RFR: JDK-8297660: x86: Redundant test+jump in C1 allocateArray

[Yi Yang]

On Fri, 25 Nov 2022 18:24:06 GMT, Thomas Stuefe <stuefe at openjdk.org> wrote:

> In `C1_MacroAssembler::initialize_body()` we test the input array length for 0. We do this again in `MacroAssembler::zero_memory()`. This results in a redundant test+jump instruction:
> 
> 
>  13442   0x00007f58a8ca2f02:   sub    $0x10,%rsi                 
>  13443   0x00007f58a8ca2f06:   je     0x00007f58a8ca2f26   <<
>  13444   0x00007f58a8ca2f0c:   test   %rsi,%rsi                     
>  13445   0x00007f58a8ca2f0f:   je     0x00007f58a8ca2f26    <<
>  13446   0x00007f58a8ca2f15:   xor    %rbx,%rbx
>  13447   0x00007f58a8ca2f18:   shr    $0x3,%rsi
>  13448   0x00007f58a8ca2f1c:   mov    %rbx,0x8(%rax,%rsi,8)
>  13449   0x00007f58a8ca2f21:   dec    %rsi
>  13450   0x00007f58a8ca2f24:   jne    0x00007f58a8ca2f1c           ;*anewarray {reexecute=0 rethrow=0 return_oop=0}
>  13451                                                             ; - java.lang.invoke.MethodHandles::<clinit>@24 (line 5109)
> 
> 
> Since `MacroAssembler::zero_memory()` is only ever called from `C1_MacroAssembler::initialize_body()`, it does not need to test for len=0, since its caller already does.
> 
> ---
> 
> Patch removes one test+jump and adds an assertion for len>0 to zero_memory.
> 
> Patch ran through SAP nightlies and GHAs.

Is it reasonable to remove such checking from the caller? Macro assembler is shared across all assemblers, zero_memory may be used in other place in the future due to its non-ad-hoc implementation.

-------------

PR: https://git.openjdk.org/jdk/pull/11372