[5.3] IRGen: Always eliminate frame pointers of leaf functions

Explanation: This change makes it such that we don't emit frame pointers
in leaf functions leading to decreased code size and improved
performance.

Scope: Should not visibly affect Swift programs other than making them
faster.

Risk: Medium. There might be unknown dependencies on having the frame
pointer available in leaf functions.

Testing: Unit test added

Reviewer: Andrew Trick

rdar://20933449

Author: aschwaighofer

include/swift/AST/IRGenOptions.h

@@ -194,6 +194,7 @@ class IRGenOptions {
   unsigned DisableLLVMSLPVectorizer : 1;
 
   /// Disable frame pointer elimination?
+  unsigned DisableFPElimLeaf : 1;
   unsigned DisableFPElim : 1;
 
   /// Special codegen for playgrounds.
@@ -322,6 +323,7 @@ class IRGenOptions {
         DisableClangModuleSkeletonCUs(false), UseJIT(false),
         IntegratedREPL(false), DisableLLVMOptzns(false),
         DisableSwiftSpecificLLVMOptzns(false), DisableLLVMSLPVectorizer(false),
+        DisableFPElimLeaf(false),
         DisableFPElim(true), Playground(false), EmitStackPromotionChecks(false),
         FunctionSections(false), PrintInlineTree(false), EmbedMode(IRGenEmbedMode::None),
         HasValueNamesSetting(false), ValueNames(false),

include/swift/Option/Options.td

@@ -1012,6 +1012,10 @@ def enable_private_imports : Flag<["-"], "enable-private-imports">,
   Flags<[FrontendOption, NoInteractiveOption, HelpHidden]>,
   HelpText<"Allows this module's internal and private API to be accessed">;
 
+def disable_leaf_frame_pointer_elim : Flag<["-"], "no-omit-leaf-frame-pointer">,
+  Flags<[FrontendOption, NoInteractiveOption, HelpHidden]>,
+  HelpText<"Don't omit the frame pointer for leaf functions">;
+
 def sanitize_EQ : CommaJoined<["-"], "sanitize=">,
   Flags<[FrontendOption, NoInteractiveOption]>, MetaVarName<"<check>">,
   HelpText<"Turn on runtime checks for erroneous behavior.">;

lib/Driver/ToolChains.cpp

@@ -289,6 +289,8 @@ void ToolChain::addCommonFrontendArgs(const OutputInfo &OI,
       arguments.push_back("-enable-anonymous-context-mangled-names");
   }
 
+  inputArgs.AddLastArg(arguments, options::OPT_disable_leaf_frame_pointer_elim);
+
   // Pass through any subsystem flags.
   inputArgs.AddAllArgs(arguments, options::OPT_Xllvm);
   inputArgs.AddAllArgs(arguments, options::OPT_Xcc);

lib/Frontend/CompilerInvocation.cpp

@@ -1477,6 +1477,9 @@ static bool ParseIRGenArgs(IRGenOptions &Opts, ArgList &Args,
         getRuntimeCompatVersion();
   }
 
+  if (Args.hasArg(OPT_disable_leaf_frame_pointer_elim))
+    Opts.DisableFPElimLeaf = true;
+
   return false;
 }
 

lib/IRGen/IRGenModule.cpp

@@ -89,6 +89,20 @@ static llvm::PointerType *createStructPointerType(IRGenModule &IGM,
   return createStructType(IGM, name, types)->getPointerTo(DefaultAS);
 };
 
+static clang::CodeGenOptions::FramePointerKind
+shouldUseFramePointer(const IRGenOptions &Opts) {
+  if (Opts.DisableFPElim) {
+    // General frame pointer elimination is disabled.
+    // Should we at least eliminate in leaf functions?
+    return Opts.DisableFPElimLeaf
+               ? clang::CodeGenOptions::FramePointerKind::All
+               : clang::CodeGenOptions::FramePointerKind::NonLeaf;
+  }
+
+  return clang::CodeGenOptions::FramePointerKind::None;
+
+}
+
 static clang::CodeGenerator *createClangCodeGenerator(ASTContext &Context,
                                                  llvm::LLVMContext &LLVMContext,
                                                       const IRGenOptions &Opts,
@@ -101,9 +115,7 @@ static clang::CodeGenerator *createClangCodeGenerator(ASTContext &Context,
 
   auto &CGO = Importer->getClangCodeGenOpts();
   CGO.OptimizationLevel = Opts.shouldOptimize() ? 3 : 0;
-  CGO.setFramePointer(Opts.DisableFPElim
-                          ? clang::CodeGenOptions::FramePointerKind::All
-                          : clang::CodeGenOptions::FramePointerKind::None);
+  CGO.setFramePointer(shouldUseFramePointer(Opts));
   CGO.DiscardValueNames = !Opts.shouldProvideValueNames();
   switch (Opts.DebugInfoLevel) {
   case IRGenDebugInfoLevel::None:
@@ -983,7 +995,8 @@ bool swift::irgen::shouldRemoveTargetFeature(StringRef feature) {
 
 void IRGenModule::setHasFramePointer(llvm::AttrBuilder &Attrs,
                                      bool HasFramePointer) {
-  Attrs.addAttribute("frame-pointer", HasFramePointer ? "all" : "none");
+  auto UseFramePointer = IRGen.Opts.DisableFPElimLeaf ? "all" : "non-leaf";
+  Attrs.addAttribute("frame-pointer", HasFramePointer ? UseFramePointer : "none");
 }
 
 void IRGenModule::setHasFramePointer(llvm::Function *F,

test/IRGen/c_globals.swift

@@ -32,4 +32,4 @@ public func testCaptureGlobal() {
 }
 
 // CHECK-DAG: attributes [[CLANG_FUNC_ATTR]] = { noinline nounwind {{.*}}"frame-pointer"="all"{{.*}}
-// CHECK-DAG: attributes [[SWIFT_FUNC_ATTR]] = { "frame-pointer"="all" {{.*}}"target-cpu"
+// CHECK-DAG: attributes [[SWIFT_FUNC_ATTR]] = { {{.*}}"frame-pointer"="non-leaf" {{.*}}"target-cpu"

test/IRGen/framepointer.sil

@@ -0,0 +1,62 @@
+// RUN: %target-swift-frontend -primary-file %s -emit-ir | %FileCheck %s --check-prefix=CHECK
+// RUN: %target-swift-frontend -primary-file %s -emit-ir -no-omit-leaf-frame-pointer| %FileCheck %s --check-prefix=CHECK-ALL
+// RUN: %target-swift-frontend -primary-file %s -S | %FileCheck %s  --check-prefix=CHECKASM --check-prefix=CHECKASM-%target-os-%target-cpu
+
+sil_stage canonical
+
+import Swift
+
+sil @leaf_function_no_frame_pointer : $@convention(thin) (Int32) -> Int32 {
+entry(%i : $Int32):
+  return %i : $Int32
+}
+
+sil @non_leaf_function_with_frame_pointer : $@convention(thin) (Int32) -> Int32 {
+entry(%i : $Int32):
+  %f = function_ref @leaf_function_no_frame_pointer : $@convention(thin) (Int32) -> Int32
+  %r = apply %f(%i) : $@convention(thin) (Int32) -> Int32
+  return %r : $Int32
+}
+
+// CHECK: define{{.*}} swiftcc i32 @leaf_function_no_frame_pointer(i32 %0) [[ATTR:#.*]] {
+// CHECK: entry:
+// CHECK:   ret i32 %0
+// CHECK: }
+
+// CHECK: define{{.*}} swiftcc i32 @non_leaf_function_with_frame_pointer(i32 %0) [[ATTR]] {
+// CHECK: entry:
+// CHECK:   %1 = call swiftcc i32 @leaf_function_no_frame_pointer(i32 %0)
+// CHECK:   ret i32 %1
+// CHECK: }
+
+// CHECK: attributes [[ATTR]] = {{{.*}}"frame-pointer"="non-leaf"
+
+// CHECK-ALL: define{{.*}} swiftcc i32 @leaf_function_no_frame_pointer(i32 %0) [[ATTR:#.*]] {
+// CHECK-ALL: entry:
+// CHECK-ALL:   ret i32 %0
+// CHECK-ALL: }
+
+// CHECK-ALL: define{{.*}} swiftcc i32 @non_leaf_function_with_frame_pointer(i32 %0) [[ATTR]] {
+// CHECK-ALL: entry:
+// CHECK-ALL:   %1 = call swiftcc i32 @leaf_function_no_frame_pointer(i32 %0)
+// CHECK-ALL:   ret i32 %1
+// CHECK-ALL: }
+
+// CHECK-ALL: attributes [[ATTR]] = {{{.*}}"frame-pointer"="all"
+
+// Silence other os-archs.
+// CHECKASM: {{.*}}
+
+// CHECKASM-macosx-x86_64-LABEL: _leaf_function_no_frame_pointer:
+// CHECKASM-macosx-x86_64-NOT: push
+// CHECKASM-macosx-x86_64: movl    %edi, %eax
+// CHECKASM-macosx-x86_64-NOT: pop
+// CHECKASM-macosx-x86_64: ret
+
+
+// CHECKASM-macosx-x86_64-LABEL: _non_leaf_function_with_frame_pointer:
+// CHECKASM-macosx-x86_64:   pushq %rbp
+// CHECKASM-macosx-x86_64:   movq %rsp, %rbp
+// CHECKASM-macosx-x86_64:   callq _leaf_function_no_frame_pointer
+// CHECKASM-macosx-x86_64:   popq %rbp
+// CHECKASM-macosx-x86_64:   ret

test/Misc/tbi.sil

@@ -19,19 +19,13 @@
 
 // NO_TBI-LABEL: .globl  _testTBI
 // NO_TBI: _testTBI
-// NO_TBI-NEXT: stp
-// NO_TBI-NEXT: mov
 // NO_TBI-NEXT: and
 // NO_TBI-NEXT: ldr
-// NO_TBI-NEXT: ldp
 // NO_TBI-NEXT: ret
 
 // TBI-LABEL: .globl  _testTBI
 // TBI: _testTBI:
-// TBI-NEXT: stp
-// TBI-NEXT: mov
 // TBI-NEXT: ldr
-// TBI-NEXT: ldp
 // TBI-NEXT: ret
 
 sil_stage canonical