207659
2014-04-30 23:29:57 +08:00
if (SrcVT == MVT::f128)
return false;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
unsigned Opc;
if (SrcVT == MVT::f64) {
if (Signed)
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
Opc = (DestVT == MVT::i32) ? AArch64::FCVTZSUWDr : AArch64::FCVTZSUXDr;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
else
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
Opc = (DestVT == MVT::i32) ? AArch64::FCVTZUUWDr : AArch64::FCVTZUUXDr;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
} else {
if (Signed)
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
Opc = (DestVT == MVT::i32) ? AArch64::FCVTZSUWSr : AArch64::FCVTZSUXSr;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
else
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
Opc = (DestVT == MVT::i32) ? AArch64::FCVTZUUWSr : AArch64::FCVTZUUXSr;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
ARM64: add constraints to various FastISel operations llvm-svn: 206284
2014-04-15 21:59:53 +08:00
unsigned ResultReg = createResultReg(
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
DestVT == MVT::i32 ? &AArch64::GPR32RegClass : &AArch64::GPR64RegClass);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
.addReg(SrcReg);
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
updateValueMap(I, ResultReg);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return true;
}
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
bool AArch64FastISel::SelectIntToFP(const Instruction *I, bool Signed) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
MVT DestVT;
if (!isTypeLegal(I->getType(), DestVT) || DestVT.isVector())
return false;
[ARM64][fast-isel] Fast-isel doesn't know how to handle f128. llvm-svn: 207659
2014-04-30 23:29:57 +08:00
assert ((DestVT == MVT::f32 || DestVT == MVT::f64) &&
"Unexpected value type.");
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
unsigned SrcReg = getRegForValue(I->getOperand(0));
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
if (!SrcReg)
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return false;
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
bool SrcIsKill = hasTrivialKill(I->getOperand(0));
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType(), true);
// Handle sign-extension.
if (SrcVT == MVT::i16 || SrcVT == MVT::i8 || SrcVT == MVT::i1) {
SrcReg =
EmitIntExt(SrcVT.getSimpleVT(), SrcReg, MVT::i32, /*isZExt*/ !Signed);
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
if (!SrcReg)
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return false;
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
SrcIsKill = true;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
unsigned Opc;
if (SrcVT == MVT::i64) {
if (Signed)
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
Opc = (DestVT == MVT::f32) ? AArch64::SCVTFUXSri : AArch64::SCVTFUXDri;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
else
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
Opc = (DestVT == MVT::f32) ? AArch64::UCVTFUXSri : AArch64::UCVTFUXDri;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
} else {
if (Signed)
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
Opc = (DestVT == MVT::f32) ? AArch64::SCVTFUWSri : AArch64::SCVTFUWDri;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
else
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
Opc = (DestVT == MVT::f32) ? AArch64::UCVTFUWSri : AArch64::UCVTFUWDri;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
unsigned ResultReg = fastEmitInst_r(Opc, TLI.getRegClassFor(DestVT), SrcReg,
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
SrcIsKill);
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
updateValueMap(I, ResultReg);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return true;
}
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
bool AArch64FastISel::fastLowerArguments() {
[FastISel][AArch64] Implement the FastLowerArguments hook. This implements basic argument lowering for AArch64 in FastISel. It only handles a small subset of the C calling convention. It supports simple arguments that can be passed in GPR and FPR registers. This should cover most of the trivial cases without falling back to SelectionDAG. This fixes <rdar://problem/17890986>. llvm-svn: 214846
2014-08-05 13:43:48 +08:00
if (!FuncInfo.CanLowerReturn)
return false;
const Function *F = FuncInfo.Fn;
if (F->isVarArg())
return false;
CallingConv::ID CC = F->getCallingConv();
if (CC != CallingConv::C)
return false;
// Only handle simple cases like i1/i8/i16/i32/i64/f32/f64 of up to 8 GPR and
// FPR each.
unsigned GPRCnt = 0;
unsigned FPRCnt = 0;
unsigned Idx = 0;
for (auto const &Arg : F->args()) {
// The first argument is at index 1.
++Idx;
if (F->getAttributes().hasAttribute(Idx, Attribute::ByVal) ||
F->getAttributes().hasAttribute(Idx, Attribute::InReg) ||
F->getAttributes().hasAttribute(Idx, Attribute::StructRet) ||
F->getAttributes().hasAttribute(Idx, Attribute::Nest))
return false;
Type *ArgTy = Arg.getType();
if (ArgTy->isStructTy() || ArgTy->isArrayTy() || ArgTy->isVectorTy())
return false;
EVT ArgVT = TLI.getValueType(ArgTy);
if (!ArgVT.isSimple()) return false;
switch (ArgVT.getSimpleVT().SimpleTy) {
default: return false;
case MVT::i1:
case MVT::i8:
case MVT::i16:
case MVT::i32:
case MVT::i64:
++GPRCnt;
break;
case MVT::f16:
case MVT::f32:
case MVT::f64:
++FPRCnt;
break;
}
if (GPRCnt > 8 || FPRCnt > 8)
return false;
}
static const MCPhysReg Registers[5][8] = {
{ AArch64::W0, AArch64::W1, AArch64::W2, AArch64::W3, AArch64::W4,
AArch64::W5, AArch64::W6, AArch64::W7 },
{ AArch64::X0, AArch64::X1, AArch64::X2, AArch64::X3, AArch64::X4,
AArch64::X5, AArch64::X6, AArch64::X7 },
{ AArch64::H0, AArch64::H1, AArch64::H2, AArch64::H3, AArch64::H4,
AArch64::H5, AArch64::H6, AArch64::H7 },
{ AArch64::S0, AArch64::S1, AArch64::S2, AArch64::S3, AArch64::S4,
AArch64::S5, AArch64::S6, AArch64::S7 },
{ AArch64::D0, AArch64::D1, AArch64::D2, AArch64::D3, AArch64::D4,
AArch64::D5, AArch64::D6, AArch64::D7 }
};
unsigned GPRIdx = 0;
unsigned FPRIdx = 0;
for (auto const &Arg : F->args()) {
MVT VT = TLI.getSimpleValueType(Arg.getType());
unsigned SrcReg;
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
const TargetRegisterClass *RC = nullptr;
[FastISel][AArch64] Implement the FastLowerArguments hook. This implements basic argument lowering for AArch64 in FastISel. It only handles a small subset of the C calling convention. It supports simple arguments that can be passed in GPR and FPR registers. This should cover most of the trivial cases without falling back to SelectionDAG. This fixes <rdar://problem/17890986>. llvm-svn: 214846
2014-08-05 13:43:48 +08:00
switch (VT.SimpleTy) {
default: llvm_unreachable("Unexpected value type.");
case MVT::i1:
case MVT::i8:
case MVT::i16: VT = MVT::i32; // fall-through
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
case MVT::i32:
SrcReg = Registers[0][GPRIdx++]; RC = &AArch64::GPR32RegClass; break;
case MVT::i64:
SrcReg = Registers[1][GPRIdx++]; RC = &AArch64::GPR64RegClass; break;
case MVT::f16:
SrcReg = Registers[2][FPRIdx++]; RC = &AArch64::FPR16RegClass; break;
case MVT::f32:
SrcReg = Registers[3][FPRIdx++]; RC = &AArch64::FPR32RegClass; break;
case MVT::f64:
SrcReg = Registers[4][FPRIdx++]; RC = &AArch64::FPR64RegClass; break;
[FastISel][AArch64] Implement the FastLowerArguments hook. This implements basic argument lowering for AArch64 in FastISel. It only handles a small subset of the C calling convention. It supports simple arguments that can be passed in GPR and FPR registers. This should cover most of the trivial cases without falling back to SelectionDAG. This fixes <rdar://problem/17890986>. llvm-svn: 214846
2014-08-05 13:43:48 +08:00
}
// Skip unused arguments.
if (Arg.use_empty()) {
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
updateValueMap(&Arg, 0);
[FastISel][AArch64] Implement the FastLowerArguments hook. This implements basic argument lowering for AArch64 in FastISel. It only handles a small subset of the C calling convention. It supports simple arguments that can be passed in GPR and FPR registers. This should cover most of the trivial cases without falling back to SelectionDAG. This fixes <rdar://problem/17890986>. llvm-svn: 214846
2014-08-05 13:43:48 +08:00
continue;
}
unsigned DstReg = FuncInfo.MF->addLiveIn(SrcReg, RC);
// FIXME: Unfortunately it's necessary to emit a copy from the livein copy.
// Without this, EmitLiveInCopies may eliminate the livein if its only
// use is a bitcast (which isn't turned into an instruction).
unsigned ResultReg = createResultReg(RC);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
TII.get(TargetOpcode::COPY), ResultReg)
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
.addReg(DstReg, getKillRegState(true));
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
updateValueMap(&Arg, ResultReg);
[FastISel][AArch64] Implement the FastLowerArguments hook. This implements basic argument lowering for AArch64 in FastISel. It only handles a small subset of the C calling convention. It supports simple arguments that can be passed in GPR and FPR registers. This should cover most of the trivial cases without falling back to SelectionDAG. This fixes <rdar://problem/17890986>. llvm-svn: 214846
2014-08-05 13:43:48 +08:00
}
return true;
}
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
bool AArch64FastISel::ProcessCallArgs(CallLoweringInfo &CLI,
SmallVectorImpl<MVT> &OutVTs,
unsigned &NumBytes) {
CallingConv::ID CC = CLI.CallConv;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
SmallVector<CCValAssign, 16> ArgLocs;
Remove the target machine from CCState. Previously it was only used to get the subtarget and that's accessible from the MachineFunction now. This helps clear the way for smaller changes where we getting a subtarget will require passing in a MachineFunction/Function as well. llvm-svn: 214988
2014-08-07 02:45:26 +08:00
CCState CCInfo(CC, false, *FuncInfo.MF, ArgLocs, *Context);
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
CCInfo.AnalyzeCallOperands(OutVTs, CLI.OutFlags, CCAssignFnForCall(CC));
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Get a count of how many bytes are to be pushed on the stack.
NumBytes = CCInfo.getNextStackOffset();
// Issue CALLSEQ_START
unsigned AdjStackDown = TII.getCallFrameSetupOpcode();
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackDown))
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
.addImm(NumBytes);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Process the args.
for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
CCValAssign &VA = ArgLocs[i];
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
const Value *ArgVal = CLI.OutVals[VA.getValNo()];
MVT ArgVT = OutVTs[VA.getValNo()];
unsigned ArgReg = getRegForValue(ArgVal);
if (!ArgReg)
return false;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Handle arg promotion: SExt, ZExt, AExt.
switch (VA.getLocInfo()) {
case CCValAssign::Full:
break;
case CCValAssign::SExt: {
MVT DestVT = VA.getLocVT();
MVT SrcVT = ArgVT;
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
ArgReg = EmitIntExt(SrcVT, ArgReg, DestVT, /*isZExt=*/false);
if (!ArgReg)
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return false;
break;
}
case CCValAssign::AExt:
// Intentional fall-through.
case CCValAssign::ZExt: {
MVT DestVT = VA.getLocVT();
MVT SrcVT = ArgVT;
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
ArgReg = EmitIntExt(SrcVT, ArgReg, DestVT, /*isZExt=*/true);
if (!ArgReg)
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return false;
break;
}
default:
llvm_unreachable("Unknown arg promotion!");
}
// Now copy/store arg to correct locations.
if (VA.isRegLoc() && !VA.needsCustom()) {
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
TII.get(TargetOpcode::COPY), VA.getLocReg()).addReg(ArgReg);
CLI.OutRegs.push_back(VA.getLocReg());
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
} else if (VA.needsCustom()) {
// FIXME: Handle custom args.
return false;
} else {
assert(VA.isMemLoc() && "Assuming store on stack.");
[FastISel][AArch64 and X86] Don't emit stores for UNDEF arguments during function call lowering. UNDEF arguments are not ment to be touched - especially for the webkit_js calling convention. This fix reproduces the already existing behavior of SelectionDAG in FastISel. llvm-svn: 214366
2014-07-31 08:11:11 +08:00
// Don't emit stores for undef values.
if (isa<UndefValue>(ArgVal))
continue;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Need to store on the stack.
AArch64: mark small types (i1, i8, i16) as promoted This means the output of LowerFormalArguments returns a lowered SDValue with the correct type (expected in SelectionDAGBuilder). Without this, an assertion under a DEBUG macro triggers when those types are passed on the stack. llvm-svn: 210102
2014-06-03 21:54:53 +08:00
unsigned ArgSize = (ArgVT.getSizeInBits() + 7) / 8;
[ARM64-BE] Teach fast-isel about how to set up sub-word stack arguments for big endian calls. SelectionDAG already knows about this, but fast-isel was ignorant. llvm-svn: 208307
2014-05-08 20:53:50 +08:00
unsigned BEAlign = 0;
if (ArgSize < 8 && !Subtarget->isLittleEndian())
BEAlign = 8 - ArgSize;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
Address Addr;
Addr.setKind(Address::RegBase);
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
Addr.setReg(AArch64::SP);
[ARM64-BE] Teach fast-isel about how to set up sub-word stack arguments for big endian calls. SelectionDAG already knows about this, but fast-isel was ignorant. llvm-svn: 208307
2014-05-08 20:53:50 +08:00
Addr.setOffset(VA.getLocMemOffset() + BEAlign);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
[FastISel][AArch64] Attach MachineMemOperands to load and store instructions. llvm-svn: 215231
2014-08-09 01:24:10 +08:00
unsigned Alignment = DL.getABITypeAlignment(ArgVal->getType());
MachineMemOperand *MMO = FuncInfo.MF->getMachineMemOperand(
MachinePointerInfo::getStack(Addr.getOffset()),
MachineMemOperand::MOStore, ArgVT.getStoreSize(), Alignment);
if (!EmitStore(ArgVT, ArgReg, Addr, MMO))
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return false;
}
}
return true;
}
[FastISel][AArch64] Fix return type in FastLowerCall. I used the wrong method to obtain the return type inside FinishCall. This fix simply uses the return type from FastLowerCall, which we already determined to be a valid type. Reduced test case from Chad. Thanks. llvm-svn: 213788
2014-07-24 04:03:13 +08:00
bool AArch64FastISel::FinishCall(CallLoweringInfo &CLI, MVT RetVT,
unsigned NumBytes) {
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
CallingConv::ID CC = CLI.CallConv;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Issue CALLSEQ_END
unsigned AdjStackUp = TII.getCallFrameDestroyOpcode();
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackUp))
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
.addImm(NumBytes).addImm(0);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Now the return value.
if (RetVT != MVT::isVoid) {
SmallVector<CCValAssign, 16> RVLocs;
Remove the target machine from CCState. Previously it was only used to get the subtarget and that's accessible from the MachineFunction now. This helps clear the way for smaller changes where we getting a subtarget will require passing in a MachineFunction/Function as well. llvm-svn: 214988
2014-08-07 02:45:26 +08:00
CCState CCInfo(CC, false, *FuncInfo.MF, RVLocs, *Context);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC));
// Only handle a single return value.
if (RVLocs.size() != 1)
return false;
// Copy all of the result registers out of their specified physreg.
MVT CopyVT = RVLocs[0].getValVT();
unsigned ResultReg = createResultReg(TLI.getRegClassFor(CopyVT));
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
TII.get(TargetOpcode::COPY), ResultReg)
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
.addReg(RVLocs[0].getLocReg());
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
CLI.InRegs.push_back(RVLocs[0].getLocReg());
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
CLI.ResultReg = ResultReg;
CLI.NumResultRegs = 1;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
return true;
}
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
bool AArch64FastISel::fastLowerCall(CallLoweringInfo &CLI) {
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
CallingConv::ID CC = CLI.CallConv;
[AArch64, fast-isel] Fall back to SelectionDAG to select tail calls. Certain functions such as objc_autoreleaseReturnValue have to be called as tail-calls even at -O0. Since normal fast-isel doesn't emit calls as tail calls, we have to fall back to SelectionDAG to select calls that are marked as tail. <rdar://problem/17991614> llvm-svn: 215600
2014-08-14 07:23:58 +08:00
bool IsTailCall = CLI.IsTailCall;
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
bool IsVarArg = CLI.IsVarArg;
const Value *Callee = CLI.Callee;
const char *SymName = CLI.SymName;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
[AArch64, fast-isel] Fall back to SelectionDAG to select tail calls. Certain functions such as objc_autoreleaseReturnValue have to be called as tail-calls even at -O0. Since normal fast-isel doesn't emit calls as tail calls, we have to fall back to SelectionDAG to select calls that are marked as tail. <rdar://problem/17991614> llvm-svn: 215600
2014-08-14 07:23:58 +08:00
// Allow SelectionDAG isel to handle tail calls.
if (IsTailCall)
return false;
[FastISel][AArch64] Add MachO large code model support for function calls. Currently the large code model for MachO uses the GOT to make function calls. Emit the required adrp and ldr instructions to load the address from the GOT. Related to <rdar://problem/17733076>. llvm-svn: 214381
2014-07-31 12:10:40 +08:00
CodeModel::Model CM = TM.getCodeModel();
// Only support the small and large code model.
if (CM != CodeModel::Small && CM != CodeModel::Large)
return false;
// FIXME: Add large code model support for ELF.
if (CM == CodeModel::Large && !Subtarget->isTargetMachO())
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return false;
// Let SDISel handle vararg functions.
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
if (IsVarArg)
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return false;
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
// FIXME: Only handle *simple* calls for now.
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
MVT RetVT;
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
if (CLI.RetTy->isVoidTy())
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
RetVT = MVT::isVoid;
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
else if (!isTypeLegal(CLI.RetTy, RetVT))
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return false;
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
for (auto Flag : CLI.OutFlags)
if (Flag.isInReg() || Flag.isSRet() || Flag.isNest() || Flag.isByVal())
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return false;
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
// Set up the argument vectors.
SmallVector<MVT, 16> OutVTs;
OutVTs.reserve(CLI.OutVals.size());
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
for (auto *Val : CLI.OutVals) {
MVT VT;
if (!isTypeLegal(Val->getType(), VT) &&
!(VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16))
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return false;
// We don't handle vector parameters yet.
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
if (VT.isVector() || VT.getSizeInBits() > 64)
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return false;
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
OutVTs.push_back(VT);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
[FastISel][AArch64] Add MachO large code model support for function calls. Currently the large code model for MachO uses the GOT to make function calls. Emit the required adrp and ldr instructions to load the address from the GOT. Related to <rdar://problem/17733076>. llvm-svn: 214381
2014-07-31 12:10:40 +08:00
Address Addr;
if (!ComputeCallAddress(Callee, Addr))
return false;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Handle the arguments now that we've gotten them.
unsigned NumBytes;
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
if (!ProcessCallArgs(CLI, OutVTs, NumBytes))
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return false;
// Issue the call.
MachineInstrBuilder MIB;
[FastISel][AArch64] Add MachO large code model support for function calls. Currently the large code model for MachO uses the GOT to make function calls. Emit the required adrp and ldr instructions to load the address from the GOT. Related to <rdar://problem/17733076>. llvm-svn: 214381
2014-07-31 12:10:40 +08:00
if (CM == CodeModel::Small) {
[FastISel][AArch64] Use the correct register class for branches. Also constrain the register class for branches. This fixes rdar://problem/18181496. llvm-svn: 216804
2014-08-30 07:48:06 +08:00
const MCInstrDesc &II = TII.get(Addr.getReg() ? AArch64::BLR : AArch64::BL);
MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II);
[FastISel][AArch64] Add MachO large code model support for function calls. Currently the large code model for MachO uses the GOT to make function calls. Emit the required adrp and ldr instructions to load the address from the GOT. Related to <rdar://problem/17733076>. llvm-svn: 214381
2014-07-31 12:10:40 +08:00
if (SymName)
MIB.addExternalSymbol(SymName, 0);
else if (Addr.getGlobalValue())
MIB.addGlobalAddress(Addr.getGlobalValue(), 0, 0);
[FastISel][AArch64] Use the correct register class for branches. Also constrain the register class for branches. This fixes rdar://problem/18181496. llvm-svn: 216804
2014-08-30 07:48:06 +08:00
else if (Addr.getReg()) {
unsigned Reg = constrainOperandRegClass(II, Addr.getReg(), 0);
MIB.addReg(Reg);
} else
[FastISel][AArch64] Add MachO large code model support for function calls. Currently the large code model for MachO uses the GOT to make function calls. Emit the required adrp and ldr instructions to load the address from the GOT. Related to <rdar://problem/17733076>. llvm-svn: 214381
2014-07-31 12:10:40 +08:00
return false;
} else {
unsigned CallReg = 0;
if (SymName) {
unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
ADRPReg)
.addExternalSymbol(SymName, AArch64II::MO_GOT | AArch64II::MO_PAGE);
CallReg = createResultReg(&AArch64::GPR64RegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::LDRXui),
CallReg)
.addReg(ADRPReg)
.addExternalSymbol(SymName, AArch64II::MO_GOT | AArch64II::MO_PAGEOFF |
AArch64II::MO_NC);
} else if (Addr.getGlobalValue()) {
CallReg = AArch64MaterializeGV(Addr.getGlobalValue());
} else if (Addr.getReg())
CallReg = Addr.getReg();
if (!CallReg)
return false;
[FastISel][AArch64] Use the correct register class for branches. Also constrain the register class for branches. This fixes rdar://problem/18181496. llvm-svn: 216804
2014-08-30 07:48:06 +08:00
const MCInstrDesc &II = TII.get(AArch64::BLR);
CallReg = constrainOperandRegClass(II, CallReg, 0);
MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II).addReg(CallReg);
[FastISel][AArch64] Add MachO large code model support for function calls. Currently the large code model for MachO uses the GOT to make function calls. Emit the required adrp and ldr instructions to load the address from the GOT. Related to <rdar://problem/17733076>. llvm-svn: 214381
2014-07-31 12:10:40 +08:00
}
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Add implicit physical register uses to the call.
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
for (auto Reg : CLI.OutRegs)
MIB.addReg(Reg, RegState::Implicit);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Add a register mask with the call-preserved registers.
// Proper defs for return values will be added by setPhysRegsDeadExcept().
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
MIB.addRegMask(TRI.getCallPreservedMask(CC));
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
[FastISel][AArch64] Add MachO large code model support for function calls. Currently the large code model for MachO uses the GOT to make function calls. Emit the required adrp and ldr instructions to load the address from the GOT. Related to <rdar://problem/17733076>. llvm-svn: 214381
2014-07-31 12:10:40 +08:00
CLI.Call = MIB;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Finish off the call including any return values.
[FastISel][AArch64] Fix return type in FastLowerCall. I used the wrong method to obtain the return type inside FinishCall. This fix simply uses the return type from FastLowerCall, which we already determined to be a valid type. Reduced test case from Chad. Thanks. llvm-svn: 213788
2014-07-24 04:03:13 +08:00
return FinishCall(CLI, RetVT, NumBytes);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
bool AArch64FastISel::IsMemCpySmall(uint64_t Len, unsigned Alignment) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
if (Alignment)
return Len / Alignment <= 4;
else
return Len < 32;
}
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
bool AArch64FastISel::TryEmitSmallMemCpy(Address Dest, Address Src,
uint64_t Len, unsigned Alignment) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Make sure we don't bloat code by inlining very large memcpy's.
if (!IsMemCpySmall(Len, Alignment))
return false;
int64_t UnscaledOffset = 0;
Address OrigDest = Dest;
Address OrigSrc = Src;
while (Len) {
MVT VT;
if (!Alignment || Alignment >= 8) {
if (Len >= 8)
VT = MVT::i64;
else if (Len >= 4)
VT = MVT::i32;
else if (Len >= 2)
VT = MVT::i16;
else {
VT = MVT::i8;
}
} else {
// Bound based on alignment.
if (Len >= 4 && Alignment == 4)
VT = MVT::i32;
else if (Len >= 2 && Alignment == 2)
VT = MVT::i16;
else {
VT = MVT::i8;
}
}
bool RV;
unsigned ResultReg;
RV = EmitLoad(VT, ResultReg, Src);
AArch64: make FastISel memcpy emission more robust. We were hitting an assert if FastISel couldn't create the load or store we requested. Currently this happens for large frame-local addresses, though CodeGen could be improved there. rdar://problem/17187463 llvm-svn: 210519
2014-06-10 17:52:40 +08:00
if (!RV)
return false;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
RV = EmitStore(VT, ResultReg, Dest);
AArch64: make FastISel memcpy emission more robust. We were hitting an assert if FastISel couldn't create the load or store we requested. Currently this happens for large frame-local addresses, though CodeGen could be improved there. rdar://problem/17187463 llvm-svn: 210519
2014-06-10 17:52:40 +08:00
if (!RV)
return false;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
int64_t Size = VT.getSizeInBits() / 8;
Len -= Size;
UnscaledOffset += Size;
// We need to recompute the unscaled offset for each iteration.
Dest.setOffset(OrigDest.getOffset() + UnscaledOffset);
Src.setOffset(OrigSrc.getOffset() + UnscaledOffset);
}
return true;
}
[FastISel][AArch64] Add branch folding support for the XALU intrinsics. This improves the code generation for the XALU intrinsics when the condition is feeding a branch instruction. This is related to <rdar://problem/17831117>. llvm-svn: 214349
2014-07-31 06:04:34 +08:00
/// \brief Check if it is possible to fold the condition from the XALU intrinsic
/// into the user. The condition code will only be updated on success.
bool AArch64FastISel::foldXALUIntrinsic(AArch64CC::CondCode &CC,
const Instruction *I,
const Value *Cond) {
if (!isa<ExtractValueInst>(Cond))
return false;
const auto *EV = cast<ExtractValueInst>(Cond);
if (!isa<IntrinsicInst>(EV->getAggregateOperand()))
return false;
const auto *II = cast<IntrinsicInst>(EV->getAggregateOperand());
MVT RetVT;
const Function *Callee = II->getCalledFunction();
Type *RetTy =
cast<StructType>(Callee->getReturnType())->getTypeAtIndex(0U);
if (!isTypeLegal(RetTy, RetVT))
return false;
if (RetVT != MVT::i32 && RetVT != MVT::i64)
return false;
AArch64CC::CondCode TmpCC;
switch (II->getIntrinsicID()) {
default: return false;
case Intrinsic::sadd_with_overflow:
case Intrinsic::ssub_with_overflow: TmpCC = AArch64CC::VS; break;
case Intrinsic::uadd_with_overflow: TmpCC = AArch64CC::HS; break;
case Intrinsic::usub_with_overflow: TmpCC = AArch64CC::LO; break;
case Intrinsic::smul_with_overflow:
case Intrinsic::umul_with_overflow: TmpCC = AArch64CC::NE; break;
}
// Check if both instructions are in the same basic block.
if (II->getParent() != I->getParent())
return false;
// Make sure nothing is in the way
BasicBlock::const_iterator Start = I;
BasicBlock::const_iterator End = II;
for (auto Itr = std::prev(Start); Itr != End; --Itr) {
// We only expect extractvalue instructions between the intrinsic and the
// instruction to be selected.
if (!isa<ExtractValueInst>(Itr))
return false;
// Check that the extractvalue operand comes from the intrinsic.
const auto *EVI = cast<ExtractValueInst>(Itr);
if (EVI->getAggregateOperand() != II)
return false;
}
CC = TmpCC;
return true;
}
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
bool AArch64FastISel::fastLowerIntrinsicCall(const IntrinsicInst *II) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// FIXME: Handle more intrinsics.
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
switch (II->getIntrinsicID()) {
[FastISel][AArch64] Add support for frameaddress intrinsic. This commit implements the frameaddress intrinsic for the AArch64 architecture in FastISel. There were two test cases that pretty much tested the same, so I combined them to a single test case. Fixes <rdar://problem/17811834> llvm-svn: 213959
2014-07-26 01:47:14 +08:00
default: return false;
case Intrinsic::frameaddress: {
MachineFrameInfo *MFI = FuncInfo.MF->getFrameInfo();
MFI->setFrameAddressIsTaken(true);
const AArch64RegisterInfo *RegInfo =
Remove the TargetMachine forwards for TargetSubtargetInfo based information and update all callers. No functional change. llvm-svn: 214781
2014-08-05 05:25:23 +08:00
static_cast<const AArch64RegisterInfo *>(
TM.getSubtargetImpl()->getRegisterInfo());
[FastISel][AArch64] Add support for frameaddress intrinsic. This commit implements the frameaddress intrinsic for the AArch64 architecture in FastISel. There were two test cases that pretty much tested the same, so I combined them to a single test case. Fixes <rdar://problem/17811834> llvm-svn: 213959
2014-07-26 01:47:14 +08:00
unsigned FramePtr = RegInfo->getFrameRegister(*(FuncInfo.MF));
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
unsigned SrcReg = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
TII.get(TargetOpcode::COPY), SrcReg).addReg(FramePtr);
[FastISel][AArch64] Add support for frameaddress intrinsic. This commit implements the frameaddress intrinsic for the AArch64 architecture in FastISel. There were two test cases that pretty much tested the same, so I combined them to a single test case. Fixes <rdar://problem/17811834> llvm-svn: 213959
2014-07-26 01:47:14 +08:00
// Recursively load frame address
// ldr x0, [fp]
// ldr x0, [x0]
// ldr x0, [x0]
// ...
unsigned DestReg;
unsigned Depth = cast<ConstantInt>(II->getOperand(0))->getZExtValue();
while (Depth--) {
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
DestReg = fastEmitInst_ri(AArch64::LDRXui, &AArch64::GPR64RegClass,
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
SrcReg, /*IsKill=*/true, 0);
assert(DestReg && "Unexpected LDR instruction emission failure.");
[FastISel][AArch64] Add support for frameaddress intrinsic. This commit implements the frameaddress intrinsic for the AArch64 architecture in FastISel. There were two test cases that pretty much tested the same, so I combined them to a single test case. Fixes <rdar://problem/17811834> llvm-svn: 213959
2014-07-26 01:47:14 +08:00
SrcReg = DestReg;
}
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
updateValueMap(II, SrcReg);
[FastISel][AArch64] Add support for frameaddress intrinsic. This commit implements the frameaddress intrinsic for the AArch64 architecture in FastISel. There were two test cases that pretty much tested the same, so I combined them to a single test case. Fixes <rdar://problem/17811834> llvm-svn: 213959
2014-07-26 01:47:14 +08:00
return true;
}
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
case Intrinsic::memcpy:
case Intrinsic::memmove: {
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
const auto *MTI = cast<MemTransferInst>(II);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Don't handle volatile.
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
if (MTI->isVolatile())
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return false;
Revert "[FastISel][AArch64] Don't fold instructions too aggressively into the memory operation." Quentin pointed out that this is not the correct approach and there is a better and easier solution. llvm-svn: 216632
2014-08-28 07:09:40 +08:00
// Disable inlining for memmove before calls to ComputeAddress. Otherwise,
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// we would emit dead code because we don't currently handle memmoves.
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
bool IsMemCpy = (II->getIntrinsicID() == Intrinsic::memcpy);
if (isa<ConstantInt>(MTI->getLength()) && IsMemCpy) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Small memcpy's are common enough that we want to do them without a call
// if possible.
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
uint64_t Len = cast<ConstantInt>(MTI->getLength())->getZExtValue();
unsigned Alignment = MTI->getAlignment();
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
if (IsMemCpySmall(Len, Alignment)) {
Address Dest, Src;
Revert "[FastISel][AArch64] Don't fold instructions too aggressively into the memory operation." Quentin pointed out that this is not the correct approach and there is a better and easier solution. llvm-svn: 216632
2014-08-28 07:09:40 +08:00
if (!ComputeAddress(MTI->getRawDest(), Dest) ||
!ComputeAddress(MTI->getRawSource(), Src))
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return false;
if (TryEmitSmallMemCpy(Dest, Src, Len, Alignment))
return true;
}
}
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
if (!MTI->getLength()->getType()->isIntegerTy(64))
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return false;
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
if (MTI->getSourceAddressSpace() > 255 || MTI->getDestAddressSpace() > 255)
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Fast instruction selection doesn't support the special
// address spaces.
return false;
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
const char *IntrMemName = isa<MemCpyInst>(II) ? "memcpy" : "memmove";
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
return lowerCallTo(II, IntrMemName, II->getNumArgOperands() - 2);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
case Intrinsic::memset: {
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
const MemSetInst *MSI = cast<MemSetInst>(II);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Don't handle volatile.
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
if (MSI->isVolatile())
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return false;
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
if (!MSI->getLength()->getType()->isIntegerTy(64))
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return false;
[FastIsel][AArch64] Add support for the FastLowerCall and FastLowerIntrinsicCall target-hooks. This commit modifies the existing call lowering functions to be used as the FastLowerCall and FastLowerIntrinsicCall target-hooks instead. This enables patchpoint intrinsic lowering for AArch64. This fixes <rdar://problem/17733076> llvm-svn: 213704
2014-07-23 07:14:58 +08:00
if (MSI->getDestAddressSpace() > 255)
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Fast instruction selection doesn't support the special
// address spaces.
return false;
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
return lowerCallTo(II, "memset", II->getNumArgOperands() - 2);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
case Intrinsic::trap: {
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::BRK))
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
.addImm(1);
return true;
}
[FastISel][AArch64] Add sqrt intrinsic support. Fixes <rdar://problem/17867067>. llvm-svn: 214388
2014-07-31 14:25:33 +08:00
case Intrinsic::sqrt: {
Type *RetTy = II->getCalledFunction()->getReturnType();
MVT VT;
if (!isTypeLegal(RetTy, VT))
return false;
unsigned Op0Reg = getRegForValue(II->getOperand(0));
if (!Op0Reg)
return false;
bool Op0IsKill = hasTrivialKill(II->getOperand(0));
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
unsigned ResultReg = fastEmit_r(VT, VT, ISD::FSQRT, Op0Reg, Op0IsKill);
[FastISel][AArch64] Add sqrt intrinsic support. Fixes <rdar://problem/17867067>. llvm-svn: 214388
2014-07-31 14:25:33 +08:00
if (!ResultReg)
return false;
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
updateValueMap(II, ResultReg);
[FastISel][AArch64] Add sqrt intrinsic support. Fixes <rdar://problem/17867067>. llvm-svn: 214388
2014-07-31 14:25:33 +08:00
return true;
}
[FastISel][AArch64] Add {s|u}{add|sub|mul}.with.overflow intrinsic support. This commit adds support for the {s|u}{add|sub|mul}.with.overflow intrinsics. The unit tests for FastISel will be enabled in a later commit, once there is also branch and select folding support. This is related to <rdar://problem/17831117>. llvm-svn: 214348
2014-07-31 06:04:31 +08:00
case Intrinsic::sadd_with_overflow:
case Intrinsic::uadd_with_overflow:
case Intrinsic::ssub_with_overflow:
case Intrinsic::usub_with_overflow:
case Intrinsic::smul_with_overflow:
case Intrinsic::umul_with_overflow: {
// This implements the basic lowering of the xalu with overflow intrinsics.
const Function *Callee = II->getCalledFunction();
auto *Ty = cast<StructType>(Callee->getReturnType());
Type *RetTy = Ty->getTypeAtIndex(0U);
MVT VT;
if (!isTypeLegal(RetTy, VT))
return false;
if (VT != MVT::i32 && VT != MVT::i64)
return false;
const Value *LHS = II->getArgOperand(0);
const Value *RHS = II->getArgOperand(1);
// Canonicalize immediate to the RHS.
if (isa<ConstantInt>(LHS) && !isa<ConstantInt>(RHS) &&
isCommutativeIntrinsic(II))
std::swap(LHS, RHS);
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
unsigned ResultReg1 = 0, ResultReg2 = 0, MulReg = 0;
[FastISel][AArch64] Add {s|u}{add|sub|mul}.with.overflow intrinsic support. This commit adds support for the {s|u}{add|sub|mul}.with.overflow intrinsics. The unit tests for FastISel will be enabled in a later commit, once there is also branch and select folding support. This is related to <rdar://problem/17831117>. llvm-svn: 214348
2014-07-31 06:04:31 +08:00
AArch64CC::CondCode CC = AArch64CC::Invalid;
switch (II->getIntrinsicID()) {
default: llvm_unreachable("Unexpected intrinsic!");
case Intrinsic::sadd_with_overflow:
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
ResultReg1 = emitAdd(VT, LHS, RHS, /*SetFlags=*/true);
CC = AArch64CC::VS;
break;
[FastISel][AArch64] Add {s|u}{add|sub|mul}.with.overflow intrinsic support. This commit adds support for the {s|u}{add|sub|mul}.with.overflow intrinsics. The unit tests for FastISel will be enabled in a later commit, once there is also branch and select folding support. This is related to <rdar://problem/17831117>. llvm-svn: 214348
2014-07-31 06:04:31 +08:00
case Intrinsic::uadd_with_overflow:
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
ResultReg1 = emitAdd(VT, LHS, RHS, /*SetFlags=*/true);
CC = AArch64CC::HS;
break;
[FastISel][AArch64] Add {s|u}{add|sub|mul}.with.overflow intrinsic support. This commit adds support for the {s|u}{add|sub|mul}.with.overflow intrinsics. The unit tests for FastISel will be enabled in a later commit, once there is also branch and select folding support. This is related to <rdar://problem/17831117>. llvm-svn: 214348
2014-07-31 06:04:31 +08:00
case Intrinsic::ssub_with_overflow:
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
ResultReg1 = emitSub(VT, LHS, RHS, /*SetFlags=*/true);
CC = AArch64CC::VS;
break;
[FastISel][AArch64] Add {s|u}{add|sub|mul}.with.overflow intrinsic support. This commit adds support for the {s|u}{add|sub|mul}.with.overflow intrinsics. The unit tests for FastISel will be enabled in a later commit, once there is also branch and select folding support. This is related to <rdar://problem/17831117>. llvm-svn: 214348
2014-07-31 06:04:31 +08:00
case Intrinsic::usub_with_overflow:
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
ResultReg1 = emitSub(VT, LHS, RHS, /*SetFlags=*/true);
CC = AArch64CC::LO;
break;
[FastISel][AArch64] Add {s|u}{add|sub|mul}.with.overflow intrinsic support. This commit adds support for the {s|u}{add|sub|mul}.with.overflow intrinsics. The unit tests for FastISel will be enabled in a later commit, once there is also branch and select folding support. This is related to <rdar://problem/17831117>. llvm-svn: 214348
2014-07-31 06:04:31 +08:00
case Intrinsic::smul_with_overflow: {
CC = AArch64CC::NE;
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
unsigned LHSReg = getRegForValue(LHS);
if (!LHSReg)
return false;
bool LHSIsKill = hasTrivialKill(LHS);
unsigned RHSReg = getRegForValue(RHS);
[FastISel][AArch64] Fix the immediate versions of the {s|u}{add|sub}.with.overflow intrinsics. ADDS and SUBS cannot encode negative immediates or immediates larger than 12bit. This fix checks if the immediate version can be used under this constraints and if we can convert ADDS to SUBS or vice versa to support negative immediates. Also update the test cases to test the immediate versions. llvm-svn: 214470
2014-08-01 09:25:55 +08:00
if (!RHSReg)
return false;
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
bool RHSIsKill = hasTrivialKill(RHS);
[FastISel][AArch64] Fix the immediate versions of the {s|u}{add|sub}.with.overflow intrinsics. ADDS and SUBS cannot encode negative immediates or immediates larger than 12bit. This fix checks if the immediate version can be used under this constraints and if we can convert ADDS to SUBS or vice versa to support negative immediates. Also update the test cases to test the immediate versions. llvm-svn: 214470
2014-08-01 09:25:55 +08:00
[FastISel][AArch64] Add {s|u}{add|sub|mul}.with.overflow intrinsic support. This commit adds support for the {s|u}{add|sub|mul}.with.overflow intrinsics. The unit tests for FastISel will be enabled in a later commit, once there is also branch and select folding support. This is related to <rdar://problem/17831117>. llvm-svn: 214348
2014-07-31 06:04:31 +08:00
if (VT == MVT::i32) {
MulReg = Emit_SMULL_rr(MVT::i64, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
[FastISel][AArch64] Fold Sign-/Zero-Extend into the shift immediate instruction. llvm-svn: 216510
2014-08-27 08:58:26 +08:00
unsigned ShiftReg = emitLSR_ri(MVT::i64, MVT::i64, MulReg,
/*IsKill=*/false, 32);
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
MulReg = fastEmitInst_extractsubreg(VT, MulReg, /*IsKill=*/true,
[FastISel][AArch64] Add {s|u}{add|sub|mul}.with.overflow intrinsic support. This commit adds support for the {s|u}{add|sub|mul}.with.overflow intrinsics. The unit tests for FastISel will be enabled in a later commit, once there is also branch and select folding support. This is related to <rdar://problem/17831117>. llvm-svn: 214348
2014-07-31 06:04:31 +08:00
AArch64::sub_32);
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
ShiftReg = fastEmitInst_extractsubreg(VT, ShiftReg, /*IsKill=*/true,
[FastISel][AArch64] Add {s|u}{add|sub|mul}.with.overflow intrinsic support. This commit adds support for the {s|u}{add|sub|mul}.with.overflow intrinsics. The unit tests for FastISel will be enabled in a later commit, once there is also branch and select folding support. This is related to <rdar://problem/17831117>. llvm-svn: 214348
2014-07-31 06:04:31 +08:00
AArch64::sub_32);
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
emitSubs_rs(VT, ShiftReg, /*IsKill=*/true, MulReg, /*IsKill=*/false,
AArch64_AM::ASR, 31, /*WantResult=*/false);
[FastISel][AArch64] Add {s|u}{add|sub|mul}.with.overflow intrinsic support. This commit adds support for the {s|u}{add|sub|mul}.with.overflow intrinsics. The unit tests for FastISel will be enabled in a later commit, once there is also branch and select folding support. This is related to <rdar://problem/17831117>. llvm-svn: 214348
2014-07-31 06:04:31 +08:00
} else {
assert(VT == MVT::i64 && "Unexpected value type.");
MulReg = Emit_MUL_rr(VT, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
unsigned SMULHReg = fastEmit_rr(VT, VT, ISD::MULHS, LHSReg, LHSIsKill,
[FastISel][AArch64] Add {s|u}{add|sub|mul}.with.overflow intrinsic support. This commit adds support for the {s|u}{add|sub|mul}.with.overflow intrinsics. The unit tests for FastISel will be enabled in a later commit, once there is also branch and select folding support. This is related to <rdar://problem/17831117>. llvm-svn: 214348
2014-07-31 06:04:31 +08:00
RHSReg, RHSIsKill);
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
emitSubs_rs(VT, SMULHReg, /*IsKill=*/true, MulReg, /*IsKill=*/false,
AArch64_AM::ASR, 63, /*WantResult=*/false);
[FastISel][AArch64] Add {s|u}{add|sub|mul}.with.overflow intrinsic support. This commit adds support for the {s|u}{add|sub|mul}.with.overflow intrinsics. The unit tests for FastISel will be enabled in a later commit, once there is also branch and select folding support. This is related to <rdar://problem/17831117>. llvm-svn: 214348
2014-07-31 06:04:31 +08:00
}
break;
}
case Intrinsic::umul_with_overflow: {
CC = AArch64CC::NE;
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
unsigned LHSReg = getRegForValue(LHS);
if (!LHSReg)
return false;
bool LHSIsKill = hasTrivialKill(LHS);
unsigned RHSReg = getRegForValue(RHS);
[FastISel][AArch64] Fix the immediate versions of the {s|u}{add|sub}.with.overflow intrinsics. ADDS and SUBS cannot encode negative immediates or immediates larger than 12bit. This fix checks if the immediate version can be used under this constraints and if we can convert ADDS to SUBS or vice versa to support negative immediates. Also update the test cases to test the immediate versions. llvm-svn: 214470
2014-08-01 09:25:55 +08:00
if (!RHSReg)
return false;
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
bool RHSIsKill = hasTrivialKill(RHS);
[FastISel][AArch64] Fix the immediate versions of the {s|u}{add|sub}.with.overflow intrinsics. ADDS and SUBS cannot encode negative immediates or immediates larger than 12bit. This fix checks if the immediate version can be used under this constraints and if we can convert ADDS to SUBS or vice versa to support negative immediates. Also update the test cases to test the immediate versions. llvm-svn: 214470
2014-08-01 09:25:55 +08:00
[FastISel][AArch64] Add {s|u}{add|sub|mul}.with.overflow intrinsic support. This commit adds support for the {s|u}{add|sub|mul}.with.overflow intrinsics. The unit tests for FastISel will be enabled in a later commit, once there is also branch and select folding support. This is related to <rdar://problem/17831117>. llvm-svn: 214348
2014-07-31 06:04:31 +08:00
if (VT == MVT::i32) {
MulReg = Emit_UMULL_rr(MVT::i64, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
emitSubs_rs(MVT::i64, AArch64::XZR, /*IsKill=*/true, MulReg,
/*IsKill=*/false, AArch64_AM::LSR, 32,
/*WantResult=*/false);
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
MulReg = fastEmitInst_extractsubreg(VT, MulReg, /*IsKill=*/true,
[FastISel][AArch64] Add {s|u}{add|sub|mul}.with.overflow intrinsic support. This commit adds support for the {s|u}{add|sub|mul}.with.overflow intrinsics. The unit tests for FastISel will be enabled in a later commit, once there is also branch and select folding support. This is related to <rdar://problem/17831117>. llvm-svn: 214348
2014-07-31 06:04:31 +08:00
AArch64::sub_32);
} else {
assert(VT == MVT::i64 && "Unexpected value type.");
MulReg = Emit_MUL_rr(VT, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
unsigned UMULHReg = fastEmit_rr(VT, VT, ISD::MULHU, LHSReg, LHSIsKill,
[FastISel][AArch64] Add {s|u}{add|sub|mul}.with.overflow intrinsic support. This commit adds support for the {s|u}{add|sub|mul}.with.overflow intrinsics. The unit tests for FastISel will be enabled in a later commit, once there is also branch and select folding support. This is related to <rdar://problem/17831117>. llvm-svn: 214348
2014-07-31 06:04:31 +08:00
RHSReg, RHSIsKill);
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
emitSubs_rr(VT, AArch64::XZR, /*IsKill=*/true, UMULHReg,
/*IsKill=*/false, /*WantResult=*/false);
[FastISel][AArch64] Add {s|u}{add|sub|mul}.with.overflow intrinsic support. This commit adds support for the {s|u}{add|sub|mul}.with.overflow intrinsics. The unit tests for FastISel will be enabled in a later commit, once there is also branch and select folding support. This is related to <rdar://problem/17831117>. llvm-svn: 214348
2014-07-31 06:04:31 +08:00
}
break;
}
}
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
if (MulReg) {
ResultReg1 = createResultReg(TLI.getRegClassFor(VT));
[FastISel][AArch64] Add {s|u}{add|sub|mul}.with.overflow intrinsic support. This commit adds support for the {s|u}{add|sub|mul}.with.overflow intrinsics. The unit tests for FastISel will be enabled in a later commit, once there is also branch and select folding support. This is related to <rdar://problem/17831117>. llvm-svn: 214348
2014-07-31 06:04:31 +08:00
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
TII.get(TargetOpcode::COPY), ResultReg1).addReg(MulReg);
}
[FastISel][AArch64] Add {s|u}{add|sub|mul}.with.overflow intrinsic support. This commit adds support for the {s|u}{add|sub|mul}.with.overflow intrinsics. The unit tests for FastISel will be enabled in a later commit, once there is also branch and select folding support. This is related to <rdar://problem/17831117>. llvm-svn: 214348
2014-07-31 06:04:31 +08:00
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
ResultReg2 = fastEmitInst_rri(AArch64::CSINCWr, &AArch64::GPR32RegClass,
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
AArch64::WZR, /*IsKill=*/true, AArch64::WZR,
/*IsKill=*/true, getInvertedCondCode(CC));
[FastISel][AArch64] Factor out ADDS/SUBS instruction emission and add support for extensions and shift folding. Factor out the ADDS/SUBS instruction emission code into helper functions and make the helper functions more clever to support most of the different ADDS/SUBS instructions the architecture support. This includes better immedediate support, shift folding, and sign-/zero-extend folding. This fixes <rdar://problem/17913111>. llvm-svn: 216033
2014-08-20 06:29:55 +08:00
assert((ResultReg1 + 1) == ResultReg2 &&
"Nonconsecutive result registers.");
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
updateValueMap(II, ResultReg1, 2);
[FastISel][AArch64] Add {s|u}{add|sub|mul}.with.overflow intrinsic support. This commit adds support for the {s|u}{add|sub|mul}.with.overflow intrinsics. The unit tests for FastISel will be enabled in a later commit, once there is also branch and select folding support. This is related to <rdar://problem/17831117>. llvm-svn: 214348
2014-07-31 06:04:31 +08:00
return true;
}
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
return false;
}
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
bool AArch64FastISel::SelectRet(const Instruction *I) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
const ReturnInst *Ret = cast<ReturnInst>(I);
const Function &F = *I->getParent()->getParent();
if (!FuncInfo.CanLowerReturn)
return false;
if (F.isVarArg())
return false;
// Build a list of return value registers.
SmallVector<unsigned, 4> RetRegs;
if (Ret->getNumOperands() > 0) {
CallingConv::ID CC = F.getCallingConv();
SmallVector<ISD::OutputArg, 4> Outs;
GetReturnInfo(F.getReturnType(), F.getAttributes(), Outs, TLI);
// Analyze operands of the call, assigning locations to each operand.
SmallVector<CCValAssign, 16> ValLocs;
Remove the target machine from CCState. Previously it was only used to get the subtarget and that's accessible from the MachineFunction now. This helps clear the way for smaller changes where we getting a subtarget will require passing in a MachineFunction/Function as well. llvm-svn: 214988
2014-08-07 02:45:26 +08:00
CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, ValLocs, I->getContext());
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
CCAssignFn *RetCC = CC == CallingConv::WebKit_JS ? RetCC_AArch64_WebKit_JS
: RetCC_AArch64_AAPCS;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
CCInfo.AnalyzeReturn(Outs, RetCC);
// Only handle a single return value for now.
if (ValLocs.size() != 1)
return false;
CCValAssign &VA = ValLocs[0];
const Value *RV = Ret->getOperand(0);
// Don't bother handling odd stuff for now.
if (VA.getLocInfo() != CCValAssign::Full)
return false;
// Only handle register returns for now.
if (!VA.isRegLoc())
return false;
unsigned Reg = getRegForValue(RV);
if (Reg == 0)
return false;
unsigned SrcReg = Reg + VA.getValNo();
unsigned DestReg = VA.getLocReg();
// Avoid a cross-class copy. This is very unlikely.
if (!MRI.getRegClass(SrcReg)->contains(DestReg))
return false;
EVT RVEVT = TLI.getValueType(RV->getType());
if (!RVEVT.isSimple())
return false;
[ARM64-BE] Fix fast-isel, and add appropriate RUN lines to appropriate tests. llvm-svn: 208200
2014-05-07 20:33:55 +08:00
// Vectors (of > 1 lane) in big endian need tricky handling.
if (RVEVT.isVector() && RVEVT.getVectorNumElements() > 1)
return false;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
MVT RVVT = RVEVT.getSimpleVT();
[ARM64][fast-isel] Fast-isel doesn't know how to handle f128. llvm-svn: 207659
2014-04-30 23:29:57 +08:00
if (RVVT == MVT::f128)
return false;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
MVT DestVT = VA.getValVT();
// Special handling for extended integers.
if (RVVT != DestVT) {
if (RVVT != MVT::i1 && RVVT != MVT::i8 && RVVT != MVT::i16)
return false;
if (!Outs[0].Flags.isZExt() && !Outs[0].Flags.isSExt())
return false;
bool isZExt = Outs[0].Flags.isZExt();
SrcReg = EmitIntExt(RVVT, SrcReg, DestVT, isZExt);
if (SrcReg == 0)
return false;
}
// Make the copy.
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
TII.get(TargetOpcode::COPY), DestReg).addReg(SrcReg);
// Add register to return instruction.
RetRegs.push_back(VA.getLocReg());
}
MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
TII.get(AArch64::RET_ReallyLR));
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
for (unsigned i = 0, e = RetRegs.size(); i != e; ++i)
MIB.addReg(RetRegs[i], RegState::Implicit);
return true;
}
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
bool AArch64FastISel::SelectTrunc(const Instruction *I) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
Type *DestTy = I->getType();
Value *Op = I->getOperand(0);
Type *SrcTy = Op->getType();
EVT SrcEVT = TLI.getValueType(SrcTy, true);
EVT DestEVT = TLI.getValueType(DestTy, true);
if (!SrcEVT.isSimple())
return false;
if (!DestEVT.isSimple())
return false;
MVT SrcVT = SrcEVT.getSimpleVT();
MVT DestVT = DestEVT.getSimpleVT();
if (SrcVT != MVT::i64 && SrcVT != MVT::i32 && SrcVT != MVT::i16 &&
SrcVT != MVT::i8)
return false;
if (DestVT != MVT::i32 && DestVT != MVT::i16 && DestVT != MVT::i8 &&
DestVT != MVT::i1)
return false;
unsigned SrcReg = getRegForValue(Op);
if (!SrcReg)
return false;
[FastISel][AArch64] Factor out ANDWri instruction generation into a helper function. NFCI. llvm-svn: 216199
2014-08-22 02:02:25 +08:00
bool SrcIsKill = hasTrivialKill(Op);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// If we're truncating from i64 to a smaller non-legal type then generate an
[FastISel][AArch64] Fix an incorrect kill flag due to a bug in SelectTrunc. When we select a trunc instruction we don't emit any code if the type is already i32 or smaller. This is because the instruction that uses the truncated value will deal with it. This behavior can incorrectly transfer a kill flag, which was meant for the result of the truncate, onto the source register. %2 = trunc i32 %1 to i16 ... = ... %2 -> ... = ... vreg1 <kill> ... = ... %1 ... = ... vreg1 This commit fixes this by emitting a COPY instruction, so that the result and source register are distinct virtual registers. This fixes rdar://problem/18178188. llvm-svn: 216750
2014-08-30 01:58:16 +08:00
// AND. Otherwise, we know the high bits are undefined and a truncate only
// generate a COPY. We cannot mark the source register also as result
// register, because this can incorrectly transfer the kill flag onto the
// source register.
unsigned ResultReg;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
if (SrcVT == MVT::i64) {
uint64_t Mask = 0;
switch (DestVT.SimpleTy) {
default:
// Trunc i64 to i32 is handled by the target-independent fast-isel.
return false;
case MVT::i1:
Mask = 0x1;
break;
case MVT::i8:
Mask = 0xff;
break;
case MVT::i16:
Mask = 0xffff;
break;
}
// Issue an extract_subreg to get the lower 32-bits.
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
unsigned Reg32 = fastEmitInst_extractsubreg(MVT::i32, SrcReg, SrcIsKill,
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
AArch64::sub_32);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Create the AND instruction which performs the actual truncation.
[FastISel][AArch64] Add target-specific lowering for logical operations. This change adds support for immediate and shift-left folding into logical operations. This fixes rdar://problem/18223183. llvm-svn: 217118
2014-09-04 09:29:18 +08:00
ResultReg = emitAnd_ri(MVT::i32, Reg32, /*IsKill=*/true, Mask);
[FastISel][AArch64] Fix an incorrect kill flag due to a bug in SelectTrunc. When we select a trunc instruction we don't emit any code if the type is already i32 or smaller. This is because the instruction that uses the truncated value will deal with it. This behavior can incorrectly transfer a kill flag, which was meant for the result of the truncate, onto the source register. %2 = trunc i32 %1 to i16 ... = ... %2 -> ... = ... vreg1 <kill> ... = ... %1 ... = ... vreg1 This commit fixes this by emitting a COPY instruction, so that the result and source register are distinct virtual registers. This fixes rdar://problem/18178188. llvm-svn: 216750
2014-08-30 01:58:16 +08:00
assert(ResultReg && "Unexpected AND instruction emission failure.");
} else {
ResultReg = createResultReg(&AArch64::GPR32RegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
TII.get(TargetOpcode::COPY), ResultReg)
.addReg(SrcReg, getKillRegState(SrcIsKill));
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
updateValueMap(I, ResultReg);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return true;
}
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
unsigned AArch64FastISel::Emiti1Ext(unsigned SrcReg, MVT DestVT, bool isZExt) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
assert((DestVT == MVT::i8 || DestVT == MVT::i16 || DestVT == MVT::i32 ||
DestVT == MVT::i64) &&
"Unexpected value type.");
// Handle i8 and i16 as i32.
if (DestVT == MVT::i8 || DestVT == MVT::i16)
DestVT = MVT::i32;
if (isZExt) {
[FastISel][AArch64] Add target-specific lowering for logical operations. This change adds support for immediate and shift-left folding into logical operations. This fixes rdar://problem/18223183. llvm-svn: 217118
2014-09-04 09:29:18 +08:00
unsigned ResultReg = emitAnd_ri(MVT::i32, SrcReg, /*TODO:IsKill=*/false, 1);
[FastISel][AArch64] Factor out ANDWri instruction generation into a helper function. NFCI. llvm-svn: 216199
2014-08-22 02:02:25 +08:00
assert(ResultReg && "Unexpected AND instruction emission failure.");
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
if (DestVT == MVT::i64) {
// We're ZExt i1 to i64. The ANDWri Wd, Ws, #1 implicitly clears the
// upper 32 bits. Emit a SUBREG_TO_REG to extend from Wd to Xd.
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
unsigned Reg64 = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
TII.get(AArch64::SUBREG_TO_REG), Reg64)
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
.addImm(0)
.addReg(ResultReg)
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
.addImm(AArch64::sub_32);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
ResultReg = Reg64;
}
return ResultReg;
} else {
if (DestVT == MVT::i64) {
// FIXME: We're SExt i1 to i64.
return 0;
}
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
return fastEmitInst_rii(AArch64::SBFMWri, &AArch64::GPR32RegClass, SrcReg,
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
/*TODO:IsKill=*/false, 0, 0);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
}
[FastISel][AArch64] Create helper functions to create the various multiplies on AArch64. llvm-svn: 214346
2014-07-31 06:04:25 +08:00
unsigned AArch64FastISel::Emit_MUL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
unsigned Op1, bool Op1IsKill) {
unsigned Opc, ZReg;
switch (RetVT.SimpleTy) {
default: return 0;
case MVT::i8:
case MVT::i16:
case MVT::i32:
RetVT = MVT::i32;
Opc = AArch64::MADDWrrr; ZReg = AArch64::WZR; break;
case MVT::i64:
Opc = AArch64::MADDXrrr; ZReg = AArch64::XZR; break;
}
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
const TargetRegisterClass *RC =
(RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
return fastEmitInst_rrr(Opc, RC, Op0, Op0IsKill, Op1, Op1IsKill,
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
/*IsKill=*/ZReg, true);
[FastISel][AArch64] Create helper functions to create the various multiplies on AArch64. llvm-svn: 214346
2014-07-31 06:04:25 +08:00
}
unsigned AArch64FastISel::Emit_SMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
unsigned Op1, bool Op1IsKill) {
if (RetVT != MVT::i64)
return 0;
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
return fastEmitInst_rrr(AArch64::SMADDLrrr, &AArch64::GPR64RegClass,
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
Op0, Op0IsKill, Op1, Op1IsKill,
AArch64::XZR, /*IsKill=*/true);
[FastISel][AArch64] Create helper functions to create the various multiplies on AArch64. llvm-svn: 214346
2014-07-31 06:04:25 +08:00
}
unsigned AArch64FastISel::Emit_UMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
unsigned Op1, bool Op1IsKill) {
if (RetVT != MVT::i64)
return 0;
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
return fastEmitInst_rrr(AArch64::UMADDLrrr, &AArch64::GPR64RegClass,
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
Op0, Op0IsKill, Op1, Op1IsKill,
AArch64::XZR, /*IsKill=*/true);
[FastISel][AArch64] Create helper functions to create the various multiplies on AArch64. llvm-svn: 214346
2014-07-31 06:04:25 +08:00
}
[FastISel][AArch64] Add support for variable shift. This adds the missing variable shift support for value type i8, i16, and i32. This fixes <rdar://problem/18095685>. llvm-svn: 216242
2014-08-22 07:06:07 +08:00
unsigned AArch64FastISel::emitLSL_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
unsigned Op1Reg, bool Op1IsKill) {
unsigned Opc = 0;
bool NeedTrunc = false;
uint64_t Mask = 0;
switch (RetVT.SimpleTy) {
default: return 0;
case MVT::i8: Opc = AArch64::LSLVWr; NeedTrunc = true; Mask = 0xff; break;
case MVT::i16: Opc = AArch64::LSLVWr; NeedTrunc = true; Mask = 0xffff; break;
case MVT::i32: Opc = AArch64::LSLVWr; break;
case MVT::i64: Opc = AArch64::LSLVXr; break;
}
const TargetRegisterClass *RC =
(RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
if (NeedTrunc) {
[FastISel][AArch64] Add target-specific lowering for logical operations. This change adds support for immediate and shift-left folding into logical operations. This fixes rdar://problem/18223183. llvm-svn: 217118
2014-09-04 09:29:18 +08:00
Op1Reg = emitAnd_ri(MVT::i32, Op1Reg, Op1IsKill, Mask);
[FastISel][AArch64] Add support for variable shift. This adds the missing variable shift support for value type i8, i16, and i32. This fixes <rdar://problem/18095685>. llvm-svn: 216242
2014-08-22 07:06:07 +08:00
Op1IsKill = true;
}
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
unsigned ResultReg = fastEmitInst_rr(Opc, RC, Op0Reg, Op0IsKill, Op1Reg,
[FastISel][AArch64] Add support for variable shift. This adds the missing variable shift support for value type i8, i16, and i32. This fixes <rdar://problem/18095685>. llvm-svn: 216242
2014-08-22 07:06:07 +08:00
Op1IsKill);
if (NeedTrunc)
[FastISel][AArch64] Add target-specific lowering for logical operations. This change adds support for immediate and shift-left folding into logical operations. This fixes rdar://problem/18223183. llvm-svn: 217118
2014-09-04 09:29:18 +08:00
ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
[FastISel][AArch64] Add support for variable shift. This adds the missing variable shift support for value type i8, i16, and i32. This fixes <rdar://problem/18095685>. llvm-svn: 216242
2014-08-22 07:06:07 +08:00
return ResultReg;
}
[FastISel][AArch64] Fold Sign-/Zero-Extend into the shift immediate instruction. llvm-svn: 216510
2014-08-27 08:58:26 +08:00
unsigned AArch64FastISel::emitLSL_ri(MVT RetVT, MVT SrcVT, unsigned Op0,
bool Op0IsKill, uint64_t Shift,
bool IsZext) {
assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
"Unexpected source/return type pair.");
assert((SrcVT == MVT::i8 || SrcVT == MVT::i16 || SrcVT == MVT::i32 ||
SrcVT == MVT::i64) && "Unexpected source value type.");
assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 ||
RetVT == MVT::i64) && "Unexpected return value type.");
bool Is64Bit = (RetVT == MVT::i64);
unsigned RegSize = Is64Bit ? 64 : 32;
unsigned DstBits = RetVT.getSizeInBits();
unsigned SrcBits = SrcVT.getSizeInBits();
[FastISel][AArch64] Add support for shift-immediate. Currently the shift-immediate versions are not supported by tblgen and hopefully this can be later removed, once the required support has been added to tblgen. llvm-svn: 214345
2014-07-31 06:04:22 +08:00
[FastISel][AArch64] Fold Sign-/Zero-Extend into the shift immediate instruction. llvm-svn: 216510
2014-08-27 08:58:26 +08:00
// Don't deal with undefined shifts.
if (Shift >= DstBits)
return 0;
// For immediate shifts we can fold the zero-/sign-extension into the shift.
// {S|U}BFM Wd, Wn, #r, #s
// Wd<32+s-r,32-r> = Wn<s:0> when r > s
// %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
// %2 = shl i16 %1, 4
// Wd<32+7-28,32-28> = Wn<7:0> <- clamp s to 7
// 0b1111_1111_1111_1111__1111_1010_1010_0000 sext
// 0b0000_0000_0000_0000__0000_0101_0101_0000 sext | zext
// 0b0000_0000_0000_0000__0000_1010_1010_0000 zext
// %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
// %2 = shl i16 %1, 8
// Wd<32+7-24,32-24> = Wn<7:0>
// 0b1111_1111_1111_1111__1010_1010_0000_0000 sext
// 0b0000_0000_0000_0000__0101_0101_0000_0000 sext | zext
// 0b0000_0000_0000_0000__1010_1010_0000_0000 zext
// %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
// %2 = shl i16 %1, 12
// Wd<32+3-20,32-20> = Wn<3:0>
// 0b1111_1111_1111_1111__1010_0000_0000_0000 sext
// 0b0000_0000_0000_0000__0101_0000_0000_0000 sext | zext
// 0b0000_0000_0000_0000__1010_0000_0000_0000 zext
unsigned ImmR = RegSize - Shift;
// Limit the width to the length of the source type.
unsigned ImmS = std::min<unsigned>(SrcBits - 1, DstBits - 1 - Shift);
static const unsigned OpcTable[2][2] = {
{AArch64::SBFMWri, AArch64::SBFMXri},
{AArch64::UBFMWri, AArch64::UBFMXri}
};
unsigned Opc = OpcTable[IsZext][Is64Bit];
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
const TargetRegisterClass *RC =
[FastISel][AArch64] Fold Sign-/Zero-Extend into the shift immediate instruction. llvm-svn: 216510
2014-08-27 08:58:26 +08:00
Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
unsigned TmpReg = MRI.createVirtualRegister(RC);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
TII.get(AArch64::SUBREG_TO_REG), TmpReg)
.addImm(0)
.addReg(Op0, getKillRegState(Op0IsKill))
.addImm(AArch64::sub_32);
Op0 = TmpReg;
Op0IsKill = true;
}
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
return fastEmitInst_rii(Opc, RC, Op0, Op0IsKill, ImmR, ImmS);
[FastISel][AArch64] Add support for shift-immediate. Currently the shift-immediate versions are not supported by tblgen and hopefully this can be later removed, once the required support has been added to tblgen. llvm-svn: 214345
2014-07-31 06:04:22 +08:00
}
[FastISel][AArch64] Add support for variable shift. This adds the missing variable shift support for value type i8, i16, and i32. This fixes <rdar://problem/18095685>. llvm-svn: 216242
2014-08-22 07:06:07 +08:00
unsigned AArch64FastISel::emitLSR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
unsigned Op1Reg, bool Op1IsKill) {
unsigned Opc = 0;
bool NeedTrunc = false;
uint64_t Mask = 0;
switch (RetVT.SimpleTy) {
default: return 0;
case MVT::i8: Opc = AArch64::LSRVWr; NeedTrunc = true; Mask = 0xff; break;
case MVT::i16: Opc = AArch64::LSRVWr; NeedTrunc = true; Mask = 0xffff; break;
case MVT::i32: Opc = AArch64::LSRVWr; break;
case MVT::i64: Opc = AArch64::LSRVXr; break;
}
const TargetRegisterClass *RC =
(RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
if (NeedTrunc) {
[FastISel][AArch64] Add target-specific lowering for logical operations. This change adds support for immediate and shift-left folding into logical operations. This fixes rdar://problem/18223183. llvm-svn: 217118
2014-09-04 09:29:18 +08:00
Op0Reg = emitAnd_ri(MVT::i32, Op0Reg, Op0IsKill, Mask);
Op1Reg = emitAnd_ri(MVT::i32, Op1Reg, Op1IsKill, Mask);
[FastISel][AArch64] Add support for variable shift. This adds the missing variable shift support for value type i8, i16, and i32. This fixes <rdar://problem/18095685>. llvm-svn: 216242
2014-08-22 07:06:07 +08:00
Op0IsKill = Op1IsKill = true;
}
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
unsigned ResultReg = fastEmitInst_rr(Opc, RC, Op0Reg, Op0IsKill, Op1Reg,
[FastISel][AArch64] Add support for variable shift. This adds the missing variable shift support for value type i8, i16, and i32. This fixes <rdar://problem/18095685>. llvm-svn: 216242
2014-08-22 07:06:07 +08:00
Op1IsKill);
if (NeedTrunc)
[FastISel][AArch64] Add target-specific lowering for logical operations. This change adds support for immediate and shift-left folding into logical operations. This fixes rdar://problem/18223183. llvm-svn: 217118
2014-09-04 09:29:18 +08:00
ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
[FastISel][AArch64] Add support for variable shift. This adds the missing variable shift support for value type i8, i16, and i32. This fixes <rdar://problem/18095685>. llvm-svn: 216242
2014-08-22 07:06:07 +08:00
return ResultReg;
}
[FastISel][AArch64] Fold Sign-/Zero-Extend into the shift immediate instruction. llvm-svn: 216510
2014-08-27 08:58:26 +08:00
unsigned AArch64FastISel::emitLSR_ri(MVT RetVT, MVT SrcVT, unsigned Op0,
bool Op0IsKill, uint64_t Shift,
bool IsZExt) {
assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
"Unexpected source/return type pair.");
assert((SrcVT == MVT::i8 || SrcVT == MVT::i16 || SrcVT == MVT::i32 ||
SrcVT == MVT::i64) && "Unexpected source value type.");
assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 ||
RetVT == MVT::i64) && "Unexpected return value type.");
bool Is64Bit = (RetVT == MVT::i64);
unsigned RegSize = Is64Bit ? 64 : 32;
unsigned DstBits = RetVT.getSizeInBits();
unsigned SrcBits = SrcVT.getSizeInBits();
// Don't deal with undefined shifts.
if (Shift >= DstBits)
return 0;
// For immediate shifts we can fold the zero-/sign-extension into the shift.
// {S|U}BFM Wd, Wn, #r, #s
// Wd<s-r:0> = Wn<s:r> when r <= s
// %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
// %2 = lshr i16 %1, 4
// Wd<7-4:0> = Wn<7:4>
// 0b0000_0000_0000_0000__0000_1111_1111_1010 sext
// 0b0000_0000_0000_0000__0000_0000_0000_0101 sext | zext
// 0b0000_0000_0000_0000__0000_0000_0000_1010 zext
// %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
// %2 = lshr i16 %1, 8
// Wd<7-7,0> = Wn<7:7>
// 0b0000_0000_0000_0000__0000_0000_1111_1111 sext
// 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
// 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
// %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
// %2 = lshr i16 %1, 12
// Wd<7-7,0> = Wn<7:7> <- clamp r to 7
// 0b0000_0000_0000_0000__0000_0000_0000_1111 sext
// 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
// 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
if (Shift >= SrcBits && IsZExt)
return AArch64MaterializeInt(ConstantInt::get(*Context, APInt(RegSize, 0)),
RetVT);
// It is not possible to fold a sign-extend into the LShr instruction. In this
// case emit a sign-extend.
if (!IsZExt) {
Op0 = EmitIntExt(SrcVT, Op0, RetVT, IsZExt);
if (!Op0)
return 0;
Op0IsKill = true;
SrcVT = RetVT;
SrcBits = SrcVT.getSizeInBits();
IsZExt = true;
[FastISel][AArch64] Add support for shift-immediate. Currently the shift-immediate versions are not supported by tblgen and hopefully this can be later removed, once the required support has been added to tblgen. llvm-svn: 214345
2014-07-31 06:04:22 +08:00
}
[FastISel][AArch64] Fold Sign-/Zero-Extend into the shift immediate instruction. llvm-svn: 216510
2014-08-27 08:58:26 +08:00
unsigned ImmR = std::min<unsigned>(SrcBits - 1, Shift);
unsigned ImmS = SrcBits - 1;
static const unsigned OpcTable[2][2] = {
{AArch64::SBFMWri, AArch64::SBFMXri},
{AArch64::UBFMWri, AArch64::UBFMXri}
};
unsigned Opc = OpcTable[IsZExt][Is64Bit];
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
const TargetRegisterClass *RC =
[FastISel][AArch64] Fold Sign-/Zero-Extend into the shift immediate instruction. llvm-svn: 216510
2014-08-27 08:58:26 +08:00
Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
unsigned TmpReg = MRI.createVirtualRegister(RC);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
TII.get(AArch64::SUBREG_TO_REG), TmpReg)
.addImm(0)
.addReg(Op0, getKillRegState(Op0IsKill))
.addImm(AArch64::sub_32);
Op0 = TmpReg;
Op0IsKill = true;
}
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
return fastEmitInst_rii(Opc, RC, Op0, Op0IsKill, ImmR, ImmS);
[FastISel][AArch64] Add support for shift-immediate. Currently the shift-immediate versions are not supported by tblgen and hopefully this can be later removed, once the required support has been added to tblgen. llvm-svn: 214345
2014-07-31 06:04:22 +08:00
}
[FastISel][AArch64] Add support for variable shift. This adds the missing variable shift support for value type i8, i16, and i32. This fixes <rdar://problem/18095685>. llvm-svn: 216242
2014-08-22 07:06:07 +08:00
unsigned AArch64FastISel::emitASR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
unsigned Op1Reg, bool Op1IsKill) {
unsigned Opc = 0;
bool NeedTrunc = false;
uint64_t Mask = 0;
switch (RetVT.SimpleTy) {
default: return 0;
case MVT::i8: Opc = AArch64::ASRVWr; NeedTrunc = true; Mask = 0xff; break;
case MVT::i16: Opc = AArch64::ASRVWr; NeedTrunc = true; Mask = 0xffff; break;
case MVT::i32: Opc = AArch64::ASRVWr; break;
case MVT::i64: Opc = AArch64::ASRVXr; break;
}
const TargetRegisterClass *RC =
(RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
if (NeedTrunc) {
Op0Reg = EmitIntExt(RetVT, Op0Reg, MVT::i32, /*IsZExt=*/false);
[FastISel][AArch64] Add target-specific lowering for logical operations. This change adds support for immediate and shift-left folding into logical operations. This fixes rdar://problem/18223183. llvm-svn: 217118
2014-09-04 09:29:18 +08:00
Op1Reg = emitAnd_ri(MVT::i32, Op1Reg, Op1IsKill, Mask);
[FastISel][AArch64] Add support for variable shift. This adds the missing variable shift support for value type i8, i16, and i32. This fixes <rdar://problem/18095685>. llvm-svn: 216242
2014-08-22 07:06:07 +08:00
Op0IsKill = Op1IsKill = true;
}
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
unsigned ResultReg = fastEmitInst_rr(Opc, RC, Op0Reg, Op0IsKill, Op1Reg,
[FastISel][AArch64] Add support for variable shift. This adds the missing variable shift support for value type i8, i16, and i32. This fixes <rdar://problem/18095685>. llvm-svn: 216242
2014-08-22 07:06:07 +08:00
Op1IsKill);
if (NeedTrunc)
[FastISel][AArch64] Add target-specific lowering for logical operations. This change adds support for immediate and shift-left folding into logical operations. This fixes rdar://problem/18223183. llvm-svn: 217118
2014-09-04 09:29:18 +08:00
ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
[FastISel][AArch64] Add support for variable shift. This adds the missing variable shift support for value type i8, i16, and i32. This fixes <rdar://problem/18095685>. llvm-svn: 216242
2014-08-22 07:06:07 +08:00
return ResultReg;
}
[FastISel][AArch64] Fold Sign-/Zero-Extend into the shift immediate instruction. llvm-svn: 216510
2014-08-27 08:58:26 +08:00
unsigned AArch64FastISel::emitASR_ri(MVT RetVT, MVT SrcVT, unsigned Op0,
bool Op0IsKill, uint64_t Shift,
bool IsZExt) {
assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
"Unexpected source/return type pair.");
assert((SrcVT == MVT::i8 || SrcVT == MVT::i16 || SrcVT == MVT::i32 ||
SrcVT == MVT::i64) && "Unexpected source value type.");
assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 ||
RetVT == MVT::i64) && "Unexpected return value type.");
bool Is64Bit = (RetVT == MVT::i64);
unsigned RegSize = Is64Bit ? 64 : 32;
unsigned DstBits = RetVT.getSizeInBits();
unsigned SrcBits = SrcVT.getSizeInBits();
[FastISel][AArch64] Add support for shift-immediate. Currently the shift-immediate versions are not supported by tblgen and hopefully this can be later removed, once the required support has been added to tblgen. llvm-svn: 214345
2014-07-31 06:04:22 +08:00
[FastISel][AArch64] Fold Sign-/Zero-Extend into the shift immediate instruction. llvm-svn: 216510
2014-08-27 08:58:26 +08:00
// Don't deal with undefined shifts.
if (Shift >= DstBits)
return 0;
// For immediate shifts we can fold the zero-/sign-extension into the shift.
// {S|U}BFM Wd, Wn, #r, #s
// Wd<s-r:0> = Wn<s:r> when r <= s
// %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
// %2 = ashr i16 %1, 4
// Wd<7-4:0> = Wn<7:4>
// 0b1111_1111_1111_1111__1111_1111_1111_1010 sext
// 0b0000_0000_0000_0000__0000_0000_0000_0101 sext | zext
// 0b0000_0000_0000_0000__0000_0000_0000_1010 zext
// %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
// %2 = ashr i16 %1, 8
// Wd<7-7,0> = Wn<7:7>
// 0b1111_1111_1111_1111__1111_1111_1111_1111 sext
// 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
// 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
// %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
// %2 = ashr i16 %1, 12
// Wd<7-7,0> = Wn<7:7> <- clamp r to 7
// 0b1111_1111_1111_1111__1111_1111_1111_1111 sext
// 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
// 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
if (Shift >= SrcBits && IsZExt)
return AArch64MaterializeInt(ConstantInt::get(*Context, APInt(RegSize, 0)),
RetVT);
unsigned ImmR = std::min<unsigned>(SrcBits - 1, Shift);
unsigned ImmS = SrcBits - 1;
static const unsigned OpcTable[2][2] = {
{AArch64::SBFMWri, AArch64::SBFMXri},
{AArch64::UBFMWri, AArch64::UBFMXri}
};
unsigned Opc = OpcTable[IsZExt][Is64Bit];
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
const TargetRegisterClass *RC =
[FastISel][AArch64] Fold Sign-/Zero-Extend into the shift immediate instruction. llvm-svn: 216510
2014-08-27 08:58:26 +08:00
Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
unsigned TmpReg = MRI.createVirtualRegister(RC);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
TII.get(AArch64::SUBREG_TO_REG), TmpReg)
.addImm(0)
.addReg(Op0, getKillRegState(Op0IsKill))
.addImm(AArch64::sub_32);
Op0 = TmpReg;
Op0IsKill = true;
}
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
return fastEmitInst_rii(Opc, RC, Op0, Op0IsKill, ImmR, ImmS);
[FastISel][AArch64] Add support for shift-immediate. Currently the shift-immediate versions are not supported by tblgen and hopefully this can be later removed, once the required support has been added to tblgen. llvm-svn: 214345
2014-07-31 06:04:22 +08:00
}
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
unsigned AArch64FastISel::EmitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT,
bool isZExt) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
assert(DestVT != MVT::i1 && "ZeroExt/SignExt an i1?");
Allow AArch64FastISel to degrade graceully in the presence of an MVT::i128 Currently AArch64FastISel crashes if it tries to extend an integer into an MVT::i128. This can happen by creating 128 bit integers like so: typedef unsigned int uint128_t __attribute__((mode(TI))); typedef int sint128_t __attribute__((mode(TI))); This patch makes EmitIntExt check for their presence and then falls back to SelectionDAG. Tests included. rdar://17516686 llvm-svn: 212492
2014-07-08 05:37:51 +08:00
Make AArch64FastISel::EmitIntExt explicitly check its source and destination types This is a follow up to r212492. There should be no functional difference, but this patch makes it clear that SrcVT must be an i1/i8/16/i32 and DestVT must be an i8/i16/i32/i64. rdar://17516686 llvm-svn: 212633
2014-07-10 01:54:32 +08:00
// FastISel does not have plumbing to deal with extensions where the SrcVT or
// DestVT are odd things, so test to make sure that they are both types we can
// handle (i1/i8/i16/i32 for SrcVT and i8/i16/i32/i64 for DestVT), otherwise
// bail out to SelectionDAG.
if (((DestVT != MVT::i8) && (DestVT != MVT::i16) &&
(DestVT != MVT::i32) && (DestVT != MVT::i64)) ||
((SrcVT != MVT::i1) && (SrcVT != MVT::i8) &&
(SrcVT != MVT::i16) && (SrcVT != MVT::i32)))
Allow AArch64FastISel to degrade graceully in the presence of an MVT::i128 Currently AArch64FastISel crashes if it tries to extend an integer into an MVT::i128. This can happen by creating 128 bit integers like so: typedef unsigned int uint128_t __attribute__((mode(TI))); typedef int sint128_t __attribute__((mode(TI))); This patch makes EmitIntExt check for their presence and then falls back to SelectionDAG. Tests included. rdar://17516686 llvm-svn: 212492
2014-07-08 05:37:51 +08:00
return 0;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
unsigned Opc;
unsigned Imm = 0;
switch (SrcVT.SimpleTy) {
default:
return 0;
case MVT::i1:
return Emiti1Ext(SrcReg, DestVT, isZExt);
case MVT::i8:
if (DestVT == MVT::i64)
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
Opc = isZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
else
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
Opc = isZExt ? AArch64::UBFMWri : AArch64::SBFMWri;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
Imm = 7;
break;
case MVT::i16:
if (DestVT == MVT::i64)
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
Opc = isZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
else
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
Opc = isZExt ? AArch64::UBFMWri : AArch64::SBFMWri;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
Imm = 15;
break;
case MVT::i32:
assert(DestVT == MVT::i64 && "IntExt i32 to i32?!?");
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
Opc = isZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
Imm = 31;
break;
}
// Handle i8 and i16 as i32.
if (DestVT == MVT::i8 || DestVT == MVT::i16)
DestVT = MVT::i32;
ARM64: make sure FastISel uses a GPR64 source in 64-bit extensions. llvm-svn: 207620
2014-04-30 17:32:01 +08:00
else if (DestVT == MVT::i64) {
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
unsigned Src64 = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
ARM64: make sure FastISel uses a GPR64 source in 64-bit extensions. llvm-svn: 207620
2014-04-30 17:32:01 +08:00
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
TII.get(AArch64::SUBREG_TO_REG), Src64)
ARM64: make sure FastISel uses a GPR64 source in 64-bit extensions. llvm-svn: 207620
2014-04-30 17:32:01 +08:00
.addImm(0)
.addReg(SrcReg)
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
.addImm(AArch64::sub_32);
ARM64: make sure FastISel uses a GPR64 source in 64-bit extensions. llvm-svn: 207620
2014-04-30 17:32:01 +08:00
SrcReg = Src64;
}
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
const TargetRegisterClass *RC =
(DestVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
return fastEmitInst_rii(Opc, RC, SrcReg, /*TODO:IsKill=*/false, 0, Imm);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
bool AArch64FastISel::SelectIntExt(const Instruction *I) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// On ARM, in general, integer casts don't involve legal types; this code
// handles promotable integers. The high bits for a type smaller than
// the register size are assumed to be undefined.
Type *DestTy = I->getType();
Value *Src = I->getOperand(0);
Type *SrcTy = Src->getType();
bool isZExt = isa<ZExtInst>(I);
unsigned SrcReg = getRegForValue(Src);
if (!SrcReg)
return false;
EVT SrcEVT = TLI.getValueType(SrcTy, true);
EVT DestEVT = TLI.getValueType(DestTy, true);
if (!SrcEVT.isSimple())
return false;
if (!DestEVT.isSimple())
return false;
MVT SrcVT = SrcEVT.getSimpleVT();
MVT DestVT = DestEVT.getSimpleVT();
[FastISel][AArch64] Don't perform sign-/zero-extension for function arguments that have already been sign-/zero-extended. llvm-svn: 214844
2014-08-05 13:43:44 +08:00
unsigned ResultReg = 0;
// Check if it is an argument and if it is already zero/sign-extended.
if (const auto *Arg = dyn_cast<Argument>(Src)) {
if ((isZExt && Arg->hasZExtAttr()) || (!isZExt && Arg->hasSExtAttr())) {
[FastIsel][AArch64] Fix previous commit r214844 (Don't perform sign-/zero-extension for function arguments that have already been sign-/zero-extended.) The original code would fail for unsupported value types like i1, i8, and i16. This fix changes the code to only create a sub-register copy for i64 value types and all other types (i1/i8/i16/i32) just use the source register without any modifications. getRegClassFor() is now guarded by the i64 value type check, that guarantees that we always request a register for a valid value type. llvm-svn: 214848
2014-08-05 15:31:30 +08:00
if (DestVT == MVT::i64) {
ResultReg = createResultReg(TLI.getRegClassFor(DestVT));
[FastISel][AArch64] Don't perform sign-/zero-extension for function arguments that have already been sign-/zero-extended. llvm-svn: 214844
2014-08-05 13:43:44 +08:00
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
TII.get(AArch64::SUBREG_TO_REG), ResultReg)
.addImm(0)
.addReg(SrcReg)
.addImm(AArch64::sub_32);
[FastIsel][AArch64] Fix previous commit r214844 (Don't perform sign-/zero-extension for function arguments that have already been sign-/zero-extended.) The original code would fail for unsupported value types like i1, i8, and i16. This fix changes the code to only create a sub-register copy for i64 value types and all other types (i1/i8/i16/i32) just use the source register without any modifications. getRegClassFor() is now guarded by the i64 value type check, that guarantees that we always request a register for a valid value type. llvm-svn: 214848
2014-08-05 15:31:30 +08:00
} else
ResultReg = SrcReg;
[FastISel][AArch64] Don't perform sign-/zero-extension for function arguments that have already been sign-/zero-extended. llvm-svn: 214844
2014-08-05 13:43:44 +08:00
}
}
if (!ResultReg)
ResultReg = EmitIntExt(SrcVT, SrcReg, DestVT, isZExt);
if (!ResultReg)
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return false;
[FastISel][AArch64] Don't perform sign-/zero-extension for function arguments that have already been sign-/zero-extended. llvm-svn: 214844
2014-08-05 13:43:44 +08:00
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
updateValueMap(I, ResultReg);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return true;
}
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
bool AArch64FastISel::SelectRem(const Instruction *I, unsigned ISDOpcode) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
EVT DestEVT = TLI.getValueType(I->getType(), true);
if (!DestEVT.isSimple())
return false;
MVT DestVT = DestEVT.getSimpleVT();
if (DestVT != MVT::i64 && DestVT != MVT::i32)
return false;
unsigned DivOpc;
bool is64bit = (DestVT == MVT::i64);
switch (ISDOpcode) {
default:
return false;
case ISD::SREM:
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
DivOpc = is64bit ? AArch64::SDIVXr : AArch64::SDIVWr;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
break;
case ISD::UREM:
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
DivOpc = is64bit ? AArch64::UDIVXr : AArch64::UDIVWr;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
break;
}
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
unsigned MSubOpc = is64bit ? AArch64::MSUBXrrr : AArch64::MSUBWrrr;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
unsigned Src0Reg = getRegForValue(I->getOperand(0));
if (!Src0Reg)
return false;
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
bool Src0IsKill = hasTrivialKill(I->getOperand(0));
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
unsigned Src1Reg = getRegForValue(I->getOperand(1));
if (!Src1Reg)
return false;
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
bool Src1IsKill = hasTrivialKill(I->getOperand(1));
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
const TargetRegisterClass *RC =
(DestVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
unsigned QuotReg = fastEmitInst_rr(DivOpc, RC, Src0Reg, /*IsKill=*/false,
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
Src1Reg, /*IsKill=*/false);
assert(QuotReg && "Unexpected DIV instruction emission failure.");
Replacing a non-ASCII character in a comment with an ASCII character. Fixes a C4819 warning in MSVC. llvm-svn: 206403
2014-04-17 01:09:20 +08:00
// The remainder is computed as numerator - (quotient * denominator) using the
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// MSUB instruction.
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
unsigned ResultReg = fastEmitInst_rrr(MSubOpc, RC, QuotReg, /*IsKill=*/true,
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
Src1Reg, Src1IsKill, Src0Reg,
Src0IsKill);
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
updateValueMap(I, ResultReg);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return true;
}
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
bool AArch64FastISel::SelectMul(const Instruction *I) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
EVT SrcEVT = TLI.getValueType(I->getOperand(0)->getType(), true);
if (!SrcEVT.isSimple())
return false;
MVT SrcVT = SrcEVT.getSimpleVT();
// Must be simple value type. Don't handle vectors.
if (SrcVT != MVT::i64 && SrcVT != MVT::i32 && SrcVT != MVT::i16 &&
SrcVT != MVT::i8)
return false;
unsigned Src0Reg = getRegForValue(I->getOperand(0));
if (!Src0Reg)
return false;
[FastISel][AArch64] Create helper functions to create the various multiplies on AArch64. llvm-svn: 214346
2014-07-31 06:04:25 +08:00
bool Src0IsKill = hasTrivialKill(I->getOperand(0));
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
unsigned Src1Reg = getRegForValue(I->getOperand(1));
if (!Src1Reg)
return false;
[FastISel][AArch64] Create helper functions to create the various multiplies on AArch64. llvm-svn: 214346
2014-07-31 06:04:25 +08:00
bool Src1IsKill = hasTrivialKill(I->getOperand(1));
unsigned ResultReg =
Emit_MUL_rr(SrcVT, Src0Reg, Src0IsKill, Src1Reg, Src1IsKill);
if (!ResultReg)
return false;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
updateValueMap(I, ResultReg);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return true;
}
[FastISel][AArch64] Add support for variable shift. This adds the missing variable shift support for value type i8, i16, and i32. This fixes <rdar://problem/18095685>. llvm-svn: 216242
2014-08-22 07:06:07 +08:00
bool AArch64FastISel::SelectShift(const Instruction *I) {
[FastISel][AArch64] Fold Sign-/Zero-Extend into the shift immediate instruction. llvm-svn: 216510
2014-08-27 08:58:26 +08:00
MVT RetVT;
[FastISel][AArch64] Use a new helper function to determine if a value type is supported. NFCI. FastISel for AArch64 supports more value types than are actually legal. Use a dedicated helper function to reflect this. It is very similar to the isLoadStoreTypeLegal function, with the exception that vector types are not supported yet. llvm-svn: 216984
2014-09-03 06:33:53 +08:00
if (!isTypeSupported(I->getType(), RetVT))
[FastISel][AArch64] Add support for shift-immediate. Currently the shift-immediate versions are not supported by tblgen and hopefully this can be later removed, once the required support has been added to tblgen. llvm-svn: 214345
2014-07-31 06:04:22 +08:00
return false;
[FastISel][AArch64] Add support for variable shift. This adds the missing variable shift support for value type i8, i16, and i32. This fixes <rdar://problem/18095685>. llvm-svn: 216242
2014-08-22 07:06:07 +08:00
if (const auto *C = dyn_cast<ConstantInt>(I->getOperand(1))) {
unsigned ResultReg = 0;
uint64_t ShiftVal = C->getZExtValue();
[FastISel][AArch64] Fold Sign-/Zero-Extend into the shift immediate instruction. llvm-svn: 216510
2014-08-27 08:58:26 +08:00
MVT SrcVT = RetVT;
bool IsZExt = (I->getOpcode() == Instruction::AShr) ? false : true;
[FastISel][AArch64] Don't fold instructions that are not in the same basic block. This fix checks first if the instruction to be folded (e.g. sign-/zero-extend, or shift) is in the same machine basic block as the instruction we are folding into. Not doing so can result in incorrect code, because the value might not be live-out of the basic block, where the value is defined. This fixes rdar://problem/18169495. llvm-svn: 216700
2014-08-29 08:19:21 +08:00
const Value *Op0 = I->getOperand(0);
[FastISel][AArch64] Fold Sign-/Zero-Extend into the shift immediate instruction. llvm-svn: 216510
2014-08-27 08:58:26 +08:00
if (const auto *ZExt = dyn_cast<ZExtInst>(Op0)) {
MVT TmpVT;
[FastISel][AArch64] Use a new helper function to determine if a value type is supported. NFCI. FastISel for AArch64 supports more value types than are actually legal. Use a dedicated helper function to reflect this. It is very similar to the isLoadStoreTypeLegal function, with the exception that vector types are not supported yet. llvm-svn: 216984
2014-09-03 06:33:53 +08:00
if (isValueAvailable(ZExt) && isTypeSupported(ZExt->getSrcTy(), TmpVT)) {
[FastISel][AArch64] Fold Sign-/Zero-Extend into the shift immediate instruction. llvm-svn: 216510
2014-08-27 08:58:26 +08:00
SrcVT = TmpVT;
IsZExt = true;
Op0 = ZExt->getOperand(0);
}
} else if (const auto *SExt = dyn_cast<SExtInst>(Op0)) {
MVT TmpVT;
[FastISel][AArch64] Use a new helper function to determine if a value type is supported. NFCI. FastISel for AArch64 supports more value types than are actually legal. Use a dedicated helper function to reflect this. It is very similar to the isLoadStoreTypeLegal function, with the exception that vector types are not supported yet. llvm-svn: 216984
2014-09-03 06:33:53 +08:00
if (isValueAvailable(SExt) && isTypeSupported(SExt->getSrcTy(), TmpVT)) {
[FastISel][AArch64] Fold Sign-/Zero-Extend into the shift immediate instruction. llvm-svn: 216510
2014-08-27 08:58:26 +08:00
SrcVT = TmpVT;
IsZExt = false;
Op0 = SExt->getOperand(0);
}
}
unsigned Op0Reg = getRegForValue(Op0);
if (!Op0Reg)
return false;
bool Op0IsKill = hasTrivialKill(Op0);
[FastISel][AArch64] Add support for variable shift. This adds the missing variable shift support for value type i8, i16, and i32. This fixes <rdar://problem/18095685>. llvm-svn: 216242
2014-08-22 07:06:07 +08:00
switch (I->getOpcode()) {
default: llvm_unreachable("Unexpected instruction.");
case Instruction::Shl:
[FastISel][AArch64] Fold Sign-/Zero-Extend into the shift immediate instruction. llvm-svn: 216510
2014-08-27 08:58:26 +08:00
ResultReg = emitLSL_ri(RetVT, SrcVT, Op0Reg, Op0IsKill, ShiftVal, IsZExt);
[FastISel][AArch64] Add support for variable shift. This adds the missing variable shift support for value type i8, i16, and i32. This fixes <rdar://problem/18095685>. llvm-svn: 216242
2014-08-22 07:06:07 +08:00
break;
case Instruction::AShr:
[FastISel][AArch64] Fold Sign-/Zero-Extend into the shift immediate instruction. llvm-svn: 216510
2014-08-27 08:58:26 +08:00
ResultReg = emitASR_ri(RetVT, SrcVT, Op0Reg, Op0IsKill, ShiftVal, IsZExt);
[FastISel][AArch64] Add support for variable shift. This adds the missing variable shift support for value type i8, i16, and i32. This fixes <rdar://problem/18095685>. llvm-svn: 216242
2014-08-22 07:06:07 +08:00
break;
case Instruction::LShr:
[FastISel][AArch64] Fold Sign-/Zero-Extend into the shift immediate instruction. llvm-svn: 216510
2014-08-27 08:58:26 +08:00
ResultReg = emitLSR_ri(RetVT, SrcVT, Op0Reg, Op0IsKill, ShiftVal, IsZExt);
[FastISel][AArch64] Add support for variable shift. This adds the missing variable shift support for value type i8, i16, and i32. This fixes <rdar://problem/18095685>. llvm-svn: 216242
2014-08-22 07:06:07 +08:00
break;
}
if (!ResultReg)
return false;
[FastISel][AArch64] Add support for shift-immediate. Currently the shift-immediate versions are not supported by tblgen and hopefully this can be later removed, once the required support has been added to tblgen. llvm-svn: 214345
2014-07-31 06:04:22 +08:00
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
updateValueMap(I, ResultReg);
[FastISel][AArch64] Add support for variable shift. This adds the missing variable shift support for value type i8, i16, and i32. This fixes <rdar://problem/18095685>. llvm-svn: 216242
2014-08-22 07:06:07 +08:00
return true;
}
[FastISel][AArch64] Fold Sign-/Zero-Extend into the shift immediate instruction. llvm-svn: 216510
2014-08-27 08:58:26 +08:00
unsigned Op0Reg = getRegForValue(I->getOperand(0));
if (!Op0Reg)
return false;
bool Op0IsKill = hasTrivialKill(I->getOperand(0));
[FastISel][AArch64] Add support for variable shift. This adds the missing variable shift support for value type i8, i16, and i32. This fixes <rdar://problem/18095685>. llvm-svn: 216242
2014-08-22 07:06:07 +08:00
unsigned Op1Reg = getRegForValue(I->getOperand(1));
if (!Op1Reg)
return false;
bool Op1IsKill = hasTrivialKill(I->getOperand(1));
unsigned ResultReg = 0;
switch (I->getOpcode()) {
default: llvm_unreachable("Unexpected instruction.");
case Instruction::Shl:
ResultReg = emitLSL_rr(RetVT, Op0Reg, Op0IsKill, Op1Reg, Op1IsKill);
break;
case Instruction::AShr:
ResultReg = emitASR_rr(RetVT, Op0Reg, Op0IsKill, Op1Reg, Op1IsKill);
break;
case Instruction::LShr:
ResultReg = emitLSR_rr(RetVT, Op0Reg, Op0IsKill, Op1Reg, Op1IsKill);
break;
[FastISel][AArch64] Add support for shift-immediate. Currently the shift-immediate versions are not supported by tblgen and hopefully this can be later removed, once the required support has been added to tblgen. llvm-svn: 214345
2014-07-31 06:04:22 +08:00
}
if (!ResultReg)
return false;
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
updateValueMap(I, ResultReg);
[FastISel][AArch64] Add support for shift-immediate. Currently the shift-immediate versions are not supported by tblgen and hopefully this can be later removed, once the required support has been added to tblgen. llvm-svn: 214345
2014-07-31 06:04:22 +08:00
return true;
}
[FastISel][AArch64] Add basic bitcast support for conversion between float and int. Fixes <rdar://problem/17867078>. llvm-svn: 214389
2014-07-31 14:25:37 +08:00
bool AArch64FastISel::SelectBitCast(const Instruction *I) {
MVT RetVT, SrcVT;
if (!isTypeLegal(I->getOperand(0)->getType(), SrcVT))
return false;
if (!isTypeLegal(I->getType(), RetVT))
return false;
unsigned Opc;
if (RetVT == MVT::f32 && SrcVT == MVT::i32)
Opc = AArch64::FMOVWSr;
else if (RetVT == MVT::f64 && SrcVT == MVT::i64)
Opc = AArch64::FMOVXDr;
else if (RetVT == MVT::i32 && SrcVT == MVT::f32)
Opc = AArch64::FMOVSWr;
else if (RetVT == MVT::i64 && SrcVT == MVT::f64)
Opc = AArch64::FMOVDXr;
else
return false;
[FastISel][AArch64] Use the correct register class to make the MI verifier happy. This is mostly achieved by providing the correct register class manually, because getRegClassFor always returns the GPR*AllRegClass for MVT::i32 and MVT::i64. Also cleanup the code to use the FastEmitInst_* method whenever possible. This makes sure that the operands' register class is properly constrained. For all the remaining cases this adds the missing constrainOperandRegClass calls for each operand. llvm-svn: 216225
2014-08-22 04:57:57 +08:00
const TargetRegisterClass *RC = nullptr;
switch (RetVT.SimpleTy) {
default: llvm_unreachable("Unexpected value type.");
case MVT::i32: RC = &AArch64::GPR32RegClass; break;
case MVT::i64: RC = &AArch64::GPR64RegClass; break;
case MVT::f32: RC = &AArch64::FPR32RegClass; break;
case MVT::f64: RC = &AArch64::FPR64RegClass; break;
}
[FastISel][AArch64] Add basic bitcast support for conversion between float and int. Fixes <rdar://problem/17867078>. llvm-svn: 214389
2014-07-31 14:25:37 +08:00
unsigned Op0Reg = getRegForValue(I->getOperand(0));
if (!Op0Reg)
return false;
bool Op0IsKill = hasTrivialKill(I->getOperand(0));
[FastISel][tblgen] Rename tblgen generated FastISel functions. NFC. This is the final round of renaming. This changes tblgen to emit lower-case function names for FastEmitInst_* and FastEmit_*, and updates all its uses in the source code. Reviewed by Eric llvm-svn: 217075
2014-09-04 04:56:59 +08:00
unsigned ResultReg = fastEmitInst_r(Opc, RC, Op0Reg, Op0IsKill);
[FastISel][AArch64] Add basic bitcast support for conversion between float and int. Fixes <rdar://problem/17867078>. llvm-svn: 214389
2014-07-31 14:25:37 +08:00
if (!ResultReg)
return false;
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
updateValueMap(I, ResultReg);
[FastISel][AArch64] Add basic bitcast support for conversion between float and int. Fixes <rdar://problem/17867078>. llvm-svn: 214389
2014-07-31 14:25:37 +08:00
return true;
}
[FastISel] Rename public visible FastISel functions. NFC. This commit renames the following public FastISel functions: LowerArguments -> lowerArguments SelectInstruction -> selectInstruction TargetSelectInstruction -> fastSelectInstruction FastLowerArguments -> fastLowerArguments FastLowerCall -> fastLowerCall FastLowerIntrinsicCall -> fastLowerIntrinsicCall FastEmitZExtFromI1 -> fastEmitZExtFromI1 FastEmitBranch -> fastEmitBranch UpdateValueMap -> updateValueMap TargetMaterializeConstant -> fastMaterializeConstant TargetMaterializeAlloca -> fastMaterializeAlloca TargetMaterializeFloatZero -> fastMaterializeFloatZero LowerCallTo -> lowerCallTo Reviewed by Eric llvm-svn: 217074
2014-09-04 04:56:52 +08:00
bool AArch64FastISel::fastSelectInstruction(const Instruction *I) {
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
switch (I->getOpcode()) {
default:
[FastISel][AArch64] Cleanup and simplify 'fastSelectInstruction'. NFC. llvm-svn: 217119
2014-09-04 09:29:21 +08:00
break;
[FastISel][AArch64] Move over to target-dependent instruction selection only. This change moves FastISel for AArch64 to target-dependent instruction selection only. This change replicates the existing target-independent behavior, therefore there are no changes to the unit tests or new tests. Future changes will take advantage of this change and update functionality and unit tests. llvm-svn: 216955
2014-09-03 05:32:54 +08:00
case Instruction::Add:
[FastISel][AArch64] Add target-dependent instruction selection for Add/Sub. There is already target-dependent instruction selection support for Adds/Subs to support compares and the intrinsics with overflow check. This takes advantage of the existing infrastructure to also support Add/Sub, which allows the folding of immediates, sign-/zero-extends, and shifts. This fixes rdar://problem/18207316. llvm-svn: 217007
2014-09-03 09:38:36 +08:00
case Instruction::Sub:
[FastISel][AArch64] Cleanup and simplify 'fastSelectInstruction'. NFC. llvm-svn: 217119
2014-09-04 09:29:21 +08:00
if (selectAddSub(I))
return true;
break;
[FastISel][AArch64] Move over to target-dependent instruction selection only. This change moves FastISel for AArch64 to target-dependent instruction selection only. This change replicates the existing target-independent behavior, therefore there are no changes to the unit tests or new tests. Future changes will take advantage of this change and update functionality and unit tests. llvm-svn: 216955
2014-09-03 05:32:54 +08:00
case Instruction::Mul:
[FastISel] Some long overdue spring cleaning of FastISel. Things got a little bit messy over the years and it is time for a little bit spring cleaning. This first commit is focused on the FastISel base class itself. It doxyfies all comments, C++11fies the code where it makes sense, renames internal methods to adhere to the coding standard, and clang-formats the files. Reviewed by Eric llvm-svn: 217060
2014-09-04 02:46:45 +08:00
if (!selectBinaryOp(I, ISD::MUL))
[FastISel][AArch64] Move over to target-dependent instruction selection only. This change moves FastISel for AArch64 to target-dependent instruction selection only. This change replicates the existing target-independent behavior, therefore there are no changes to the unit tests or new tests. Future changes will take advantage of this change and update functionality and unit tests. llvm-svn: 216955
2014-09-03 05:32:54 +08:00
return SelectMul(I);
return true;
case Instruction::SRem:
[FastISel] Some long overdue spring cleaning of FastISel. Things got a little bit messy over the years and it is time for a little bit spring cleaning. This first commit is focused on the FastISel base class itself. It doxyfies all comments, C++11fies the code where it makes sense, renames internal methods to adhere to the coding standard, and clang-formats the files. Reviewed by Eric llvm-svn: 217060
2014-09-04 02:46:45 +08:00
if (!selectBinaryOp(I, ISD::SREM))
[FastISel][AArch64] Move over to target-dependent instruction selection only. This change moves FastISel for AArch64 to target-dependent instruction selection only. This change replicates the existing target-independent behavior, therefore there are no changes to the unit tests or new tests. Future changes will take advantage of this change and update functionality and unit tests. llvm-svn: 216955
2014-09-03 05:32:54 +08:00
return SelectRem(I, ISD::SREM);
return true;
case Instruction::URem:
[FastISel] Some long overdue spring cleaning of FastISel. Things got a little bit messy over the years and it is time for a little bit spring cleaning. This first commit is focused on the FastISel base class itself. It doxyfies all comments, C++11fies the code where it makes sense, renames internal methods to adhere to the coding standard, and clang-formats the files. Reviewed by Eric llvm-svn: 217060
2014-09-04 02:46:45 +08:00
if (!selectBinaryOp(I, ISD::UREM))
[FastISel][AArch64] Move over to target-dependent instruction selection only. This change moves FastISel for AArch64 to target-dependent instruction selection only. This change replicates the existing target-independent behavior, therefore there are no changes to the unit tests or new tests. Future changes will take advantage of this change and update functionality and unit tests. llvm-svn: 216955
2014-09-03 05:32:54 +08:00
return SelectRem(I, ISD::UREM);
return true;
case Instruction::Shl:
case Instruction::LShr:
case Instruction::AShr:
[FastISel][AArch64] Cleanup and simplify 'fastSelectInstruction'. NFC. llvm-svn: 217119
2014-09-04 09:29:21 +08:00
if (SelectShift(I))
return true;
break;
[FastISel][AArch64] Move over to target-dependent instruction selection only. This change moves FastISel for AArch64 to target-dependent instruction selection only. This change replicates the existing target-independent behavior, therefore there are no changes to the unit tests or new tests. Future changes will take advantage of this change and update functionality and unit tests. llvm-svn: 216955
2014-09-03 05:32:54 +08:00
case Instruction::And:
[FastISel][AArch64] Add support for non-native types for logical ops. Extend the logical ops selection to also support non-native types such as i1, i8, and i16. Fixes rdar://problem/18330589. llvm-svn: 217732
2014-09-14 07:46:28 +08:00
if (selectLogicalOp(I, ISD::AND))
return true;
break;
[FastISel][AArch64] Move over to target-dependent instruction selection only. This change moves FastISel for AArch64 to target-dependent instruction selection only. This change replicates the existing target-independent behavior, therefore there are no changes to the unit tests or new tests. Future changes will take advantage of this change and update functionality and unit tests. llvm-svn: 216955
2014-09-03 05:32:54 +08:00
case Instruction::Or:
[FastISel][AArch64] Add support for non-native types for logical ops. Extend the logical ops selection to also support non-native types such as i1, i8, and i16. Fixes rdar://problem/18330589. llvm-svn: 217732
2014-09-14 07:46:28 +08:00
if (selectLogicalOp(I, ISD::OR))
return true;
break;
[FastISel][AArch64] Move over to target-dependent instruction selection only. This change moves FastISel for AArch64 to target-dependent instruction selection only. This change replicates the existing target-independent behavior, therefore there are no changes to the unit tests or new tests. Future changes will take advantage of this change and update functionality and unit tests. llvm-svn: 216955
2014-09-03 05:32:54 +08:00
case Instruction::Xor:
[FastISel][AArch64] Add support for non-native types for logical ops. Extend the logical ops selection to also support non-native types such as i1, i8, and i16. Fixes rdar://problem/18330589. llvm-svn: 217732
2014-09-14 07:46:28 +08:00
if (selectLogicalOp(I, ISD::XOR))
[FastISel][AArch64] Cleanup and simplify 'fastSelectInstruction'. NFC. llvm-svn: 217119
2014-09-04 09:29:21 +08:00
return true;
break;
[FastISel][AArch64] Move unconditional branch handling into 'SelectBranch'. NFC. llvm-svn: 217054
2014-09-04 01:58:10 +08:00
case Instruction::Br:
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
return SelectBranch(I);
case Instruction::IndirectBr:
return SelectIndirectBr(I);
[FastISel][AArch64] Move over to target-dependent instruction selection only. This change moves FastISel for AArch64 to target-dependent instruction selection only. This change replicates the existing target-independent behavior, therefore there are no changes to the unit tests or new tests. Future changes will take advantage of this change and update functionality and unit tests. llvm-svn: 216955
2014-09-03 05:32:54 +08:00
case Instruction::BitCast:
[FastISel] Some long overdue spring cleaning of FastISel. Things got a little bit messy over the years and it is time for a little bit spring cleaning. This first commit is focused on the FastISel base class itself. It doxyfies all comments, C++11fies the code where it makes sense, renames internal methods to adhere to the coding standard, and clang-formats the files. Reviewed by Eric llvm-svn: 217060
2014-09-04 02:46:45 +08:00
if (!FastISel::selectBitCast(I))
[FastISel][AArch64] Move over to target-dependent instruction selection only. This change moves FastISel for AArch64 to target-dependent instruction selection only. This change replicates the existing target-independent behavior, therefore there are no changes to the unit tests or new tests. Future changes will take advantage of this change and update functionality and unit tests. llvm-svn: 216955
2014-09-03 05:32:54 +08:00
return SelectBitCast(I);
return true;
case Instruction::FPToSI:
[FastISel] Some long overdue spring cleaning of FastISel. Things got a little bit messy over the years and it is time for a little bit spring cleaning. This first commit is focused on the FastISel base class itself. It doxyfies all comments, C++11fies the code where it makes sense, renames internal methods to adhere to the coding standard, and clang-formats the files. Reviewed by Eric llvm-svn: 217060
2014-09-04 02:46:45 +08:00
if (!selectCast(I, ISD::FP_TO_SINT))
[FastISel][AArch64] Move over to target-dependent instruction selection only. This change moves FastISel for AArch64 to target-dependent instruction selection only. This change replicates the existing target-independent behavior, therefore there are no changes to the unit tests or new tests. Future changes will take advantage of this change and update functionality and unit tests. llvm-svn: 216955
2014-09-03 05:32:54 +08:00
return SelectFPToInt(I, /*Signed=*/true);
return true;
case Instruction::FPToUI:
return SelectFPToInt(I, /*Signed=*/false);
case Instruction::ZExt:
[FastISel] Some long overdue spring cleaning of FastISel. Things got a little bit messy over the years and it is time for a little bit spring cleaning. This first commit is focused on the FastISel base class itself. It doxyfies all comments, C++11fies the code where it makes sense, renames internal methods to adhere to the coding standard, and clang-formats the files. Reviewed by Eric llvm-svn: 217060
2014-09-04 02:46:45 +08:00
if (!selectCast(I, ISD::ZERO_EXTEND))
[FastISel][AArch64] Move over to target-dependent instruction selection only. This change moves FastISel for AArch64 to target-dependent instruction selection only. This change replicates the existing target-independent behavior, therefore there are no changes to the unit tests or new tests. Future changes will take advantage of this change and update functionality and unit tests. llvm-svn: 216955
2014-09-03 05:32:54 +08:00
return SelectIntExt(I);
return true;
case Instruction::SExt:
[FastISel] Some long overdue spring cleaning of FastISel. Things got a little bit messy over the years and it is time for a little bit spring cleaning. This first commit is focused on the FastISel base class itself. It doxyfies all comments, C++11fies the code where it makes sense, renames internal methods to adhere to the coding standard, and clang-formats the files. Reviewed by Eric llvm-svn: 217060
2014-09-04 02:46:45 +08:00
if (!selectCast(I, ISD::SIGN_EXTEND))
[FastISel][AArch64] Move over to target-dependent instruction selection only. This change moves FastISel for AArch64 to target-dependent instruction selection only. This change replicates the existing target-independent behavior, therefore there are no changes to the unit tests or new tests. Future changes will take advantage of this change and update functionality and unit tests. llvm-svn: 216955
2014-09-03 05:32:54 +08:00
return SelectIntExt(I);
return true;
case Instruction::Trunc:
[FastISel] Some long overdue spring cleaning of FastISel. Things got a little bit messy over the years and it is time for a little bit spring cleaning. This first commit is focused on the FastISel base class itself. It doxyfies all comments, C++11fies the code where it makes sense, renames internal methods to adhere to the coding standard, and clang-formats the files. Reviewed by Eric llvm-svn: 217060
2014-09-04 02:46:45 +08:00
if (!selectCast(I, ISD::TRUNCATE))
[FastISel][AArch64] Move over to target-dependent instruction selection only. This change moves FastISel for AArch64 to target-dependent instruction selection only. This change replicates the existing target-independent behavior, therefore there are no changes to the unit tests or new tests. Future changes will take advantage of this change and update functionality and unit tests. llvm-svn: 216955
2014-09-03 05:32:54 +08:00
return SelectTrunc(I);
return true;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
case Instruction::FPExt:
return SelectFPExt(I);
case Instruction::FPTrunc:
return SelectFPTrunc(I);
case Instruction::SIToFP:
[FastISel] Some long overdue spring cleaning of FastISel. Things got a little bit messy over the years and it is time for a little bit spring cleaning. This first commit is focused on the FastISel base class itself. It doxyfies all comments, C++11fies the code where it makes sense, renames internal methods to adhere to the coding standard, and clang-formats the files. Reviewed by Eric llvm-svn: 217060
2014-09-04 02:46:45 +08:00
if (!selectCast(I, ISD::SINT_TO_FP))
[FastISel][AArch64] Move over to target-dependent instruction selection only. This change moves FastISel for AArch64 to target-dependent instruction selection only. This change replicates the existing target-independent behavior, therefore there are no changes to the unit tests or new tests. Future changes will take advantage of this change and update functionality and unit tests. llvm-svn: 216955
2014-09-03 05:32:54 +08:00
return SelectIntToFP(I, /*Signed=*/true);
return true;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
case Instruction::UIToFP:
return SelectIntToFP(I, /*Signed=*/false);
[FastISel][AArch64] Move over to target-dependent instruction selection only. This change moves FastISel for AArch64 to target-dependent instruction selection only. This change replicates the existing target-independent behavior, therefore there are no changes to the unit tests or new tests. Future changes will take advantage of this change and update functionality and unit tests. llvm-svn: 216955
2014-09-03 05:32:54 +08:00
case Instruction::Load:
return SelectLoad(I);
case Instruction::Store:
return SelectStore(I);
case Instruction::FCmp:
case Instruction::ICmp:
return SelectCmp(I);
case Instruction::Select:
return SelectSelect(I);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
case Instruction::Ret:
return SelectRet(I);
}
[FastISel][AArch64] Move over to target-dependent instruction selection only. This change moves FastISel for AArch64 to target-dependent instruction selection only. This change replicates the existing target-independent behavior, therefore there are no changes to the unit tests or new tests. Future changes will take advantage of this change and update functionality and unit tests. llvm-svn: 216955
2014-09-03 05:32:54 +08:00
[FastISel][AArch64] Cleanup and simplify 'fastSelectInstruction'. NFC. llvm-svn: 217119
2014-09-04 09:29:21 +08:00
// fall-back to target-independent instruction selection.
return selectOperator(I, I->getOpcode());
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
// Silence warnings.
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
(void)&CC_AArch64_DarwinPCS_VarArg;
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
namespace llvm {
AArch64/ARM64: move ARM64 into AArch64's place This commit starts with a "git mv ARM64 AArch64" and continues out from there, renaming the C++ classes, intrinsics, and other target-local objects for consistency. "ARM64" test directories are also moved, and tests that began their life in ARM64 use an arm64 triple, those from AArch64 use an aarch64 triple. Both should be equivalent though. This finishes the AArch64 merge, and everyone should feel free to continue committing as normal now. llvm-svn: 209577
2014-05-24 20:50:23 +08:00
llvm::FastISel *AArch64::createFastISel(FunctionLoweringInfo &funcInfo,
const TargetLibraryInfo *libInfo) {
return new AArch64FastISel(funcInfo, libInfo);
ARM64: initial backend import This adds a second implementation of the AArch64 architecture to LLVM, accessible in parallel via the "arm64" triple. The plan over the coming weeks & months is to merge the two into a single backend, during which time thorough code review should naturally occur. Everything will be easier with the target in-tree though, hence this commit. llvm-svn: 205090
2014-03-29 18:18:08 +08:00
}
}