[llvm-commits] [llvm] r139995 - in /llvm/trunk: lib/Target/X86/X86ISelLowering.cpp test/CodeGen/X86/trunc-ext-ld-st.ll

[Eli Friedman]

On Sun, Sep 18, 2011 at 3:39 AM, Nadav Rotem <nadav.rotem at intel.com> wrote:
> +    // Attempt to load the original value using a single load op.
> +    // Find a scalar type which is equal to the loaded word size.
> +    MVT SclrLoadTy = MVT::i8;
> +    for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
> +         tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
> +      MVT Tp = (MVT::SimpleValueType)tp;
> +      if (TLI.isTypeLegal(Tp) &&  Tp.getSizeInBits() == MemSz) {
> +        SclrLoadTy = Tp;
> +        break;
> +      }
> +    }
> +
> +    // Proceed if a load word is found.
> +    if (SclrLoadTy.getSizeInBits() != MemSz) return SDValue();

This approach to finding a load misses one somewhat important case:
the 64-bit case on x86-32 and ARM.  Do you have any ideas here?

> @@ -13479,9 +13563,7 @@
>
>     // From, To sizes and ElemCount must be pow of two
>     if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
> -    // We are going to use the original vector elt for storing.
> -    // accumulated smaller vector elements must be a multiple of bigger size.
> -    if (0 != (NumElems * ToSz) % FromSz) return SDValue();
> +

What is this change supposed to do?

-Eli