The TBB and TBH instructions in Thumb-2 allow jump tables to be compressed into sequences of bytes or shorts respectively. These instructions do not exist in Thumb-1, however it is possible to synthesize them out of a sequence of other instructions.
It turns out this sequence is so short that it's almost never a lose for performance and is ALWAYS a significant win for code size.
TBB example:
Before: lsls r0, r0, #2 After: add r0, pc
adr r1, .LJTI0_0 ldrb r0, [r0, #6]
ldr r0, [r0, r1] lsls r0, r0, #1
mov pc, r0 add pc, r0
=> No change in prologue code size or dynamic instruction count. Jump table shrunk by a factor of 4.
The only case that can increase dynamic instruction count is the TBH case:
Before: lsls r0, r4, #2 After: lsls r4, r4, #1
adr r1, .LJTI0_0 add r4, pc
ldr r0, [r0, r1] ldrh r4, [r4, #6]
mov pc, r0 lsls r4, r4, #1
add pc, r4
=> 1 more instruction in prologue. Jump table shrunk by a factor of 2.
So there is an argument that this should be disabled when optimizing for performance (and a TBH needs to be generated). I'm not so sure about that in practice, because on small cores with Thumb-1 performance is often tied to code size. But I'm willing to turn it off when optimizing for performance if people want (also note that TBHs are fairly rare in practice!)
llvm-svn: 284580
.addReg(0));
return;
}
+ case ARM::tTBB_JT:
+ case ARM::tTBH_JT: {
+
+ bool Is8Bit = MI->getOpcode() == ARM::tTBB_JT;
+ unsigned Base = MI->getOperand(0).getReg();
+ unsigned Idx = MI->getOperand(1).getReg();
+ assert(MI->getOperand(1).isKill() && "We need the index register as scratch!");
+
+ // Multiply up idx if necessary.
+ if (!Is8Bit)
+ EmitToStreamer(*OutStreamer, MCInstBuilder(ARM::tLSLri)
+ .addReg(Idx)
+ .addReg(ARM::CPSR)
+ .addReg(Idx)
+ .addImm(1)
+ // Add predicate operands.
+ .addImm(ARMCC::AL)
+ .addReg(0));
+
+ if (Base == ARM::PC) {
+ // TBB [base, idx] =
+ // ADDS idx, idx, base
+ // LDRB idx, [idx, #4] ; or LDRH if TBH
+ // LSLS idx, #1
+ // ADDS pc, pc, idx
+
+ EmitToStreamer(*OutStreamer, MCInstBuilder(ARM::tADDhirr)
+ .addReg(Idx)
+ .addReg(Idx)
+ .addReg(Base)
+ // Add predicate operands.
+ .addImm(ARMCC::AL)
+ .addReg(0));
+
+ unsigned Opc = Is8Bit ? ARM::tLDRBi : ARM::tLDRHi;
+ EmitToStreamer(*OutStreamer, MCInstBuilder(Opc)
+ .addReg(Idx)
+ .addReg(Idx)
+ .addImm(Is8Bit ? 4 : 2)
+ // Add predicate operands.
+ .addImm(ARMCC::AL)
+ .addReg(0));
+ } else {
+ // TBB [base, idx] =
+ // LDRB idx, [base, idx] ; or LDRH if TBH
+ // LSLS idx, #1
+ // ADDS pc, pc, idx
+
+ unsigned Opc = Is8Bit ? ARM::tLDRBr : ARM::tLDRHr;
+ EmitToStreamer(*OutStreamer, MCInstBuilder(Opc)
+ .addReg(Idx)
+ .addReg(Base)
+ .addReg(Idx)
+ // Add predicate operands.
+ .addImm(ARMCC::AL)
+ .addReg(0));
+ }
+
+ EmitToStreamer(*OutStreamer, MCInstBuilder(ARM::tLSLri)
+ .addReg(Idx)
+ .addReg(ARM::CPSR)
+ .addReg(Idx)
+ .addImm(1)
+ // Add predicate operands.
+ .addImm(ARMCC::AL)
+ .addReg(0));
+
+ OutStreamer->EmitLabel(GetCPISymbol(MI->getOperand(3).getImm()));
+ EmitToStreamer(*OutStreamer, MCInstBuilder(ARM::tADDhirr)
+ .addReg(ARM::PC)
+ .addReg(ARM::PC)
+ .addReg(Idx)
+ // Add predicate operands.
+ .addImm(ARMCC::AL)
+ .addReg(0));
+ return;
+ }
case ARM::tBR_JTr:
case ARM::BR_JTr: {
// Lower and emit the instruction itself, then the jump table following it.
CPMaxIteration("arm-constant-island-max-iteration", cl::Hidden, cl::init(30),
cl::desc("The max number of iteration for converge"));
+static cl::opt<bool> SynthesizeThumb1TBB(
+ "arm-synthesize-thumb-1-tbb", cl::Hidden, cl::init(true),
+ cl::desc("Use compressed jump tables in Thumb-1 by synthesizing an "
+ "equivalent to the TBB/TBH instructions"));
+
namespace {
/// ARMConstantIslands - Due to limited PC-relative displacements, ARM
/// requires constant pool entries to be scattered among the instructions
bool isThumb;
bool isThumb1;
bool isThumb2;
+ bool isPositionIndependentOrROPI;
public:
static char ID;
ARMConstantIslands() : MachineFunctionPass(ID) {}
STI = &static_cast<const ARMSubtarget &>(MF->getSubtarget());
TII = STI->getInstrInfo();
+ isPositionIndependentOrROPI =
+ STI->getTargetLowering()->isPositionIndependent() || STI->isROPI();
AFI = MF->getInfo<ARMFunctionInfo>();
isThumb = AFI->isThumbFunction();
isThumb2 = AFI->isThumb2Function();
HasFarJump = false;
+ bool GenerateTBB = isThumb2 || (isThumb1 && SynthesizeThumb1TBB);
// This pass invalidates liveness information when it splits basic blocks.
MF->getRegInfo().invalidateLiveness();
// Try to reorder and otherwise adjust the block layout to make good use
// of the TB[BH] instructions.
bool MadeChange = false;
- if (isThumb2 && AdjustJumpTableBlocks) {
+ if (GenerateTBB && AdjustJumpTableBlocks) {
scanFunctionJumpTables();
MadeChange |= reorderThumb2JumpTables();
// Data is out of date, so clear it. It'll be re-computed later.
MadeChange |= optimizeThumb2Branches();
// Optimize jump tables using TBB / TBH.
- if (isThumb2)
+ if (GenerateTBB)
MadeChange |= optimizeThumb2JumpTables();
// After a while, this might be made debug-only, but it is not expensive.
case ARM::t2BR_JT:
JTOpcode = ARM::JUMPTABLE_INSTS;
break;
+ case ARM::tTBB_JT:
case ARM::t2TBB_JT:
JTOpcode = ARM::JUMPTABLE_TBB;
break;
+ case ARM::tTBH_JT:
case ARM::t2TBH_JT:
JTOpcode = ARM::JUMPTABLE_TBH;
break;
void ARMConstantIslands::scanFunctionJumpTables() {
for (MachineBasicBlock &MBB : *MF) {
for (MachineInstr &I : MBB)
- if (I.isBranch() && I.getOpcode() == ARM::t2BR_JT)
+ if (I.isBranch() &&
+ (I.getOpcode() == ARM::t2BR_JT || I.getOpcode() == ARM::tBR_JTr))
T2JumpTables.push_back(&I);
}
}
default:
continue; // Ignore other JT branches
case ARM::t2BR_JT:
+ case ARM::tBR_JTr:
T2JumpTables.push_back(&I);
continue; // Does not get an entry in ImmBranches
case ARM::Bcc:
if (RemovableAdd) {
RemovableAdd->eraseFromParent();
- DeadSize += 4;
+ DeadSize += isThumb2 ? 4 : 2;
} else if (BaseReg == EntryReg) {
// The add wasn't removable, but clobbered the base for the TBB. So we can't
// preserve it.
if (!ByteOk && !HalfWordOk)
continue;
+ CPUser &User = CPUsers[JumpTableUserIndices[JTI]];
MachineBasicBlock *MBB = MI->getParent();
if (!MI->getOperand(0).isKill()) // FIXME: needed now?
continue;
- unsigned IdxReg = MI->getOperand(1).getReg();
- bool IdxRegKill = MI->getOperand(1).isKill();
- CPUser &User = CPUsers[JumpTableUserIndices[JTI]];
unsigned DeadSize = 0;
bool CanDeleteLEA = false;
bool BaseRegKill = false;
- bool PreservedBaseReg =
+
+ unsigned IdxReg = ~0U;
+ bool IdxRegKill = true;
+ if (isThumb2) {
+ IdxReg = MI->getOperand(1).getReg();
+ IdxRegKill = MI->getOperand(1).isKill();
+
+ bool PreservedBaseReg =
preserveBaseRegister(MI, User.MI, DeadSize, CanDeleteLEA, BaseRegKill);
+ if (!jumpTableFollowsTB(MI, User.CPEMI) && !PreservedBaseReg)
+ continue;
+ } else {
+ // We're in thumb-1 mode, so we must have something like:
+ // %idx = tLSLri %idx, 2
+ // %base = tLEApcrelJT
+ // %t = tLDRr %idx, %base
+ unsigned BaseReg = User.MI->getOperand(0).getReg();
+
+ MachineInstr *Shift = User.MI->getPrevNode();
+ if (Shift->getOpcode() != ARM::tLSLri ||
+ Shift->getOperand(3).getImm() != 2 ||
+ !Shift->getOperand(2).isKill())
+ continue;
+ IdxReg = Shift->getOperand(2).getReg();
+ unsigned ShiftedIdxReg = Shift->getOperand(0).getReg();
- if (!jumpTableFollowsTB(MI, User.CPEMI) && !PreservedBaseReg)
- continue;
+ MachineInstr *Load = User.MI->getNextNode();
+ if (Load->getOpcode() != ARM::tLDRr)
+ continue;
+ if (Load->getOperand(1).getReg() != ShiftedIdxReg ||
+ Load->getOperand(2).getReg() != BaseReg ||
+ !Load->getOperand(1).isKill())
+ continue;
+ // If we're in PIC mode, there should be another ADD following.
+ if (isPositionIndependentOrROPI) {
+ MachineInstr *Add = Load->getNextNode();
+ if (Add->getOpcode() != ARM::tADDrr ||
+ Add->getOperand(2).getReg() != Load->getOperand(0).getReg() ||
+ Add->getOperand(3).getReg() != BaseReg ||
+ !Add->getOperand(2).isKill())
+ continue;
+ if (Add->getOperand(0).getReg() != MI->getOperand(0).getReg())
+ continue;
+
+ Add->eraseFromParent();
+ DeadSize += 2;
+ } else {
+ if (Load->getOperand(0).getReg() != MI->getOperand(0).getReg())
+ continue;
+ }
+
+
+ // Now safe to delete the load and lsl. The LEA will be removed later.
+ CanDeleteLEA = true;
+ Shift->eraseFromParent();
+ Load->eraseFromParent();
+ DeadSize += 4;
+ }
+
DEBUG(dbgs() << "Shrink JT: " << *MI);
MachineInstr *CPEMI = User.CPEMI;
unsigned Opc = ByteOk ? ARM::t2TBB_JT : ARM::t2TBH_JT;
+ if (!isThumb2)
+ Opc = ByteOk ? ARM::tTBB_JT : ARM::tTBH_JT;
+
MachineBasicBlock::iterator MI_JT = MI;
MachineInstr *NewJTMI =
BuildMI(*MBB, MI_JT, MI->getDebugLoc(), TII->get(Opc))
if (CanDeleteLEA) {
User.MI->eraseFromParent();
- DeadSize += 4;
+ DeadSize += isThumb2 ? 4 : 2;
// The LEA was eliminated, the TBB instruction becomes the only new user
// of the jump table.
// Add an unconditional branch from NewBB to BB.
// There doesn't seem to be meaningful DebugInfo available; this doesn't
// correspond directly to anything in the source.
- assert (isThumb2 && "Adjusting for TB[BH] but not in Thumb2?");
- BuildMI(NewBB, DebugLoc(), TII->get(ARM::t2B)).addMBB(BB)
- .addImm(ARMCC::AL).addReg(0);
+ if (isThumb2)
+ BuildMI(NewBB, DebugLoc(), TII->get(ARM::t2B))
+ .addMBB(BB)
+ .addImm(ARMCC::AL)
+ .addReg(0);
+ else
+ BuildMI(NewBB, DebugLoc(), TII->get(ARM::tB))
+ .addMBB(BB)
+ .addImm(ARMCC::AL)
+ .addReg(0);
// Update internal data structures to account for the newly inserted MBB.
MF->RenumberBlocks(NewBB);
(ins i32imm:$label, pred:$p),
2, IIC_iALUi, []>, Sched<[WriteALU]>;
+// Thumb-1 doesn't have the TBB or TBH instructions, but we can synthesize them
+// and make use of the same compressed jump table format as Thumb-2.
+let Size = 2 in {
+def tTBB_JT : tPseudoInst<(outs),
+ (ins tGPR:$base, tGPR:$index, i32imm:$jt, i32imm:$pclbl), 0, IIC_Br, []>,
+ Sched<[WriteBr]>;
+
+def tTBH_JT : tPseudoInst<(outs),
+ (ins tGPR:$base, tGPR:$index, i32imm:$jt, i32imm:$pclbl), 0, IIC_Br, []>,
+ Sched<[WriteBr]>;
+}
+
//===----------------------------------------------------------------------===//
// TLS Instructions
//
; RUN: llc -relocation-model=ropi -mtriple=thumbv7m--none-eabi -disable-block-placement < %s | FileCheck %s --check-prefix=CHECK --check-prefix=THUMB2
; RUN: llc -relocation-model=ropi-rwpi -mtriple=thumbv7m--none-eabi -disable-block-placement < %s | FileCheck %s --check-prefix=CHECK --check-prefix=THUMB2
-; RUN: llc -relocation-model=static -mtriple=thumbv6m--none-eabi -disable-block-placement < %s | FileCheck %s --check-prefix=CHECK --check-prefix=THUMB1 --check-prefix=THUMB1_ABS
-; RUN: llc -relocation-model=ropi -mtriple=thumbv6m--none-eabi -disable-block-placement < %s | FileCheck %s --check-prefix=CHECK --check-prefix=THUMB1 --check-prefix=THUMB1_PC
-; RUN: llc -relocation-model=ropi-rwpi -mtriple=thumbv6m--none-eabi -disable-block-placement < %s | FileCheck %s --check-prefix=CHECK --check-prefix=THUMB1 --check-prefix=THUMB1_PC
+; RUN: llc -relocation-model=static -mtriple=thumbv6m--none-eabi -disable-block-placement < %s | FileCheck %s --check-prefix=CHECK --check-prefix=THUMB1
+; RUN: llc -relocation-model=ropi -mtriple=thumbv6m--none-eabi -disable-block-placement < %s | FileCheck %s --check-prefix=CHECK --check-prefix=THUMB1
+; RUN: llc -relocation-model=ropi-rwpi -mtriple=thumbv6m--none-eabi -disable-block-placement < %s | FileCheck %s --check-prefix=CHECK --check-prefix=THUMB1
declare void @exit0()
; THUMB2: [[LBB4]]
; THUMB2-NEXT: b exit4
-; THUMB1: lsls r[[R_TAB_INDEX:[0-9]+]], r{{[0-9]+}}, #2
-; THUMB1: adr r[[R_TAB_BASE:[0-9]+]], [[LJTI:\.LJTI[0-9]+_[0-9]+]]
-; THUMB1: ldr r[[R_BB_ADDR:[0-9]+]], [r[[R_TAB_INDEX]], r[[R_TAB_BASE]]]
-; THUMB1_PC: adds r[[R_BB_ADDR]], r[[R_BB_ADDR]], r[[R_TAB_BASE]]
-; THUMB1: mov pc, r[[R_BB_ADDR]]
-; THUMB1: [[LJTI]]
-; THUMB1_ABS: .long [[LBB1:\.LBB[0-9]+_[0-9]+]]+1
-; THUMB1_ABS: .long [[LBB2:\.LBB[0-9]+_[0-9]+]]+1
-; THUMB1_ABS: .long [[LBB3:\.LBB[0-9]+_[0-9]+]]+1
-; THUMB1_ABS: .long [[LBB4:\.LBB[0-9]+_[0-9]+]]+1
-; THUMB1_PC: .long [[LBB1:\.LBB[0-9]+_[0-9]+]]-[[LJTI]]
-; THUMB1_PC: .long [[LBB2:\.LBB[0-9]+_[0-9]+]]-[[LJTI]]
-; THUMB1_PC: .long [[LBB3:\.LBB[0-9]+_[0-9]+]]-[[LJTI]]
-; THUMB1_PC: .long [[LBB4:\.LBB[0-9]+_[0-9]+]]-[[LJTI]]
+
+; THUMB1: add r[[x:[0-9]+]], pc
+; THUMB1: ldrb r[[x]], [r[[x]], #4]
+; THUMB1: lsls r[[x]], r[[x]], #1
+; THUMB1: [[LCPI:\.LCPI[0-9]+_[0-9]+]]:
+; THUMB1: add pc, r[[x]]
+; THUMB1: .byte ([[LBB1:\.LBB[0-9]+_[0-9]+]]-([[LCPI]]+4))/2
+; THUMB1: .byte ([[LBB2:\.LBB[0-9]+_[0-9]+]]-([[LCPI]]+4))/2
+; THUMB1: .byte ([[LBB3:\.LBB[0-9]+_[0-9]+]]-([[LCPI]]+4))/2
+; THUMB1: .byte ([[LBB4:\.LBB[0-9]+_[0-9]+]]-([[LCPI]]+4))/2
; THUMB1: [[LBB1]]
; THUMB1-NEXT: bl exit1
-; THUMB1-NEXT: pop
; THUMB1: [[LBB2]]
; THUMB1-NEXT: bl exit2
-; THUMB1-NEXT: pop
; THUMB1: [[LBB3]]
; THUMB1-NEXT: bl exit3
-; THUMB1-NEXT: pop
; THUMB1: [[LBB4]]
; THUMB1-NEXT: bl exit4
-; THUMB1-NEXT: pop
}
--- /dev/null
+; RUN: llc -mtriple=thumbv7m-linux-gnu -o - %s | FileCheck %s --check-prefix=T2
+; RUN: llc -mtriple=thumbv6m-linux-gnu -o - %s | FileCheck %s --check-prefix=T1
+
+declare void @foo(double)
+declare i32 @llvm.arm.space(i32, i32)
+
+define i32 @test_tbh(i1 %tst, i32 %sw, i32 %l) {
+ br label %complex
+
+; T2-LABEL: test_tbh:
+; T2: [[ANCHOR:.LCPI[0-9_]+]]:
+; T2: tbh [pc, r{{[0-9]+}}, lsl #1]
+; T2-NEXT: @ BB#1
+; T2-NEXT: LJTI
+; T2-NEXT: .short (.LBB0_[[x:[0-9]+]]-([[ANCHOR]]+4))/2
+; T2-NEXT: .short (.LBB0_{{[0-9]+}}-([[ANCHOR]]+4))/2
+; T2-NEXT: .short (.LBB0_{{[0-9]+}}-([[ANCHOR]]+4))/2
+; T2-NEXT: .short (.LBB0_[[x]]-([[ANCHOR]]+4))/2
+
+; T1-LABEL: test_tbh:
+; T1: lsls [[x:r[0-9]+]], r4, #1
+; T1: add [[x]], pc
+; T1: ldrh [[x]], {{\[}}[[x]], #4]
+; T1: lsls [[x]], [[x]], #1
+; T1: [[ANCHOR:.LCPI[0-9_]+]]:
+; T1: add pc, [[x]]
+; T1-NEXT: @ BB#2
+; T1-NEXT: LJTI
+; T1-NEXT: .short (.LBB0_[[x:[0-9]+]]-([[ANCHOR]]+4))/2
+; T1-NEXT: .short (.LBB0_{{[0-9]+}}-([[ANCHOR]]+4))/2
+; T1-NEXT: .short (.LBB0_{{[0-9]+}}-([[ANCHOR]]+4))/2
+; T1-NEXT: .short (.LBB0_[[x]]-([[ANCHOR]]+4))/2
+
+complex:
+ call void @foo(double 12345.0)
+ switch i32 %sw, label %second [ i32 0, label %other
+ i32 1, label %third
+ i32 2, label %end
+ i32 3, label %other ]
+
+second:
+ ret i32 43
+third:
+ call i32 @llvm.arm.space(i32 970, i32 undef)
+ ret i32 0
+
+other:
+ call void @bar()
+ unreachable
+
+end:
+ ret i32 42
+}
+
+declare void @bar()
; RUN: llc -mtriple=thumb-eabi -mcpu=arm1156t2-s -mattr=+thumb2 -arm-adjust-jump-tables=0 %s -o - | FileCheck %s
+; RUN: llc -mtriple=thumbv6-eabi -mcpu=cortex-m0 -arm-adjust-jump-tables=0 %s -o - | FileCheck %s
; Do not use tbb / tbh if any destination is before the jumptable.
; rdar://7102917
; RUN: llc < %s -mtriple=thumbv7-apple-darwin | FileCheck %s
; RUN: llc < %s -mtriple=thumbv7-apple-darwin -relocation-model=pic | FileCheck %s
+; RUN: llc < %s -mtriple=thumbv6m-apple-darwin -relocation-model=static | FileCheck %s --check-prefix=THUMB1
+; RUN: llc < %s -mtriple=thumbv6m-apple-darwin -relocation-model=pic | FileCheck %s --check-prefix=THUMB1
define void @bar(i32 %n.u) {
entry:
; CHECK: .end_data_region
; CHECK-NEXT: .p2align 1
+; THUMB1-LABEL: bar:
+; THUMB1: add pc, r0
+; THUMB1: .data_region jt8
+; THUMB1: .byte (LBB0_3-(LCPI0_0+4))/2
+; THUMB1: .end_data_region
+; THUMB1-NEXT: .p2align 1
+
switch i32 %n.u, label %bb12 [i32 1, label %bb i32 2, label %bb6 i32 4, label %bb7 i32 5, label %bb8 i32 6, label %bb10 i32 7, label %bb1 i32 8, label %bb3 i32 9, label %bb4 i32 10, label %bb9 i32 11, label %bb2 i32 12, label %bb5 i32 13, label %bb11 ]
bb:
tail call void(...) @foo1()
-; RUN: llc < %s -mtriple=thumbv7-apple-darwin -relocation-model=pic | FileCheck %s
+; RUN: llc < %s -mtriple=thumbv7-apple-darwin -relocation-model=pic | FileCheck %s --check-prefix=CHECK --check-prefix=T2
+; RUN: llc < %s -mtriple=thumbv6m-apple-darwin -relocation-model=pic | FileCheck %s --check-prefix=CHECK --check-prefix=T1
+; RUN: llc < %s -mtriple=thumbv6m-apple-darwin -relocation-model=static | FileCheck %s --check-prefix=CHECK --check-prefix=T1
; Thumb2 target should reorder the bb's in order to use tbb / tbh.
define i32 @main(i32 %argc, i8** nocapture %argv) nounwind {
; CHECK-LABEL: main:
; CHECK-NOT: adr {{r[0-9]+}}, LJTI
+; T1: lsls r[[x:[0-9]+]], {{r[0-9]+}}, #1
; CHECK: [[PCREL_ANCHOR:LCPI[0-9]+_[0-9]+]]:
-; CHECK-NEXT: tbb [pc, {{r[0-9]+}}]
+; T2-NEXT: tbb [pc, {{r[0-9]+}}]
+; T1-NEXT: add pc, r[[x]]
; CHECK: LJTI0_0:
; CHECK-NEXT: .data_region jt8