summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorEvan Cheng <evan.cheng@apple.com>2008-06-25 01:16:38 +0000
committerEvan Cheng <evan.cheng@apple.com>2008-06-25 01:16:38 +0000
commit601ca4b434f5c67503a30575cc36b688b0d959e6 (patch)
treec976edbbf80af613dde058cb8b554fc47d6d2a5c
parente4f6ee590b65d7d164d11c4f185ac7e0f9fab648 (diff)
downloadllvm-601ca4b434f5c67503a30575cc36b688b0d959e6.tar.gz
llvm-601ca4b434f5c67503a30575cc36b688b0d959e6.tar.bz2
llvm-601ca4b434f5c67503a30575cc36b688b0d959e6.tar.xz
Enable two-address remat by default.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52701 91177308-0d34-0410-b5e6-96231b3b80d8
-rw-r--r--lib/CodeGen/TwoAddressInstructionPass.cpp68
-rw-r--r--lib/Target/X86/X86Instr64bit.td2
-rw-r--r--test/CodeGen/X86/2007-11-30-LoadFolding-Bug.ll2
-rw-r--r--test/CodeGen/X86/twoaddr-remat.ll67
4 files changed, 102 insertions, 37 deletions
diff --git a/lib/CodeGen/TwoAddressInstructionPass.cpp b/lib/CodeGen/TwoAddressInstructionPass.cpp
index 0d71589213..54adc96e4b 100644
--- a/lib/CodeGen/TwoAddressInstructionPass.cpp
+++ b/lib/CodeGen/TwoAddressInstructionPass.cpp
@@ -53,10 +53,6 @@ STATISTIC(NumConvertedTo3Addr, "Number of instructions promoted to 3-address");
STATISTIC(Num3AddrSunk, "Number of 3-address instructions sunk");
STATISTIC(NumReMats, "Number of instructions re-materialized");
-static cl::opt<bool>
-EnableReMat("two-addr-remat", cl::init(false), cl::Hidden,
- cl::desc("Two-addr conversion should remat when possible."));
-
namespace {
class VISIBILITY_HIDDEN TwoAddressInstructionPass
: public MachineFunctionPass {
@@ -71,8 +67,8 @@ namespace {
bool isSafeToReMat(unsigned DstReg, MachineInstr *MI);
bool isProfitableToReMat(unsigned Reg, const TargetRegisterClass *RC,
- MachineInstr *MI, unsigned Loc,
- MachineInstr *DefMI, MachineBasicBlock *MBB,
+ MachineInstr *MI, MachineInstr *DefMI,
+ MachineBasicBlock *MBB, unsigned Loc,
DenseMap<MachineInstr*, unsigned> &DistanceMap);
public:
static char ID; // Pass identification, replacement for typeid
@@ -248,12 +244,9 @@ static bool isTwoAddrUse(MachineInstr *UseMI, unsigned Reg) {
bool
TwoAddressInstructionPass::isProfitableToReMat(unsigned Reg,
const TargetRegisterClass *RC,
- MachineInstr *MI, unsigned Loc,
- MachineInstr *DefMI, MachineBasicBlock *MBB,
- DenseMap<MachineInstr*, unsigned> &DistanceMap) {
- if (DefMI->getParent() != MBB)
- return true;
- // If earlier uses in MBB are not two-address uses, then don't remat.
+ MachineInstr *MI, MachineInstr *DefMI,
+ MachineBasicBlock *MBB, unsigned Loc,
+ DenseMap<MachineInstr*, unsigned> &DistanceMap){
bool OtherUse = false;
for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(Reg),
UE = MRI->use_end(); UI != UE; ++UI) {
@@ -261,18 +254,26 @@ TwoAddressInstructionPass::isProfitableToReMat(unsigned Reg,
if (!UseMO.isUse())
continue;
MachineInstr *UseMI = UseMO.getParent();
- if (UseMI->getParent() != MBB)
- continue;
- DenseMap<MachineInstr*, unsigned>::iterator DI = DistanceMap.find(UseMI);
- if (DI != DistanceMap.end() && DI->second == Loc)
- continue; // Current use.
- OtherUse = true;
- // There is at least one other use in the MBB that will clobber the
- // register.
- if (isTwoAddrUse(UseMI, Reg))
- return true;
+ MachineBasicBlock *UseMBB = UseMI->getParent();
+ if (UseMBB == MBB) {
+ DenseMap<MachineInstr*, unsigned>::iterator DI = DistanceMap.find(UseMI);
+ if (DI != DistanceMap.end() && DI->second == Loc)
+ continue; // Current use.
+ OtherUse = true;
+ // There is at least one other use in the MBB that will clobber the
+ // register.
+ if (isTwoAddrUse(UseMI, Reg))
+ return true;
+ }
}
- return !OtherUse;
+
+ // If other uses in MBB are not two-address uses, then don't remat.
+ if (OtherUse)
+ return false;
+
+ // No other uses in the same block, remat if it's defined in the same
+ // block so it does not unnecessarily extend the live range.
+ return MBB == DefMI->getParent();
}
/// runOnMachineFunction - Reduce two-address instructions to two operands.
@@ -428,9 +429,9 @@ bool TwoAddressInstructionPass::runOnMachineFunction(MachineFunction &MF) {
MachineInstr *DefMI = MRI->getVRegDef(regB);
// If it's safe and profitable, remat the definition instead of
// copying it.
- if (EnableReMat && DefMI &&
+ if (DefMI &&
isSafeToReMat(regB, DefMI) &&
- isProfitableToReMat(regB, rc, mi, Dist, DefMI, mbbi,DistanceMap)){
+ isProfitableToReMat(regB, rc, mi, DefMI, mbbi, Dist,DistanceMap)){
DEBUG(cerr << "2addr: REMATTING : " << *DefMI << "\n");
TII->reMaterialize(*mbbi, mi, regA, DefMI);
ReMatRegs.set(regB);
@@ -473,17 +474,14 @@ bool TwoAddressInstructionPass::runOnMachineFunction(MachineFunction &MF) {
}
}
- if (EnableReMat) {
- // Some remat'ed instructions are dead.
- int VReg = ReMatRegs.find_first();
- while (VReg != -1) {
- if (MRI->use_empty(VReg)) {
- MachineInstr *DefMI = MRI->getVRegDef(VReg);
- DefMI->eraseFromParent();
- }
- VReg = ReMatRegs.find_next(VReg);
+ // Some remat'ed instructions are dead.
+ int VReg = ReMatRegs.find_first();
+ while (VReg != -1) {
+ if (MRI->use_empty(VReg)) {
+ MachineInstr *DefMI = MRI->getVRegDef(VReg);
+ DefMI->eraseFromParent();
}
-
+ VReg = ReMatRegs.find_next(VReg);
}
return MadeChange;
diff --git a/lib/Target/X86/X86Instr64bit.td b/lib/Target/X86/X86Instr64bit.td
index 382a27c88e..8398e9aa2a 100644
--- a/lib/Target/X86/X86Instr64bit.td
+++ b/lib/Target/X86/X86Instr64bit.td
@@ -199,7 +199,7 @@ let neverHasSideEffects = 1 in
def MOV64rr : RI<0x89, MRMDestReg, (outs GR64:$dst), (ins GR64:$src),
"mov{q}\t{$src, $dst|$dst, $src}", []>;
-let isReMaterializable = 1 in {
+let isReMaterializable = 1, isAsCheapAsAMove = 1 in {
def MOV64ri : RIi64<0xB8, AddRegFrm, (outs GR64:$dst), (ins i64imm:$src),
"movabs{q}\t{$src, $dst|$dst, $src}",
[(set GR64:$dst, imm:$src)]>;
diff --git a/test/CodeGen/X86/2007-11-30-LoadFolding-Bug.ll b/test/CodeGen/X86/2007-11-30-LoadFolding-Bug.ll
index c123eccef3..21dff69373 100644
--- a/test/CodeGen/X86/2007-11-30-LoadFolding-Bug.ll
+++ b/test/CodeGen/X86/2007-11-30-LoadFolding-Bug.ll
@@ -3,7 +3,7 @@
declare fastcc void @rdft(i32, i32, double*, i32*, double*)
-define fastcc void @mp_sqrt(i32 %n, i32 %radix, i32* %in, i32* %out, i32* %tmp1, i32* %tmp2, i32 %nfft, double* %tmp1fft, double* %tmp2fft, i32* %ip, double* %w) {
+define fastcc void @mp_sqrt(i32 %n, i32 %radix, i32* %in, i32* %out, i32* %tmp1, i32* %tmp2, i32 %nfft, double* %tmp1fft, double* %tmp2fft, i32* %ip, double* %w) nounwind {
entry:
br label %bb.i5
diff --git a/test/CodeGen/X86/twoaddr-remat.ll b/test/CodeGen/X86/twoaddr-remat.ll
new file mode 100644
index 0000000000..b74b70cedb
--- /dev/null
+++ b/test/CodeGen/X86/twoaddr-remat.ll
@@ -0,0 +1,67 @@
+; RUN: llvm-as < %s | llc -march=x86 | grep 59796 | count 3
+
+ %Args = type %Value*
+ %Exec = type opaque*
+ %Identifier = type opaque*
+ %JSFunction = type %Value (%Exec, %Scope, %Value, %Args)
+ %PropertyNameArray = type opaque*
+ %Scope = type opaque*
+ %Value = type opaque*
+
+declare i1 @X1(%Exec) readonly
+
+declare %Value @X2(%Exec)
+
+declare i32 @X3(%Exec, %Value)
+
+declare %Value @X4(i32) readnone
+
+define internal %Value @fast3bitlookup(%Exec %exec, %Scope %scope, %Value %this, %Args %args) nounwind {
+prologue:
+ %eh_check = tail call i1 @X1( %Exec %exec ) readonly ; <i1> [#uses=1]
+ br i1 %eh_check, label %exception, label %no_exception
+
+exception: ; preds = %no_exception, %prologue
+ %rethrow_result = tail call %Value @X2( %Exec %exec ) ; <%Value> [#uses=1]
+ ret %Value %rethrow_result
+
+no_exception: ; preds = %prologue
+ %args_intptr = bitcast %Args %args to i32* ; <i32*> [#uses=1]
+ %argc_val = load i32* %args_intptr ; <i32> [#uses=1]
+ %cmpParamArgc = icmp sgt i32 %argc_val, 0 ; <i1> [#uses=1]
+ %arg_ptr = getelementptr %Args %args, i32 1 ; <%Args> [#uses=1]
+ %arg_val = load %Args %arg_ptr ; <%Value> [#uses=1]
+ %ext_arg_val = select i1 %cmpParamArgc, %Value %arg_val, %Value inttoptr (i32 5 to %Value) ; <%Value> [#uses=1]
+ %toInt325 = tail call i32 @X3( %Exec %exec, %Value %ext_arg_val ) ; <i32> [#uses=3]
+ %eh_check6 = tail call i1 @X1( %Exec %exec ) readonly ; <i1> [#uses=1]
+ br i1 %eh_check6, label %exception, label %no_exception7
+
+no_exception7: ; preds = %no_exception
+ %shl_tmp_result = shl i32 %toInt325, 1 ; <i32> [#uses=1]
+ %rhs_masked13 = and i32 %shl_tmp_result, 14 ; <i32> [#uses=1]
+ %ashr_tmp_result = lshr i32 59796, %rhs_masked13 ; <i32> [#uses=1]
+ %and_tmp_result15 = and i32 %ashr_tmp_result, 3 ; <i32> [#uses=1]
+ %ashr_tmp_result3283 = lshr i32 %toInt325, 2 ; <i32> [#uses=1]
+ %rhs_masked38 = and i32 %ashr_tmp_result3283, 14 ; <i32> [#uses=1]
+ %ashr_tmp_result39 = lshr i32 59796, %rhs_masked38 ; <i32> [#uses=1]
+ %and_tmp_result41 = and i32 %ashr_tmp_result39, 3 ; <i32> [#uses=1]
+ %addconv = add i32 %and_tmp_result15, %and_tmp_result41 ; <i32> [#uses=1]
+ %ashr_tmp_result6181 = lshr i32 %toInt325, 5 ; <i32> [#uses=1]
+ %rhs_masked67 = and i32 %ashr_tmp_result6181, 6 ; <i32> [#uses=1]
+ %ashr_tmp_result68 = lshr i32 59796, %rhs_masked67 ; <i32> [#uses=1]
+ %and_tmp_result70 = and i32 %ashr_tmp_result68, 3 ; <i32> [#uses=1]
+ %addconv82 = add i32 %addconv, %and_tmp_result70 ; <i32> [#uses=3]
+ %rangetmp = add i32 %addconv82, 536870912 ; <i32> [#uses=1]
+ %rangecmp = icmp ult i32 %rangetmp, 1073741824 ; <i1> [#uses=1]
+ br i1 %rangecmp, label %NumberLiteralIntFast, label %NumberLiteralIntSlow
+
+NumberLiteralIntFast: ; preds = %no_exception7
+ %imm_shift = shl i32 %addconv82, 2 ; <i32> [#uses=1]
+ %imm_or = or i32 %imm_shift, 3 ; <i32> [#uses=1]
+ %imm_val = inttoptr i32 %imm_or to %Value ; <%Value> [#uses=1]
+ ret %Value %imm_val
+
+NumberLiteralIntSlow: ; preds = %no_exception7
+ %toVal = call %Value @X4( i32 %addconv82 ) ; <%Value> [#uses=1]
+ ret %Value %toVal
+}