llvm dev - Oct 2013 - [LLVMdev] Exploiting 'unreachable' for optimization purposes

Hi,

When clang/llvm compiles the following sample (with -O2) it optimizes 
away the second comparison in test1 but not in test2. Is this handling 
of 'unreachable' by purpose, or is this just a shortcoming of the 
current optimization passes? GCC and MSVC (with the equivalent code 
using the __assume intrinsic) both optimize away the comparison in test2.


void f1();

void f2();

void abort() __attribute__((noreturn));

void test1(int x) {
   if (x < 0) abort();
   // the following comparison is optimized away
   if (x < 0) f1(); else f2();
}

void test2(int x) {
   if (x < 0) __builtin_unreachable();
   // the following comparison is NOT optimized away
   if (x < 0) f1(); else f2();
}

- Stephan

On Oct 24, 2013, at 5:22 AM, Stephan Tolksdorf <st at quanttec.com> wrote:
> Hi,
> 
> When clang/llvm compiles the following sample (with -O2) it optimizes away
the second comparison in test1 but not in test2. Is this handling of
'unreachable' by purpose, or is this just a shortcoming of the current
optimization passes? GCC and MSVC (with the equivalent code using the __assume
intrinsic) both optimize away the comparison in test2.
Hi Stephan,

I think this is just a shortcoming. Can you file a bug and assign it to me? I
have another limitation related to unreachable that I need to take a look at,
and when I do so I will take a look at this as well.

Thanks,

Mark
> 
> 
> void f1();
> 
> void f2();
> 
> void abort() __attribute__((noreturn));
> 
> void test1(int x) {
>  if (x < 0) abort();
>  // the following comparison is optimized away
>  if (x < 0) f1(); else f2();
> }
> 
> void test2(int x) {
>  if (x < 0) __builtin_unreachable();
>  // the following comparison is NOT optimized away
>  if (x < 0) f1(); else f2();
> }
> 
> - Stephan
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev

Hi Mark,

I just dug out a bug from 2006 which since 2009 seems to also have 
covered efforts to make better use of unreachable for optimization 
purposes: http://llvm.org/bugs/show_bug.cgi?id=810

I've added a comment and assigned it to you as requested.

Thanks,
   Stephan

On 24.10.13 19:05, Mark Lacey wrote:
>
> On Oct 24, 2013, at 5:22 AM, Stephan Tolksdorf <st at quanttec.com>
wrote:
>
>> Hi,
>>
>> When clang/llvm compiles the following sample (with -O2) it optimizes
away the second comparison in test1 but not in test2. Is this handling of
'unreachable' by purpose, or is this just a shortcoming of the current
optimization passes? GCC and MSVC (with the equivalent code using the __assume
intrinsic) both optimize away the comparison in test2.
>
> Hi Stephan,
>
> I think this is just a shortcoming. Can you file a bug and assign it to me?
I have another limitation related to unreachable that I need to take a look at,
and when I do so I will take a look at this as well.
>
> Thanks,
>
> Mark
>
>>
>>
>> void f1();
>>
>> void f2();
>>
>> void abort() __attribute__((noreturn));
>>
>> void test1(int x) {
>>   if (x < 0) abort();
>>   // the following comparison is optimized away
>>   if (x < 0) f1(); else f2();
>> }
>>
>> void test2(int x) {
>>   if (x < 0) __builtin_unreachable();
>>   // the following comparison is NOT optimized away
>>   if (x < 0) f1(); else f2();
>> }
>>
>> - Stephan
>> _______________________________________________
>> LLVM Developers mailing list
>> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>