RFR: 8204479: Bitwise AND on byte value sometimes produces wrong result

[Tobias Hartmann]

Hi Aleksey,

this looks good to me (and I think the change is trivial because it reverts a previous one).

Best regards,
Tobias

On 08.06.2018 09:59, Aleksey Shipilev wrote:
> Bug:
>   https://bugs.openjdk.java.net/browse/JDK-8204479
> 
> Fix:
>   http://cr.openjdk.java.net/~shade/8204479/webrev.03/
> 
> The trouble with unsigned comparisons is that original code would have generate:
> 
>  movzbl (mem), %r
>  cmp %r, $imm
>  jl ...
> 
> ...while the new code with compUB_mem_imm matcher generates:
> 
>  cmpb (mem), $imm
>  jl ...
> 
> Original code loaded non-sign-extended byte into int register and then safely did signed comparison.
> New code does signed byte comparison against memory, which is incorrect. It is still safe when we do
> proper LoadB, but not LoadUB.
> 
> Vladimir thinks the second matcher (testUB_mem_imm) suffers from the same issue. I don't think it is
> ever exposed in our code, because comparison with zero usually conditionally jumps on ZF. But testb
> does set SF to the MSB of and-ed operands, so we can accidentally use (incorrect) SF for unsigned
> byte in conditional jump and fail? I haven't been able to reproduce that in the test though.
> 
> This partially reverts JDK-8203628 change:
>   http://hg.openjdk.java.net/jdk/jdk/rev/197ee9d8e228
> 
> Testing: regression test, failing Nashorn tests
> 
> Thanks,
> -Aleksey
>