Only emit empty inline asm instructions for cond_fail when optimizing.

These inline asm instructions are used to block branch folding when
optimizing, to ensure that we have unique trap locations (and associated
debug info) for each cond_fail.

We only need these when we're optimizing, and emitting them can block
fast isel, so only emit them when optimizing.

Author: rudkx

lib/IRGen/IRGenSIL.cpp

@@ -4629,19 +4629,21 @@ void IRGenSILFunction::visitCondFailInst(swift::CondFailInst *i) {
   Builder.CreateCondBr(cond, failBB, contBB);
   Builder.emitBlock(failBB);
 
-  // Emit unique side-effecting inline asm calls in order to eliminate
-  // the possibility that an LLVM optimization or code generation pass
-  // will merge these blocks back together again. We emit an empty asm
-  // string with the side-effect flag set, and with a unique integer
-  // argument for each cond_fail we see in the function.
-  llvm::IntegerType *asmArgTy = IGM.Int32Ty;
-  llvm::Type *argTys = { asmArgTy };
-  llvm::FunctionType *asmFnTy =
-    llvm::FunctionType::get(IGM.VoidTy, argTys, false /* = isVarArg */);
-  llvm::InlineAsm *inlineAsm =
-    llvm::InlineAsm::get(asmFnTy, "", "n", true /* = SideEffects */);
-  Builder.CreateCall(inlineAsm,
-                     llvm::ConstantInt::get(asmArgTy, NumCondFails++));
+  if (IGM.IRGen.Opts.Optimize) {
+    // Emit unique side-effecting inline asm calls in order to eliminate
+    // the possibility that an LLVM optimization or code generation pass
+    // will merge these blocks back together again. We emit an empty asm
+    // string with the side-effect flag set, and with a unique integer
+    // argument for each cond_fail we see in the function.
+    llvm::IntegerType *asmArgTy = IGM.Int32Ty;
+    llvm::Type *argTys = { asmArgTy };
+    llvm::FunctionType *asmFnTy =
+      llvm::FunctionType::get(IGM.VoidTy, argTys, false /* = isVarArg */);
+    llvm::InlineAsm *inlineAsm =
+      llvm::InlineAsm::get(asmFnTy, "", "n", true /* = SideEffects */);
+    Builder.CreateCall(inlineAsm,
+                       llvm::ConstantInt::get(asmArgTy, NumCondFails++));
+  }
 
   // Emit the trap instruction.
   llvm::Function *trapIntrinsic =

test/IRGen/condfail.sil

@@ -1,12 +1,21 @@
-// RUN: %target-swift-frontend -primary-file %s -O -g -S  | FileCheck %s --check-prefix=CHECK-%target-cpu
+// RUN: %target-swift-frontend -primary-file %s -O -g -S  | FileCheck %s --check-prefix CHECK --check-prefix CHECK-%target-cpu --check-prefix CHECK-OPT-%target-os
+// RUN: %target-swift-frontend -primary-file %s -g -S  | FileCheck %s --check-prefix CHECK --check-prefix CHECK-%target-cpu --check-prefix CHECK-NOOPT-%target-os
 
 import Builtin
 import Swift
 
 // Make sure we emit two traps.
 
-// CHECK-LABEL:       _test_cond_fail:
+// CHECK-LABEL:       test_cond_fail:
 // CHECK:             .cfi_startproc
+// CHECK-OPT-macosx:  ## InlineAsm Start
+// CHECK-OPT-macosx:  ## InlineAsm End
+// CHECK-OPT-linux:   ##APP
+// CHECK-OPT-linux:   ##NO_APP
+// CHECK-NOOPT-macosx-NOT: ## InlineAsm Start
+// CHECK-NOOPT-macosx-NOT: ## InlineAsm End
+// CHECK-NOOPT-linux-NOT:  ##APP
+// CHECK-NOOPT-linux-NOT:  ##NO_APP
 // CHECK-x86_64:      ud2
 // CHECK-i386:        ud2
 // CHECK-arm64:       brk
@@ -16,6 +25,14 @@ import Swift
 // CHECK-powerpc64le: trap
 // CHECK-s390x:       trap
 // CHECK-NOT:         .cfi_endproc
+// CHECK-OPT-macosx:  ## InlineAsm Start
+// CHECK-OPT-macosx:  ## InlineAsm End
+// CHECK-OPT-linux:   ##APP
+// CHECK-OPT-linux:   ##NO_APP
+// CHECK-NOOPT-macosx-NOT: ## InlineAsm Start
+// CHECK-NOOPT-macosx-NOT: ## InlineAsm End
+// CHECK-NOOPT-linux-NOT:  ##APP
+// CHECK-NOOPT-linux-NOT:  ##NO_APP
 // CHECK-x86_64:      ud2
 // CHECK-i386:        ud2
 // CHECK-arm64:       brk