ARM: don't return by popping PC if we have to adjust the stack afterwards.

[t.p.northover]

Summary

In mandatory tail calling conventions we might have to deallocate stack space used by our arguments before return. This happens after popping CSRs, so the pop cannot be turned into the return itself in this case.

The else branch here was already a nop, so removing it as a tidy-up.

[t.p.northover]

Ping.

[Gerolf]

Nice cleanup, too. LGTM.

[t.p.northover]

Thanks Gerolf, committed as 19d2e42be2cd.

llvm/lib/Target/ARM/ARMFrameLowering.cpp

while (i != 0) {
     for (; i != 0; --i) {
       CalleeSavedInfo &Info = CSI[i-1];
       unsigned Reg = Info.getReg();
       if (!(Func)(Reg, STI.splitFramePushPop(MF))) continue;
 
       // The aligned reloads from area DPRCS2 are not inserted here.
       if (Reg >= ARM::D8 && Reg < ARM::D8 + NumAlignedDPRCS2Regs)
         continue;
-
       if (Reg == ARM::LR && !isTailCall && !isVarArg && !isInterrupt &&
-          !isCmseEntry && !isTrap && STI.hasV5TOps()) {
-        if (MBB.succ_empty()) {
-          Reg = ARM::PC;
-          // Fold the return instruction into the LDM.
-          DeleteRet = true;
-          LdmOpc = AFI->isThumbFunction() ? ARM::t2LDMIA_RET : ARM::LDMIA_RET;
-          // We 'restore' LR into PC so it is not live out of the return block:
-          // Clear Restored bit.
-          Info.setRestored(false);
-        } else
-          LdmOpc = AFI->isThumbFunction() ? ARM::t2LDMIA_UPD : ARM::LDMIA_UPD;
+          !isCmseEntry && !isTrap && AFI->getArgumentStackToRestore() == 0 &&
+          STI.hasV5TOps() && MBB.succ_empty()) {
       }
 
       // If NoGap is true, pop consecutive registers and then leave the rest
       // for other instructions. e.g.
       // vpop {d8, d10, d11} -> vpop {d8}, vpop {d10, d11}
       if (NoGap && LastReg && LastReg != Reg-1)
         break;
 

llvm/test/CodeGen/ARM/tailcc-call.ll

 ; COMMON: str [[TMP]], [sp]{{$}}
 ; COMMON: bl _Ccallee_stack4
 ; COMMON-NOT: sub sp,
 
   call void @Ccallee_stack4([4 x i32] undef, i32 42)
   call void @Ccallee_stack4([4 x i32] undef, i32 42)
   ret void
 }
+
+; Don't try to return by popping pc if there's stack to reclaim.
+define tailcc void @notail_stackclean([4 x i32], i32) {
+; COMMON-LABEL: notail_stackclean:
+; COMMON: pop {r7, lr}
+; COMMON: add sp, #8
+; COMMON: bx lr
+
+
+  call void @callee_stack0()
+  ret void
+}