[JIT] Add `ccmp` and enable conditional compares for X64 #110826

anthonycanino · 2024-12-18T23:03:12Z

Overview

This PR is built on top of #108796.

This PR adds APX new ccmp instruction to the X86 backend, and enables some of the existing if-conversion functionality for X86.

Design

Currently, the if-conversion optimization is hidden behind a flag DOTNET_JitEnableApxIfConv and defaults to 0.

For reference, there is a unique extended evex encoding for ccmp:

where SC0 - SC3 encode the condition for ccmp to conditionally execute on (please see SDM Vol 1, Appendix B). If the status codes fail to satisfy the condition encoded by SC0 - SC3, no compare will be performed, and the OF, SF, ZF, and CF flags will be set to the default flag value (DFV) fields of, sf, zf and cf.

Testing

Note: The testing plan for APX work has been discussed in #106557, please refer to that PR for details, only results and comments will be posted in this PR. Results posted below.

Update comments. Merge the REX2 changes into the original legacy emit path bug fix: Set REX2.W with correct mask code. register encoding and prefix emitting logics. Add REX2 prefix emit logic bug fixes Add Stress mode for REX2 encoding and some bug fixes resolve comments: 1. add assertion check for UD opcodes. 2. add checks for EGPRs. Add REX2 to emitOutputAM, and let LEA to be REX2 compatible. Add REX2.X encoding for SIB byte But fixes: add REX2 prefix on the path in RI where MOV is specially handled. Enable REX2 encoding for `movups` fixed bugs in REX2 prefix emitting logic when working with map 1 instructions, and enabled REX2 for POPCNT legacy map index-er bug fixes some clean-up Adding initial APX unit testing path. Adding a coredistools dll that has LLVM APX disasm capability. It must be coppied into a CORE_ROOT manually. clean up work for REX2 narrow the REX2 scope to `sub` only some clean up based on the comments. bug fix resolve comment

- SV path is mostly for debugging purposes Added encoding unit tests for instructions with immediates

Code refactoring: AddX86PrefixIfNeeded.

… missing in JIT, may indicate these instructions are not being used in JIT, drop them for now.

…lled before adding any prefix.

Refactor REX2 encoding stress logics.

(this will have side effect that the estimated code will go up and mismatch with actual code size.)

…operly handled.

…om LOCK prefixd instructions.

…otion from LOCK prefixd instructions." This reverts commit 1be4b12.

…ill be prefixed.

dotnet-policy-service · 2024-12-18T23:03:57Z

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

anthonycanino · 2024-12-18T23:11:37Z

1. Emitter unit tests

The left is output from JitDisasm, and right from JitLateDisasm.

2. Intel SDE testing

Run the suite in SDE with JitEnableApxIfConv off:

Run the suite in SDE with JitenableApxIfConv on:

#3. SuperPMI results

Base is with JitEnableApxIfConv off. Diff is with JitenableApxIfConv on.

Diffs are based on 2,623,457 contexts (1,043,127 MinOpts, 1,580,330 FullOpts).

MISSED contexts: 2,983 (0.11%)

Base JIT options: JitBypassApxCheck=1

Diff JIT options: JitBypassApxCheck=1;JitEnableApxIfConv=1

Overall (-165,684 bytes)

Collection	Base size (bytes)	Diff size (bytes)	PerfScore in Diffs
aspnet.run.windows.x64.checked.mch	42,216,437	-6,257	-9.13%
benchmarks.run.windows.x64.checked.mch	8,860,704	-21,878	-4.99%
benchmarks.run_pgo.windows.x64.checked.mch	35,294,983	-25,089	-8.82%
benchmarks.run_tiered.windows.x64.checked.mch	12,613,813	-20,816	-4.88%
coreclr_tests.run.windows.x64.checked.mch	388,693,881	-10,675	-8.48%
libraries.crossgen2.windows.x64.checked.mch	44,941,069	+1,343	-8.64%
libraries.pmi.windows.x64.checked.mch	60,078,628	-7,782	-9.54%
libraries_tests.run.windows.x64.Release.mch	317,477,436	-54,130	-9.82%
libraries_tests_no_tiered_compilation.run.windows.x64.Release.mch	147,744,535	-20,077	-6.96%
realworld.run.windows.x64.checked.mch	10,242,976	-451	-6.47%
smoke_tests.nativeaot.windows.x64.checked.mch	4,496,305	+128	-6.51%

FullOpts (-165,684 bytes)

Collection	Base size (bytes)	Diff size (bytes)	PerfScore in Diffs
aspnet.run.windows.x64.checked.mch	23,294,229	-6,257	-9.13%
benchmarks.run.windows.x64.checked.mch	8,860,282	-21,878	-4.99%
benchmarks.run_pgo.windows.x64.checked.mch	20,646,767	-25,089	-8.82%
benchmarks.run_tiered.windows.x64.checked.mch	3,214,524	-20,816	-4.88%
coreclr_tests.run.windows.x64.checked.mch	117,503,708	-10,675	-8.48%
libraries.crossgen2.windows.x64.checked.mch	44,939,354	+1,343	-8.64%
libraries.pmi.windows.x64.checked.mch	59,965,735	-7,782	-9.54%
libraries_tests.run.windows.x64.Release.mch	128,808,465	-54,130	-9.82%
libraries_tests_no_tiered_compilation.run.windows.x64.Release.mch	137,094,840	-20,077	-6.96%
realworld.run.windows.x64.checked.mch	10,018,142	-451	-6.47%
smoke_tests.nativeaot.windows.x64.checked.mch	4,495,216	+128	-6.51%

jakobbotsch · 2024-12-19T07:08:11Z

Happy to see this.

I have one request. Can you please split this PR into two? A first PR that adds the necessary support to codegen/emit to generate the new instructions. The second PR to enable the existing support for emitting conditional compares in the middle-end/lowering? All the work should be in the former PR, which should be free of diffs, while the latter PR should mostly be unifying and enabling the existing support that exists for arm64.

Also, the title is inaccurate. If-conversion is enabled for x64 already. The use of conditional compares is an orthogonal optimization to if-conversion.

anthonycanino · 2024-12-19T20:01:37Z

Happy to see this.

I have one request. Can you please split this PR into two? A first PR that adds the necessary support to codegen/emit to generate the new instructions. The second PR to enable the existing support for emitting conditional compares in the middle-end/lowering? All the work should be in the former PR, which should be free of diffs, while the latter PR should mostly be unifying and enabling the existing support that exists for arm64.

Also, the title is inaccurate. If-conversion is enabled for x64 already. The use of conditional compares is an orthogonal optimization to if-conversion.

Will do. I plan to get this PR in a clean build/run shape and then will split.

Major changes include: 1. Adding `ccmp` (and promoted cmp) logic into emitter backend. 2. Hooking `TryLowerAndOrToCCMP` from ARM pathway into Intel pathway. 3. Hooking `optimizeCompareChainCondBlock` from ARM pathway into Intel pathway.

Ruihan-Yin added 30 commits November 19, 2024 12:05

resolve comments

d1afc68

refactor register encoding for REX2

2335aa3

merge REX2 path to legacy path

6578c58

Enable REX2 in more instructions.

01eeb80

Avoid repeatedly estimate the size of REX2 prefix

690aee3

Enable REX2 encoding on RI and SV path

31d7fb4

- SV path is mostly for debugging purposes Added encoding unit tests for instructions with immediates

Add rex2 support to rotate and shift.

a995878

CR session.

74aacf6

Testing infra updates: assert REX2 is enabled.

c330927

Code refactoring: AddX86PrefixIfNeeded.

revert rcl_N and rcr_N, tp and latency data for these instructions is…

fbf20d1

… missing in JIT, may indicate these instructions are not being used in JIT, drop them for now.

partially enable REX2 on emitOutputAM, case covered: R_AR and AR_R.

ea02e70

Adding unit tests.

c74b801

push, pop, inc, dec, neg, not, xadd, shld, shrd, cmpxchg, setcc, bswap.

34980b4

bug fix for bswap

2ffdbeb

bt

3a729bb

xchg, idiv

d943b03

Make sure add REX2 prefix if register encoding for EGPRs are being ca…

c8fee9c

…lled before adding any prefix.

Ensure code size is correctly computed in R_R_I path.

6ec0e97

clean up

1d01003

Change all AddSimdPrefix to AddX86Prefix

1acc219

Refactor REX2 encoding stress logics.

div, mulEAX

87ad443

filter out test from REX2 encoding when using ACC form.

bb9905a

(this will have side effect that the estimated code will go up and mismatch with actual code size.)

Make sure REX prefix will not be added when emitting with REX2.

86083b2

resolve comments.

dfe8760

make sure the APX debug knob is only available under debug build.

64761cd

clean up some out-dated code.

f1aba62

enable movsxd

f5cc5a8

Enable "Call"

7ca8433

Enable "JMP"

bc4d225

Ruihan-Yin added 19 commits November 20, 2024 12:24

Add Evex.nf to instrDesc

25d66bf

{nf} add reg, reg

a19da9e

Enable EVEX.NF in more instructions

2e8d714

more instructions

df59342

comments.

226fabb

lzcnt, tzcnt, popcnt

36c6631

Exclude ACC form from EVEX promotion.

5f8a01d

BMI instructions.

0453630

bug fixes

07868bc

Tweak the code size calculation to make sure REX2 and APX-EVEX are pr…

69f7e8b

…operly handled.

bug fixes for stress mode

1c1a894

Add idEvexNoPromotion to emitter to exclude the APX-EVEX promotion fr…

1be4b12

…om LOCK prefixd instructions.

resolve merge error

bfb06c7

fix merge error

9541a99

Revert "Add idEvexNoPromotion to emitter to exclude the APX-EVEX prom…

543d949

…otion from LOCK prefixd instructions." This reverts commit 1be4b12.

bug fix

a879019

introduce _no_evex suffix for some instructions for cases when LOCK w…

55cbda6

…ill be prefixed.

Merge remote-tracking branch 'origin/main' into apx-evex-nf-nov

a9a3d5c

fix merge errors.

69c7a29

dotnet-issue-labeler bot added the area-CodeGen-coreclr CLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI label Dec 18, 2024

dotnet-policy-service bot added the community-contribution Indicates that the PR has been added by a community member label Dec 18, 2024

anthonycanino force-pushed the apx-if-conv branch from 984c510 to df08843 Compare December 19, 2024 00:35

anthonycanino changed the title ~~[JIT] Add ccmp and enable if-conversion for X86~~ [JIT] Add ccmp and enable conditional compares for X64 Dec 19, 2024

anthonycanino force-pushed the apx-if-conv branch from df08843 to c5fae2e Compare December 19, 2024 20:18

Adds if-conversion lowering for Intel platform.

519e290

Major changes include: 1. Adding `ccmp` (and promoted cmp) logic into emitter backend. 2. Hooking `TryLowerAndOrToCCMP` from ARM pathway into Intel pathway. 3. Hooking `optimizeCompareChainCondBlock` from ARM pathway into Intel pathway.

anthonycanino force-pushed the apx-if-conv branch from c5fae2e to 519e290 Compare December 19, 2024 21:33

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[JIT] Add `ccmp` and enable conditional compares for X64 #110826

[JIT] Add `ccmp` and enable conditional compares for X64 #110826

anthonycanino commented Dec 18, 2024

dotnet-policy-service bot commented Dec 18, 2024

anthonycanino commented Dec 18, 2024 •

edited

Loading

jakobbotsch commented Dec 19, 2024 •

edited

Loading

anthonycanino commented Dec 19, 2024

[JIT] Add ccmp and enable conditional compares for X64 #110826

Are you sure you want to change the base?

[JIT] Add ccmp and enable conditional compares for X64 #110826

Conversation

anthonycanino commented Dec 18, 2024

Overview

Design

Testing

dotnet-policy-service bot commented Dec 18, 2024

anthonycanino commented Dec 18, 2024 • edited Loading

1. Emitter unit tests

2. Intel SDE testing

jakobbotsch commented Dec 19, 2024 • edited Loading

anthonycanino commented Dec 19, 2024

[JIT] Add `ccmp` and enable conditional compares for X64 #110826

[JIT] Add `ccmp` and enable conditional compares for X64 #110826

anthonycanino commented Dec 18, 2024 •

edited

Loading

jakobbotsch commented Dec 19, 2024 •

edited

Loading