Thanks, David, I understand. Then, is there a way of disabling generating
the llvm. intrinsics? opt seems to have an option called
-disable-simplify-libcalls. However, in my case, it does not remove the
llvm.memcpy instruction from the bitcode.
On Thu, Feb 11, 2016 at 6:04 PM, David Blaikie <dblai...@gmail.com> wrote:
> There probably is a rule, but I don't know what it is - I would imagine
> memcpy is used when storing a whole aggregate (but then you'll get into ABI
> issues, etc - maybe if the struct contains only a single primitive type it
> just switches to a store, etc).
>
> On Thu, Feb 11, 2016 at 8:44 AM, Simona Simona <other.dev.sim...@gmail.com
> > wrote:
>
>> Thanks, David, this is useful.
>>
>> So sometimes the front-end generates llvm.memcpy instead of store
>> instructions.
>> Is there a rule in generating llvm.memcpy instructions instead of stores?
>> I would have the same question for other instrinsics, such as memset and
>> memmove.
>>
>> Thanks,
>> Simona
>>
>> On Thu, Feb 11, 2016 at 5:24 PM, David Blaikie <dblai...@gmail.com>
>> wrote:
>>
>>>
>>>
>>> On Thu, Feb 11, 2016 at 7:25 AM, Simona Simona via cfe-users <
>>> cfe-users@lists.llvm.org> wrote:
>>>
>>>> Hi,
>>>>
>>>> I'm using clang 3.4 to generate the bitcode of a C source file.
>>>> The source file is the following:
>>>>
>>>> typedef struct __attribute__ ((__packed__)) { float x, y; } myType;
>>>> myType make_float2(float x, float y) { myType f = { x, y }; return f; }
>>>>
>>>> int main(int argc, char* argv[])
>>>> {
>>>> myType myVar[5];
>>>>
>>>> for(int i=0;i<5;i++)
>>>> myVar[i] = make_float2(i,i);
>>>>
>>>> return(myVar[1].x);
>>>> }
>>>>
>>>> The bitcode is generated using the following command:
>>>> clang -c -emit-llvm -O0 -fno-vectorize -fno-slp-vectorize
>>>> -fno-lax-vector-conversions main.c -o main.bc
>>>>
>>>> target triple = "x86_64-unknown-linux-gnu"
>>>>
>>>> %struct.myType = type <{ float, float }>
>>>>
>>>> ; Function Attrs: nounwind uwtable
>>>> define <2 x float> @_Z11make_float2ff(float %x, float %y) #0 {
>>>> entry:
>>>> %retval = alloca %struct.myType, align 1
>>>> %x1 = getelementptr inbounds %struct.myType* %retval, i32 0, i32 0
>>>> store float %x, float* %x1, align 1
>>>> %y2 = getelementptr inbounds %struct.myType* %retval, i32 0, i32 1
>>>> store float %y, float* %y2, align 1
>>>> %0 = bitcast %struct.myType* %retval to <2 x float>*
>>>> %1 = load <2 x float>* %0, align 1
>>>> ret <2 x float> %1
>>>> }
>>>>
>>>> ; Function Attrs: nounwind uwtable
>>>> define i32 @main(i32 %argc, i8** %argv) #0 {
>>>> entry:
>>>> %myVar = alloca [100 x %struct.myType], align 16
>>>>
>>>
>>> Looks like your IR corresponds to an array of length 100, not 5 as in
>>> your source, but that's not too important
>>>
>>>
>>>> * %ref.tmp = alloca %struct.myType, align 1*
>>>> br label %for.cond
>>>>
>>>> for.cond: ; preds = %for.inc,
>>>> %entry
>>>> %i.0 = phi i32 [ 0, %entry ], [ %inc, %for.inc ]
>>>> %cmp = icmp slt i32 %i.0, 5
>>>> br i1 %cmp, label %for.body, label %for.end
>>>>
>>>> for.body: ; preds = %for.cond
>>>> %idxprom = sext i32 %i.0 to i64
>>>> %arrayidx = getelementptr inbounds [100 x %struct.myType]* %myVar,
>>>> i32 0, i64 %idxprom
>>>> %conv = sitofp i32 %i.0 to float
>>>> %conv1 = sitofp i32 %i.0 to float
>>>> * %call = call <2 x float> @_Z11make_float2ff(float %conv, float
>>>> %conv1)*
>>>> * %0 = bitcast %struct.myType* %ref.tmp to <2 x float>**
>>>> * store <2 x float> %call, <2 x float>* %0, align 1*
>>>> %1 = bitcast %struct.myType* %arrayidx to i8*
>>>> %2 = bitcast %struct.myType* %ref.tmp to i8*
>>>> call void @llvm.memcpy.p0i8.p0i8.i64(i8* %1, i8* %2, i64 8, i32 1, i1
>>>> false)
>>>>
>>>
>>> Here is the store ^ into your array (%1 is the destination, a bitcast of
>>> %arrayidx, which is the pointer into your array at index %idxprom, which is
>>> %i.0, etc) using the memcpy intrinsic, rather than a store instruction.
>>>
>>>
>>>> br label %for.inc
>>>>
>>>> for.inc: ; preds = %for.body
>>>> %inc = add nsw i32 %i.0, 1
>>>> br label %for.cond
>>>>
>>>> for.end: ; preds = %for.cond
>>>> %arrayidx2 = getelementptr inbounds [100 x %struct.myType]* %myVar,
>>>> i32 0, i64 1
>>>> %x = getelementptr inbounds %struct.myType* %arrayidx2, i32 0, i32 0
>>>> %3 = load float* %x, align 1
>>>> %conv3 = fptosi float %3 to i32
>>>> ret i32 %conv3
>>>> }
>>>>
>>>> Looking at the C source code there should be 5 store instructions
>>>> corresponding to the 5 assignments of myVar[0], myVar[1], myVar[2],
>>>> myVar[3] and myVar[4].
>>>> When I look at the bitcode however, I see 5 instances of *store <2 x
>>>> float> %call, <2 x float>* %0, align 1 *which correspond to 5 stores
>>>> at the same address
>>>> of %0 (which is actually %ref.tmp defined as *%ref.tmp = alloca
>>>> %struct.myType, align 1*).
>>>>
>>>> I would appreciate it if anyone could let me know how the 5 memory
>>>> accesses at the 5 *different* memory addresses are implemented in the
>>>> bitcode.
>>>>
>>>> Thanks,
>>>> Simona
>>>>
>>>>
>>>> _______________________________________________
>>>> cfe-users mailing list
>>>> cfe-users@lists.llvm.org
>>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-users
>>>>
>>>>
>>>
>>
>
_______________________________________________
cfe-users mailing list
cfe-users@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-users
