[ARM] Always lower direct calls as direct when the outliner is enabled (#66434)

The indirect lowering hinders the outliner's ability to see that
sequences are in fact common, since the sequence similarity is rendered
opaque by the register callee. The size savings from making them
indirect seems to be dwarfed by the outliner's savings from
de-duplication.

rdar://115178034
rdar://115459865

Author: jroelofs

llvm/lib/CodeGen/TargetPassConfig.cpp

@@ -1241,6 +1241,9 @@ void TargetPassConfig::addMachinePasses() {
   addPass(&LiveDebugValuesID);
   addPass(&MachineSanitizerBinaryMetadataID);
 
+  if (EnableMachineOutliner == RunOutliner::NeverOutline)
+    TM->Options.EnableMachineOutliner = false;
+
   if (TM->Options.EnableMachineOutliner &&
       getOptLevel() != CodeGenOptLevel::None &&
       EnableMachineOutliner != RunOutliner::NeverOutline) {

llvm/lib/Target/ARM/ARMISelLowering.cpp

@@ -2395,6 +2395,14 @@ ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
                          return isa<Instruction>(U) &&
                                 cast<Instruction>(U)->getParent() == BB;
                        }) > 2;
+      // The indirect call lowering hinders the MachineOutliner's ability to
+      // recognize common sequences. The resulting indirect calls all have the
+      // same target, but the outliner can't tell this a priori, since the
+      // branch target is turned into a register operand, and those can't (yet?)
+      // be assumed to have the same value at runtime.
+      const TargetOptions &Options = DAG.getTarget().Options;
+      if (Options.EnableMachineOutliner)
+        PreferIndirect = false;
     }
   }
   if (isTailCall) {

llvm/test/CodeGen/ARM/minsize-call-cse.ll

@@ -1,12 +1,17 @@
-; RUN: llc < %s | FileCheck %s
+; RUN: llc < %s | FileCheck %s --check-prefixes=OUTLINER,CHECK
+; RUN: llc -enable-machine-outliner=always < %s | FileCheck %s --check-prefixes=OUTLINER,CHECK
+; RUN: llc -enable-machine-outliner=never < %s  | FileCheck %s --check-prefixes=INDIRECT,CHECK
 
 target datalayout = "e-m:e-p:32:32-i64:64-v128:64:128-a:0:32-n32-S64"
 target triple = "thumbv7m-arm-none-eabi"
 
 ; CHECK-LABEL: f:
-; CHECK: blx r
-; CHECK: blx r
-; CHECK: blx r
+; OUTLINER: bl g
+; OUTLINER: bl g
+; OUTLINER: bl g
+; INDIRECT: blx r
+; INDIRECT: blx r
+; INDIRECT: blx r
 define void @f() minsize optsize {
 entry:
   call void @g(i32 45, i32 66)

llvm/test/CodeGen/ARM/pr42062.ll

@@ -1,5 +1,5 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc -o - %s 2>&1 | FileCheck %s --implicit-check-not=error
+; RUN: llc -o - %s -enable-machine-outliner=never 2>&1 | FileCheck %s --implicit-check-not=error
 target triple = "thumbv8m.base-arm-none-eabi"
 @foo = external global i8
 declare i32 @bar(ptr nocapture, i32, i32, ptr nocapture)