Re: [PATCH] llvm: fix output_op_[ptr]cast()

[Dibyendu Majumdar <mobile@majumdar.org.uk>]

Perhaps we need these fixes?

https://patchwork.kernel.org/patch/8175541/

On 3 March 2017 at 18:06, Dibyendu Majumdar <mobile@majumdar.org.uk> wrote:
> On 3 March 2017 at 07:37, Dibyendu Majumdar <mobile@majumdar.org.uk> wrote:
>> On 3 March 2017 at 05:24, Luc Van Oostenryck
>> <luc.vanoostenryck@gmail.com> wrote:
>>> This should fox the problems of mixed types in casts.
>>> Without much testing but should be essentialy OK.
>>>
>>> CC: Dibyendu Majumdar <mobile@majumdar.org.uk>
>>> ---
>>>  sparse-llvm.c | 40 +++++++++++++++++++++++++++++++++++-----
>>>  1 file changed, 35 insertions(+), 5 deletions(-)
>>>
>>> diff --git a/sparse-llvm.c b/sparse-llvm.c
>>> index edd0615ec..92ad26845 100644
>>> --- a/sparse-llvm.c
>>> +++ b/sparse-llvm.c
>>> @@ -763,6 +763,8 @@ static void output_op_phi(struct function *fn, struct instruction *insn)
>>>  static void output_op_ptrcast(struct function *fn, struct instruction *insn)
>>>  {
>>>         LLVMValueRef src, target;
>>> +       LLVMTypeRef dtype;
>>> +       LLVMOpcode op;
>>>         char target_name[64];
>>>
>>>         src = insn->src->priv;
>>> @@ -773,15 +775,31 @@ static void output_op_ptrcast(struct function *fn, struct instruction *insn)
>>>
>>>         assert(!symbol_is_fp_type(insn->type));
>>>
>>> -       target = LLVMBuildBitCast(fn->builder, src, insn_symbol_type(fn->module, insn), target_name);
>>> +       dtype = insn_symbol_type(fn->module, insn);
>>> +       switch (LLVMGetTypeKind(LLVMTypeOf(src))) {
>>> +       case LLVMPointerTypeKind:
>>> +               op = LLVMBitCast;
>>> +               break;
>>> +       case LLVMIntegerTypeKind:
>>> +               op = LLVMIntToPtr;
>>> +               break;
>>> +       default:
>>> +               assert(0);
>>> +       }
>>>
>>> +       target = LLVMBuildCast(fn->builder, op, src, dtype, target_name);
>>>         insn->target->priv = target;
>>>  }
>>>
>>>  static void output_op_cast(struct function *fn, struct instruction *insn, LLVMOpcode op)
>>>  {
>>>         LLVMValueRef src, target;
>>> +       LLVMTypeRef dtype;
>>>         char target_name[64];
>>> +       unsigned int width;
>>> +
>>> +       if (is_ptr_type(insn->type))
>>> +               return output_op_ptrcast(fn, insn);
>>>
>>>         src = insn->src->priv;
>>>         if (!src)
>>> @@ -791,11 +809,23 @@ static void output_op_cast(struct function *fn, struct instruction *insn, LLVMOp
>>>
>>>         assert(!symbol_is_fp_type(insn->type));
>>>
>>> -       if (insn->size < LLVMGetIntTypeWidth(LLVMTypeOf(src)))
>>> -               target = LLVMBuildTrunc(fn->builder, src, insn_symbol_type(fn->module, insn), target_name);
>>> -       else
>>> -               target = LLVMBuildCast(fn->builder, op, src, insn_symbol_type(fn->module, insn), target_name);
>>> +       dtype = insn_symbol_type(fn->module, insn);
>>> +       switch (LLVMGetTypeKind(LLVMTypeOf(src))) {
>>> +       case LLVMPointerTypeKind:
>>> +               op = LLVMPtrToInt;
>>> +               break;
>>> +       case LLVMIntegerTypeKind:
>>> +               width = LLVMGetIntTypeWidth(LLVMTypeOf(src));
>>> +               if (insn->size < width)
>>> +                       op = LLVMTrunc;
>>> +               else if (insn->size == width)
>>> +                       op = LLVMBitCast;
>>> +               break;
>>> +       default:
>>> +               assert(0);
>>> +       }
>>>
>>> +       target = LLVMBuildCast(fn->builder, op, src, dtype, target_name);
>>>         insn->target->priv = target;
>>>  }
>>>
>>
>>
>> This doesn't quite work. The problem is that in op_cast, the pointer
>> is being cast to int, but subsequent operations expect a pointer.
>>
>
> The problem occurs in this sequence:
>
>        ptrcast.64  %r26 <- (64) %r20
>        add.64      %r27 <- %r26, %r23
>        cast.64     %r28 <- (64) %r27
>        store.64    %r28 -> 16[%arg1]
>
> The last cast finds that the instruction type in an integer and does a
> cast to Integer, so that causes the store to fail as it expects a
> pointer.
>
> Question in my mind is what is the result type of an add operation
> when one of the arguments is a pointer?
>
> Regards
> Dibyendu