RFR: 8307351: (CmpI/L(AndI/L reg1 reg2)) on x86 can be optimized

[Tobias Hotz]

On Fri, 21 Apr 2023 17:30:39 GMT, Tobias Hotz <duke at openjdk.org> wrote:

> This patch aims to optimize a case where a And-Node followed by a Cmp-Node would not be converted into a single "test" instruction. The Architecture Description file currently only handles the cases where the right operand of the And-Node is a constant, but not if both are a register.
> Before this patch, a "and" followed by a "test" would be emitted, so the removed "and" means 2 bytes less have to be emitted.
> I've attached a JMH Benchmark to demonstrate the performance improvements. Here are the numbers of my Windows 11 machine:
> Before:
> 
> Benchmark                                                   Mode  Cnt   Score   Error  Units
> AndCmpTestInstruction.benchmarkOpaqueAndCmpEqualsInt        avgt    8  26,736 ± 0,131  ns/op
> AndCmpTestInstruction.benchmarkOpaqueAndCmpEqualsLong       avgt    8  24,305 ± 0,610  ns/op
> AndCmpTestInstruction.benchmarkStaticAndCmpEqualsInt        avgt    8  33,052 ± 0,056  ns/op
> AndCmpTestInstruction.benchmarkStaticLargeAndCmpEqualsLong  avgt    8  18,355 ± 0,030  ns/op
> AndCmpTestInstruction.benchmarkStaticSmallAndCmpEqualsLong  avgt    8  25,587 ± 0,107  ns/op
> 
> After:
> 
> Benchmark                                                   Mode  Cnt   Score   Error  Units  Improvement
> AndCmpTestInstruction.benchmarkOpaqueAndCmpEqualsInt        avgt    8  22,665 ± 0,170  ns/op  (~18%)
> AndCmpTestInstruction.benchmarkOpaqueAndCmpEqualsLong       avgt    8  18,880 ± 0,123  ns/op  (~29%)
> AndCmpTestInstruction.benchmarkStaticAndCmpEqualsInt        avgt    8  33,198 ± 0,126  ns/op  (unchanged)
> AndCmpTestInstruction.benchmarkStaticLargeAndCmpEqualsLong  avgt    8  18,427 ± 0,079  ns/op  (unchanged)
> AndCmpTestInstruction.benchmarkStaticSmallAndCmpEqualsLong  avgt    8  25,641 ± 0,168  ns/op  (unchanged)
> 
> As you can see, the cases with a small static mask have not improved as they have already been covered by another match rule. The test with the large static mask (benchmarkStaticLargeAndCmpEqualsLong) has a sorted instruction sequence as the value is moved to a register and it is not being used directly in the instruction, but there is no measurable performance uplift here.
> I've tested my changes using the Tier1 jtreg Tests on Windows.

Regarding operations on memory: I am not sure if this would be benificial, as (if I understand it correctly) these operations cannot be marco-fused. C++ compilers like clang or gcc also don't seem to emit test instructions with memory operands and immediates. So I would like to omit that from the pull request and limit it to register case, which never result in performance degradation.

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

PR Comment: https://git.openjdk.org/jdk/pull/13587#issuecomment-1539768370