[X86][Win64] Avoid statepoints in trailing call position

[zero9178]

Summary

The "avoid trailing call pass" makes sure that no function ends with a call instruction for the purpose of the unwinder.
It starts of by skipping over any non real instruction, which is approximated via the Pseudo and Meta property. This sadly leads to issues when the last machine instruction is a STATEPOINT, as it is skipped despite it lowering to a call.

This patch fixes the use of a statepoint in the trailing call position by making sure it is not skipped.

[reames]

LGTM in the sense that this doesn't look obviously wrong, but that I don't understand the Win64 code here. You should probably confirm with someone more knowledgeable of that area.

llvm/test/CodeGen/X86/win64-eh-trailing-statepoint.ll

4

Can you use update_lit_test here? It's really hard to see what the resulting assembly actually is here.

[zero9178]

Address reviewers comments:

• Make use of update_llc_test_checks.py script to generate the output checks

[rnk]

lgtm

I have some other changes I want to make, but just go ahead and land this and I'll do them as a follow-up.

llvm/lib/Target/X86/X86AvoidTrailingCall.cpp

70–71

This comment about being "conservative" isn't correct anymore. I'll follow-up and fix it.

72–73

So, statepoints are pseudo instructions that are later lowered to calls. Does the MI.isCall() predicate return true for statepoints? Could we use this condition?

return !MI.isCall() && !MI.isPseudo() && !MI.isMetaInstruction();

[zero9178]

llvm/lib/Target/X86/X86AvoidTrailingCall.cpp

72–73

It does so using that property would probably be a good idea and be more generic! I am assuming you mean
MI.isCall() || (!MI.isPseudo() && !MI.isMetaInstruction()) though, so that the code below would find the first real or call instruction

[zero9178]

Makes use of the isCall property of a machine instruction instead of specializing for statepoint in particular.
That way the code fixes the issue for any pseudo instruction that lowers to a call.

llvm/lib/Target/X86/X86AvoidTrailingCall.cpp

 
 FunctionPass *llvm::createX86AvoidTrailingCallPass() {
   return new X86AvoidTrailingCallPass();
 }
 
 INITIALIZE_PASS(X86AvoidTrailingCallPass, AVOIDCALL_NAME, AVOIDCALL_DESC, false, false)
 
 // A real instruction is a non-meta, non-pseudo instruction.  Some pseudos
 // expand to nothing, and some expand to code. This logic conservatively assumes
 // they might expand to nothing.
-static bool isRealInstruction(MachineInstr &MI) {
-  return !MI.isPseudo() && !MI.isMetaInstruction();
+static bool isCallOrRealInstruction(MachineInstr &MI) {
+  return MI.isCall() || (!MI.isPseudo() && !MI.isMetaInstruction());
 }
 
 // Return true if this is a call instruction, but not a tail call.
 static bool isCallInstruction(const MachineInstr &MI) {
   return MI.isCall() && !MI.isReturn();
 }
 
 bool X86AvoidTrailingCallPass::runOnMachineFunction(MachineFunction &MF) {
bool X86AvoidTrailingCallPass::runOnMachineFunction(MachineFunction &MF) {
   for (MachineBasicBlock &MBB : MF) {
     // Look for basic blocks that precede funclet entries or are at the end of
     // the function.
     MachineBasicBlock *NextMBB = MBB.getNextNode();
     if (NextMBB && !NextMBB->isEHFuncletEntry())
       continue;
 
     // Find the last real instruction in this block.
-    auto LastRealInstr = llvm::find_if(reverse(MBB), isRealInstruction);
+    auto LastRealInstr = llvm::find_if(reverse(MBB), isCallOrRealInstruction);
 
     // If the block is empty or the last real instruction is a call instruction,
     // insert an int3. If there is a call instruction, insert the int3 between
     // the call and any labels or other meta instructions. If the block is
     // empty, insert at block end.
     bool IsEmpty = LastRealInstr == MBB.rend();
     bool IsCall = !IsEmpty && isCallInstruction(*LastRealInstr);
     if (IsEmpty || IsCall) {

llvm/test/CodeGen/X86/win64-eh-trailing-statepoint.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=x86_64-windows-gnu %s -o - | FileCheck %s
+
+define void @foo() gc "statepoint-example" personality i8* bitcast (i32 (...)* @__gxx_personality_seh0 to i8*) {
+; CHECK-LABEL: foo:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    subq $40, %rsp
+; CHECK-NEXT:    .seh_stackalloc 40
+; CHECK-NEXT:    .seh_endprologue
+; CHECK-NEXT:    callq raise
+; CHECK-NEXT:  .Ltmp0:
+; CHECK-NEXT:    int3
+; CHECK-NEXT:    .seh_endproc
+    %statepoint_token = call token (i64, i32, void ()*, i32, i32, ...) @llvm.experimental.gc.statepoint.p0f_isVoidf(i64 2882400000, i32 0, void ()* elementtype(void ()) @raise, i32 0, i32 0, i32 0, i32 0)
+    unreachable
+}
+
+declare void @raise()
+
+declare dso_local i32 @__gxx_personality_seh0(...)
+
+declare token @llvm.experimental.gc.statepoint.p0f_isVoidf(i64 immarg, i32 immarg, void ()*, i32 immarg, i32 immarg, ...)