RFR: 8279225: C1 longs comparison operation destroys argument registers

[Andrew Haley]

On Mon, 27 Dec 2021 08:07:37 GMT, Hao Sun <haosun at openjdk.org> wrote:

>> Several regression tests are failed on arm32 CPU if tiered compilation is enabled.
>> 
>> The list includes
>> java/math/BigDecimal/DivideMcTests
>> java/util/Arrays/Sorting.java
>> java/util/Arrays/SortingNearlySortedPrimitive.java
>> java/util/concurrent/tck/JSR166TestCase
>> java/util/stream/SliceOpTest.java
>> etc
>> 
>> It appears C1 comp_op for long operands destroys arguments registers:
>> void LIR_Assembler::comp_op(LIR_Condition condition, LIR_Opr opr1, LIR_Opr opr2, LIR_Op2* op) {
>> ....
>>       Register ylo = opr2->as_register_lo();
>>       Register yhi = opr2->as_register_hi();
>>       if (condition == lir_cond_equal || condition == lir_cond_notEqual) {
>>         __ teq(xhi, yhi);
>>         __ teq(xlo, ylo, eq);
>>       } else {
>>         __ subs(xlo, xlo, ylo); // <<< incorrect
>>         __ sbcs(xhi, xhi, yhi); // <<< incorrect
>>       }
>> ...
>> } 
>> 
>> Tested with hotspot_tier2, jdk_tier3 on linux_arm
>
> src/hotspot/cpu/arm/c1_LIRAssembler_arm.cpp line 1823:
> 
>> 1821:         __ teq(xlo, ylo, eq);
>> 1822:       } else {
>> 1823:         __ cmp(xlo, ylo);
> 
> nit: we may want to use `__ subs(Rtemp, xlo, ylo);` here to align with the usage in match rules in arm.ad, e.g., `compL_reg_reg_LEGT`. It's okay to me if you use `cmp` anyway.

I agree.

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

PR: https://git.openjdk.java.net/jdk/pull/6934