[NVPTX] Fix PR41651

[hliao]

Summary

• Use the reliable way to retrive the IR module from callsite. As the indirect callsite doesn't have a called function, it would crash the compiler. Retrive the module from CS's call instruction instead.

[tra]

llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp

1296 ↗

(On Diff #212407)

Can we just use DAG.getDataLayout() as we do everywhere else?

[hliao]

llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp

1296 ↗

(On Diff #212407)

Unfortunately, either DAG or CLI (CallLoweringInfo) is not passed as parameter. Without changing prototype of that function, get data layout from CS is the trivial way.

[hliao]

llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp

1296 ↗

(On Diff #212407)

Wait, DL is passed as parameter, I will simplify the patch

[hliao]

simplify the patch

[tra]

llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp

1296 ↗

(On Diff #212407)

Even better! :-)

llvm/trunk/lib/Target/NVPTX/NVPTXISelLowering.cpp

if (retTy->isFloatingPointTy() || (retTy->isIntegerTy() && !retTy->isIntegerTy(128))) {
       // bits in size.  fp16 normally uses .b16 as its storage type in
       // PTX, so its size must be adjusted here, too.
       if (size < 32)
         size = 32;
 
       O << ".param .b" << size << " _";
     } else if (isa<PointerType>(retTy)) {
       O << ".param .b" << PtrVT.getSizeInBits() << " _";
-    } else if (retTy->isAggregateType() || retTy->isVectorTy() || retTy->isIntegerTy(128)) {
-      auto &DL = CS.getCalledFunction()->getParent()->getDataLayout();
+    } else if (retTy->isAggregateType() || retTy->isVectorTy() ||
+               retTy->isIntegerTy(128)) {
       O << ".param .align " << retAlignment << " .b8 _["
         << DL.getTypeAllocSize(retTy) << "]";
     } else {
       llvm_unreachable("Unknown return type");
     }
     O << ") ";
   }
   O << "_ (";

llvm/trunk/test/CodeGen/NVPTX/bug41651.ll

+; RUN: llc -filetype=asm -o - %s | FileCheck %s
+target datalayout = "e-i64:64-i128:128-v16:16-v32:32-n16:32:64"
+target triple = "nvptx64-nvidia-cuda"
+
+%func = type { i32 (i32, i32)** }
+
+; CHECK: foo
+; CHECK: call
+; CHECK: ret
+define void @foo() {
+  %call = call %func undef(i32 0, i32 1)
+  ret void
+}