[ARM] Update IsRestored for LR based on all returns (#82745)

PR #75527 fixed ARMFrameLowering to set the IsRestored flag for LR based
on all of the return instructions in the function, not just one.
However, there is also code in ARMLoadStoreOptimizer which changes
return instructions, but it set IsRestored based on the one instruction
it changed, not the whole function.

The fix is to factor out the code added in #75527, and also call it from
ARMLoadStoreOptimizer if it made a change to return instructions.

Fixes #80287.

(cherry picked from commit 749384c08e042739342c88b521c8ba5dac1b9276)

4 files changed, 27 insertions, 22 deletions

diff --git a/llvm/lib/Target/ARM/ARMFrameLowering.cpp b/llvm/lib/Target/ARM/ARMFrameLowering.cpp
index eeb7f64..9b54dd4 100644
--- a/llvm/lib/Target/ARM/ARMFrameLowering.cpp
+++ b/llvm/lib/Target/ARM/ARMFrameLowering.cpp
@@ -2781,10 +2781,7 @@ void ARMFrameLowering::determineCalleeSaves(MachineFunction &MF,
   AFI->setLRIsSpilled(SavedRegs.test(ARM::LR));
 }
 
-void ARMFrameLowering::processFunctionBeforeFrameFinalized(
-    MachineFunction &MF, RegScavenger *RS) const {
-  TargetFrameLowering::processFunctionBeforeFrameFinalized(MF, RS);
-
+void ARMFrameLowering::updateLRRestored(MachineFunction &MF) {
   MachineFrameInfo &MFI = MF.getFrameInfo();
   if (!MFI.isCalleeSavedInfoValid())
     return;
@@ -2808,6 +2805,12 @@ void ARMFrameLowering::processFunctionBeforeFrameFinalized(
   }
 }
 
+void ARMFrameLowering::processFunctionBeforeFrameFinalized(
+    MachineFunction &MF, RegScavenger *RS) const {
+  TargetFrameLowering::processFunctionBeforeFrameFinalized(MF, RS);
+  updateLRRestored(MF);
+}
+
 void ARMFrameLowering::getCalleeSaves(const MachineFunction &MF,
                                       BitVector &SavedRegs) const {
   TargetFrameLowering::getCalleeSaves(MF, SavedRegs);
diff --git a/llvm/lib/Target/ARM/ARMFrameLowering.h b/llvm/lib/Target/ARM/ARMFrameLowering.h
index 8d2b8be..3c7358d 100644
--- a/llvm/lib/Target/ARM/ARMFrameLowering.h
+++ b/llvm/lib/Target/ARM/ARMFrameLowering.h
@@ -59,6 +59,10 @@ public:
   void determineCalleeSaves(MachineFunction &MF, BitVector &SavedRegs,
                             RegScavenger *RS) const override;
 
+  /// Update the IsRestored flag on LR if it is spilled, based on the return
+  /// instructions.
+  static void updateLRRestored(MachineFunction &MF);
+
   void processFunctionBeforeFrameFinalized(
       MachineFunction &MF, RegScavenger *RS = nullptr) const override;
 
diff --git a/llvm/lib/Target/ARM/ARMLoadStoreOptimizer.cpp b/llvm/lib/Target/ARM/ARMLoadStoreOptimizer.cpp
index ed9d30c..6121055 100644
--- a/llvm/lib/Target/ARM/ARMLoadStoreOptimizer.cpp
+++ b/llvm/lib/Target/ARM/ARMLoadStoreOptimizer.cpp
@@ -2062,17 +2062,6 @@ bool ARMLoadStoreOpt::MergeReturnIntoLDM(MachineBasicBlock &MBB) {
       MO.setReg(ARM::PC);
       PrevMI.copyImplicitOps(*MBB.getParent(), *MBBI);
       MBB.erase(MBBI);
-      // We now restore LR into PC so it is not live-out of the return block
-      // anymore: Clear the CSI Restored bit.
-      MachineFrameInfo &MFI = MBB.getParent()->getFrameInfo();
-      // CSI should be fixed after PrologEpilog Insertion
-      assert(MFI.isCalleeSavedInfoValid() && "CSI should be valid");
-      for (CalleeSavedInfo &Info : MFI.getCalleeSavedInfo()) {
-        if (Info.getReg() == ARM::LR) {
-          Info.setRestored(false);
-          break;
-        }
-      }
       return true;
     }
   }
@@ -2120,14 +2109,22 @@ bool ARMLoadStoreOpt::runOnMachineFunction(MachineFunction &Fn) {
   isThumb2 = AFI->isThumb2Function();
   isThumb1 = AFI->isThumbFunction() && !isThumb2;
 
-  bool Modified = false;
+  bool Modified = false, ModifiedLDMReturn = false;
   for (MachineBasicBlock &MBB : Fn) {
     Modified |= LoadStoreMultipleOpti(MBB);
     if (STI->hasV5TOps() && !AFI->shouldSignReturnAddress())
-      Modified |= MergeReturnIntoLDM(MBB);
+      ModifiedLDMReturn |= MergeReturnIntoLDM(MBB);
     if (isThumb1)
       Modified |= CombineMovBx(MBB);
   }
+  Modified |= ModifiedLDMReturn;
+
+  // If we merged a BX instruction into an LDM, we need to re-calculate whether
+  // LR is restored. This check needs to consider the whole function, not just
+  // the instruction(s) we changed, because there may be other BX returns which
+  // still need LR to be restored.
+  if (ModifiedLDMReturn)
+    ARMFrameLowering::updateLRRestored(Fn);
 
   Allocator.DestroyAll();
   return Modified;
diff --git a/llvm/test/CodeGen/ARM/ldst-opt-lr-restored.ll b/llvm/test/CodeGen/ARM/ldst-opt-lr-restored.ll
index 883a812..9494880 100644
--- a/llvm/test/CodeGen/ARM/ldst-opt-lr-restored.ll
+++ b/llvm/test/CodeGen/ARM/ldst-opt-lr-restored.ll
@@ -10,16 +10,17 @@ define i32 @foo(ptr %ctx) {
 ; CHECK:       @ %bb.0: @ %entry
 ; CHECK-NEXT:    cbz r0, .LBB0_2
 ; CHECK-NEXT:  @ %bb.1: @ %if.end
+; CHECK-NEXT:    movw r12, :lower16:val2
+; CHECK-NEXT:    movw r3, :lower16:val1
 ; CHECK-NEXT:    movw r2, :lower16:val0
 ; CHECK-NEXT:    mov r1, r0
 ; CHECK-NEXT:    movs r0, #0
-; CHECK-NEXT:    movw r12, :lower16:val2
-; CHECK-NEXT:    movw r3, :lower16:val1
-; CHECK-NEXT:    movt r2, :upper16:val0
-; CHECK-NEXT:    add.w lr, r1, #4
 ; CHECK-NEXT:    movt r12, :upper16:val2
 ; CHECK-NEXT:    movt r3, :upper16:val1
-; CHECK-NEXT:    stm.w lr, {r2, r3, r12}
+; CHECK-NEXT:    movt r2, :upper16:val0
+; CHECK-NEXT:    str r2, [r1, #4]
+; CHECK-NEXT:    str r3, [r1, #8]
+; CHECK-NEXT:    str.w r12, [r1, #12]
 ; CHECK-NEXT:    str r0, [r1, #16]
 ; CHECK-NEXT:    bx lr
 ; CHECK-NEXT:  .LBB0_2: @ %if.then