NotificationsYou must be signed in to change notification settings
Fork5.2k
Star17.2k

JIT: Add scalar evolution analysis and do IV widening based on it#97865

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to ourterms of service andprivacy statement. We’ll occasionally send you account related emails.

Already on GitHub?Sign in to your account

Jump to bottom

Merged

jakobbotsch merged 73 commits intodotnet:mainfromjakobbotsch:simplify-ivs

Feb 27, 2024

Merged

JIT: Add scalar evolution analysis and do IV widening based on it#97865

jakobbotsch merged 73 commits intodotnet:mainfromjakobbotsch:simplify-ivs

Feb 27, 2024

Conversation

Copy link

Member

jakobbotsch commentedFeb 2, 2024•
edited
Loading

This adds a new phase meant for optimizing induction variables. It adds infrastructure for SSA-based analysis of induction variables (scalar evolution analysis), and uses it to do induction variable widening.
For example, with this optimization, codegen for

[MethodImpl(MethodImplOptions.NoInlining)]staticintFoo(int[]arr){intsum=0;for(inti=0;i<arr.Length;i++){sum+=arr[i];}returnsum;}

goes from

; Assembly listing for method ConsoleApp34.Program:Foo(int[]):int (FullOpts); Emitting BLENDED_CODE for X64 with AVX - Windows; FullOpts code; optimized code; rsp based frame; fully interruptible; No PGO data; Final local variable assignments;;  V00 arg0         [V00,T02] (  4,  7   )     ref  ->  rcx         class-hnd single-def <int[]>;  V01 loc0         [V01,T01] (  4, 10   )     int  ->  rax;  V02 loc1         [V02,T00] (  5, 17   )     int  ->  rdx;  V03 OutArgs      [V03    ] (  1,  1   )  struct (32) [rsp+0x00]  do-not-enreg[XS] addr-exposed "OutgoingArgSpace";  V04 cse0         [V04,T03] (  3,  6   )     int  ->   r8         "CSE - aggressive";; Lcl frame size = 40G_M8112_IG01:subrsp,40;; size=4 bbWeight=1 PerfScore 0.25G_M8112_IG02:xoreax,eaxxoredx,edxmovr8d, dword ptr[rcx+0x08]testr8d,r8djle      SHORT G_M8112_IG04align[0 bytes for IG03];; size=13 bbWeight=1 PerfScore 3.75G_M8112_IG03:movr10d,edxaddeax, dword ptr[rcx+4*r10+0x10]incedxcmpr8d,edxjg       SHORT G_M8112_IG03;; size=15 bbWeight=4 PerfScore 19.00G_M8112_IG04:addrsp,40ret;; size=5 bbWeight=1 PerfScore 1.25; Total bytes of code 37, prolog size 4, PerfScore 24.25, instruction count 14, allocated bytes for code 37 (MethodHash=d1cce04f) for method ConsoleApp34.Program:Foo(int[]):int (FullOpts); ============================================================

; Assembly listing for method ConsoleApp34.Program:Foo(int[]):int (FullOpts); Emitting BLENDED_CODE for X64 with AVX - Windows; FullOpts code; optimized code; rsp based frame; fully interruptible; No PGO data; Final local variable assignments;;  V00 arg0         [V00,T02] (  4,  7   )     ref  ->  rcx         class-hnd single-def <int[]>;  V01 loc0         [V01,T01] (  4, 10   )     int  ->  rax;* V02 loc1         [V02,T04] (  0,  0   )     int  ->  zero-ref;  V03 OutArgs      [V03    ] (  1,  1   )  struct (32) [rsp+0x00]  do-not-enreg[XS] addr-exposed "OutgoingArgSpace";  V04 tmp1         [V04,T00] (  5, 17   )    long  ->   r8         "Widened primary induction variable";  V05 cse0         [V05,T03] (  3,  6   )     int  ->  rdx         "CSE - aggressive";; Lcl frame size = 40G_M8112_IG01:  ;; offset=0x0000subrsp,40;; size=4 bbWeight=1 PerfScore 0.25G_M8112_IG02:  ;; offset=0x0004xoreax,eaxmovedx, dword ptr[rcx+0x08]testedx,edxjle      SHORT G_M8112_IG04xorr8d,r8dalign[0 bytes for IG03];; size=12 bbWeight=1 PerfScore 3.75G_M8112_IG03:  ;; offset=0x0010addeax, dword ptr[rcx+4*r8+0x10]incr8dcmpedx,r8djg       SHORT G_M8112_IG03;; size=13 bbWeight=4 PerfScore 18.00G_M8112_IG04:  ;; offset=0x001Daddrsp,40ret;; size=5 bbWeight=1 PerfScore 1.25; Total bytes of code 34, prolog size 4, PerfScore 23.25, instruction count 13, allocated bytes for code 34 (MethodHash=d1cce04f) for method ConsoleApp34.Program:Foo(int[]):int (FullOpts)

where we were able to drop a zero extension of the index inside the loop. In the future I plan to build strength reduction on top of the same analysis package.

The analysis is inspired by [1] and by LLVM's scalar evolution package. It provides a small IR that represents the evolving value of IR nodes inside loops. At the core of this IR is the notion of an "add recurrence", which describes a changing value as<loop, start, step>; the value of such an add recurrence is$start + N * step$, where N is the iteration index. Currently only simple add recurrences are supported where the start and step are either constants or invariant locals, but the framework generalizes nicely to allow chains of recurrences if we wish to support that. The IR also supports constants, invariant locals and operators on top of these (casts, adds, multiplications and shifts).

Analysis example

For the IR for the above, the analysis produces the following:

Analyzing scalar evolution inL00header:BB03Members (1):BB03Entry:BB02->BB03Exit:BB03->BB04Back:BB03->BB03BB03 [0001] [006..016)->BB03,BB04 (cond), preds={BB02,BB03} succs={BB04,BB03}STMT00009 (??? ...??? )N004 (0,0) [000045]DA---------▌STORE_LCL_VAR intV01 loc0         d:3N003 (0,0) [000044]-----------└──▌PHI       intN001 (0,0) [000050]----------- predBB03├──▌PHI_ARG   intV01 loc0         u:4N002 (0,0) [000047]----------- predBB02└──▌PHI_ARG   intV01 loc0         u:2[000047]=>0STMT00007 (??? ...??? )N004 (0,0) [000041]DA---------▌STORE_LCL_VAR intV02 loc1         d:3N003 (0,0) [000040]-----------└──▌PHI       intN001 (0,0) [000051]----------- predBB03├──▌PHI_ARG   intV02 loc1         u:4N002 (0,0) [000048]----------- predBB02└──▌PHI_ARG   intV02 loc1         u:2[000051]=> <L00, 1, 1>[000048]=>0[000041]=> <L00, 0, 1>

For example, here it was able to show thatV02, the index, is a primary induction variable; it is an add recurrence that starts at 0 and steps by 1 every iteration of the loop. It also showed that the value that comes from the backedge is similarly an add recurrence, except that it starts at 1 in the first loop.

STMT00003 (0x006[E-] ...0x00B )N015 (8,9) [000015]DA--GO-----▌STORE_LCL_VAR intV01 loc0         d:4N014 (8,9) [000014]----GO-----└──▌ADD       intN012 (6,7) [000033]----GO-N---├──▌COMMA     intN001 (0,0) [000025]-----------│├──▌NOP       voidN011 (6,7) [000034] n---GO-----│└──▌IND       intN010 (3,5) [000032]-----O-----│└──▌ARR_ADDR  byref int[]N009 (3,5) [000031]-------N---│└──▌ADD       byrefN002 (1,1) [000022]-----------│├──▌LCL_VAR   refV00 arg0         u:1N008 (4,5) [000030]-------N---│└──▌ADD       longN006 (3,4) [000028]-------N---│├──▌LSH       longN004 (2,3) [000026]---------U-││├──▌CAST      long<- uintN003 (1,1) [000023]-----------│││└──▌LCL_VAR   intV02 loc1         u:3N005 (1,1) [000027]-------N---││└──▌CNS_INT   long2N007 (1,1) [000029]-----------│└──▌CNS_INT   long16N013 (1,1) [000009]-----------└──▌LCL_VAR   intV01 loc0         u:3 (last use)[000022]=>V00.1[000023]=> <L00, 0, 1>[000026]=> <L00, 0, 1>[000027]=>2[000028]=> <L00, 0, 4>[000029]=>16[000030]=> <L00, 16, 4>[000031]=> <L00, (V00.1 + 16), 4>[000032]=> <L00, (V00.1 + 16), 4>

This one is more interesting since we can see hints of how strength reduction is going to utilize the information. In particular, the analysis was able to show that the address[000032] is also an add recurrence; it starts at value(V00.1 + 16) (the address of the first array element) and steps by4 in every iteration.

STMT00004 (0x00C[E-] ...0x00F )N004 (3,3) [000019]DA---------▌STORE_LCL_VAR intV02 loc1         d:4N003 (3,3) [000018]-----------└──▌ADD       intN001 (1,1) [000016]-----------├──▌LCL_VAR   intV02 loc1         u:3 (last use)N002 (1,1) [000017]-----------└──▌CNS_INT   int1[000016]=> <L00, 0, 1>[000017]=>1[000018]=> <L00, 1, 1>[000019]=> <L00, 1, 1>STMT00002 (0x010[E-] ...0x014 )N005 (7,7) [000008]---X-------▌JTRUE     voidN004 (5,5) [000007]J--X---N---└──▌GT        intN002 (3,3) [000006]---X-------├──▌ARR_LENGTH intN001 (1,1) [000005]-----------│└──▌LCL_VAR   refV00 arg0         u:1N003 (1,1) [000004]-----------└──▌LCL_VAR   intV02 loc1         u:4[000005]=>V00.1[000004]=> <L00, 1, 1>

Here the analysis shows that the array object is invariant (otherwise analysis would fail) and that the compared index is an add recurrence that starts at 1.

When the induction variable has uses after the loop the widening pass will store the widened version back to the old local. This is only possible if all exits where the old local is live-in are not critical blocks in the sense that all their preds must come from inside the loop.
Exit canonicalization ensure sthat this is usually the case, but RBO/assertion prop may have uncovered new natural loops, so we still have to repeat the check.

[1] Michael Wolfe. 1992. Beyond induction variables. In Proceedings of the ACM SIGPLAN 1992 conference on Programming language design and implementation (PLDI '92). Association for Computing Machinery, New York, NY, USA, 162–174.https://doi.org/10.1145/143095.143131

jakobbotsch added11 commits

January 30, 2024 14:51

WIP

c86effc

WIP

f8f899e

Check invariance on IR instead of on SCEV to avoid cycles

e2bca3a

Some work on recursive scevs

f400f03

Simplify for now

05562bb

IV widening, fix bugs

3036826

Merge branch 'main' of github.com:dotnet/runtime into simplify-ivs

8f0230e

Bail for non-integer constants

974e7f9

Add costing

c510054

Run jit-format

96d5bec

Clean up

0592111

ghost assignedjakobbotsch

Feb 2, 2024

ghost added the area-CodeGen-coreclrCLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI label

Feb 2, 2024

Copy link

ghost commentedFeb 2, 2024

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

Issue Details

[MethodImpl(MethodImplOptions.NoInlining)]staticintFoo(int[]arr){intsum=0;foreach(intvinarr){sum+=v;}returnsum;}

goes from

; Assembly listing for method ConsoleApp34.Program:Foo(int[]):int (FullOpts); Emitting BLENDED_CODE for X64 with AVX - Windows; FullOpts code; optimized code; rsp based frame; fully interruptible; No PGO data; Final local variable assignments;;  V00 arg0         [V00,T02] (  4,  7   )     ref  ->  rcx         class-hnd single-def <int[]>;  V01 loc0         [V01,T01] (  4, 10   )     int  ->  rax;  V02 loc1         [V02,T00] (  5, 17   )     int  ->  rdx;  V03 OutArgs      [V03    ] (  1,  1   )  struct (32) [rsp+0x00]  do-not-enreg[XS] addr-exposed "OutgoingArgSpace";  V04 cse0         [V04,T03] (  3,  6   )     int  ->   r8         "CSE - aggressive";; Lcl frame size = 40G_M8112_IG01:subrsp,40;; size=4 bbWeight=1 PerfScore 0.25G_M8112_IG02:xoreax,eaxxoredx,edxmovr8d, dword ptr[rcx+0x08]testr8d,r8djle      SHORT G_M8112_IG04align[0 bytes for IG03];; size=13 bbWeight=1 PerfScore 3.75G_M8112_IG03:movr10d,edxaddeax, dword ptr[rcx+4*r10+0x10]incedxcmpr8d,edxjg       SHORT G_M8112_IG03;; size=15 bbWeight=4 PerfScore 19.00G_M8112_IG04:addrsp,40ret;; size=5 bbWeight=1 PerfScore 1.25; Total bytes of code 37, prolog size 4, PerfScore 24.25, instruction count 14, allocated bytes for code 37 (MethodHash=d1cce04f) for method ConsoleApp34.Program:Foo(int[]):int (FullOpts); ============================================================

;  V00 arg0         [V00,T02] (  4,  7   )     ref  ->  rcx         class-hnd single-def <int[]>;  V01 loc0         [V01,T01] (  4, 10   )     int  ->  rax;* V02 loc1         [V02,T04] (  0,  0   )     int  ->  zero-ref;  V03 OutArgs      [V03    ] (  1,  1   )  struct (32) [rsp+0x00]  do-not-enreg[XS] addr-exposed "OutgoingArgSpace";  V04 tmp1         [V04,T00] (  5, 17   )    long  ->   r8         "Widened primary induction variable";  V05 cse0         [V05,T03] (  3,  6   )     int  ->  rdx         "CSE - aggressive";; Lcl frame size = 40G_M8112_IG01:subrsp,40;; size=4 bbWeight=1 PerfScore 0.25G_M8112_IG02:xoreax,eaxmovedx, dword ptr[rcx+0x08]testedx,edxjle      SHORT G_M8112_IG04xorr8d,r8djmp      SHORT G_M8112_IG03align[0 bytes for IG03];; size=14 bbWeight=1 PerfScore 5.75G_M8112_IG03:addeax, dword ptr[rcx+4*r8+0x10]incr8dcmpedx,r8djg       SHORT G_M8112_IG03;; size=13 bbWeight=4 PerfScore 18.00G_M8112_IG04:addrsp,40ret;; size=5 bbWeight=1 PerfScore 1.25; Total bytes of code 36, prolog size 4, PerfScore 25.25, instruction count 14, allocated bytes for code 36 (MethodHash=d1cce04f) for method ConsoleApp34.Program:Foo(int[]):int (FullOpts)

where we were able to drop a zero extension of the index inside the loop. In the future I plan to build strength reduction on top of the same analysis package.

Analysis example

For the IR for the above, the analysis produces the following:

Analyzing scalar evolution inL00header:BB03Members (1):BB03Entry:BB02->BB03Exit:BB03->BB04Back:BB03->BB03BB03 [0001] [006..016)->BB03,BB04 (cond), preds={BB02,BB03} succs={BB04,BB03}STMT00009 (??? ...??? )N004 (0,0) [000045]DA---------▌STORE_LCL_VAR intV01 loc0         d:3N003 (0,0) [000044]-----------└──▌PHI       intN001 (0,0) [000050]----------- predBB03├──▌PHI_ARG   intV01 loc0         u:4N002 (0,0) [000047]----------- predBB02└──▌PHI_ARG   intV01 loc0         u:2[000047]=>0STMT00007 (??? ...??? )N004 (0,0) [000041]DA---------▌STORE_LCL_VAR intV02 loc1         d:3N003 (0,0) [000040]-----------└──▌PHI       intN001 (0,0) [000051]----------- predBB03├──▌PHI_ARG   intV02 loc1         u:4N002 (0,0) [000048]----------- predBB02└──▌PHI_ARG   intV02 loc1         u:2[000051]=> <L00, 1, 1>[000048]=>0[000041]=> <L00, 0, 1>

STMT00003 (0x006[E-] ...0x00B )N015 (8,9) [000015]DA--GO-----▌STORE_LCL_VAR intV01 loc0         d:4N014 (8,9) [000014]----GO-----└──▌ADD       intN012 (6,7) [000033]----GO-N---├──▌COMMA     intN001 (0,0) [000025]-----------│├──▌NOP       voidN011 (6,7) [000034] n---GO-----│└──▌IND       intN010 (3,5) [000032]-----O-----│└──▌ARR_ADDR  byref int[]N009 (3,5) [000031]-------N---│└──▌ADD       byrefN002 (1,1) [000022]-----------│├──▌LCL_VAR   refV00 arg0         u:1N008 (4,5) [000030]-------N---│└──▌ADD       longN006 (3,4) [000028]-------N---│├──▌LSH       longN004 (2,3) [000026]---------U-││├──▌CAST      long<- uintN003 (1,1) [000023]-----------│││└──▌LCL_VAR   intV02 loc1         u:3N005 (1,1) [000027]-------N---││└──▌CNS_INT   long2N007 (1,1) [000029]-----------│└──▌CNS_INT   long16N013 (1,1) [000009]-----------└──▌LCL_VAR   intV01 loc0         u:3 (last use)[000022]=>V00.1[000023]=> <L00, 0, 1>[000026]=> <L00, 0, 1>[000027]=>2[000028]=> <L00, 0, 4>[000029]=>16[000030]=> <L00, 16, 4>[000031]=> <L00, (V00.1 + 16), 4>[000032]=> <L00, (V00.1 + 16), 4>

STMT00004 (0x00C[E-] ...0x00F )N004 (3,3) [000019]DA---------▌STORE_LCL_VAR intV02 loc1         d:4N003 (3,3) [000018]-----------└──▌ADD       intN001 (1,1) [000016]-----------├──▌LCL_VAR   intV02 loc1         u:3 (last use)N002 (1,1) [000017]-----------└──▌CNS_INT   int1[000016]=> <L00, 0, 1>[000017]=>1[000018]=> <L00, 1, 1>[000019]=> <L00, 1, 1>STMT00002 (0x010[E-] ...0x014 )N005 (7,7) [000008]---X-------▌JTRUE     voidN004 (5,5) [000007]J--X---N---└──▌GT        intN002 (3,3) [000006]---X-------├──▌ARR_LENGTH intN001 (1,1) [000005]-----------│└──▌LCL_VAR   refV00 arg0         u:1N003 (1,1) [000004]-----------└──▌LCL_VAR   intV02 loc1         u:4[000005]=>V00.1[000004]=> <L00, 1, 1>

I do not think we can enable this by default yet, for two reasons:

It breaks range check in some cases since range check does not handleTYP_LONG locals well.
Currently it rebuilds SSA fully from scratch whenever it widened an IV widening. There is some tension in this phase since it is SSA-based but also creates new locals that we really want to put in SSA. My plan is to add the ability to incrementally update SSA to get rid of this rebuilding.

Also, there is work necessary to properly prove that some analysis and folding is ok in the face of potential overflows. Still, I do not see why the use in this PR for IV widening should not be correct, so I want to see what CI finds. The implementation makes use of the fact that

Almost all our extensions are zero extensions, and it is free for our targeted platforms to do a zero extension when storing from a 32-bit value; thus, the actual operations themselves are not widened and there is no risk of semantics changing overflow
The 64-bit backends already supportTYP_INT access ofTYP_LONG locals, so no truncations need to be inserted.

Author:	jakobbotsch
Assignees:	jakobbotsch
Labels:	`area-CodeGen-coreclr`
Milestone:	-

jakobbotsch added2 commits

February 2, 2024 13:35

Fix linux build

2a02a48

Clean up

99cca89

Copy link

ryujit-bot commentedFeb 2, 2024

Diff results for#97865

Assembly diffs

Assembly diffs for linux/arm64 ran on windows/x64

Diffs are based on2,507,286 contexts (1,007,092 MinOpts,1,500,194 FullOpts).

MISSED contexts: base:27 (0.00%), diff:32 (0.00%)

Overall (+1,719,060 bytes)

Collection	Base size (bytes)	Diff size (bytes)
benchmarks.run.linux.arm64.checked.mch	15,557,856	+27,216
benchmarks.run_pgo.linux.arm64.checked.mch	80,097,444	+360,784
benchmarks.run_tiered.linux.arm64.checked.mch	24,601,848	+14,124
coreclr_tests.run.linux.arm64.checked.mch	508,729,840	+334,212
libraries.crossgen2.linux.arm64.checked.mch	55,844,176	+78,032
libraries.pmi.linux.arm64.checked.mch	76,294,380	+89,360
libraries_tests.run.linux.arm64.Release.mch	395,716,116	+669,880
libraries_tests_no_tiered_compilation.run.linux.arm64.Release.mch	165,002,020	+115,260
realworld.run.linux.arm64.checked.mch	15,904,164	+20,608
smoke_tests.nativeaot.linux.arm64.checked.mch	2,946,828	+9,584

FullOpts (+1,719,060 bytes)

Collection	Base size (bytes)	Diff size (bytes)
benchmarks.run.linux.arm64.checked.mch	15,252,928	+27,216
benchmarks.run_pgo.linux.arm64.checked.mch	54,162,660	+360,784
benchmarks.run_tiered.linux.arm64.checked.mch	4,863,412	+14,124
coreclr_tests.run.linux.arm64.checked.mch	160,581,688	+334,212
libraries.crossgen2.linux.arm64.checked.mch	55,842,540	+78,032
libraries.pmi.linux.arm64.checked.mch	76,174,396	+89,360
libraries_tests.run.linux.arm64.Release.mch	180,584,164	+669,880
libraries_tests_no_tiered_compilation.run.linux.arm64.Release.mch	151,504,556	+115,260
realworld.run.linux.arm64.checked.mch	15,323,240	+20,608
smoke_tests.nativeaot.linux.arm64.checked.mch	2,945,840	+9,584

Assembly diffs for linux/x64 ran on windows/x64

Diffs are based on2,517,881 contexts (991,070 MinOpts,1,526,811 FullOpts).

MISSED contexts:28 (0.00%)

Overall (+572,703 bytes)

Collection	Base size (bytes)	Diff size (bytes)
benchmarks.run.linux.x64.checked.mch	14,336,747	+1,856
benchmarks.run_pgo.linux.x64.checked.mch	71,580,068	+153,953
benchmarks.run_tiered.linux.x64.checked.mch	21,436,595	-2,845
coreclr_tests.run.linux.x64.checked.mch	403,712,034	-51,428
libraries.crossgen2.linux.x64.checked.mch	38,727,480	+18,381
libraries.pmi.linux.x64.checked.mch	60,420,078	+23,143
libraries_tests.run.linux.x64.Release.mch	337,125,301	+391,429
libraries_tests_no_tiered_compilation.run.linux.x64.Release.mch	132,553,319	+29,341
realworld.run.linux.x64.checked.mch	13,174,530	+8,436
smoke_tests.nativeaot.linux.x64.checked.mch	4,234,684	+437

FullOpts (+572,703 bytes)

Collection	Base size (bytes)	Diff size (bytes)
benchmarks.run.linux.x64.checked.mch	14,037,444	+1,856
benchmarks.run_pgo.linux.x64.checked.mch	47,780,079	+153,953
benchmarks.run_tiered.linux.x64.checked.mch	3,695,815	-2,845
coreclr_tests.run.linux.x64.checked.mch	123,957,574	-51,428
libraries.crossgen2.linux.x64.checked.mch	38,726,282	+18,381
libraries.pmi.linux.x64.checked.mch	60,307,221	+23,143
libraries_tests.run.linux.x64.Release.mch	153,365,692	+391,429
libraries_tests_no_tiered_compilation.run.linux.x64.Release.mch	121,935,551	+29,341
realworld.run.linux.x64.checked.mch	12,788,646	+8,436
smoke_tests.nativeaot.linux.x64.checked.mch	4,233,735	+437

Assembly diffs for osx/arm64 ran on windows/x64

Diffs are based on2,270,841 contexts (932,669 MinOpts,1,338,172 FullOpts).

MISSED contexts: base:26 (0.00%), diff:29 (0.00%)

Overall (+1,124,576 bytes)

Collection	Base size (bytes)	Diff size (bytes)
benchmarks.run.osx.arm64.checked.mch	11,186,636	+27,356
benchmarks.run_pgo.osx.arm64.checked.mch	34,442,660	+59,268
benchmarks.run_tiered.osx.arm64.checked.mch	15,517,000	+13,372
coreclr_tests.run.osx.arm64.checked.mch	486,415,276	+312,820
libraries.crossgen2.osx.arm64.checked.mch	55,725,640	+77,884
libraries.pmi.osx.arm64.checked.mch	80,218,788	+99,136
libraries_tests.run.osx.arm64.Release.mch	324,601,508	+399,340
libraries_tests_no_tiered_compilation.run.osx.arm64.Release.mch	162,572,640	+114,848
realworld.run.osx.arm64.checked.mch	15,061,556	+20,552

FullOpts (+1,124,576 bytes)

Collection	Base size (bytes)	Diff size (bytes)
benchmarks.run.osx.arm64.checked.mch	11,186,100	+27,356
benchmarks.run_pgo.osx.arm64.checked.mch	18,141,360	+59,268
benchmarks.run_tiered.osx.arm64.checked.mch	4,012,516	+13,372
coreclr_tests.run.osx.arm64.checked.mch	153,802,352	+312,820
libraries.crossgen2.osx.arm64.checked.mch	55,724,012	+77,884
libraries.pmi.osx.arm64.checked.mch	80,097,660	+99,136
libraries_tests.run.osx.arm64.Release.mch	120,885,660	+399,340
libraries_tests_no_tiered_compilation.run.osx.arm64.Release.mch	149,418,912	+114,848
realworld.run.osx.arm64.checked.mch	14,497,600	+20,552

Assembly diffs for windows/arm64 ran on windows/x64

Diffs are based on2,341,080 contexts (938,449 MinOpts,1,402,631 FullOpts).

MISSED contexts: base:34 (0.00%), diff:37 (0.00%)

Overall (+1,320,572 bytes)

Collection	Base size (bytes)	Diff size (bytes)
benchmarks.run.windows.arm64.checked.mch	10,966,144	+26,900
benchmarks.run_pgo.windows.arm64.checked.mch	45,576,512	+150,544
benchmarks.run_tiered.windows.arm64.checked.mch	15,587,900	+14,344
coreclr_tests.run.windows.arm64.checked.mch	495,268,304	+323,640
libraries.crossgen2.windows.arm64.checked.mch	59,069,380	+83,104
libraries.pmi.windows.arm64.checked.mch	79,846,048	+94,408
libraries_tests.run.windows.arm64.Release.mch	330,812,748	+472,236
libraries_tests_no_tiered_compilation.run.windows.arm64.Release.mch	171,579,600	+120,020
realworld.run.windows.arm64.checked.mch	15,905,252	+21,532
smoke_tests.nativeaot.windows.arm64.checked.mch	3,970,272	+13,844

FullOpts (+1,320,572 bytes)

Collection	Base size (bytes)	Diff size (bytes)
benchmarks.run.windows.arm64.checked.mch	10,965,608	+26,900
benchmarks.run_pgo.windows.arm64.checked.mch	29,565,116	+150,544
benchmarks.run_tiered.windows.arm64.checked.mch	4,410,712	+14,344
coreclr_tests.run.windows.arm64.checked.mch	156,579,200	+323,640
libraries.crossgen2.windows.arm64.checked.mch	59,067,744	+83,104
libraries.pmi.windows.arm64.checked.mch	79,726,064	+94,408
libraries_tests.run.windows.arm64.Release.mch	127,379,152	+472,236
libraries_tests_no_tiered_compilation.run.windows.arm64.Release.mch	158,425,852	+120,020
realworld.run.windows.arm64.checked.mch	15,341,272	+21,532
smoke_tests.nativeaot.windows.arm64.checked.mch	3,969,260	+13,844

Assembly diffs for windows/x64 ran on windows/x64

Diffs are based on2,512,182 contexts (997,391 MinOpts,1,514,791 FullOpts).

MISSED contexts: base:29 (0.00%), diff:30 (0.00%)

Overall (-18,030 bytes)

Collection	Base size (bytes)	Diff size (bytes)
aspnet.run.windows.x64.checked.mch	47,040,393	+5,552
benchmarks.run.windows.x64.checked.mch	8,742,476	-2,838
benchmarks.run_pgo.windows.x64.checked.mch	36,233,314	-2,927
benchmarks.run_tiered.windows.x64.checked.mch	12,416,932	-4,780
coreclr_tests.run.windows.x64.checked.mch	393,194,802	-71,902
libraries.crossgen2.windows.x64.checked.mch	39,486,541	-1,798
libraries.pmi.windows.x64.checked.mch	61,663,361	-2,020
libraries_tests.run.windows.x64.Release.mch	282,135,537	+68,806
libraries_tests_no_tiered_compilation.run.windows.x64.Release.mch	137,061,874	-5,098
realworld.run.windows.x64.checked.mch	14,130,826	+1,095
smoke_tests.nativeaot.windows.x64.checked.mch	5,083,179	-2,120

FullOpts (-18,030 bytes)

Collection	Base size (bytes)	Diff size (bytes)
aspnet.run.windows.x64.checked.mch	28,549,344	+5,552
benchmarks.run.windows.x64.checked.mch	8,742,113	-2,838
benchmarks.run_pgo.windows.x64.checked.mch	22,062,658	-2,927
benchmarks.run_tiered.windows.x64.checked.mch	3,317,693	-4,780
coreclr_tests.run.windows.x64.checked.mch	120,406,222	-71,902
libraries.crossgen2.windows.x64.checked.mch	39,485,354	-1,798
libraries.pmi.windows.x64.checked.mch	61,549,840	-2,020
libraries_tests.run.windows.x64.Release.mch	106,277,303	+68,806
libraries_tests_no_tiered_compilation.run.windows.x64.Release.mch	126,442,768	-5,098
realworld.run.windows.x64.checked.mch	13,744,217	+1,095
smoke_tests.nativeaot.windows.x64.checked.mch	5,082,232	-2,120

Detailshere

Throughput diffs

Throughput diffs for linux/arm64 ran on windows/x64

Overall (+0.78% to+2.55%)

Collection	PDIFF
benchmarks.run.linux.arm64.checked.mch	+1.23%
benchmarks.run_pgo.linux.arm64.checked.mch	+2.23%
benchmarks.run_tiered.linux.arm64.checked.mch	+0.90%
coreclr_tests.run.linux.arm64.checked.mch	+1.15%
libraries.crossgen2.linux.arm64.checked.mch	+0.99%
libraries.pmi.linux.arm64.checked.mch	+1.04%
libraries_tests.run.linux.arm64.Release.mch	+2.55%
libraries_tests_no_tiered_compilation.run.linux.arm64.Release.mch	+0.78%
realworld.run.linux.arm64.checked.mch	+1.32%
smoke_tests.nativeaot.linux.arm64.checked.mch	+1.36%

MinOpts (-0.00% to+0.01%)

Collection	PDIFF
libraries.pmi.linux.arm64.checked.mch	+0.01%

FullOpts (+0.80% to+3.29%)

Collection	PDIFF
benchmarks.run.linux.arm64.checked.mch	+1.23%
benchmarks.run_pgo.linux.arm64.checked.mch	+2.52%
benchmarks.run_tiered.linux.arm64.checked.mch	+1.86%
coreclr_tests.run.linux.arm64.checked.mch	+1.93%
libraries.crossgen2.linux.arm64.checked.mch	+0.99%
libraries.pmi.linux.arm64.checked.mch	+1.04%
libraries_tests.run.linux.arm64.Release.mch	+3.29%
libraries_tests_no_tiered_compilation.run.linux.arm64.Release.mch	+0.80%
realworld.run.linux.arm64.checked.mch	+1.33%
smoke_tests.nativeaot.linux.arm64.checked.mch	+1.36%

Throughput diffs for linux/x64 ran on windows/x64

Overall (+0.82% to+2.31%)

Collection	PDIFF
benchmarks.run.linux.x64.checked.mch	+1.19%
benchmarks.run_pgo.linux.x64.checked.mch	+2.29%
benchmarks.run_tiered.linux.x64.checked.mch	+0.90%
coreclr_tests.run.linux.x64.checked.mch	+1.23%
libraries.crossgen2.linux.x64.checked.mch	+1.08%
libraries.pmi.linux.x64.checked.mch	+1.10%
libraries_tests.run.linux.x64.Release.mch	+2.31%
libraries_tests_no_tiered_compilation.run.linux.x64.Release.mch	+0.82%
realworld.run.linux.x64.checked.mch	+1.39%
smoke_tests.nativeaot.linux.x64.checked.mch	+1.50%

FullOpts (+0.84% to+2.94%)

Collection	PDIFF
benchmarks.run.linux.x64.checked.mch	+1.20%
benchmarks.run_pgo.linux.x64.checked.mch	+2.57%
benchmarks.run_tiered.linux.x64.checked.mch	+1.86%
coreclr_tests.run.linux.x64.checked.mch	+2.06%
libraries.crossgen2.linux.x64.checked.mch	+1.08%
libraries.pmi.linux.x64.checked.mch	+1.10%
libraries_tests.run.linux.x64.Release.mch	+2.94%
libraries_tests_no_tiered_compilation.run.linux.x64.Release.mch	+0.84%
realworld.run.linux.x64.checked.mch	+1.39%
smoke_tests.nativeaot.linux.x64.checked.mch	+1.50%

Throughput diffs for osx/arm64 ran on windows/x64

Overall (+0.78% to+2.27%)

Collection	PDIFF
benchmarks.run.osx.arm64.checked.mch	+1.58%
benchmarks.run_pgo.osx.arm64.checked.mch	+2.05%
benchmarks.run_tiered.osx.arm64.checked.mch	+1.15%
coreclr_tests.run.osx.arm64.checked.mch	+1.14%
libraries.crossgen2.osx.arm64.checked.mch	+0.99%
libraries.pmi.osx.arm64.checked.mch	+1.06%
libraries_tests.run.osx.arm64.Release.mch	+2.27%
libraries_tests_no_tiered_compilation.run.osx.arm64.Release.mch	+0.78%
realworld.run.osx.arm64.checked.mch	+1.37%

FullOpts (+0.80% to+3.18%)

Collection	PDIFF
benchmarks.run.osx.arm64.checked.mch	+1.58%
benchmarks.run_pgo.osx.arm64.checked.mch	+2.52%
benchmarks.run_tiered.osx.arm64.checked.mch	+1.99%
coreclr_tests.run.osx.arm64.checked.mch	+1.92%
libraries.crossgen2.osx.arm64.checked.mch	+0.99%
libraries.pmi.osx.arm64.checked.mch	+1.07%
libraries_tests.run.osx.arm64.Release.mch	+3.18%
libraries_tests_no_tiered_compilation.run.osx.arm64.Release.mch	+0.80%
realworld.run.osx.arm64.checked.mch	+1.38%

Throughput diffs for windows/arm64 ran on windows/x64

Overall (+0.77% to+2.53%)

Collection	PDIFF
benchmarks.run.windows.arm64.checked.mch	+1.58%
benchmarks.run_pgo.windows.arm64.checked.mch	+2.53%
benchmarks.run_tiered.windows.arm64.checked.mch	+1.16%
coreclr_tests.run.windows.arm64.checked.mch	+1.14%
libraries.crossgen2.windows.arm64.checked.mch	+1.00%
libraries.pmi.windows.arm64.checked.mch	+1.05%
libraries_tests.run.windows.arm64.Release.mch	+2.21%
libraries_tests_no_tiered_compilation.run.windows.arm64.Release.mch	+0.77%
realworld.run.windows.arm64.checked.mch	+1.36%
smoke_tests.nativeaot.windows.arm64.checked.mch	+1.48%

FullOpts (+0.79% to+3.08%)

Collection	PDIFF
benchmarks.run.windows.arm64.checked.mch	+1.58%
benchmarks.run_pgo.windows.arm64.checked.mch	+2.88%
benchmarks.run_tiered.windows.arm64.checked.mch	+1.92%
coreclr_tests.run.windows.arm64.checked.mch	+1.93%
libraries.crossgen2.windows.arm64.checked.mch	+1.00%
libraries.pmi.windows.arm64.checked.mch	+1.06%
libraries_tests.run.windows.arm64.Release.mch	+3.08%
libraries_tests_no_tiered_compilation.run.windows.arm64.Release.mch	+0.79%
realworld.run.windows.arm64.checked.mch	+1.37%
smoke_tests.nativeaot.windows.arm64.checked.mch	+1.48%

Throughput diffs for windows/x64 ran on windows/x64

Overall (+0.83% to+2.44%)

Collection	PDIFF
aspnet.run.windows.x64.checked.mch	+1.80%
benchmarks.run.windows.x64.checked.mch	+1.59%
benchmarks.run_pgo.windows.x64.checked.mch	+2.44%
benchmarks.run_tiered.windows.x64.checked.mch	+1.14%
coreclr_tests.run.windows.x64.checked.mch	+1.23%
libraries.crossgen2.windows.x64.checked.mch	+1.09%
libraries.pmi.windows.x64.checked.mch	+1.11%
libraries_tests.run.windows.x64.Release.mch	+2.05%
libraries_tests_no_tiered_compilation.run.windows.x64.Release.mch	+0.83%
realworld.run.windows.x64.checked.mch	+1.35%
smoke_tests.nativeaot.windows.x64.checked.mch	+1.54%

FullOpts (+0.85% to+2.80%)

Collection	PDIFF
aspnet.run.windows.x64.checked.mch	+2.06%
benchmarks.run.windows.x64.checked.mch	+1.59%
benchmarks.run_pgo.windows.x64.checked.mch	+2.80%
benchmarks.run_tiered.windows.x64.checked.mch	+1.84%
coreclr_tests.run.windows.x64.checked.mch	+2.04%
libraries.crossgen2.windows.x64.checked.mch	+1.09%
libraries.pmi.windows.x64.checked.mch	+1.11%
libraries_tests.run.windows.x64.Release.mch	+2.80%
libraries_tests_no_tiered_compilation.run.windows.x64.Release.mch	+0.85%
realworld.run.windows.x64.checked.mch	+1.35%
smoke_tests.nativeaot.windows.x64.checked.mch	+1.54%

Detailshere

Throughput diffs for linux/arm64 ran on linux/x64

Overall (+0.68% to+2.26%)

Collection	PDIFF
coreclr_tests.run.linux.arm64.checked.mch	+0.92%
benchmarks.run_pgo.linux.arm64.checked.mch	+2.03%
benchmarks.run_tiered.linux.arm64.checked.mch	+0.78%
libraries.pmi.linux.arm64.checked.mch	+0.92%
libraries_tests.run.linux.arm64.Release.mch	+2.26%
libraries_tests_no_tiered_compilation.run.linux.arm64.Release.mch	+0.68%
smoke_tests.nativeaot.linux.arm64.checked.mch	+1.22%
libraries.crossgen2.linux.arm64.checked.mch	+0.88%
benchmarks.run.linux.arm64.checked.mch	+1.06%
realworld.run.linux.arm64.checked.mch	+1.15%

FullOpts (+0.70% to+2.91%)

Collection	PDIFF
coreclr_tests.run.linux.arm64.checked.mch	+1.59%
benchmarks.run_pgo.linux.arm64.checked.mch	+2.29%
benchmarks.run_tiered.linux.arm64.checked.mch	+1.58%
libraries.pmi.linux.arm64.checked.mch	+0.92%
libraries_tests.run.linux.arm64.Release.mch	+2.91%
libraries_tests_no_tiered_compilation.run.linux.arm64.Release.mch	+0.70%
smoke_tests.nativeaot.linux.arm64.checked.mch	+1.22%
libraries.crossgen2.linux.arm64.checked.mch	+0.88%
benchmarks.run.linux.arm64.checked.mch	+1.07%
realworld.run.linux.arm64.checked.mch	+1.16%

Throughput diffs for linux/x64 ran on linux/x64

Overall (+0.72% to+2.06%)

Collection	PDIFF
libraries.pmi.linux.x64.checked.mch	+0.96%
libraries.crossgen2.linux.x64.checked.mch	+0.95%
benchmarks.run_pgo.linux.x64.checked.mch	+2.06%
smoke_tests.nativeaot.linux.x64.checked.mch	+1.31%
libraries_tests.run.linux.x64.Release.mch	+2.02%
benchmarks.run_tiered.linux.x64.checked.mch	+0.76%
realworld.run.linux.x64.checked.mch	+1.20%
libraries_tests_no_tiered_compilation.run.linux.x64.Release.mch	+0.72%
coreclr_tests.run.linux.x64.checked.mch	+0.96%
benchmarks.run.linux.x64.checked.mch	+1.02%

FullOpts (+0.74% to+2.58%)

Collection	PDIFF
libraries.pmi.linux.x64.checked.mch	+0.96%
libraries.crossgen2.linux.x64.checked.mch	+0.95%
benchmarks.run_pgo.linux.x64.checked.mch	+2.31%
smoke_tests.nativeaot.linux.x64.checked.mch	+1.31%
libraries_tests.run.linux.x64.Release.mch	+2.58%
benchmarks.run_tiered.linux.x64.checked.mch	+1.56%
realworld.run.linux.x64.checked.mch	+1.21%
libraries_tests_no_tiered_compilation.run.linux.x64.Release.mch	+0.74%
coreclr_tests.run.linux.x64.checked.mch	+1.67%
benchmarks.run.linux.x64.checked.mch	+1.03%

Detailshere

Throughput diffs for linux/arm ran on windows/x86

Overall (+0.31% to+1.70%)

Collection	PDIFF
benchmarks.run.linux.arm.checked.mch	+0.65%
benchmarks.run_pgo.linux.arm.checked.mch	+0.31%
benchmarks.run_tiered.linux.arm.checked.mch	+0.62%
coreclr_tests.run.linux.arm.checked.mch	+1.70%
libraries.crossgen2.linux.arm.checked.mch	+0.57%
libraries.pmi.linux.arm.checked.mch	+0.60%
libraries_tests.run.linux.arm.Release.mch	+0.77%
libraries_tests_no_tiered_compilation.run.linux.arm.Release.mch	+0.47%
realworld.run.linux.arm.checked.mch	+0.93%

FullOpts (+0.33% to+2.77%)

Collection	PDIFF
benchmarks.run.linux.arm.checked.mch	+0.65%
benchmarks.run_pgo.linux.arm.checked.mch	+0.33%
benchmarks.run_tiered.linux.arm.checked.mch	+0.75%
coreclr_tests.run.linux.arm.checked.mch	+2.77%
libraries.crossgen2.linux.arm.checked.mch	+0.57%
libraries.pmi.linux.arm.checked.mch	+0.60%
libraries_tests.run.linux.arm.Release.mch	+0.97%
libraries_tests_no_tiered_compilation.run.linux.arm.Release.mch	+0.49%
realworld.run.linux.arm.checked.mch	+0.94%

Throughput diffs for windows/x86 ran on windows/x86

Overall (+0.48% to+1.86%)

Collection	PDIFF
benchmarks.run.windows.x86.checked.mch	+1.03%
benchmarks.run_pgo.windows.x86.checked.mch	+0.48%
benchmarks.run_tiered.windows.x86.checked.mch	+1.12%
coreclr_tests.run.windows.x86.checked.mch	+1.86%
libraries.crossgen2.windows.x86.checked.mch	+0.68%
libraries.pmi.windows.x86.checked.mch	+0.66%
libraries_tests.run.windows.x86.Release.mch	+0.95%
libraries_tests_no_tiered_compilation.run.windows.x86.Release.mch	+0.50%
realworld.run.windows.x86.checked.mch	+0.81%

FullOpts (+0.50% to+2.77%)

Collection	PDIFF
benchmarks.run.windows.x86.checked.mch	+1.03%
benchmarks.run_pgo.windows.x86.checked.mch	+0.50%
benchmarks.run_tiered.windows.x86.checked.mch	+1.34%
coreclr_tests.run.windows.x86.checked.mch	+2.77%
libraries.crossgen2.windows.x86.checked.mch	+0.68%
libraries.pmi.windows.x86.checked.mch	+0.66%
libraries_tests.run.windows.x86.Release.mch	+1.19%
libraries_tests_no_tiered_compilation.run.windows.x86.Release.mch	+0.51%
realworld.run.windows.x86.checked.mch	+0.81%

Detailshere

Copy link

MemberAuthor

jakobbotsch commentedFeb 2, 2024

/azp run runtime-coreclr jitstress, runtime-coreclr libraries-jitstress, Fuzzlyn

Copy link

azure-pipelinesbot commentedFeb 2, 2024

Azure Pipelines successfully started running 3 pipeline(s).

Merge branch 'main' of github.com:dotnet/runtime into simplify-ivs

1dc2673

Copy link

MemberAuthor

jakobbotsch commentedFeb 2, 2024

Test failures look like#97892.
Regressions are much larger outside win-x64. Need to understand why, when I look at the examples almost all of them are range check related for the reason I mentioned above. I don't understand why that wouldn't affect win-x64 equally, though.
Some of the arm64 diffs look to be because we do not support the zero-costSTORE_LCL_VAR(CAST<long <- int>(data)) pattern there, but that should be implementable. Also, on arm64 the IV widening is proportionally less useful since address modes can zero extend as part of the instruction, so fewer direct benefits from IV widening are to be expected.

This was referencedFeb 2, 2024

[jitstress] Assertion failed '!"Inconsistent profile data"'#97892

Closed

Test failure _ParallelCrashTester::ParallelCrashTester.ParallelCrashMainThreadAndWorkerThreads()#94393

Closed

jakobbotsch mentioned this pull request

Feb 3, 2024

JIT: Support more scaled addressing modes on arm64#97921

Merged

jakobbotsch added3 commits

February 3, 2024 17:39

Do not swap addrecs unnecessarily when folding

0df9a03

Add some todos

cf31ff6

JIT: Support more scaled addressing modes on arm64

c498884

We currently support scaled addressing modes when the index also needsan extension through contained `BFIZ` nodes. However, we did not supportscaled addressing modes if the index was 64 bits. This adds thatsupport as a natural extension to the `GT_LEA`.

Copy link

ryujit-bot commentedFeb 3, 2024

Diff results for#97865

Throughput diffs

Throughput diffs for linux/arm64 ran on windows/x64

Overall (+0.30% to+1.63%)

Collection	PDIFF
benchmarks.run.linux.arm64.checked.mch	+0.56%
benchmarks.run_pgo.linux.arm64.checked.mch	+1.40%
benchmarks.run_tiered.linux.arm64.checked.mch	+0.30%
coreclr_tests.run.linux.arm64.checked.mch	+0.39%
libraries.crossgen2.linux.arm64.checked.mch	+0.47%
libraries.pmi.linux.arm64.checked.mch	+0.54%
libraries_tests.run.linux.arm64.Release.mch	+1.63%
libraries_tests_no_tiered_compilation.run.linux.arm64.Release.mch	+0.41%
realworld.run.linux.arm64.checked.mch	+0.67%
smoke_tests.nativeaot.linux.arm64.checked.mch	+0.61%

MinOpts (-0.03% to+0.00%)

Collection	PDIFF
benchmarks.run.linux.arm64.checked.mch	-0.02%
benchmarks.run_pgo.linux.arm64.checked.mch	-0.02%
benchmarks.run_tiered.linux.arm64.checked.mch	-0.02%
libraries.pmi.linux.arm64.checked.mch	-0.03%
libraries_tests.run.linux.arm64.Release.mch	-0.01%
libraries_tests_no_tiered_compilation.run.linux.arm64.Release.mch	-0.01%
smoke_tests.nativeaot.linux.arm64.checked.mch	-0.01%

FullOpts (+0.42% to+2.10%)

Collection	PDIFF
benchmarks.run.linux.arm64.checked.mch	+0.56%
benchmarks.run_pgo.linux.arm64.checked.mch	+1.59%
benchmarks.run_tiered.linux.arm64.checked.mch	+0.65%
coreclr_tests.run.linux.arm64.checked.mch	+0.66%
libraries.crossgen2.linux.arm64.checked.mch	+0.47%
libraries.pmi.linux.arm64.checked.mch	+0.54%
libraries_tests.run.linux.arm64.Release.mch	+2.10%
libraries_tests_no_tiered_compilation.run.linux.arm64.Release.mch	+0.42%
realworld.run.linux.arm64.checked.mch	+0.67%
smoke_tests.nativeaot.linux.arm64.checked.mch	+0.61%

Throughput diffs for linux/x64 ran on windows/x64

Overall (+0.33% to+1.50%)

Collection	PDIFF
benchmarks.run.linux.x64.checked.mch	+0.57%
benchmarks.run_pgo.linux.x64.checked.mch	+1.49%
benchmarks.run_tiered.linux.x64.checked.mch	+0.33%
coreclr_tests.run.linux.x64.checked.mch	+0.44%
libraries.crossgen2.linux.x64.checked.mch	+0.56%
libraries.pmi.linux.x64.checked.mch	+0.60%
libraries_tests.run.linux.x64.Release.mch	+1.50%
libraries_tests_no_tiered_compilation.run.linux.x64.Release.mch	+0.46%
realworld.run.linux.x64.checked.mch	+0.74%
smoke_tests.nativeaot.linux.x64.checked.mch	+0.74%

MinOpts (+0.00% to+0.01%)

Collection	PDIFF
libraries.pmi.linux.x64.checked.mch	+0.01%

FullOpts (+0.47% to+1.91%)

Collection	PDIFF
benchmarks.run.linux.x64.checked.mch	+0.58%
benchmarks.run_pgo.linux.x64.checked.mch	+1.67%
benchmarks.run_tiered.linux.x64.checked.mch	+0.68%
coreclr_tests.run.linux.x64.checked.mch	+0.74%
libraries.crossgen2.linux.x64.checked.mch	+0.56%
libraries.pmi.linux.x64.checked.mch	+0.60%
libraries_tests.run.linux.x64.Release.mch	+1.91%
libraries_tests_no_tiered_compilation.run.linux.x64.Release.mch	+0.47%
realworld.run.linux.x64.checked.mch	+0.74%
smoke_tests.nativeaot.linux.x64.checked.mch	+0.74%

Throughput diffs for osx/arm64 ran on windows/x64

Overall (+0.37% to+1.43%)

Collection	PDIFF
benchmarks.run.osx.arm64.checked.mch	+0.73%
benchmarks.run_pgo.osx.arm64.checked.mch	+0.95%
benchmarks.run_tiered.osx.arm64.checked.mch	+0.39%
coreclr_tests.run.osx.arm64.checked.mch	+0.37%
libraries.crossgen2.osx.arm64.checked.mch	+0.47%
libraries.pmi.osx.arm64.checked.mch	+0.56%
libraries_tests.run.osx.arm64.Release.mch	+1.43%
libraries_tests_no_tiered_compilation.run.osx.arm64.Release.mch	+0.41%
realworld.run.osx.arm64.checked.mch	+0.69%

MinOpts (-0.04% to+0.01%)

Collection	PDIFF
benchmarks.run_pgo.osx.arm64.checked.mch	-0.03%
benchmarks.run_tiered.osx.arm64.checked.mch	-0.03%
libraries.crossgen2.osx.arm64.checked.mch	-0.01%
libraries.pmi.osx.arm64.checked.mch	-0.04%
libraries_tests.run.osx.arm64.Release.mch	-0.01%
libraries_tests_no_tiered_compilation.run.osx.arm64.Release.mch	-0.01%
realworld.run.osx.arm64.checked.mch	+0.01%

FullOpts (+0.42% to+2.02%)

Collection	PDIFF
benchmarks.run.osx.arm64.checked.mch	+0.73%
benchmarks.run_pgo.osx.arm64.checked.mch	+1.18%
benchmarks.run_tiered.osx.arm64.checked.mch	+0.70%
coreclr_tests.run.osx.arm64.checked.mch	+0.63%
libraries.crossgen2.osx.arm64.checked.mch	+0.47%
libraries.pmi.osx.arm64.checked.mch	+0.56%
libraries_tests.run.osx.arm64.Release.mch	+2.02%
libraries_tests_no_tiered_compilation.run.osx.arm64.Release.mch	+0.42%
realworld.run.osx.arm64.checked.mch	+0.70%

Throughput diffs for windows/arm64 ran on windows/x64

Overall (+0.39% to+1.50%)

Collection	PDIFF
benchmarks.run.windows.arm64.checked.mch	+0.73%
benchmarks.run_pgo.windows.arm64.checked.mch	+1.50%
benchmarks.run_tiered.windows.arm64.checked.mch	+0.40%
coreclr_tests.run.windows.arm64.checked.mch	+0.39%
libraries.crossgen2.windows.arm64.checked.mch	+0.47%
libraries.pmi.windows.arm64.checked.mch	+0.55%
libraries_tests.run.windows.arm64.Release.mch	+1.40%
libraries_tests_no_tiered_compilation.run.windows.arm64.Release.mch	+0.41%
realworld.run.windows.arm64.checked.mch	+0.69%
smoke_tests.nativeaot.windows.arm64.checked.mch	+0.65%

MinOpts (-0.03% to+0.00%)

Collection	PDIFF
benchmarks.run_pgo.windows.arm64.checked.mch	-0.03%
benchmarks.run_tiered.windows.arm64.checked.mch	-0.03%
libraries.pmi.windows.arm64.checked.mch	-0.03%
libraries_tests.run.windows.arm64.Release.mch	-0.01%
libraries_tests_no_tiered_compilation.run.windows.arm64.Release.mch	-0.01%
smoke_tests.nativeaot.windows.arm64.checked.mch	-0.01%

FullOpts (+0.42% to+1.96%)

Collection	PDIFF
benchmarks.run.windows.arm64.checked.mch	+0.73%
benchmarks.run_pgo.windows.arm64.checked.mch	+1.72%
benchmarks.run_tiered.windows.arm64.checked.mch	+0.69%
coreclr_tests.run.windows.arm64.checked.mch	+0.66%
libraries.crossgen2.windows.arm64.checked.mch	+0.47%
libraries.pmi.windows.arm64.checked.mch	+0.55%
libraries_tests.run.windows.arm64.Release.mch	+1.96%
libraries_tests_no_tiered_compilation.run.windows.arm64.Release.mch	+0.42%
realworld.run.windows.arm64.checked.mch	+0.70%
smoke_tests.nativeaot.windows.arm64.checked.mch	+0.65%

Throughput diffs for windows/x64 ran on windows/x64

Overall (+0.43% to+1.31%)

Collection	PDIFF
aspnet.run.windows.x64.checked.mch	+1.19%
benchmarks.run.windows.x64.checked.mch	+0.78%
benchmarks.run_pgo.windows.x64.checked.mch	+1.28%
benchmarks.run_tiered.windows.x64.checked.mch	+0.43%
coreclr_tests.run.windows.x64.checked.mch	+0.44%
libraries.crossgen2.windows.x64.checked.mch	+0.56%
libraries.pmi.windows.x64.checked.mch	+0.60%
libraries_tests.run.windows.x64.Release.mch	+1.31%
libraries_tests_no_tiered_compilation.run.windows.x64.Release.mch	+0.46%
realworld.run.windows.x64.checked.mch	+0.72%
smoke_tests.nativeaot.windows.x64.checked.mch	+0.76%

MinOpts (+0.00% to+0.01%)

Collection	PDIFF
libraries.pmi.windows.x64.checked.mch	+0.01%

FullOpts (+0.47% to+1.79%)

Collection	PDIFF
aspnet.run.windows.x64.checked.mch	+1.36%
benchmarks.run.windows.x64.checked.mch	+0.78%
benchmarks.run_pgo.windows.x64.checked.mch	+1.47%
benchmarks.run_tiered.windows.x64.checked.mch	+0.70%
coreclr_tests.run.windows.x64.checked.mch	+0.74%
libraries.crossgen2.windows.x64.checked.mch	+0.56%
libraries.pmi.windows.x64.checked.mch	+0.60%
libraries_tests.run.windows.x64.Release.mch	+1.79%
libraries_tests_no_tiered_compilation.run.windows.x64.Release.mch	+0.47%
realworld.run.windows.x64.checked.mch	+0.72%
smoke_tests.nativeaot.windows.x64.checked.mch	+0.76%

Detailshere

Throughput diffs for linux/arm ran on windows/x86

MinOpts (-0.02% to-0.00%)

Collection	PDIFF
benchmarks.run.linux.arm.checked.mch	-0.01%
benchmarks.run_pgo.linux.arm.checked.mch	-0.01%
benchmarks.run_tiered.linux.arm.checked.mch	-0.01%
libraries.pmi.linux.arm.checked.mch	-0.02%
libraries_tests.run.linux.arm.Release.mch	-0.01%

Throughput diffs for windows/x86 ran on windows/x86

Overall (+0.01%)

Collection	PDIFF
benchmarks.run.windows.x86.checked.mch	+0.01%
benchmarks.run_pgo.windows.x86.checked.mch	+0.01%
benchmarks.run_tiered.windows.x86.checked.mch	+0.01%
coreclr_tests.run.windows.x86.checked.mch	+0.01%
libraries.crossgen2.windows.x86.checked.mch	+0.01%
libraries.pmi.windows.x86.checked.mch	+0.01%
libraries_tests.run.windows.x86.Release.mch	+0.01%
libraries_tests_no_tiered_compilation.run.windows.x86.Release.mch	+0.01%
realworld.run.windows.x86.checked.mch	+0.01%

MinOpts (+0.00% to+0.02%)

Collection	PDIFF
benchmarks.run_pgo.windows.x86.checked.mch	+0.01%
benchmarks.run_tiered.windows.x86.checked.mch	+0.01%
libraries.pmi.windows.x86.checked.mch	+0.02%
libraries_tests.run.windows.x86.Release.mch	+0.01%
libraries_tests_no_tiered_compilation.run.windows.x86.Release.mch	+0.01%

FullOpts (+0.01%)

Collection	PDIFF
benchmarks.run.windows.x86.checked.mch	+0.01%
benchmarks.run_pgo.windows.x86.checked.mch	+0.01%
benchmarks.run_tiered.windows.x86.checked.mch	+0.01%
coreclr_tests.run.windows.x86.checked.mch	+0.01%
libraries.crossgen2.windows.x86.checked.mch	+0.01%
libraries.pmi.windows.x86.checked.mch	+0.01%
libraries_tests.run.windows.x86.Release.mch	+0.01%
libraries_tests_no_tiered_compilation.run.windows.x86.Release.mch	+0.01%
realworld.run.windows.x86.checked.mch	+0.01%

Detailshere

Copy link

ryujit-bot commentedFeb 3, 2024

Diff results for#97865

Assembly diffs

Assembly diffs for linux/arm64 ran on windows/x64

Diffs are based on2,505,877 contexts (1,007,092 MinOpts,1,498,785 FullOpts).

MISSED contexts: base:1,433 (0.06%), diff:1,441 (0.06%)

Overall (+904,416 bytes)

Collection	Base size (bytes)	Diff size (bytes)
benchmarks.run.linux.arm64.checked.mch	15,545,232	+13,448
benchmarks.run_pgo.linux.arm64.checked.mch	79,735,896	+232,112
benchmarks.run_tiered.linux.arm64.checked.mch	24,597,876	+1,620
coreclr_tests.run.linux.arm64.checked.mch	508,521,576	+166,136
libraries.crossgen2.linux.arm64.checked.mch	55,834,212	+40,568
libraries.pmi.linux.arm64.checked.mch	76,274,408	+50,684
libraries_tests.run.linux.arm64.Release.mch	394,024,440	+327,664
libraries_tests_no_tiered_compilation.run.linux.arm64.Release.mch	164,978,292	+57,440
realworld.run.linux.arm64.checked.mch	15,900,008	+10,512
smoke_tests.nativeaot.linux.arm64.checked.mch	2,829,664	+4,232

MinOpts (-22,888 bytes)

Collection	Base size (bytes)	Diff size (bytes)
benchmarks.run_pgo.linux.arm64.checked.mch	25,934,784	-5,292
benchmarks.run_tiered.linux.arm64.checked.mch	19,738,436	-3,656
coreclr_tests.run.linux.arm64.checked.mch	348,148,152	-2,712
libraries_tests.run.linux.arm64.Release.mch	215,131,952	-10,636
libraries_tests_no_tiered_compilation.run.linux.arm64.Release.mch	13,497,464	-592

FullOpts (+927,304 bytes)

Collection	Base size (bytes)	Diff size (bytes)
benchmarks.run.linux.arm64.checked.mch	15,240,304	+13,448
benchmarks.run_pgo.linux.arm64.checked.mch	53,801,112	+237,404
benchmarks.run_tiered.linux.arm64.checked.mch	4,859,440	+5,276
coreclr_tests.run.linux.arm64.checked.mch	160,373,424	+168,848
libraries.crossgen2.linux.arm64.checked.mch	55,832,576	+40,568
libraries.pmi.linux.arm64.checked.mch	76,154,424	+50,684
libraries_tests.run.linux.arm64.Release.mch	178,892,488	+338,300
libraries_tests_no_tiered_compilation.run.linux.arm64.Release.mch	151,480,828	+58,032
realworld.run.linux.arm64.checked.mch	15,319,084	+10,512
smoke_tests.nativeaot.linux.arm64.checked.mch	2,828,676	+4,232

Assembly diffs for linux/x64 ran on windows/x64

Diffs are based on2,516,325 contexts (991,070 MinOpts,1,525,255 FullOpts).

MISSED contexts:1,584 (0.06%)

Overall (+565,740 bytes)

Collection	Base size (bytes)	Diff size (bytes)
benchmarks.run.linux.x64.checked.mch	14,336,053	+1,904
benchmarks.run_pgo.linux.x64.checked.mch	71,262,198	+157,649
benchmarks.run_tiered.linux.x64.checked.mch	21,436,449	-2,851
coreclr_tests.run.linux.x64.checked.mch	403,514,426	-53,512
libraries.crossgen2.linux.x64.checked.mch	38,726,572	+18,689
libraries.pmi.linux.x64.checked.mch	60,417,780	+23,461
libraries_tests.run.linux.x64.Release.mch	335,675,120	+382,346
libraries_tests_no_tiered_compilation.run.linux.x64.Release.mch	132,550,474	+29,541
realworld.run.linux.x64.checked.mch	13,174,358	+8,391
smoke_tests.nativeaot.linux.x64.checked.mch	4,064,942	+122

FullOpts (+565,740 bytes)

Collection	Base size (bytes)	Diff size (bytes)
benchmarks.run.linux.x64.checked.mch	14,036,750	+1,904
benchmarks.run_pgo.linux.x64.checked.mch	47,462,209	+157,649
benchmarks.run_tiered.linux.x64.checked.mch	3,695,669	-2,851
coreclr_tests.run.linux.x64.checked.mch	123,759,966	-53,512
libraries.crossgen2.linux.x64.checked.mch	38,725,374	+18,689
libraries.pmi.linux.x64.checked.mch	60,304,923	+23,461
libraries_tests.run.linux.x64.Release.mch	151,915,511	+382,346
libraries_tests_no_tiered_compilation.run.linux.x64.Release.mch	121,932,706	+29,541
realworld.run.linux.x64.checked.mch	12,788,474	+8,391
smoke_tests.nativeaot.linux.x64.checked.mch	4,063,993	+122

Assembly diffs for osx/arm64 ran on windows/x64

Diffs are based on2,270,092 contexts (932,669 MinOpts,1,337,423 FullOpts).

MISSED contexts: base:772 (0.03%), diff:778 (0.03%)

Overall (+539,840 bytes)

Collection	Base size (bytes)	Diff size (bytes)
benchmarks.run.osx.arm64.checked.mch	11,173,700	+13,580
benchmarks.run_pgo.osx.arm64.checked.mch	34,346,112	+21,404
benchmarks.run_tiered.osx.arm64.checked.mch	15,512,636	+1,296
coreclr_tests.run.osx.arm64.checked.mch	486,269,368	+155,172
libraries.crossgen2.osx.arm64.checked.mch	55,715,716	+40,472
libraries.pmi.osx.arm64.checked.mch	80,194,908	+61,468
libraries_tests.run.osx.arm64.Release.mch	323,385,308	+178,832
libraries_tests_no_tiered_compilation.run.osx.arm64.Release.mch	162,548,024	+57,088
realworld.run.osx.arm64.checked.mch	15,058,824	+10,528

MinOpts (-21,812 bytes)

Collection	Base size (bytes)	Diff size (bytes)
benchmarks.run_pgo.osx.arm64.checked.mch	16,301,300	-4,868
benchmarks.run_tiered.osx.arm64.checked.mch	11,504,484	-3,636
coreclr_tests.run.osx.arm64.checked.mch	332,612,924	-2,076
libraries_tests.run.osx.arm64.Release.mch	203,715,848	-10,640
libraries_tests_no_tiered_compilation.run.osx.arm64.Release.mch	13,153,728	-592

FullOpts (+561,652 bytes)

Collection	Base size (bytes)	Diff size (bytes)
benchmarks.run.osx.arm64.checked.mch	11,173,164	+13,580
benchmarks.run_pgo.osx.arm64.checked.mch	18,044,812	+26,272
benchmarks.run_tiered.osx.arm64.checked.mch	4,008,152	+4,932
coreclr_tests.run.osx.arm64.checked.mch	153,656,444	+157,248
libraries.crossgen2.osx.arm64.checked.mch	55,714,088	+40,472
libraries.pmi.osx.arm64.checked.mch	80,073,780	+61,468
libraries_tests.run.osx.arm64.Release.mch	119,669,460	+189,472
libraries_tests_no_tiered_compilation.run.osx.arm64.Release.mch	149,394,296	+57,680
realworld.run.osx.arm64.checked.mch	14,494,868	+10,528

Assembly diffs for windows/arm64 ran on windows/x64

Diffs are based on2,339,802 contexts (938,449 MinOpts,1,401,353 FullOpts).

MISSED contexts: base:1,309 (0.06%), diff:1,315 (0.06%)

Overall (+659,716 bytes)

Collection	Base size (bytes)	Diff size (bytes)
benchmarks.run.windows.arm64.checked.mch	10,957,488	+13,160
benchmarks.run_pgo.windows.arm64.checked.mch	45,442,912	+82,132
benchmarks.run_tiered.windows.arm64.checked.mch	15,586,420	+1,528
coreclr_tests.run.windows.arm64.checked.mch	495,080,320	+159,760
libraries.crossgen2.windows.arm64.checked.mch	59,059,452	+43,292
libraries.pmi.windows.arm64.checked.mch	79,827,256	+49,804
libraries_tests.run.windows.arm64.Release.mch	329,452,088	+233,324
libraries_tests_no_tiered_compilation.run.windows.arm64.Release.mch	171,557,220	+59,636
realworld.run.windows.arm64.checked.mch	15,902,524	+11,028
smoke_tests.nativeaot.windows.arm64.checked.mch	3,798,548	+6,052

MinOpts (-21,684 bytes)

Collection	Base size (bytes)	Diff size (bytes)
benchmarks.run_pgo.windows.arm64.checked.mch	16,011,396	-4,920
benchmarks.run_tiered.windows.arm64.checked.mch	11,177,188	-3,600
coreclr_tests.run.windows.arm64.checked.mch	338,689,104	-2,184
libraries_tests.run.windows.arm64.Release.mch	203,433,596	-10,388
libraries_tests_no_tiered_compilation.run.windows.arm64.Release.mch	13,153,748	-592

FullOpts (+681,400 bytes)

Collection	Base size (bytes)	Diff size (bytes)
benchmarks.run.windows.arm64.checked.mch	10,956,952	+13,160
benchmarks.run_pgo.windows.arm64.checked.mch	29,431,516	+87,052
benchmarks.run_tiered.windows.arm64.checked.mch	4,409,232	+5,128
coreclr_tests.run.windows.arm64.checked.mch	156,391,216	+161,944
libraries.crossgen2.windows.arm64.checked.mch	59,057,816	+43,292
libraries.pmi.windows.arm64.checked.mch	79,707,272	+49,804
libraries_tests.run.windows.arm64.Release.mch	126,018,492	+243,712
libraries_tests_no_tiered_compilation.run.windows.arm64.Release.mch	158,403,472	+60,228
realworld.run.windows.arm64.checked.mch	15,338,544	+11,028
smoke_tests.nativeaot.windows.arm64.checked.mch	3,797,536	+6,052

Assembly diffs for windows/x64 ran on windows/x64

Diffs are based on2,510,841 contexts (997,391 MinOpts,1,513,450 FullOpts).

MISSED contexts: base:1,370 (0.05%), diff:1,371 (0.05%)

Overall (-20,210 bytes)

Collection	Base size (bytes)	Diff size (bytes)
aspnet.run.windows.x64.checked.mch	47,016,272	+5,893
benchmarks.run.windows.x64.checked.mch	8,741,804	-2,717
benchmarks.run_pgo.windows.x64.checked.mch	36,134,420	-2,823
benchmarks.run_tiered.windows.x64.checked.mch	12,416,717	-4,767
coreclr_tests.run.windows.x64.checked.mch	393,035,946	-72,969
libraries.crossgen2.windows.x64.checked.mch	39,485,796	-1,523
libraries.pmi.windows.x64.checked.mch	61,660,326	-1,714
libraries_tests.run.windows.x64.Release.mch	280,853,081	+65,922
libraries_tests_no_tiered_compilation.run.windows.x64.Release.mch	137,059,674	-5,007
realworld.run.windows.x64.checked.mch	14,130,524	+1,107
smoke_tests.nativeaot.windows.x64.checked.mch	4,880,142	-1,612

FullOpts (-20,210 bytes)

Collection	Base size (bytes)	Diff size (bytes)
aspnet.run.windows.x64.checked.mch	28,525,223	+5,893
benchmarks.run.windows.x64.checked.mch	8,741,441	-2,717
benchmarks.run_pgo.windows.x64.checked.mch	21,963,764	-2,823
benchmarks.run_tiered.windows.x64.checked.mch	3,317,478	-4,767
coreclr_tests.run.windows.x64.checked.mch	120,247,366	-72,969
libraries.crossgen2.windows.x64.checked.mch	39,484,609	-1,523
libraries.pmi.windows.x64.checked.mch	61,546,805	-1,714
libraries_tests.run.windows.x64.Release.mch	104,994,847	+65,922
libraries_tests_no_tiered_compilation.run.windows.x64.Release.mch	126,440,568	-5,007
realworld.run.windows.x64.checked.mch	13,743,915	+1,107
smoke_tests.nativeaot.windows.x64.checked.mch	4,879,195	-1,612

Detailshere

Throughput diffs

Throughput diffs for linux/arm64 ran on windows/x64

Overall (+0.30% to+1.63%)

Collection	PDIFF
benchmarks.run.linux.arm64.checked.mch	+0.56%
benchmarks.run_pgo.linux.arm64.checked.mch	+1.40%
benchmarks.run_tiered.linux.arm64.checked.mch	+0.30%
coreclr_tests.run.linux.arm64.checked.mch	+0.39%
libraries.crossgen2.linux.arm64.checked.mch	+0.47%
libraries.pmi.linux.arm64.checked.mch	+0.54%
libraries_tests.run.linux.arm64.Release.mch	+1.63%
libraries_tests_no_tiered_compilation.run.linux.arm64.Release.mch	+0.41%
realworld.run.linux.arm64.checked.mch	+0.67%
smoke_tests.nativeaot.linux.arm64.checked.mch	+0.61%

MinOpts (-0.03% to+0.00%)

Collection	PDIFF
benchmarks.run.linux.arm64.checked.mch	-0.02%
benchmarks.run_pgo.linux.arm64.checked.mch	-0.02%
benchmarks.run_tiered.linux.arm64.checked.mch	-0.02%
libraries.pmi.linux.arm64.checked.mch	-0.03%
libraries_tests.run.linux.arm64.Release.mch	-0.01%
libraries_tests_no_tiered_compilation.run.linux.arm64.Release.mch	-0.01%
smoke_tests.nativeaot.linux.arm64.checked.mch	-0.01%

FullOpts (+0.42% to+2.10%)

Collection	PDIFF
benchmarks.run.linux.arm64.checked.mch	+0.56%
benchmarks.run_pgo.linux.arm64.checked.mch	+1.59%
benchmarks.run_tiered.linux.arm64.checked.mch	+0.65%
coreclr_tests.run.linux.arm64.checked.mch	+0.66%
libraries.crossgen2.linux.arm64.checked.mch	+0.47%
libraries.pmi.linux.arm64.checked.mch	+0.54%
libraries_tests.run.linux.arm64.Release.mch	+2.10%
libraries_tests_no_tiered_compilation.run.linux.arm64.Release.mch	+0.42%
realworld.run.linux.arm64.checked.mch	+0.67%
smoke_tests.nativeaot.linux.arm64.checked.mch	+0.61%

Throughput diffs for linux/x64 ran on windows/x64

Overall (+0.33% to+1.50%)

Collection	PDIFF
benchmarks.run.linux.x64.checked.mch	+0.57%
benchmarks.run_pgo.linux.x64.checked.mch	+1.49%
benchmarks.run_tiered.linux.x64.checked.mch	+0.33%
coreclr_tests.run.linux.x64.checked.mch	+0.44%
libraries.crossgen2.linux.x64.checked.mch	+0.56%
libraries.pmi.linux.x64.checked.mch	+0.60%
libraries_tests.run.linux.x64.Release.mch	+1.50%
libraries_tests_no_tiered_compilation.run.linux.x64.Release.mch	+0.46%
realworld.run.linux.x64.checked.mch	+0.74%
smoke_tests.nativeaot.linux.x64.checked.mch	+0.74%

MinOpts (+0.00% to+0.01%)

Collection	PDIFF
libraries.pmi.linux.x64.checked.mch	+0.01%

FullOpts (+0.47% to+1.91%)

Collection	PDIFF
benchmarks.run.linux.x64.checked.mch	+0.58%
benchmarks.run_pgo.linux.x64.checked.mch	+1.67%
benchmarks.run_tiered.linux.x64.checked.mch	+0.68%
coreclr_tests.run.linux.x64.checked.mch	+0.74%
libraries.crossgen2.linux.x64.checked.mch	+0.56%
libraries.pmi.linux.x64.checked.mch	+0.60%
libraries_tests.run.linux.x64.Release.mch	+1.91%
libraries_tests_no_tiered_compilation.run.linux.x64.Release.mch	+0.47%
realworld.run.linux.x64.checked.mch	+0.74%
smoke_tests.nativeaot.linux.x64.checked.mch	+0.74%

Throughput diffs for osx/arm64 ran on windows/x64

Overall (+0.37% to+1.43%)

Collection	PDIFF
benchmarks.run.osx.arm64.checked.mch	+0.73%
benchmarks.run_pgo.osx.arm64.checked.mch	+0.95%
benchmarks.run_tiered.osx.arm64.checked.mch	+0.39%
coreclr_tests.run.osx.arm64.checked.mch	+0.37%
libraries.crossgen2.osx.arm64.checked.mch	+0.47%
libraries.pmi.osx.arm64.checked.mch	+0.56%
libraries_tests.run.osx.arm64.Release.mch	+1.43%
libraries_tests_no_tiered_compilation.run.osx.arm64.Release.mch	+0.41%
realworld.run.osx.arm64.checked.mch	+0.69%

MinOpts (-0.04% to+0.01%)

Collection	PDIFF
benchmarks.run_pgo.osx.arm64.checked.mch	-0.03%
benchmarks.run_tiered.osx.arm64.checked.mch	-0.03%
libraries.crossgen2.osx.arm64.checked.mch	-0.01%
libraries.pmi.osx.arm64.checked.mch	-0.04%
libraries_tests.run.osx.arm64.Release.mch	-0.01%
libraries_tests_no_tiered_compilation.run.osx.arm64.Release.mch	-0.01%
realworld.run.osx.arm64.checked.mch	+0.01%

FullOpts (+0.42% to+2.02%)

Collection	PDIFF
benchmarks.run.osx.arm64.checked.mch	+0.73%
benchmarks.run_pgo.osx.arm64.checked.mch	+1.18%
benchmarks.run_tiered.osx.arm64.checked.mch	+0.70%
coreclr_tests.run.osx.arm64.checked.mch	+0.63%
libraries.crossgen2.osx.arm64.checked.mch	+0.47%
libraries.pmi.osx.arm64.checked.mch	+0.56%
libraries_tests.run.osx.arm64.Release.mch	+2.02%
libraries_tests_no_tiered_compilation.run.osx.arm64.Release.mch	+0.42%
realworld.run.osx.arm64.checked.mch	+0.70%

Throughput diffs for windows/arm64 ran on windows/x64

Overall (+0.39% to+1.50%)

Collection	PDIFF
benchmarks.run.windows.arm64.checked.mch	+0.73%
benchmarks.run_pgo.windows.arm64.checked.mch	+1.50%
benchmarks.run_tiered.windows.arm64.checked.mch	+0.40%
coreclr_tests.run.windows.arm64.checked.mch	+0.39%
libraries.crossgen2.windows.arm64.checked.mch	+0.47%
libraries.pmi.windows.arm64.checked.mch	+0.55%
libraries_tests.run.windows.arm64.Release.mch	+1.40%
libraries_tests_no_tiered_compilation.run.windows.arm64.Release.mch	+0.41%
realworld.run.windows.arm64.checked.mch	+0.69%
smoke_tests.nativeaot.windows.arm64.checked.mch	+0.65%

MinOpts (-0.03% to+0.00%)

Collection	PDIFF
benchmarks.run_pgo.windows.arm64.checked.mch	-0.03%
benchmarks.run_tiered.windows.arm64.checked.mch	-0.03%
libraries.pmi.windows.arm64.checked.mch	-0.03%
libraries_tests.run.windows.arm64.Release.mch	-0.01%
libraries_tests_no_tiered_compilation.run.windows.arm64.Release.mch	-0.01%
smoke_tests.nativeaot.windows.arm64.checked.mch	-0.01%

FullOpts (+0.42% to+1.96%)

Collection	PDIFF
benchmarks.run.windows.arm64.checked.mch	+0.73%
benchmarks.run_pgo.windows.arm64.checked.mch	+1.72%
benchmarks.run_tiered.windows.arm64.checked.mch	+0.69%
coreclr_tests.run.windows.arm64.checked.mch	+0.66%
libraries.crossgen2.windows.arm64.checked.mch	+0.47%
libraries.pmi.windows.arm64.checked.mch	+0.55%
libraries_tests.run.windows.arm64.Release.mch	+1.96%
libraries_tests_no_tiered_compilation.run.windows.arm64.Release.mch	+0.42%
realworld.run.windows.arm64.checked.mch	+0.70%
smoke_tests.nativeaot.windows.arm64.checked.mch	+0.65%

Throughput diffs for windows/x64 ran on windows/x64

Overall (+0.43% to+1.31%)

Collection	PDIFF
aspnet.run.windows.x64.checked.mch	+1.19%
benchmarks.run.windows.x64.checked.mch	+0.78%
benchmarks.run_pgo.windows.x64.checked.mch	+1.28%
benchmarks.run_tiered.windows.x64.checked.mch	+0.43%
coreclr_tests.run.windows.x64.checked.mch	+0.44%
libraries.crossgen2.windows.x64.checked.mch	+0.56%
libraries.pmi.windows.x64.checked.mch	+0.60%
libraries_tests.run.windows.x64.Release.mch	+1.31%
libraries_tests_no_tiered_compilation.run.windows.x64.Release.mch	+0.46%
realworld.run.windows.x64.checked.mch	+0.72%
smoke_tests.nativeaot.windows.x64.checked.mch	+0.76%

MinOpts (+0.00% to+0.01%)

Collection	PDIFF
libraries.pmi.windows.x64.checked.mch	+0.01%

FullOpts (+0.47% to+1.79%)

Collection	PDIFF
aspnet.run.windows.x64.checked.mch	+1.36%
benchmarks.run.windows.x64.checked.mch	+0.78%
benchmarks.run_pgo.windows.x64.checked.mch	+1.47%
benchmarks.run_tiered.windows.x64.checked.mch	+0.70%
coreclr_tests.run.windows.x64.checked.mch	+0.74%
libraries.crossgen2.windows.x64.checked.mch	+0.56%
libraries.pmi.windows.x64.checked.mch	+0.60%
libraries_tests.run.windows.x64.Release.mch	+1.79%
libraries_tests_no_tiered_compilation.run.windows.x64.Release.mch	+0.47%
realworld.run.windows.x64.checked.mch	+0.72%
smoke_tests.nativeaot.windows.x64.checked.mch	+0.76%

Detailshere

Throughput diffs for linux/arm ran on windows/x86

MinOpts (-0.02% to-0.00%)

Collection	PDIFF
benchmarks.run.linux.arm.checked.mch	-0.01%
benchmarks.run_pgo.linux.arm.checked.mch	-0.01%
benchmarks.run_tiered.linux.arm.checked.mch	-0.01%
libraries.pmi.linux.arm.checked.mch	-0.02%
libraries_tests.run.linux.arm.Release.mch	-0.01%

Throughput diffs for windows/x86 ran on windows/x86

Overall (+0.01%)

Collection	PDIFF
benchmarks.run.windows.x86.checked.mch	+0.01%
benchmarks.run_pgo.windows.x86.checked.mch	+0.01%
benchmarks.run_tiered.windows.x86.checked.mch	+0.01%
coreclr_tests.run.windows.x86.checked.mch	+0.01%
libraries.crossgen2.windows.x86.checked.mch	+0.01%
libraries.pmi.windows.x86.checked.mch	+0.01%
libraries_tests.run.windows.x86.Release.mch	+0.01%
libraries_tests_no_tiered_compilation.run.windows.x86.Release.mch	+0.01%
realworld.run.windows.x86.checked.mch	+0.01%

MinOpts (+0.00% to+0.02%)

Collection	PDIFF
benchmarks.run_pgo.windows.x86.checked.mch	+0.01%
benchmarks.run_tiered.windows.x86.checked.mch	+0.01%
libraries.pmi.windows.x86.checked.mch	+0.02%
libraries_tests.run.windows.x86.Release.mch	+0.01%
libraries_tests_no_tiered_compilation.run.windows.x86.Release.mch	+0.01%

FullOpts (+0.01%)

Collection	PDIFF
benchmarks.run.windows.x86.checked.mch	+0.01%
benchmarks.run_pgo.windows.x86.checked.mch	+0.01%
benchmarks.run_tiered.windows.x86.checked.mch	+0.01%
coreclr_tests.run.windows.x86.checked.mch	+0.01%
libraries.crossgen2.windows.x86.checked.mch	+0.01%
libraries.pmi.windows.x86.checked.mch	+0.01%
libraries_tests.run.windows.x86.Release.mch	+0.01%
libraries_tests_no_tiered_compilation.run.windows.x86.Release.mch	+0.01%
realworld.run.windows.x86.checked.mch	+0.01%

Detailshere

Throughput diffs for linux/arm64 ran on linux/x64

Overall (+0.32% to+1.53%)

Collection	PDIFF
benchmarks.run_pgo.linux.arm64.checked.mch	+1.38%
libraries_tests.run.linux.arm64.Release.mch	+1.53%
smoke_tests.nativeaot.linux.arm64.checked.mch	+0.61%
coreclr_tests.run.linux.arm64.checked.mch	+0.38%
benchmarks.run_tiered.linux.arm64.checked.mch	+0.32%
libraries.crossgen2.linux.arm64.checked.mch	+0.48%
benchmarks.run.linux.arm64.checked.mch	+0.55%
libraries_tests_no_tiered_compilation.run.linux.arm64.Release.mch	+0.41%
realworld.run.linux.arm64.checked.mch	+0.66%
libraries.pmi.linux.arm64.checked.mch	+0.55%

MinOpts (+0.01% to+0.08%)

Collection	PDIFF
benchmarks.run_pgo.linux.arm64.checked.mch	+0.01%
libraries_tests.run.linux.arm64.Release.mch	+0.02%
smoke_tests.nativeaot.linux.arm64.checked.mch	+0.02%
coreclr_tests.run.linux.arm64.checked.mch	+0.01%
benchmarks.run_tiered.linux.arm64.checked.mch	+0.01%
libraries.crossgen2.linux.arm64.checked.mch	+0.01%
benchmarks.run.linux.arm64.checked.mch	+0.05%
libraries_tests_no_tiered_compilation.run.linux.arm64.Release.mch	+0.01%
realworld.run.linux.arm64.checked.mch	+0.01%
libraries.pmi.linux.arm64.checked.mch	+0.08%

FullOpts (+0.42% to+1.96%)

Collection	PDIFF
benchmarks.run_pgo.linux.arm64.checked.mch	+1.55%
libraries_tests.run.linux.arm64.Release.mch	+1.96%
smoke_tests.nativeaot.linux.arm64.checked.mch	+0.61%
coreclr_tests.run.linux.arm64.checked.mch	+0.65%
benchmarks.run_tiered.linux.arm64.checked.mch	+0.64%
libraries.crossgen2.linux.arm64.checked.mch	+0.48%
benchmarks.run.linux.arm64.checked.mch	+0.56%
libraries_tests_no_tiered_compilation.run.linux.arm64.Release.mch	+0.42%
realworld.run.linux.arm64.checked.mch	+0.66%
libraries.pmi.linux.arm64.checked.mch	+0.55%

Throughput diffs for linux/x64 ran on linux/x64

Overall (+0.31% to+1.43%)

Collection	PDIFF
libraries.crossgen2.linux.x64.checked.mch	+0.54%
benchmarks.run_pgo.linux.x64.checked.mch	+1.43%
coreclr_tests.run.linux.x64.checked.mch	+0.40%
realworld.run.linux.x64.checked.mch	+0.70%
smoke_tests.nativeaot.linux.x64.checked.mch	+0.71%
benchmarks.run.linux.x64.checked.mch	+0.54%
libraries.pmi.linux.x64.checked.mch	+0.56%
libraries_tests.run.linux.x64.Release.mch	+1.37%
benchmarks.run_tiered.linux.x64.checked.mch	+0.31%
libraries_tests_no_tiered_compilation.run.linux.x64.Release.mch	+0.43%

MinOpts (+0.00% to+0.01%)

Collection	PDIFF
libraries.pmi.linux.x64.checked.mch	+0.01%

FullOpts (+0.44% to+1.75%)

Collection	PDIFF
libraries.crossgen2.linux.x64.checked.mch	+0.54%
benchmarks.run_pgo.linux.x64.checked.mch	+1.60%
coreclr_tests.run.linux.x64.checked.mch	+0.70%
realworld.run.linux.x64.checked.mch	+0.70%
smoke_tests.nativeaot.linux.x64.checked.mch	+0.71%
benchmarks.run.linux.x64.checked.mch	+0.54%
libraries.pmi.linux.x64.checked.mch	+0.56%
libraries_tests.run.linux.x64.Release.mch	+1.75%
benchmarks.run_tiered.linux.x64.checked.mch	+0.64%
libraries_tests_no_tiered_compilation.run.linux.x64.Release.mch	+0.44%

Detailshere

build-analysisbot mentioned this pull request

Feb 3, 2024

System.Net.Security.Tests.SslStreamCertificateContextOcspLinuxTests.RefreshOcspResponse_BeforeExpiration test failure#97779

Closed

Copy link

MemberAuthor

jakobbotsch commentedFeb 22, 2024

/azp run runtime-coreclr jitstress, runtime-coreclr libraries-jitstress

Copy link

azure-pipelinesbot commentedFeb 22, 2024

Azure Pipelines successfully started running 2 pipeline(s).

Copy link

MemberAuthor

jakobbotsch commentedFeb 22, 2024•
edited
Loading

cc @dotnet/jit-contrib PTAL@BruceForstall @AndyAyersMS

Diffs. I spent a long time analyzing these, see#97865 (comment) and#97865 (comment).

There were a bunch of regressions when IV widening in cases where the same local is the primary IV of multiple loops. This particularly affects cloning. I've disabled this case for now; I want to look closer at the case and figure out how to enable it separately.

The initial implementation always placed the initialization of the widened IV in the preheader. However, if the reaching def from outside the loop comes from a block much further back, then that means the lifetime of the widened IV is shorter than the old IV. In a lot of cases that resulted in suboptimal register allocation of the widened IV (because other live variables ended up taking the "good" register before we got to the initialization of the widened IV). So this implementation tries to place the initialization of the widened IV in the same spot as the old narrow IV, and then does a best-effort pass to replace uses of the narrow IV with the wide IV on the way to the loop. That seems to work well, but it's a bit of a hack -- I hope to reevaluate it in the future, perhaps with some LSRA work.

Throughput impact is in the 0.2-0.3% range for most collections, with benchmarks.run_pgo being a bit worse.

jakobbotsch requested review fromAndyAyersMS andBruceForstall

February 22, 2024 16:26

AndyAyersMS reviewed

Feb 22, 2024

View reviewed changes

Copy link

Member

AndyAyersMS left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others.Learn more.

Did a first review pass and it looks good overall. Will look again later.

I wonder if you might save some TP if you kept track of the uses during the locate phase and used that to drive replacement.

More generally if we want to recognize secondary IVs it seems like walking forward from the primary IV appearances might be a good plan. Perhaps when we get to strength reduction?

src/coreclr/jit/inductionvariableopts.cpp OutdatedShow resolvedHide resolved

src/coreclr/jit/inductionvariableopts.cpp Outdated

		// => <L, 0, <L, 1, 1>>
		//
		// The main issue is that it requires cache invalidation afterwards
		// and turning the recursive result into an addrec.

Copy link

Member

AndyAyersMSFeb 22, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others.Learn more.

Any idea how prevalent these more complicated IVs might be?

Copy link

MemberAuthor

jakobbotschFeb 22, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others.Learn more.

I haven't checked, but I have some (hacky) code lying around for it that I was planning to look at the impact of later.

There are some other notable missing things in the SCEV IR, like a subtraction operator, so loops that subtract by an invariant local are not representable (subtractions with a constant are usually represented asGT_ADD so those are ok).

src/coreclr/jit/inductionvariableopts.cpp OutdatedShow resolvedHide resolved

src/coreclr/jit/inductionvariableopts.cpp Outdated

		}

		return nullptr;
		}

Copy link

Member

AndyAyersMSFeb 22, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others.Learn more.

Hoisting has a more refined notion of loop invariance that can handle more tree shapes; any thought of using something like that here?

Copy link

MemberAuthor

jakobbotschFeb 22, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others.Learn more.

Do you mean by switching to use VN? I suppose that would be possible (although somewhat annoying since the VN wouldn't give me things like "reaching def from outside the loop" that I'm making use of here). I think it would mainly get us the ability to reason about loop invariant loads (so you could have e.g.<L00, 0, IND(x)>, provided we also started representing loads in the SCEV IR).

Optimize throughput

6a70c29

Copy link

MemberAuthor

jakobbotsch commentedFeb 23, 2024

I wonder if you might save some TP if you kept track of the uses during the locate phase and used that to drive replacement.
More generally if we want to recognize secondary IVs it seems like walking forward from the primary IV appearances might be a good plan. Perhaps when we get to strength reduction?

I pushed a change that switches the profitability check to use the tree list and also saves all statements in which appearences were found such that the replacement step doesn't need to walk all the loop IR again.

The savings are relatively modest (TP diff) -- I think most TP is eaten by the DFS and loop finding. But indeed it seems like we can reuse it for strength reduction as well, to limit what nodes we invoke the scalar evolution analysis on.

jakobbotsch added2 commits

February 23, 2024 11:16

Clean up, address some feedback

410c5c2

More feedback

f6efa21

AndyAyersMS approved these changes

Feb 26, 2024

View reviewed changes

Copy link

Member

AndyAyersMS left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others.Learn more.

LGTM.

Will be interesting to see how this evolves when you start to look for secondary IVs.

Copy link

Contributor

BruceForstall commentedFeb 26, 2024

cc @dotnet/jit-contrib PTAL@BruceForstall @AndyAyersMS

I've started looking at this.

BruceForstall approved these changes

Feb 27, 2024

View reviewed changes

Copy link

Contributor

BruceForstall left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others.Learn more.

Some comments:

It would be valuable for this code to have more design documentation. E.g., a Markdown document (in the JIT "doc" tree) on SCEV, on the IV widening algorithm design, etc. This could also take the form of more extensive design-oriented comments.
Would it make sense to extract the SCEV code to separate scev.h/cpp files? That is, separate the analysis and optimization code.
Add cScev/dScev debug dumpers?
Is there any useful dumper to be added to clrjt.natvis?
To match LLVM conventions, should the SCEV display the addition operator (even though that is currently the only one), e.g.,<loop, start, +, step> or<start, +, step>(loop)?

src/coreclr/jit/inductionvariableopts.cpp Outdated

		// Parameters:
		// scev - The scev node.
		//
		void ScalarEvolutionContext::DumpScev(Scev* scev)

Copy link

Contributor

BruceForstallFeb 27, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others.Learn more.

I presume this can't bestatic for some reason? If it was static it would be easier to call from a debugger, perhaps.

Copy link

MemberAuthor

jakobbotschFeb 27, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others.Learn more.

The main thing was that I wanted to display the specific loop forScevAddRec, andScevAddRec itself doesn't have the loop field. However, I just added that loop field as a debug-only field, and made the dumping a member method onScev. So now we havecScev anddScev.

src/coreclr/jit/inductionvariableopts.cpp Outdated

		}

		//------------------------------------------------------------------------
		// optBestEffortReplaceNarrowIVUsesWith: Try to find and replace uses of the specified

Copy link

Contributor

BruceForstallFeb 27, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others.Learn more.

What makes this "best effort"?

It seems awkward to end the function name with "With".

Copy link

MemberAuthor

jakobbotschFeb 27, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others.Learn more.

Renamed it tooptBestEffortReplaceNarrowIVUses, and added a remarks:

// Remarks://   This function is best effort; it might not find all uses of the provided//   SSA num, particularly because it does not follow into joins. Note that we//   only use this to replace uses of the narrow IV outside the loop; inside//   the loop we do ensure that all uses/defs are replaced.//   Keeping it best-effort outside the loop is ok; there is no correctness//   issue since we do not invalidate the value of the old narrow IV in any//   way, but it may mean we end up leaving the narrow IV live concurrently//   with the new widened IV, increasing register pressure.

jakobbotsch added3 commits

February 27, 2024 12:17

Split out into multiple files

054ad94

Address some feedback

2700cda

Add some docs

0f8d2b2

Copy link

MemberAuthor

jakobbotsch commentedFeb 27, 2024

Some comments:
It would be valuable for this code to have more design documentation. E.g., a Markdown document (in the JIT "doc" tree) on SCEV, on the IV widening algorithm design, etc. This could also take the form of more extensive design-oriented comments.

Makes sense, but personally I prefer to have it in the relevant files, so I added some documentation there.

Would it make sense to extract the SCEV code to separate scev.h/cpp files? That is, separate the analysis and optimization code.

Yeah, seems good to me. I split it into scev.h, scev.cpp and inductionvariableopts.cpp.

Add cScev/dScev debug dumpers?

Done.

Is there any useful dumper to be added to clrjt.natvis?

Added some basic display of the SCEV oper/type.

To match LLVM conventions, should the SCEV display the addition operator (even though that is currently the only one), e.g.,<loop, start, +, step> or<start, +, step>(loop)?

I actually matched the notation from the paper (I think LLVM uses curly braces as well, and has a different representation of chains of recurrences). If we end up with more complex recurrences then we can consider changing the notation.

jakobbotsch added2 commits

February 27, 2024 13:25

More remarks

3e3759e

Expand ryujit markdown docs with new phase

3ec083f

build-analysisbot mentioned this pull request

Feb 27, 2024

Tracking issue for CI build timeouts#76454

Closed

Copy link

MemberAuthor

jakobbotsch commentedFeb 27, 2024

SPMI 0 sized diffs look like#98996. All CI jobs were green last time it ran, and since then I only reorganized some things without any diffs, so the failures/timeouts look unrelated.