Merge pull request swiftlang#34115 from eeckstein/enum-optimizations

eeckstein · web-flow · commit f6c008fb48aa · 2020-09-30T20:44:27.000+02:00
SILOptimizer: two optimization improvements for address-only enums
diff --git a/lib/SILOptimizer/SILCombiner/SILCombinerMiscVisitors.cpp b/lib/SILOptimizer/SILCombiner/SILCombinerMiscVisitors.cpp
@@ -453,6 +453,10 @@ bool SILCombiner::optimizeStackAllocatedEnum(AllocStackInst *AS) {
     return false;
   
   EnumElementDecl *element = nullptr;
+  unsigned numInits =0;
+  unsigned numTakes = 0;
+  SILBasicBlock *initBlock = nullptr;
+  SILBasicBlock *takeBlock = nullptr;
   SILType payloadType;
   
   // First step: check if the stack location is only used to hold one specific
@@ -463,6 +467,8 @@ bool SILCombiner::optimizeStackAllocatedEnum(AllocStackInst *AS) {
       case SILInstructionKind::DebugValueAddrInst:
       case SILInstructionKind::DestroyAddrInst:
       case SILInstructionKind::DeallocStackInst:
+      case SILInstructionKind::InjectEnumAddrInst:
+        // We'll check init_enum_addr below.
         break;
       case SILInstructionKind::InitEnumDataAddrInst: {
         auto *ieda = cast<InitEnumDataAddrInst>(user);
@@ -472,20 +478,17 @@ bool SILCombiner::optimizeStackAllocatedEnum(AllocStackInst *AS) {
         element = el;
         assert(!payloadType || payloadType == ieda->getType());
         payloadType = ieda->getType();
-        break;
-      }
-      case SILInstructionKind::InjectEnumAddrInst: {
-        auto *el = cast<InjectEnumAddrInst>(user)->getElement();
-        if (element && el != element)
-          return false;
-        element = el;
+        numInits++;
+        initBlock = user->getParent();
         break;
       }
       case SILInstructionKind::UncheckedTakeEnumDataAddrInst: {
         auto *el = cast<UncheckedTakeEnumDataAddrInst>(user)->getElement();
         if (element && el != element)
           return false;
         element = el;
+        numTakes++;
+        takeBlock = user->getParent();
         break;
       }
       default:
@@ -495,6 +498,24 @@ bool SILCombiner::optimizeStackAllocatedEnum(AllocStackInst *AS) {
   if (!element || !payloadType)
     return false;
 
+  // If the enum has a single init-take pair in a single block, we know that
+  // the enum cannot contain any valid payload outside that init-take pair.
+  //
+  // This also means that we can ignore any inject_enum_addr of another enum
+  // case, because this can only inject a case without a payload.
+  bool singleInitTakePair =
+    (numInits == 1 && numTakes == 1 && initBlock == takeBlock);
+  if (!singleInitTakePair) {
+    // No single init-take pair: We cannot ignore inject_enum_addrs with a
+    // mismatching case.
+    for (auto *use : AS->getUses()) {
+      if (auto *inject = dyn_cast<InjectEnumAddrInst>(use->getUser())) {
+        if (inject->getElement() != element)
+          return false;
+      }
+    }
+  }
+
   // Second step: replace the enum alloc_stack with a payload alloc_stack.
   auto *newAlloc = Builder.createAllocStack(
       AS->getLoc(), payloadType, AS->getVarInfo(), AS->hasDynamicLifetime());
@@ -508,6 +529,22 @@ bool SILCombiner::optimizeStackAllocatedEnum(AllocStackInst *AS) {
         eraseInstFromFunction(*user);
         break;
       case SILInstructionKind::DestroyAddrInst:
+        if (singleInitTakePair) {
+          // It's not possible that the enum has a payload at the destroy_addr,
+          // because it must have already been taken by the take of the
+          // single init-take pair.
+          // We _have_ to remove the destroy_addr, because we also remove all
+          // inject_enum_addrs which might inject a payload-less case before
+          // the destroy_addr.
+          eraseInstFromFunction(*user);
+        } else {
+          // The enum payload can still be valid at the destroy_addr, so we have
+          // to keep the destroy_addr. Just replace the enum with the payload
+          // (and because it's not a singleInitTakePair, we can be sure that the
+          // enum cannot have any other case than the payload case).
+          use->set(newAlloc);
+        }
+        break;
       case SILInstructionKind::DeallocStackInst:
         use->set(newAlloc);
         break;
diff --git a/lib/SILOptimizer/Transforms/SimplifyCFG.cpp b/lib/SILOptimizer/Transforms/SimplifyCFG.cpp
@@ -916,16 +916,57 @@ static bool couldRemoveRelease(SILBasicBlock *SrcBB, SILValue SrcV,
   return IsReleaseOfDest;
 }
 
+/// Returns true if any instruction in \p block may write memory.
+static bool blockMayWriteMemory(SILBasicBlock *block) {
+  for (auto instAndIdx : llvm::enumerate(*block)) {
+    if (instAndIdx.value().mayWriteToMemory())
+      return true;
+    // Only look at the first 20 instructions to avoid compile time problems for
+    // corner cases of very large blocks without memory writes.
+    // 20 instructions is more than enough.
+    if (instAndIdx.index() > 50)
+      return true;
+  }
+  return false;
+}
+
+// Returns true if \p block contains an injected an enum case into \p enumAddr
+// which is valid at the end of the block.
+static bool hasInjectedEnumAtEndOfBlock(SILBasicBlock *block, SILValue enumAddr) {
+  for (auto instAndIdx : llvm::enumerate(llvm::reverse(*block))) {
+    SILInstruction &inst = instAndIdx.value();
+    if (auto *injectInst = dyn_cast<InjectEnumAddrInst>(&inst)) {
+      return injectInst->getOperand() == enumAddr;
+    }
+    if (inst.mayWriteToMemory())
+      return false;
+    // Only look at the first 20 instructions to avoid compile time problems for
+    // corner cases of very large blocks without memory writes.
+    // 20 instructions is more than enough.
+    if (instAndIdx.index() > 50)
+      return false;
+  }
+  return false;
+}
+
 /// tryJumpThreading - Check to see if it looks profitable to duplicate the
 /// destination of an unconditional jump into the bottom of this block.
 bool SimplifyCFG::tryJumpThreading(BranchInst *BI) {
   auto *DestBB = BI->getDestBB();
   auto *SrcBB = BI->getParent();
+  TermInst *destTerminator = DestBB->getTerminator();
   // If the destination block ends with a return, we don't want to duplicate it.
   // We want to maintain the canonical form of a single return where possible.
-  if (DestBB->getTerminator()->isFunctionExiting())
+  if (destTerminator->isFunctionExiting())
     return false;
 
+  // Jump threading only makes sense if there is an argument on the branch
+  // (which is reacted on in the DestBB), or if this goes through a memory
+  // location (switch_enum_addr is the only adress-instruction which we
+  // currently handle).
+  if (BI->getArgs().empty() && !isa<SwitchEnumAddrInst>(destTerminator))
+    return false;
+      
   // We don't have a great cost model at the SIL level, so we don't want to
   // blissly duplicate tons of code with a goal of improved performance (we'll
   // leave that to LLVM).  However, doing limited code duplication can lead to
@@ -956,11 +997,29 @@ bool SimplifyCFG::tryJumpThreading(BranchInst *BI) {
     }
   }
 
-  if (ThreadingBudget == 0 && isa<CondBranchInst>(DestBB->getTerminator())) {
-    for (auto V : BI->getArgs()) {
-      if (isa<IntegerLiteralInst>(V) || isa<FloatLiteralInst>(V)) {
+  if (ThreadingBudget == 0) {
+    if (isa<CondBranchInst>(destTerminator)) {
+      for (auto V : BI->getArgs()) {
+        if (isa<IntegerLiteralInst>(V) || isa<FloatLiteralInst>(V)) {
+          ThreadingBudget = 4;
+          break;
+        }
+      }
+    } else if (auto *SEA = dyn_cast<SwitchEnumAddrInst>(destTerminator)) {
+      // If the branch-block injects a certain enum case and the destination
+      // switches on that enum, it's worth jump threading. E.g.
+      //
+      //   inject_enum_addr %enum : $*Optional<T>, #Optional.some
+      //   ... // no memory writes here
+      //   br DestBB
+      // DestBB:
+      //   ... // no memory writes here
+      //   switch_enum_addr %enum : $*Optional<T>, case #Optional.some ...
+      //
+      SILValue enumAddr = SEA->getOperand();
+      if (!blockMayWriteMemory(DestBB) &&
+          hasInjectedEnumAtEndOfBlock(SrcBB, enumAddr)) {
         ThreadingBudget = 4;
-        break;
       }
     }
   }
@@ -976,7 +1035,7 @@ bool SimplifyCFG::tryJumpThreading(BranchInst *BI) {
   // control flow. Still, we make an exception for switch_enum.
   bool DestIsLoopHeader = (LoopHeaders.count(DestBB) != 0);
   if (DestIsLoopHeader) {
-    if (!isa<SwitchEnumInst>(DestBB->getTerminator()))
+    if (!isa<SwitchEnumInst>(destTerminator))
       return false;
   }
 
@@ -1286,8 +1345,7 @@ bool SimplifyCFG::simplifyBranchBlock(BranchInst *BI) {
   // If this unconditional branch has BBArgs, check to see if duplicating the
   // destination would allow it to be simplified.  This is a simple form of jump
   // threading.
-  if (!isVeryLargeFunction && !BI->getArgs().empty() &&
-      tryJumpThreading(BI))
+  if (!isVeryLargeFunction && tryJumpThreading(BI))
     return true;
 
   return Simplified;
diff --git a/test/SILOptimizer/cast_optimizer_conditional_conformance.sil b/test/SILOptimizer/cast_optimizer_conditional_conformance.sil
@@ -98,8 +98,6 @@ bb6:
 // CHECK: [[IEDA:%.*]] = init_enum_data_addr [[OPT]] : $*Optional<HasFoo>, #Optional.some!enumelt
 // CHECK: checked_cast_addr_br take_always S1<Int8> in [[VAL]] : $*S1<Int8> to HasFoo in [[IEDA]] : $*HasFoo, bb1, bb2
 // bbs...
-// CHECK: switch_enum_addr [[OPT]] : $*Optional<HasFoo>, case #Optional.some!enumelt: bb4, case #Optional.none!enumelt: bb5
-// bbs...
 // CHECK: [[UNWRAP:%.*]] = unchecked_take_enum_data_addr [[OPT]] : $*Optional<HasFoo>, #Optional.some!enumelt
 // CHECK: copy_addr [take] [[UNWRAP]] to [initialization] [[EXIS]] : $*HasFoo
 // CHECK: [[OPEN:%.*]] = open_existential_addr immutable_access [[EXIS]] : $*HasFoo to $*@opened("4E16CBC0-FD9F-11E8-A311-D0817AD9F6DD") HasFoo
diff --git a/test/SILOptimizer/enum_jump_thread.sil b/test/SILOptimizer/enum_jump_thread.sil
@@ -49,3 +49,83 @@ bb5(%7 : $E2):
   return %7 : $E2
 }
 
+// CHECK-LABEL: sil @test_enum_addr
+sil @test_enum_addr : $@convention(thin) () -> Builtin.Int32 {
+bb0:
+  %2 = alloc_stack $E
+  cond_br undef, bb1, bb2
+
+// CHECK:       bb1:
+// CHECK-NEXT:    inject_enum_addr
+// CHECK-NEXT:    switch_enum_addr
+bb1:
+  inject_enum_addr %2 : $*E, #E.A!enumelt
+  br bb3
+
+// CHECK:       bb2:
+// CHECK-NEXT:    inject_enum_addr
+// CHECK-NEXT:    switch_enum_addr
+bb2:
+  inject_enum_addr %2 : $*E, #E.B!enumelt
+  br bb3
+
+bb3:
+  switch_enum_addr %2 : $*E, case #E.A!enumelt: bb4, case #E.B!enumelt: bb5
+
+bb4:
+  %10 = integer_literal $Builtin.Int32, 1
+  br bb6(%10 : $Builtin.Int32)
+
+bb5:
+  %11 = integer_literal $Builtin.Int32, 2
+  br bb6(%11 : $Builtin.Int32)
+
+bb6(%12 : $Builtin.Int32):
+  dealloc_stack %2 : $*E
+  return %12 : $Builtin.Int32
+// CHECK: } // end sil function 'test_enum_addr'
+}
+
+// CHECK-LABEL: sil @dont_jumpthread_enum_addr
+sil @dont_jumpthread_enum_addr : $@convention(thin) (E) -> Builtin.Int32 {
+bb0(%0 : $E):
+  %2 = alloc_stack $E
+  %3 = alloc_stack $E
+  cond_br undef, bb1, bb2
+
+// CHECK:       bb1:
+// CHECK-NEXT:    inject_enum_addr
+// CHECK-NEXT:    store
+// CHECK-NEXT:    br bb3
+bb1:
+  inject_enum_addr %2 : $*E, #E.A!enumelt
+  store %0 to %2 : $*E
+  br bb3
+
+// CHECK:       bb2:
+// CHECK-NEXT:    inject_enum_addr
+// CHECK-NEXT:    br bb3
+bb2:
+  inject_enum_addr %3 : $*E, #E.A!enumelt
+  br bb3
+
+// CHECK:       bb3:
+// CHECK-NEXT:    switch_enum_addr
+bb3:
+  switch_enum_addr %2 : $*E, case #E.A!enumelt: bb4, case #E.B!enumelt: bb5
+
+bb4:
+  %10 = integer_literal $Builtin.Int32, 1
+  br bb6(%10 : $Builtin.Int32)
+
+bb5:
+  %11 = integer_literal $Builtin.Int32, 2
+  br bb6(%11 : $Builtin.Int32)
+
+bb6(%12 : $Builtin.Int32):
+  dealloc_stack %3 : $*E
+  dealloc_stack %2 : $*E
+  return %12 : $Builtin.Int32
+// CHECK: } // end sil function 'dont_jumpthread_enum_addr'
+}
+
diff --git a/test/SILOptimizer/generic_loop.swift b/test/SILOptimizer/generic_loop.swift
@@ -0,0 +1,17 @@
+// RUN: %target-swift-frontend -primary-file %s -O -sil-verify-all -module-name=test -emit-sil | %FileCheck %s
+
+// REQUIRES: swift_stdlib_no_asserts,optimized_stdlib
+
+
+// Check that we can eliminate all optionals from a loop which is iterating
+// over an array of address-only elements.
+
+// CHECK-LABEL: sil @$s4test0A18_no_optionals_usedyySayxGlF : $@convention(thin) <T> (@guaranteed Array<T>) -> () {
+// CHECK-NOT: Optional
+// CHECK: } // end sil function '$s4test0A18_no_optionals_usedyySayxGlF'
+public func test_no_optionals_used<T>(_ items: [T]) {
+  for i in items {
+    print(i)
+  }
+}
+
diff --git a/test/SILOptimizer/sil_combine_concrete_existential.swift b/test/SILOptimizer/sil_combine_concrete_existential.swift
@@ -95,10 +95,11 @@ struct SS: PPP {
 
 // The first apply has been devirtualized and inlined. The second remains unspecialized.
 // CHECK-LABEL: sil @$s32sil_combine_concrete_existential37testWitnessReturnOptionalIndirectSelfyyF : $@convention(thin) () -> () {
+// CHECK: [[O1:%.*]] = open_existential_addr immutable_access %{{.*}} : $*PPP to $*@opened("{{.*}}") PPP
 // CHECK: switch_enum_addr %{{.*}} : $*Optional<@opened("{{.*}}") PPP>, case #Optional.some!enumelt: bb{{.*}}, case #Optional.none!enumelt: bb{{.*}}
-// CHECK: [[O:%.*]] = open_existential_addr immutable_access %{{.*}} : $*PPP to $*@opened("{{.*}}") PPP
-// CHECK: [[W:%.*]] = witness_method $@opened("{{.*}}") PPP, #PPP.returnsOptionalIndirect : <Self where Self : PPP> (Self) -> () -> Self?, [[O]] : $*@opened("{{.*}}") PPP : $@convention(witness_method: PPP) <τ_0_0 where τ_0_0 : PPP> (@in_guaranteed τ_0_0) -> @out Optional<τ_0_0>
-// CHECK: apply [[W]]<@opened("{{.*}}") PPP>(%{{.*}}, [[O]]) : $@convention(witness_method: PPP) <τ_0_0 where τ_0_0 : PPP> (@in_guaranteed τ_0_0) -> @out Optional<τ_0_0>
+// CHECK: [[O2:%.*]] = open_existential_addr immutable_access %{{.*}} : $*PPP to $*@opened("{{.*}}") PPP
+// CHECK: [[W:%.*]] = witness_method $@opened("{{.*}}") PPP, #PPP.returnsOptionalIndirect : <Self where Self : PPP> (Self) -> () -> Self?, [[O1]] : $*@opened("{{.*}}") PPP : $@convention(witness_method: PPP) <τ_0_0 where τ_0_0 : PPP> (@in_guaranteed τ_0_0) -> @out Optional<τ_0_0>
+// CHECK: apply [[W]]<@opened("{{.*}}") PPP>(%{{.*}}, [[O2]]) : $@convention(witness_method: PPP) <τ_0_0 where τ_0_0 : PPP> (@in_guaranteed τ_0_0) -> @out Optional<τ_0_0>
 // CHECK-LABEL: } // end sil function '$s32sil_combine_concrete_existential37testWitnessReturnOptionalIndirectSelfyyF'
 public func testWitnessReturnOptionalIndirectSelf() {
   let p: PPP = S()
diff --git a/test/SILOptimizer/sil_combine_enums.sil b/test/SILOptimizer/sil_combine_enums.sil
