[StackMaps] Avoid crash when a call follows a patchpoint

[jmolloy]

Summary

A call after a patchpoint will cause the patchpoint to implicitly have SSP as liveout. SSP doesn't have a DWARF regnum or a register class, so we blow up in getDwarfRegNum().

Guard this by only trying to emit stackmap information for allocatable registers.

[aprantl]

llvm/lib/CodeGen/StackMaps.cpp

264

Disclaimer: I know nothing about LLVM's StackMap implementation.

Is it safe to just ignore it here? In CodeGen/AsmPrinter/DwarfExpression.cpp we have code to compose a register without a Dwarf number out of sub- or super-registers.

llvm/lib/CodeGen/StackMaps.cpp

 StackMaps::parseRegisterLiveOutMask(const uint32_t *Mask) const {
   assert(Mask && "No register mask specified");
   const TargetRegisterInfo *TRI = AP.MF->getSubtarget().getRegisterInfo();
   LiveOutVec LiveOuts;
 
   // Create a LiveOutReg for each bit that is set in the register mask.
   for (unsigned Reg = 0, NumRegs = TRI->getNumRegs(); Reg != NumRegs; ++Reg)
     if ((Mask[Reg / 32] >> (Reg % 32)) & 1)
+      if (TRI->isInAllocatableClass(Reg))
         LiveOuts.push_back(createLiveOutReg(Reg, TRI));
 
   // We don't need to keep track of a register if its super-register is already
   // in the list. Merge entries that refer to the same dwarf register and use
   // the maximum size that needs to be spilled.
 
   llvm::sort(LiveOuts, [](const LiveOutReg &LHS, const LiveOutReg &RHS) {
     // Only sort by the dwarf register number.
     return LHS.DwarfRegNum < RHS.DwarfRegNum;

llvm/test/CodeGen/X86/stackmap-shadow-optimization.ll

 ; CHECK-NOT:    nop
   call void @bar()
   tail call void (i64, i32, ...) @llvm.experimental.stackmap(i64 0, i32 8)
   call void @bar()
   call void @bar()
   ret void
 }
 
+; Check no crash.
+; CHECK-LABEL: shadow_optimization_test2:
+define void @shadow_optimization_test2() {
+entry:
+  call void @bar()
+  tail call void (i64, i32, i8*, i32, ...) @llvm.experimental.patchpoint.void(i64 0, i32 8, i8* null, i32 0)
+  call void @bar()
+  call void @bar()
+  ret void
+}
+
 declare void @llvm.experimental.stackmap(i64, i32, ...)
+declare void @llvm.experimental.patchpoint.void(i64, i32, i8*, i32, ...)
 declare void @bar()