RFR: 8346038: [REDO] - [C1] LIR Operations with one input should be implemented as LIR_Op1

[Martin Doerr]

On Thu, 12 Dec 2024 13:12:45 GMT, Martin Doerr <mdoerr at openjdk.org> wrote:

> 1st commit: Same as https://github.com/openjdk/jdk/commit/a21d21f4d7b74e21f68b6bf9c5dc9ba7d3f9963c
> 2nd commit: Removal of special "Knights Landing" code for `lir_abs` and `lir_neg` from C1.
> 
> Testing:
> make run-test TEST="test/hotspot/jtreg/compiler" JTREG="VM_OPTIONS=-XX:UseAVX=3 -XX:+UnlockDiagnosticVMOptions -XX:+UseKNLSetting"
> All passed.
> 
> This is how this PR changes the C1 code on Knights Landing CPUs (emulated by -XX:+UseKNLSetting):
> 
> `lir_abs` without this patch:
> 
> vmovsd -0x63(%rip),%xmm1
> vpandnd %zmm0,%zmm1,%zmm0
> 
> 
> `lir_neg` without this patch:
> 
> vmovsd -0x63(%rip),%xmm1
> vpxord %zmm0,%zmm1,%zmm0
> 
> 
> (The `vmovsd` loads the `LIR_OprFact::doubleConst(-0.0)`.)
> 
> `lir_abs` with this patch:
> 
> vandpd 0xa1b213d(%rip),%xmm0,%xmm0
> 
> 
> `lir_neg` with this patch:
> 
> vxorpd 0xa12585d(%rip),%xmm0,%xmm0
> 
> 
> New code is faster on our machine (using -XX:+UseKNLSetting).

@vnkozlov: This is now the version with updates from @sviswa7. Could you help finding a reviewer and making sure it gets properly tested, please? I think this cleanup makes the code better readable (and also a bit faster). I also prefer handling `UseKNLSetting` in macroAssembler_x86.cpp instead of in multiple C1 files.

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

PR Comment: https://git.openjdk.org/jdk/pull/22709#issuecomment-2582973107