[NVPTX] Use MBB.begin() instead MBB.front() in NVPTXFrameLowering::emitPrologue

[xgupta]

Summary

The second argument of NVPTXFrameLowering::emitPrologue(MachineFunction &MF, MachineBasicBlock &MBB) is the first MBB of the MF. In that function, it assumes the first MBB always contains instructions, so it gets the first instruction by MachineInstr *MI = &MBB.front();. However, with the reproducer/test case attached, all instructions in the first MBB is cleared in a previous pass for stack coloring. As a consequence, MBB.front() triggers the assertion that the first node is actually a sentinel node. Hence we are using MachineBasicBlock::iterator to iterate over MBB.

Fix #52623.

[xgupta]

I will have to reduce the test case attached in bug report and add that here, so patch is WIP, any suggestion on that would be very helpful. Thanks.

[tra]

The changes look reasonable to me.

[xgupta]

Add reduced test case of crash

[xgupta]

Add CHECK line in test case

[tra]

LGTM with a test nit.

llvm/test/CodeGen/NVPTX/bug52623.ll

10

Considering that there are no meaningful CHECK lines in the test, it may be worth adding a comment describing what it is that we're testing for here.

[tianshilei1992]

llvm/test/CodeGen/NVPTX/bug52623.ll

2

Do we need to run the FileCheck here?

[xgupta]

Add descriptive comment about the issue in test case and remove FileCheck.

[tianshilei1992]

LGTM

llvm/lib/Target/NVPTX/NVPTXFrameLowering.cpp

NVPTXFrameLowering::NVPTXFrameLowering()
     : TargetFrameLowering(TargetFrameLowering::StackGrowsUp, Align(8), 0) {}
 
 bool NVPTXFrameLowering::hasFP(const MachineFunction &MF) const { return true; }
 
 void NVPTXFrameLowering::emitPrologue(MachineFunction &MF,
                                       MachineBasicBlock &MBB) const {
   if (MF.getFrameInfo().hasStackObjects()) {
     assert(&MF.front() == &MBB && "Shrink-wrapping not yet supported");
-    MachineInstr *MI = &MBB.front();
+    MachineBasicBlock::iterator MBBI = MBB.begin();
     MachineRegisterInfo &MR = MF.getRegInfo();
 
     const NVPTXRegisterInfo *NRI =
         MF.getSubtarget<NVPTXSubtarget>().getRegisterInfo();
 
     // This instruction really occurs before first instruction
     // in the BB, so giving it no debug location.
     DebugLoc dl = DebugLoc();
 
     // Emits
     //   mov %SPL, %depot;
     //   cvta.local %SP, %SPL;
     // for local address accesses in MF.
     bool Is64Bit =
         static_cast<const NVPTXTargetMachine &>(MF.getTarget()).is64Bit();
     unsigned CvtaLocalOpcode =
         (Is64Bit ? NVPTX::cvta_local_yes_64 : NVPTX::cvta_local_yes);
     unsigned MovDepotOpcode =
         (Is64Bit ? NVPTX::MOV_DEPOT_ADDR_64 : NVPTX::MOV_DEPOT_ADDR);
     if (!MR.use_empty(NRI->getFrameRegister(MF))) {
       // If %SP is not used, do not bother emitting "cvta.local %SP, %SPL".
-      MI = BuildMI(MBB, MI, dl,
+      MBBI = BuildMI(MBB, MBBI, dl,
                      MF.getSubtarget().getInstrInfo()->get(CvtaLocalOpcode),
                      NRI->getFrameRegister(MF))
                  .addReg(NRI->getFrameLocalRegister(MF));
     }
-    BuildMI(MBB, MI, dl, MF.getSubtarget().getInstrInfo()->get(MovDepotOpcode),
+    BuildMI(MBB, MBBI, dl,
+            MF.getSubtarget().getInstrInfo()->get(MovDepotOpcode),
             NRI->getFrameLocalRegister(MF))
         .addImm(MF.getFunctionNumber());
   }
 }
 
 StackOffset
 NVPTXFrameLowering::getFrameIndexReference(const MachineFunction &MF, int FI,
                                            Register &FrameReg) const {

llvm/test/CodeGen/NVPTX/bug52623.ll

+; RUN: llc < %s -march=nvptx -mcpu=sm_75 -verify-machineinstrs
+; RUN: %if ptxas %{ llc < %s -march=nvptx -mcpu=sm_75 | %ptxas-verify %}
+
+; Check that llc will not crash even when first MBB doesn't contain
+; any instruction.
+
+target datalayout = "e-i64:64-i128:128-v16:16-v32:32-n16:32:64"
+target triple = "nvptx64"
+
+%printf_args.0.8 = type { i8* }
+
+define internal i32 @__kmpc_get_hardware_thread_id_in_block(i1 %0) {
+  %2 = alloca %printf_args.0.8, i32 0, align 8
+  %3 = bitcast %printf_args.0.8* %2 to i8*
+  br i1 true, label %._crit_edge1, label %._crit_edge
+
+._crit_edge:                                      ; preds = %1, %._crit_edge
+  %4 = call i32 null(i8* null, i8* %3)
+  br i1 %0, label %._crit_edge, label %._crit_edge1
+
+._crit_edge1:                                     ; preds = %._crit_edge, %1
+  ret i32 0
+
+; uselistorder directives
+  uselistorder label %._crit_edge, { 1, 0 }
+}