[llvm-dev] BUGS n code generated for target i386 compiling __bswapdi3, and for target x86-64 compiling __bswapsi2()

[Stefan Kanthak via llvm-dev]

"Craig Topper" <craig.topper at gmail.com> wrote:

> bswapsi2 on the x86-64 isn't using the bswap instruction because "unsigned
> long" is 64-bits on x86-64 linux.

Believe me: I KNOW THIS!
Now take one more THOROUGH look at the C source of _bswapsi2():
1. ALL 4  & operations clear the higher 32 bits;
2. the 4 >> operations BSWAP the lower 32 bits.

Assuming that the argument was loaded into RDI,  this is a
    BSWAP EDI
in Intel x86-64 machine code!
Remember that in "long mode", writing to a 32-bit register
clears the higher 32 bits of the respective 64-bit register.

AGAIN: why doesn't LLVM generate the BSWAP operation here?

The expected code is

__bswapsi2: # @__bswapsi2
     bswap edi
     mov   rax, rdi
     ret

not amused
Stefan Kanthak

> But its 32-bits on x86-64 msvc.
> 
> Not sure about the bswapdi2 i386 case.
> 
> 
> ~Craig
> 
> 
> On Sun, Nov 25, 2018 at 8:03 AM Stefan Kanthak via llvm-dev <
> llvm-dev at lists.llvm.org> wrote:
> 
>> Hi @ll,
>>
>> targetting i386, LLVM/clang generates wrong code for the following
>> functions:
>>
>> unsigned long __bswapsi2 (unsigned long ul)
>> {
>>     return (((ul) & 0xff000000ul) >> 3 * 8)
>>          | (((ul) & 0x00ff0000ul) >>     8)
>>          | (((ul) & 0x0000ff00ul) <<     8)
>>          | (((ul) & 0x000000fful) << 3 * 8);
>> }
>>
>> unsigned long long __bswapdi2(unsigned long long ull)
>> {
>>     return ((ull & 0xff00000000000000ull) >> 7 * 8)
>>          | ((ull & 0x00ff000000000000ull) >> 5 * 8)
>>          | ((ull & 0x0000ff0000000000ull) >> 3 * 8)
>>          | ((ull & 0x000000ff00000000ull) >>     8)
>>          | ((ull & 0x00000000ff000000ull) <<     8)
>>          | ((ull & 0x0000000000ff0000ull) << 3 * 8)
>>          | ((ull & 0x000000000000ff00ull) << 5 * 8)
>>          | ((ull & 0x00000000000000ffull) << 7 * 8);
>> }
>>
>> You can find these sources in "compiler-rt/lib/builtins/bswapsi2.c"
>> and "compiler-rt/lib/builtins/bswapdi2.c", for example!
>>
>>
>> Compiled with "-O3 -target i386" this yields the following code
>> (see <https://godbolt.org/z/F4UIl4>):
>>
>> __bswapsi2: # @__bswapsi2
>>     push  ebp
>>     mov   ebp, esp
>>     mov   eax, dword ptr [ebp + 8]
>>     bswap eax
>>     pop   ebp
>>     ret
>>
>> __bswapdi2: # @__bswapdi2
>>     push  ebp
>>     mov   ebp, esp
>>     mov   edx, dword ptr [ebp + 8]
>>     mov   eax, dword ptr [ebp + 12]
>>     bswap eax
>>     bswap edx
>>     pop   ebp
>>     ret
>>
>> __bswapsi2() is correct, but __bswapdi2() NOT: swapping just the
>> halves of a "long long" is OBVIOUSLY WRONG!
>>
>> From the C source, the expected result for the input value
>> 0x0123456789ABCDEF is 0xEFCDAB8967452301; the compiled code but
>> produces 0x67452301EFCDAB89
>>
>>
>> And compiled for x86-64 this yields the following code (see
   ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>> <https://godbolt.org/z/uM9nvN>):
>>
>> __bswapsi2: # @__bswapsi2
>>     mov   eax, edi
>>     shr   eax, 24
>>     mov   rcx, rdi
>>     shr   rcx, 8
>>     and   ecx, 65280
>>     or    rax, rcx
>>     mov   rcx, rdi
>>     shl   rcx, 8
>>     and   ecx, 16711680
>>     or    rax, rcx
>>     and   rdi, 255
>>     shl   rdi, 24
>>     or    rax, rdi
>>     ret
>>
>> __bswapdi2: # @__bswapdi2
>>     bswap rdi
>>     mov   rax, rdi
>>     ret
>>
>> Both are correct, but __bswapsi2() should of course use BSWAP too!
>>
>>
>> Stefan Kanthak
>>
>> PS: for comparision with another compiler, take a look at
>>     <https://skanthak.homepage.t-online.de/msvc.html#example5>
>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>
>