Modified: trunk/Source/_javascript_Core/b3/B3LowerToAir.cpp (207038 => 207039)
--- trunk/Source/_javascript_Core/b3/B3LowerToAir.cpp 2016-10-11 02:55:26 UTC (rev 207038)
+++ trunk/Source/_javascript_Core/b3/B3LowerToAir.cpp 2016-10-11 03:39:05 UTC (rev 207039)
@@ -399,7 +399,8 @@
// short, you should avoid this by using the pattern matcher to match patterns.
void commitInternal(Value* value)
{
- m_locked.add(value);
+ if (value)
+ m_locked.add(value);
}
bool crossesInterference(Value* value)
@@ -1798,6 +1799,148 @@
},
false);
}
+
+ bool tryAppendLea()
+ {
+ Air::Opcode leaOpcode = tryOpcodeForType(Lea32, Lea64, m_value->type());
+ if (!isValidForm(leaOpcode, Arg::Index, Arg::Tmp))
+ return false;
+
+ // This lets us turn things like this:
+ //
+ // Add(Add(@x, Shl(@y, $2)), $100)
+ //
+ // Into this:
+ //
+ // lea 100(%rdi,%rsi,4), %rax
+ //
+ // We have a choice here between committing the internal bits of an index or sharing
+ // them. There are solid arguments for both.
+ //
+ // Sharing: The word on the street is that the cost of a lea is one cycle no matter
+ // what it does. Every experiment I've ever seen seems to confirm this. So, sharing
+ // helps us in situations where Wasm input did this:
+ //
+ // x = a[i].x;
+ // y = a[i].y;
+ //
+ // With sharing we would do:
+ //
+ // leal (%a,%i,4), %tmp
+ // cmp (%size, %tmp)
+ // ja _fail
+ // movl (%base, %tmp), %x
+ // leal 4(%a,%i,4), %tmp
+ // cmp (%size, %tmp)
+ // ja _fail
+ // movl (%base, %tmp), %y
+ //
+ // In the absence of sharing, we may find ourselves needing separate registers for
+ // the innards of the index. That's relatively unlikely to be a thing due to other
+ // optimizations that we already have, but it could happen
+ //
+ // Committing: The worst case is that there is a complicated graph of additions and
+ // shifts, where each value has multiple uses. In that case, it's better to compute
+ // each one separately from the others since that way, each calculation will use a
+ // relatively nearby tmp as its input. That seems uncommon, but in those cases,
+ // committing is a clear winner: it would result in a simple interference graph
+ // while sharing would result in a complex one. Interference sucks because it means
+ // more time in IRC and it means worse code.
+ //
+ // It's not super clear if any of these corner cases would ever arise. Committing
+ // has the benefit that it's easier to reason about, and protects a much darker
+ // corner case (more interference).
+
+ // Here are the things we want to match:
+ // Add(Add(@x, @y), $c)
+ // Add(Shl(@x, $c), @y)
+ // Add(@x, Shl(@y, $c))
+ // Add(Add(@x, Shl(@y, $c)), $d)
+ // Add(Add(Shl(@x, $c), @y), $d)
+ //
+ // Note that if you do Add(Shl(@x, $c), $d) then we will treat $d as a non-constant and
+ // force it to materialize. You'll get something like this:
+ //
+ // movl $d, %tmp
+ // leal (%tmp,%x,1<<c), %result
+ //
+ // Which is pretty close to optimal and has the nice effect of being able to handle large
+ // constants gracefully.
+
+ Value* innerAdd = nullptr;
+
+ Value* value = m_value;
+
+ // We're going to consume Add(Add(_), $c). If we succeed at consuming it then we have these
+ // patterns left (i.e. in the Add(_)):
+ //
+ // Add(Add(@x, @y), $c)
+ // Add(Add(@x, Shl(@y, $c)), $d)
+ // Add(Add(Shl(@x, $c), @y), $d)
+ //
+ // Otherwise we are looking at these patterns:
+ //
+ // Add(Shl(@x, $c), @y)
+ // Add(@x, Shl(@y, $c))
+ //
+ // This means that the subsequent code only has to worry about three patterns:
+ //
+ // Add(Shl(@x, $c), @y)
+ // Add(@x, Shl(@y, $c))
+ // Add(@x, @y) (only if offset != 0)
+ int32_t offset = 0;
+ if (value->child(1)->isRepresentableAs<int32_t>()
+ && canBeInternal(value->child(0))
+ && value->child(0)->opcode() == Add) {
+ innerAdd = value->child(0);
+ offset = value->child(1)->asInt32();
+ value = value->child(0);
+ }
+
+ auto tryShl = [&] (Value* shl, Value* other) -> bool {
+ if (shl->opcode() != Shl)
+ return false;
+ if (!canBeInternal(shl))
+ return false;
+ if (!shl->child(1)->hasInt32())
+ return false;
+ unsigned logScale = shl->child(1)->asInt32();
+ if (m_value->type() == Int32)
+ logScale &= 31;
+ else
+ logScale &= 63;
+ // Use 64-bit math to perform the shift so that <<32 does the right thing.
+ int64_t bigScale = static_cast<uint64_t>(1) << static_cast<uint64_t>(logScale);
+ if (!isRepresentableAs<int32_t>(bigScale))
+ return false;
+ unsigned scale = static_cast<int32_t>(bigScale);
+ if (!Arg::isValidIndexForm(scale, offset))
+ return false;
+
+ ASSERT(!m_locked.contains(shl->child(0)));
+ ASSERT(!m_locked.contains(other));
+
+ append(leaOpcode, Arg::index(tmp(other), tmp(shl->child(0)), scale, offset), tmp(m_value));
+ commitInternal(innerAdd);
+ commitInternal(shl);
+ return true;
+ };
+
+ if (tryShl(value->child(0), value->child(1)))
+ return true;
+ if (tryShl(value->child(1), value->child(0)))
+ return true;
+
+ // The remaining pattern is just:
+ // Add(@x, @y) (only if offset != 0)
+ if (!offset)
+ return false;
+ ASSERT(!m_locked.contains(value->child(0)));
+ ASSERT(!m_locked.contains(value->child(1)));
+ append(leaOpcode, Arg::index(tmp(value->child(0)), tmp(value->child(1)), 1, offset), tmp(m_value));
+ commitInternal(innerAdd);
+ return true;
+ }
void lower()
{
@@ -1834,9 +1977,11 @@
}
case Add: {
+ if (tryAppendLea())
+ return;
+
Air::Opcode multiplyAddOpcode = tryOpcodeForType(MultiplyAdd32, MultiplyAdd64, m_value->type());
- if (multiplyAddOpcode != Air::Oops
- && isValidForm(multiplyAddOpcode, Arg::Tmp, Arg::Tmp, Arg::Tmp, Arg::Tmp)) {
+ if (isValidForm(multiplyAddOpcode, Arg::Tmp, Arg::Tmp, Arg::Tmp, Arg::Tmp)) {
Value* left = m_value->child(0);
Value* right = m_value->child(1);
if (!imm(right) || m_valueToTmp[right]) {
@@ -1846,7 +1991,7 @@
Value* multiplyLeft = left->child(0);
Value* multiplyRight = left->child(1);
- if (m_locked.contains(multiplyLeft) || m_locked.contains(multiplyRight))
+ if (canBeInternal(multiplyLeft) || canBeInternal(multiplyRight))
return false;
append(multiplyAddOpcode, tmp(multiplyLeft), tmp(multiplyRight), tmp(right), tmp(m_value));
@@ -2215,7 +2360,7 @@
case SlotBase: {
append(
- Lea,
+ pointerType() == Int64 ? Lea64 : Lea32,
Arg::stack(m_stackToStack.get(m_value->as<SlotBaseValue>()->slot())),
tmp(m_value));
return;
Modified: trunk/Source/_javascript_Core/b3/testb3.cpp (207038 => 207039)
--- trunk/Source/_javascript_Core/b3/testb3.cpp 2016-10-11 02:55:26 UTC (rev 207038)
+++ trunk/Source/_javascript_Core/b3/testb3.cpp 2016-10-11 03:39:05 UTC (rev 207039)
@@ -154,28 +154,37 @@
Air::validate(proc.code());
}
-void checkUsesInstruction(Compilation& compilation, const char* text)
+template<typename Func>
+void checkDisassembly(Compilation& compilation, const Func& func, CString failText)
{
CString disassembly = compilation.disassembly();
- if (strstr(disassembly.data(), text))
+ if (func(disassembly.data()))
return;
-
+
crashLock.lock();
- dataLog("Bad lowering! Expected to find ", text, " but didn't:\n");
+ dataLog("Bad lowering! ", failText, "\n");
dataLog(disassembly);
CRASH();
}
+void checkUsesInstruction(Compilation& compilation, const char* text)
+{
+ checkDisassembly(
+ compilation,
+ [&] (const char* disassembly) -> bool {
+ return strstr(disassembly, text);
+ },
+ toCString("Expected to find ", text, " but didnt!"));
+}
+
void checkDoesNotUseInstruction(Compilation& compilation, const char* text)
{
- CString disassembly = compilation.disassembly();
- if (!strstr(disassembly.data(), text))
- return;
-
- crashLock.lock();
- dataLog("Bad lowering! Did not expected to find ", text, " but it's there:\n");
- dataLog(disassembly);
- CRASH();
+ checkDisassembly(
+ compilation,
+ [&] (const char* disassembly) -> bool {
+ return !strstr(disassembly, text);
+ },
+ toCString("Did not expected to find ", text, " but it's there!"));
}
template<typename Type>
@@ -13381,6 +13390,275 @@
go(false);
}
+void testX86LeaAddAddShlLeft()
+{
+ // Add(Add(Shl(@x, $c), @y), $d)
+ Procedure proc;
+ BasicBlock* root = proc.addBlock();
+ Value* result = root->appendNew<Value>(
+ proc, Add, Origin(),
+ root->appendNew<Value>(
+ proc, Add, Origin(),
+ root->appendNew<Value>(
+ proc, Shl, Origin(),
+ root->appendNew<ArgumentRegValue>(proc, Origin(), GPRInfo::argumentGPR1),
+ root->appendNew<Const32Value>(proc, Origin(), 2)),
+ root->appendNew<ArgumentRegValue>(proc, Origin(), GPRInfo::argumentGPR0)),
+ root->appendNew<ConstPtrValue>(proc, Origin(), 100));
+ root->appendNew<Value>(proc, Return, Origin(), result);
+
+ auto code = compile(proc);
+ checkUsesInstruction(*code, "lea 0x64(%rdi,%rsi,4), %rax");
+ CHECK_EQ(invoke<intptr_t>(*code, 1, 2), (1 + (2 << 2)) + 100);
+}
+
+void testX86LeaAddAddShlRight()
+{
+ // Add(Add(@x, Shl(@y, $c)), $d)
+ Procedure proc;
+ BasicBlock* root = proc.addBlock();
+ Value* result = root->appendNew<Value>(
+ proc, Add, Origin(),
+ root->appendNew<Value>(
+ proc, Add, Origin(),
+ root->appendNew<ArgumentRegValue>(proc, Origin(), GPRInfo::argumentGPR0),
+ root->appendNew<Value>(
+ proc, Shl, Origin(),
+ root->appendNew<ArgumentRegValue>(proc, Origin(), GPRInfo::argumentGPR1),
+ root->appendNew<Const32Value>(proc, Origin(), 2))),
+ root->appendNew<ConstPtrValue>(proc, Origin(), 100));
+ root->appendNew<Value>(proc, Return, Origin(), result);
+
+ auto code = compile(proc);
+ checkUsesInstruction(*code, "lea 0x64(%rdi,%rsi,4), %rax");
+ CHECK_EQ(invoke<intptr_t>(*code, 1, 2), (1 + (2 << 2)) + 100);
+}
+
+void testX86LeaAddAdd()
+{
+ // Add(Add(@x, @y), $c)
+ Procedure proc;
+ BasicBlock* root = proc.addBlock();
+ Value* result = root->appendNew<Value>(
+ proc, Add, Origin(),
+ root->appendNew<Value>(
+ proc, Add, Origin(),
+ root->appendNew<ArgumentRegValue>(proc, Origin(), GPRInfo::argumentGPR1),
+ root->appendNew<ArgumentRegValue>(proc, Origin(), GPRInfo::argumentGPR0)),
+ root->appendNew<ConstPtrValue>(proc, Origin(), 100));
+ root->appendNew<Value>(proc, Return, Origin(), result);
+
+ auto code = compile(proc);
+ checkDisassembly(
+ *code,
+ [&] (const char* disassembly) -> bool {
+ return strstr(disassembly, "lea 0x64(%rdi,%rsi), %rax")
+ || strstr(disassembly, "lea 0x64(%rsi,%rdi), %rax");
+ },
+ "Expected to find something like lea 0x64(%rdi,%rsi), %rax but didn't!");
+ CHECK_EQ(invoke<intptr_t>(*code, 1, 2), (1 + 2) + 100);
+}
+
+void testX86LeaAddShlRight()
+{
+ // Add(Shl(@x, $c), @y)
+ Procedure proc;
+ BasicBlock* root = proc.addBlock();
+ Value* result = root->appendNew<Value>(
+ proc, Add, Origin(),
+ root->appendNew<ArgumentRegValue>(proc, Origin(), GPRInfo::argumentGPR0),
+ root->appendNew<Value>(
+ proc, Shl, Origin(),
+ root->appendNew<ArgumentRegValue>(proc, Origin(), GPRInfo::argumentGPR1),
+ root->appendNew<Const32Value>(proc, Origin(), 2)));
+ root->appendNew<Value>(proc, Return, Origin(), result);
+
+ auto code = compile(proc);
+ checkUsesInstruction(*code, "lea (%rdi,%rsi,4), %rax");
+ CHECK_EQ(invoke<intptr_t>(*code, 1, 2), 1 + (2 << 2));
+}
+
+void testX86LeaAddShlLeftScale1()
+{
+ // Add(Shl(@x, $c), @y)
+ Procedure proc;
+ BasicBlock* root = proc.addBlock();
+ Value* result = root->appendNew<Value>(
+ proc, Add, Origin(),
+ root->appendNew<ArgumentRegValue>(proc, Origin(), GPRInfo::argumentGPR0),
+ root->appendNew<Value>(
+ proc, Shl, Origin(),
+ root->appendNew<ArgumentRegValue>(proc, Origin(), GPRInfo::argumentGPR1),
+ root->appendNew<Const32Value>(proc, Origin(), 0)));
+ root->appendNew<Value>(proc, Return, Origin(), result);
+
+ auto code = compile(proc);
+ checkDisassembly(
+ *code,
+ [&] (const char* disassembly) -> bool {
+ return strstr(disassembly, "lea (%rdi,%rsi), %rax")
+ || strstr(disassembly, "lea (%rsi,%rdi), %rax");
+ },
+ "Expected to find something like lea (%rdi,%rsi), %rax but didn't!");
+ CHECK_EQ(invoke<intptr_t>(*code, 1, 2), 1 + 2);
+}
+
+void testX86LeaAddShlLeftScale2()
+{
+ // Add(Shl(@x, $c), @y)
+ Procedure proc;
+ BasicBlock* root = proc.addBlock();
+ Value* result = root->appendNew<Value>(
+ proc, Add, Origin(),
+ root->appendNew<ArgumentRegValue>(proc, Origin(), GPRInfo::argumentGPR0),
+ root->appendNew<Value>(
+ proc, Shl, Origin(),
+ root->appendNew<ArgumentRegValue>(proc, Origin(), GPRInfo::argumentGPR1),
+ root->appendNew<Const32Value>(proc, Origin(), 1)));
+ root->appendNew<Value>(proc, Return, Origin(), result);
+
+ auto code = compile(proc);
+ checkUsesInstruction(*code, "lea (%rdi,%rsi,2), %rax");
+ CHECK_EQ(invoke<intptr_t>(*code, 1, 2), 1 + (2 << 1));
+}
+
+void testX86LeaAddShlLeftScale4()
+{
+ // Add(Shl(@x, $c), @y)
+ Procedure proc;
+ BasicBlock* root = proc.addBlock();
+ Value* result = root->appendNew<Value>(
+ proc, Add, Origin(),
+ root->appendNew<Value>(
+ proc, Shl, Origin(),
+ root->appendNew<ArgumentRegValue>(proc, Origin(), GPRInfo::argumentGPR1),
+ root->appendNew<Const32Value>(proc, Origin(), 2)),
+ root->appendNew<ArgumentRegValue>(proc, Origin(), GPRInfo::argumentGPR0));
+ root->appendNew<Value>(proc, Return, Origin(), result);
+
+ auto code = compile(proc);
+ checkUsesInstruction(*code, "lea (%rdi,%rsi,4), %rax");
+ CHECK_EQ(invoke<intptr_t>(*code, 1, 2), 1 + (2 << 2));
+}
+
+void testX86LeaAddShlLeftScale8()
+{
+ // Add(Shl(@x, $c), @y)
+ Procedure proc;
+ BasicBlock* root = proc.addBlock();
+ Value* result = root->appendNew<Value>(
+ proc, Add, Origin(),
+ root->appendNew<ArgumentRegValue>(proc, Origin(), GPRInfo::argumentGPR0),
+ root->appendNew<Value>(
+ proc, Shl, Origin(),
+ root->appendNew<ArgumentRegValue>(proc, Origin(), GPRInfo::argumentGPR1),
+ root->appendNew<Const32Value>(proc, Origin(), 3)));
+ root->appendNew<Value>(proc, Return, Origin(), result);
+
+ auto code = compile(proc);
+ checkUsesInstruction(*code, "lea (%rdi,%rsi,8), %rax");
+ CHECK_EQ(invoke<intptr_t>(*code, 1, 2), 1 + (2 << 3));
+}
+
+void testAddShl32()
+{
+ // Add(Shl(@x, $c), @y)
+ Procedure proc;
+ BasicBlock* root = proc.addBlock();
+ Value* result = root->appendNew<Value>(
+ proc, Add, Origin(),
+ root->appendNew<ArgumentRegValue>(proc, Origin(), GPRInfo::argumentGPR0),
+ root->appendNew<Value>(
+ proc, Shl, Origin(),
+ root->appendNew<ArgumentRegValue>(proc, Origin(), GPRInfo::argumentGPR1),
+ root->appendNew<Const32Value>(proc, Origin(), 32)));
+ root->appendNew<Value>(proc, Return, Origin(), result);
+
+ auto code = compile(proc);
+ CHECK_EQ(invoke<intptr_t>(*code, 1, 2), 1 + (static_cast<intptr_t>(2) << static_cast<intptr_t>(32)));
+}
+
+void testAddShl64()
+{
+ // Add(Shl(@x, $c), @y)
+ Procedure proc;
+ BasicBlock* root = proc.addBlock();
+ Value* result = root->appendNew<Value>(
+ proc, Add, Origin(),
+ root->appendNew<ArgumentRegValue>(proc, Origin(), GPRInfo::argumentGPR0),
+ root->appendNew<Value>(
+ proc, Shl, Origin(),
+ root->appendNew<ArgumentRegValue>(proc, Origin(), GPRInfo::argumentGPR1),
+ root->appendNew<Const32Value>(proc, Origin(), 64)));
+ root->appendNew<Value>(proc, Return, Origin(), result);
+
+ auto code = compile(proc);
+ CHECK_EQ(invoke<intptr_t>(*code, 1, 2), 1 + 2);
+}
+
+void testAddShl65()
+{
+ // Add(Shl(@x, $c), @y)
+ Procedure proc;
+ BasicBlock* root = proc.addBlock();
+ Value* result = root->appendNew<Value>(
+ proc, Add, Origin(),
+ root->appendNew<ArgumentRegValue>(proc, Origin(), GPRInfo::argumentGPR0),
+ root->appendNew<Value>(
+ proc, Shl, Origin(),
+ root->appendNew<ArgumentRegValue>(proc, Origin(), GPRInfo::argumentGPR1),
+ root->appendNew<Const32Value>(proc, Origin(), 65)));
+ root->appendNew<Value>(proc, Return, Origin(), result);
+
+ auto code = compile(proc);
+ CHECK_EQ(invoke<intptr_t>(*code, 1, 2), 1 + (2 << 1));
+}
+
+void testReduceStrengthReassociation(bool flip)
+{
+ // Add(Add(@x, $c), @y) -> Add(Add(@x, @y), $c)
+ // and
+ // Add(@y, Add(@x, $c)) -> Add(Add(@x, @y), $c)
+ Procedure proc;
+ BasicBlock* root = proc.addBlock();
+ Value* arg1 = root->appendNew<ArgumentRegValue>(proc, Origin(), GPRInfo::argumentGPR0);
+ Value* arg2 = root->appendNew<ArgumentRegValue>(proc, Origin(), GPRInfo::argumentGPR1);
+
+ Value* innerAdd = root->appendNew<Value>(
+ proc, Add, Origin(), arg1,
+ root->appendNew<ConstPtrValue>(proc, Origin(), 42));
+
+ Value* outerAdd;
+ if (flip)
+ outerAdd = root->appendNew<Value>(proc, Add, Origin(), arg2, innerAdd);
+ else
+ outerAdd = root->appendNew<Value>(proc, Add, Origin(), innerAdd, arg2);
+
+ root->appendNew<Value>(proc, Return, Origin(), outerAdd);
+
+ proc.resetReachability();
+
+ if (shouldBeVerbose()) {
+ dataLog("IR before reduceStrength:\n");
+ dataLog(proc);
+ }
+
+ reduceStrength(proc);
+
+ if (shouldBeVerbose()) {
+ dataLog("IR after reduceStrength:\n");
+ dataLog(proc);
+ }
+
+ CHECK_EQ(root->last()->opcode(), Return);
+ CHECK_EQ(root->last()->child(0)->opcode(), Add);
+ CHECK(root->last()->child(0)->child(1)->isIntPtr(42));
+ CHECK_EQ(root->last()->child(0)->child(0)->opcode(), Add);
+ CHECK(
+ (root->last()->child(0)->child(0)->child(0) == arg1 && root->last()->child(0)->child(0)->child(1) == arg2) ||
+ (root->last()->child(0)->child(0)->child(0) == arg2 && root->last()->child(0)->child(0)->child(1) == arg1));
+}
+
// Make sure the compiler does not try to optimize anything out.
NEVER_INLINE double zero()
{
@@ -14801,6 +15079,23 @@
RUN(testPatchpointTerminalReturnValue(true));
RUN(testPatchpointTerminalReturnValue(false));
+ RUN(testMemoryFence());
+ RUN(testStoreFence());
+ RUN(testLoadFence());
+ RUN(testTrappingLoad());
+ RUN(testTrappingStore());
+ RUN(testTrappingLoadAddStore());
+ RUN(testTrappingLoadDCE());
+ RUN(testTrappingStoreElimination());
+ RUN(testMoveConstants());
+ RUN(testPCOriginMapDoesntInsertNops());
+ RUN(testPinRegisters());
+ RUN(testReduceStrengthReassociation(true));
+ RUN(testReduceStrengthReassociation(false));
+ RUN(testAddShl32());
+ RUN(testAddShl64());
+ RUN(testAddShl65());
+
if (isX86()) {
RUN(testBranchBitAndImmFusion(Identity, Int64, 1, Air::BranchTest32, Air::Arg::Tmp));
RUN(testBranchBitAndImmFusion(Identity, Int64, 0xff, Air::BranchTest32, Air::Arg::Tmp));
@@ -14810,7 +15105,14 @@
RUN(testBranchBitAndImmFusion(Load8Z, Int32, 1, Air::BranchTest8, Air::Arg::Addr));
RUN(testBranchBitAndImmFusion(Load, Int32, 1, Air::BranchTest32, Air::Arg::Addr));
RUN(testBranchBitAndImmFusion(Load, Int64, 1, Air::BranchTest32, Air::Arg::Addr));
-
+ RUN(testX86LeaAddAddShlLeft());
+ RUN(testX86LeaAddAddShlRight());
+ RUN(testX86LeaAddAdd());
+ RUN(testX86LeaAddShlRight());
+ RUN(testX86LeaAddShlLeftScale1());
+ RUN(testX86LeaAddShlLeftScale2());
+ RUN(testX86LeaAddShlLeftScale4());
+ RUN(testX86LeaAddShlLeftScale8());
}
if (isARM64()) {
@@ -14818,18 +15120,6 @@
RUN(testTernarySubInstructionSelection(Trunc, Int32, Air::Sub32));
}
- RUN(testMemoryFence());
- RUN(testStoreFence());
- RUN(testLoadFence());
- RUN(testTrappingLoad());
- RUN(testTrappingStore());
- RUN(testTrappingLoadAddStore());
- RUN(testTrappingLoadDCE());
- RUN(testTrappingStoreElimination());
- RUN(testMoveConstants());
- RUN(testPCOriginMapDoesntInsertNops());
- RUN(testPinRegisters());
-
if (tasks.isEmpty())
usage();
