Bug fix on function epilog optimization (ARM backend)

[rOptimizer]

Summary

To save a 'add sp,#val' instruction by adding registers to the final pop instruction,
the first register transferred by this pop instruction need to be found.
If the function to be optimized has a non-void return value, the operand list contains
r0 (implicit) which prevents the optimization to take place.
Therefore implicit register references should be skipped in the search loop,
because this registers are never popped from the stack.

Merge branch 'master' of https://github.com/llvm/llvm-project

[lebedev.ri]

test missing

[rOptimizer]

• Bug fix on function epilog optimization (ARM backend)

Added regression test

[ostannard]

LGTM

[rOptimizer]

I don't have commit rights.
Can anybody do this?

[ostannard]

Yep, I can commit it for you.

llvm/trunk/lib/Target/ARM/ARMBaseInstrInfo.cpp

bool llvm::tryFoldSPUpdateIntoPushPop(const ARMSubtarget &Subtarget,
   // instruction, which won't necessarily be the first register in the list.
   unsigned FirstRegEnc = -1;
 
   const TargetRegisterInfo *TRI = MF.getRegInfo().getTargetRegisterInfo();
   for (int i = MI->getNumOperands() - 1; i >= RegListIdx; --i) {
     MachineOperand &MO = MI->getOperand(i);
     RegList.push_back(MO);
 
-    if (MO.isReg() && TRI->getEncodingValue(MO.getReg()) < FirstRegEnc)
+    if (MO.isReg() && !MO.isImplicit() &&
+        TRI->getEncodingValue(MO.getReg()) < FirstRegEnc)
       FirstRegEnc = TRI->getEncodingValue(MO.getReg());
   }
 
   const MCPhysReg *CSRegs = TRI->getCalleeSavedRegs(&MF);
 
   // Now try to find enough space in the reglist to allocate NumBytes.
   for (int CurRegEnc = FirstRegEnc - 1; CurRegEnc >= 0 && RegsNeeded;
        --CurRegEnc) {
     unsigned CurReg = RegClass->getRegister(CurRegEnc);
-    if (IsT1PushPop && CurReg > ARM::R7)
+    if (IsT1PushPop && CurRegEnc > TRI->getEncodingValue(ARM::R7))
       continue;
     if (!IsPop) {
       // Pushing any register is completely harmless, mark the register involved
       // as undef since we don't care about its value and must not restore it
       // during stack unwinding.
       RegList.push_back(MachineOperand::CreateReg(CurReg, false, false,
                                                   false, false, true));
       --RegsNeeded;

llvm/trunk/test/CodeGen/ARM/fold-stack-adjust.ll

 ; CHECK-NOT: add sp,
 ; CHEC: pop {r3, r4, r5, r6, r7, pc}
 
   %var = alloca i8, i32 16
   call void @bar(i8* %var)
   ret void
 }
 
+define i32 @check_simple_ret() minsize {
+; CHECK-FNSTART-LABEL: check_simple_ret:
+; CHECK: push {r5, r6, r7, lr}
+; CHECK-NOT: sub sp,
+; ...
+; CHECK-NOT: add sp,
+; CHECK: pop {r2, r3, r7, pc}
+
+  %var = alloca i8, i32 8
+  call void @bar(i8* %var)
+  ret i32 0
+}
+
 define void @check_simple_too_big() minsize {
 ; CHECK-FNSTART-LABEL: check_simple_too_big:
 ; CHECK: push {r7, lr}
 ; CHECK: sub sp,
 ; ...
 ; CHECK: add sp,
 ; CHECK: pop {r7, pc}
   %var = alloca i8, i32 64