[WebAssembly] Clear frame base vreg in explicit-locals when stack pointer is dead

[dschuff]

Summary

Having an alloca in a function causes the stack pointer to be generated in the prolog, but if it's unused other than for debug info, explicit-locals will drop it and not allocate a local. In this case we need to reset the FrameBaseVreg.

[dschuff]

the IR to reproduce this is trivial but i'm still trying to reduce the debug info in the test. Basically this:

define hidden %"struct.foo"* @"foo"() #1 !dbg !2 {
fentry:
  %0 = alloca %"struct.bar", align 1
  call void @llvm.dbg.declare(metadata %"struct.bar"* %0, metadata !21513, metadata !DIExpression()), !dbg !21514
  ret %"struct.bar"* @"globl_", !dbg !21515
}

...plus just enough debug info to cause generation of a CU and subprogram with a frame base.

But this came up in the openCV demo so it must not be that uncommon.

[dschuff]

(FTR, the failure mode is the same assertion in MFI::getFrameBaseLocal() that we saw before)

[aardappel]

You mean alloca as in the C function, not as in the LLVM concept, I guess.

And yes, this is all very brittle code, having to potentially call these functions upon each register change. I wonder if there's a better way, like making the concept of FrameBase be a flag on a register, such that when the register gets deleted it is automatic. But that would probably still require special purpose when renumbered, when lowered to a local, etc etc.

[dschuff]

No I mean alloca as in the LLVM instruction. Just having an address-taken local in C is enough to cause an alloca in the IR.

And yeah you are right, I think it would still require the same ways to deal with it.
I think a possible alternative would be to go back to modeling SP as a physical register, and exempting it from the code that touches vregs. That would have some complications of its own (and probably result in slightly worse code in some cases), but may or may not be cleaner.

Even in this case the final output code has a prolog that calculates a value for SP but then immediately drops it because it's unused. This CL keeps it from crashing but clearly we should do better and either not emit or optimize away the prolog entirely.

llvm/lib/Target/WebAssembly/WebAssemblyExplicitLocals.cpp

for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end(); I != E;) {
           auto InsertPt = std::next(MI.getIterator());
           if (UseEmpty[Register::virtReg2Index(OldReg)]) {
             unsigned Opc = getDropOpcode(RC);
             MachineInstr *Drop =
                 BuildMI(MBB, InsertPt, MI.getDebugLoc(), TII->get(Opc))
                     .addReg(NewReg);
             // After the drop instruction, this reg operand will not be used
             Drop->getOperand(0).setIsKill();
+            if (MFI.isFrameBaseVirtual() && OldReg == MFI.getFrameBaseVreg())
+              MFI.clearFrameBaseVreg();
           } else {
             unsigned LocalId = getLocalId(Reg2Local, MFI, CurLocal, OldReg);
             unsigned Opc = getLocalSetOpcode(RC);
 
             WebAssemblyDebugValueManager(&MI).replaceWithLocal(LocalId);
 
             BuildMI(MBB, InsertPt, MI.getDebugLoc(), TII->get(Opc))
                 .addImm(LocalId)