[NVPTX] Fix integer overflow affecting array size calculation.

[tra]

Summary

It turns out, 4GB+ large arrays are a thing and a 32-bit integer is just not
enough to handle them.

[tra]

.

[efriedma]

Please don't use size_t to represent anything other than the size of a buffer allocated by the compiler itself; LLVM supports 32-bit hosts. We generally use uint64_t for the size of LLVM IR types.

[tra]

Use uint64_t instead of size_t.

[tra]

In D146767#4217998, @efriedma wrote:

Please don't use size_t to represent anything other than the size of a buffer allocated by the compiler itself; LLVM supports 32-bit hosts. We generally use uint64_t for the size of LLVM IR types.

Good point. Fixed.

[yaxunl]

LGTM.

llvm/lib/Target/NVPTX/NVPTXAsmPrinter.cpp

if (GVar->hasInitializer()) {
             !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;
+    uint64_t ElementSize = 0;
 
     // Although PTX has direct support for struct type and array type and
     // LLVM IR is very similar to PTX, the LLVM CodeGen does not support for
     // targets that support these high level field accesses. Structs, arrays
     // and vectors are lowered into arrays of bytes.
     switch (ETy->getTypeID()) {
     case Type::IntegerTyID: // Integers larger than 64 bits
     case Type::StructTyID:

llvm/test/CodeGen/NVPTX/global-variable-big.ll

 ; RUN: llc < %s | FileCheck %s
 ; RUN: %if ptxas %{ llc < %s | %ptxas-verify %}
 
 target datalayout = "e-i64:64-v16:16-v32:32-n16:32:64"
 target triple = "nvptx64-nvidia-cuda"
 
 ; Check that we can handle global variables of large integer type.
 
 ; (lsb) 0x0102'0304'0506...0F10 (msb)
 @gv = addrspace(1) externally_initialized global i128 21345817372864405881847059188222722561, align 16
 ; CHECK: .visible .global .align 16 .b8 gv[16] = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16};
+
+; Make sure that we do not overflow on large number of elements.
+; CHECK-LABEL large_data
+; CHECK: .visible .global .align 1 .b8 large_data[4831838208]
+@large_data = global [4831838208 x i8] zeroinitializer
+
+