[PATCH] TargetInfo: Use DataLayout to compute pointer width in getPointerWidth()

[Reid Kleckner]

Feel free to commit with some comments on the test.

On Wed, Jun 25, 2014 at 1:31 PM, Tom Stellard <thomas.stellard at amd.com>
wrote:

> ---
>  include/clang/Basic/TargetInfo.h            |  4 +---
>  lib/Basic/TargetInfo.cpp                    |  9 +++++++++
>  test/CodeGen/record-lowering-non-zero-as.cl | 14 ++++++++++++++
>  3 files changed, 24 insertions(+), 3 deletions(-)
>  create mode 100644 test/CodeGen/record-lowering-non-zero-as.cl
>
> diff --git a/include/clang/Basic/TargetInfo.h
> b/include/clang/Basic/TargetInfo.h
> index e1d0116..65dc101 100644
> --- a/include/clang/Basic/TargetInfo.h
> +++ b/include/clang/Basic/TargetInfo.h
> @@ -790,9 +790,7 @@ public:
>    }
>
>  protected:
> -  virtual uint64_t getPointerWidthV(unsigned AddrSpace) const {
> -    return PointerWidth;
> -  }
> +  virtual uint64_t getPointerWidthV(unsigned AddrSpace) const;
>    virtual uint64_t getPointerAlignV(unsigned AddrSpace) const {
>      return PointerAlign;
>    }
> diff --git a/lib/Basic/TargetInfo.cpp b/lib/Basic/TargetInfo.cpp
> index 71e39dd..d0f6306 100644
> --- a/lib/Basic/TargetInfo.cpp
> +++ b/lib/Basic/TargetInfo.cpp
> @@ -17,6 +17,7 @@
>  #include "clang/Basic/LangOptions.h"
>  #include "llvm/ADT/APFloat.h"
>  #include "llvm/ADT/STLExtras.h"
> +#include "llvm/IR/DataLayout.h"
>  #include "llvm/Support/ErrorHandling.h"
>  #include <cstdlib>
>  using namespace clang;
> @@ -242,6 +243,14 @@ bool TargetInfo::isTypeSigned(IntType T) {
>    };
>  }
>
> +uint64_t TargetInfo::getPointerWidthV(unsigned AddrSpace) const {
> +  if (!DescriptionString)
> +    return PointerWidth;
> +
> +  llvm::DataLayout DL(DescriptionString);
> +  return DL.getPointerSizeInBits(AddrSpace);
> +}
>

Seems reasonable.


> +
>  /// setForcedLangOptions - Set forced language options.
>  /// Apply changes to the target information with respect to certain
>  /// language options which change the target configuration.
> diff --git a/test/CodeGen/record-lowering-non-zero-as.cl b/test/CodeGen/
> record-lowering-non-zero-as.cl
> new file mode 100644
> index 0000000..a96756f
> --- /dev/null
> +++ b/test/CodeGen/record-lowering-non-zero-as.cl
> @@ -0,0 +1,14 @@
> +// REQUIRES: r600-registered-target
> +// RUN: %clang -target r600 -mcpu=verde -S -emit-llvm -o - %s
>

IMO this should be a Sema test.  CodeGen was working as intended, it's the
output of sizeof() that was wrong.  Can you make this a Sema test like:

struct test_as0 { int *ptr; };
int as0[sizeof(struct test_as0) == 4 ? 1 : -1];
struct test_as1 { global int *ptr; };
int as1[sizeof(struct test_as1) == 8 ? 1 : -1];
// ... and so on, for other interesting address spaces.

If you also want a CodeGen test, please FileCheck for the correct LLVM IR
struct type.

+
> +// Record lowering was crashing on SI and newer targets, because it
> +// was using the wrong size for test::ptr.  Since global memory
> +// has 64-bit pointers, sizeof(test::ptr) should be 8.
> +
> +struct test {
> +  global int *ptr;
> +};
> +
> +void func(struct test t, global int *ptr) {
> +  *ptr = 0;
> +}
> --
> 1.8.1.5
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20140625/cfdeff51/attachment.html>