DAG: Check if deoptimize call is available before emitting it

[arsenm]

No description provided.

[arsenm]

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

• DAG: Check if deoptimize call is available before emitting it #170940 👈 (View in Graphite)
• AMDGPU: Avoid crashing on statepoint-like pseudoinstructions #170657
• AMDGPU: Create a dummy call sequence when emitting call error #170656
• AMDGPU: Use correct chain when emitting error on a call #170645
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[llvmbot]

@llvm/pr-subscribers-llvm-selectiondag

@llvm/pr-subscribers-backend-amdgpu

Author: Matt Arsenault (arsenm)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/170940.diff

2 Files Affected:

• (modified) llvm/lib/CodeGen/SelectionDAG/StatepointLowering.cpp (+10-2)
• (modified) llvm/test/CodeGen/AMDGPU/llvm.deoptimize.ll (+2-6)

diff --git a/llvm/lib/CodeGen/SelectionDAG/StatepointLowering.cpp b/llvm/lib/CodeGen/SelectionDAG/StatepointLowering.cpp
index 5b8cd343557fa..d34f3211dc463 100644
--- a/llvm/lib/CodeGen/SelectionDAG/StatepointLowering.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/StatepointLowering.cpp
@@ -1307,9 +1307,17 @@ void SelectionDAGBuilder::visitGCRelocate(const GCRelocateInst &Relocate) {
 
 void SelectionDAGBuilder::LowerDeoptimizeCall(const CallInst *CI) {
   const auto &TLI = DAG.getTargetLoweringInfo();
-  SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(RTLIB::DEOPTIMIZE),
-                                         TLI.getPointerTy(DAG.getDataLayout()));
 
+  RTLIB::LibcallImpl DeoptImpl = TLI.getLibcallImpl(RTLIB::DEOPTIMIZE);
+  if (DeoptImpl == RTLIB::Unsupported) {
+    DAG.getContext()->emitError("no deoptimize libcall available");
+    return;
+  }
+
+  SDValue Callee =
+      DAG.getExternalSymbol(DeoptImpl, TLI.getPointerTy(DAG.getDataLayout()));
+
+  // FIXME: Should pass in the calling convention for the LibcallImpl.
   // We don't lower calls to __llvm_deoptimize as varargs, but as a regular
   // call.  We also do not lower the return value to any virtual register, and
   // change the immediately following return to a trap instruction.
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.deoptimize.ll b/llvm/test/CodeGen/AMDGPU/llvm.deoptimize.ll
index ef8c00bb3b4b0..53446b644445a 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.deoptimize.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.deoptimize.ll
@@ -1,10 +1,6 @@
-; RUN: not llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s 2> %t.err | FileCheck %s
-; RUN: FileCheck -check-prefix=ERR %s < %t.err
+; RUN: not llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s 2>&1 | FileCheck %s
 
-; ERR: error: <unknown>:0:0: in function caller_0 i32 (): unsupported external symbol
-; ERR: error: unhandled statepoint-like instruction
-
-; CHECK: ;unsupported statepoint/stackmap/patchpoint
+; CHECK: error: no deoptimize libcall available
 declare i32 @llvm.experimental.deoptimize.i32(...)
 declare i8  @llvm.experimental.deoptimize.i8(...)
 

[dantrushin]

Looks good, though I wonder if deoptimize is unavailable, where it comes from then?
(Why it is generated in IR?)

[arsenm]

Looks good, though I wonder if deoptimize is unavailable, where it comes from then? (Why it is generated in IR?)

Nothing is stopping you from writing IR with an intrinsic on it. And nothing may be keeping the optimize libcall alive if it depends on codegen.