[NVPTX] Allow undef value as global initializer

[wengxt]

Summary

shared variable may now emit undef value as initializer, do not
throw error on that.

Test Plan: test/CodeGen/NVPTX/global-addrspace.ll

Patch by Xuetian Weng

[jingyue]

+Artem

[jingyue]

I should commit this one before your Clang change, right?

lib/Target/NVPTX/NVPTXAsmPrinter.cpp

1185–1186

Please update the comments.

[wengxt]

update comments.

lib/Target/NVPTX/NVPTXAsmPrinter.cpp

if (GVar->hasInitializer()) {
       if ((PTy->getAddressSpace() == llvm::ADDRESS_SPACE_GLOBAL) ||
           (PTy->getAddressSpace() == llvm::ADDRESS_SPACE_CONST)) {
         const Constant *Initializer = GVar->getInitializer();
         // 'undef' is treated as there is no value specified.
         if (!Initializer->isNullValue() && !isa<UndefValue>(Initializer)) {
           O << " = ";
           printScalarConstant(Initializer, O);
         }
       } else {
-        // The frontend adds zero-initializer to variables that don't have an
-        // initial value, so skip warning for this case.
-        if (!GVar->getInitializer()->isNullValue()) {
+        // The frontend adds zero-initializer to device and constant variables
+        // that don't have an initial value, and UndefValue to shared
+        // variables, so skip warning for this case.
+        if (!GVar->getInitializer()->isNullValue() &&
+            !isa<UndefValue>(GVar->getInitializer())) {
           report_fatal_error("initial value of '" + GVar->getName() +
                              "' is not allowed in addrspace(" +
                              Twine(PTy->getAddressSpace()) + ")");
         }
       }
     }
   } else {
     unsigned int ElementSize = 0;

test/CodeGen/NVPTX/global-addrspace.ll

+; RUN: llc < %s -march=nvptx -mcpu=sm_20 | FileCheck %s --check-prefix=PTX32
+; RUN: llc < %s -march=nvptx64 -mcpu=sm_20 | FileCheck %s --check-prefix=PTX64
+
+; PTX32: .visible .global .align 4 .u32 i;
+; PTX32: .visible .const .align 4 .u32 j;
+; PTX32: .visible .shared .align 4 .u32 k;
+; PTX64: .visible .global .align 4 .u32 i;
+; PTX64: .visible .const .align 4 .u32 j;
+; PTX64: .visible .shared .align 4 .u32 k;
+@i = addrspace(1) externally_initialized global i32 0, align 4
+@j = addrspace(4) externally_initialized global i32 0, align 4
+@k = addrspace(3) global i32 undef, align 4