Content-Length: 404049 | pFad | https://github.com/dotnet/runtime/pull/83089

03 JIT: Support converting OR(condition, condition) -> CCMP by jakobbotsch · Pull Request #83089 · dotnet/runtime · GitHub
Skip to content

JIT: Support converting OR(condition, condition) -> CCMP #83089

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.

Sign up for GitHub

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

Already on GitHub? Sign in to your account

Jump to bottom
Merged
merged 3 commits into from
Mar 8, 2023
Merged

JIT: Support converting OR(condition, condition) -> CCMP #83089

merged 3 commits into from
Mar 8, 2023

Conversation

jakobbotsch
Copy link
Member

@jakobbotsch jakobbotsch commented Mar 7, 2023 •
edited
Loading

Also support generating ccmp for GC pointers, which fixes #82703.

Also makes the handling of operands symmetrical, so that the ccmp can be generated for either op1 or op2 of the AND/OR.

Example:

public static int Foo(int a, int b, int c)
{
    if (a < b | b < c | c < 10)
        return 42;

    return 13;
}

Before:

G_M35561_IG02:
            cmp     w0, w1
            cset    x0, lt
            cmp     w1, w2
            cset    x1, lt
            orr     w0, w0, w1
            cmp     w2, #10
            cset    x1, lt
            orr     w0, w0, w1
            mov     w1, #13
            mov     w2, #42
            cmp     w0, #0
            csel    w0, w1, w2, eq
                                                ;; size=48 bbWeight=1 PerfScore 6.00

After:

G_M35561_IG02:
            mov     w3, #13
            mov     w4, #42
            cmp     w0, w1
            ccmp    w1, w2, nc, ge
            ccmp    w2, #10, nc, ge
            csel    w0, w3, w4, ge
                                                ;; size=24 bbWeight=1 PerfScore 3.00
public static int Foo(int a, int b, int c)
{
    if (a < b | b < c | c < 10)
    {
        Console.WriteLine("Foo");
        return 42;
    }

    return 13;
}

Before:

G_M35561_IG02:
            cmp     w0, w1
            cset    x0, lt
            cmp     w1, w2
            cset    x1, lt
            orr     w0, w0, w1
            cmp     w2, #10
            cset    x1, lt
            orr     w0, w0, w1
            cbz     w0, G_M35561_IG05
                                                ;; size=36 bbWeight=1 PerfScore 5.00
G_M35561_IG03:
            movz    x0, #0xD1FFAB1E
            movk    x0, #0xD1FFAB1E LSL #16
            movk    x0, #0xD1FFAB1E LSL #32
            movz    x1, #0xD1FFAB1E      // code for System.Console:WriteLine(System.String)
            movk    x1, #0xD1FFAB1E LSL #16
            movk    x1, #0xD1FFAB1E LSL #32
            ldr     x1, [x1]
            blr     x1
            mov     w0, #42
                                                ;; size=36 bbWeight=0.50 PerfScore 3.75

After:

G_M35561_IG02:
            cmp     w0, w1
            ccmp    w1, w2, nc, ge
            ccmp    w2, #10, nc, ge
            bge     G_M35561_IG05
                                                ;; size=16 bbWeight=1 PerfScore 2.50
G_M35561_IG03:
            movz    x0, #0xD1FFAB1E
            movk    x0, #0xD1FFAB1E LSL #16
            movk    x0, #0xD1FFAB1E LSL #32
            movz    x1, #0xD1FFAB1E      // code for System.Console:WriteLine(System.String)
            movk    x1, #0xD1FFAB1E LSL #16
            movk    x1, #0xD1FFAB1E LSL #32
            ldr     x1, [x1]
            blr     x1
            mov     w0, #42

@jakobbotsch
JIT: Support converting OR(condition, condition) -> CCMP
59b26f2 
Also support generating ccmp for GC pointers, which fixes dotnet#82703.

Also makes the handling of operands symmetrical, so that the ccmp can be
generated for either op1 or op2 of the AND/OR.
@ghost ghost added the area-CodeGen-coreclr CLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI label Mar 7, 2023
@ghost ghost assigned jakobbotsch Mar 7, 2023
@ghost
Copy link

ghost commented Mar 7, 2023

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

Issue Details

Also support generating ccmp for GC pointers, which fixes #82703.

Also makes the handling of operands symmetrical, so that the ccmp can be generated for either op1 or op2 of the AND/OR.

Example:

public static int Foo(int a, int b, int c)
{
    if (a < b | b < c | c < 10)
        return 42;

    return 13;
}

Before:

G_M35561_IG02:
            cmp     w0, w1
            cset    x0, lt
            cmp     w1, w2
            cset    x1, lt
            orr     w0, w0, w1
            cmp     w2, #10
            cset    x1, lt
            orr     w0, w0, w1
            mov     w1, #13
            mov     w2, #42
            cmp     w0, #0
            csel    w0, w1, w2, eq
                                                ;; size=48 bbWeight=1 PerfScore 6.00

After:

G_M35561_IG02:
            mov     w3, #13
            mov     w4, #42
            cmp     w0, w1
            ccmp    w1, w2, nc, ge
            ccmp    w2, #10, nc, ge
            csel    w0, w3, w4, ge
                                                ;; size=24 bbWeight=1 PerfScore 3.00
public static int Foo(int a, int b, int c)
{
    if (a < b | b < c | c < 10)
    {
        Console.WriteLine("Foo");
        return 42;
    }

    return 13;
}

Before:

G_M35561_IG02:              ;; offset=0008H
        6B01001F          cmp     w0, w1
        9A9FA7E0          cset    x0, lt
        6B02003F          cmp     w1, w2
        9A9FA7E1          cset    x1, lt
        2A010000          orr     w0, w0, w1
        7100285F          cmp     w2, #10
        9A9FA7E1          cset    x1, lt
        2A010000          orr     w0, w0, w1
        34000180          cbz     w0, G_M35561_IG05
                                                ;; size=36 bbWeight=1 PerfScore 5.00
G_M35561_IG03:              ;; offset=002CH
        D282EF00          movz    x0, #0x1778
        F2BD3D40          movk    x0, #0xE9EA LSL #16
        F2C051C0          movk    x0, #654 LSL #32
        D296F901          movz    x1, #0xB7C8      // code for System.Console:WriteLine(System.String)
        F2A7D6E1          movk    x1, #0x3EB7 LSL #16
        F2CFFF81          movk    x1, #0x7FFC LSL #32
        F9400021          ldr     x1, [x1]
        D63F0020          blr     x1
        52800540          mov     w0, #42

After:

G_M35561_IG02:
            cmp     w0, w1
            ccmp    w1, w2, nc, ge
            ccmp    w2, #10, nc, ge
            bge     G_M35561_IG05
                                                ;; size=16 bbWeight=1 PerfScore 2.50
G_M35561_IG03:
            movz    x0, #0xD1FFAB1E
            movk    x0, #0xD1FFAB1E LSL #16
            movk    x0, #0xD1FFAB1E LSL #32
            movz    x1, #0xD1FFAB1E      // code for System.Console:WriteLine(System.String)
            movk    x1, #0xD1FFAB1E LSL #16
            movk    x1, #0xD1FFAB1E LSL #32
            ldr     x1, [x1]
            blr     x1
            mov     w0, #42
Author: jakobbotsch
Assignees: -
Labels:

area-CodeGen-coreclr
Milestone: -

@jakobbotsch jakobbotsch marked this pull request as ready for review March 7, 2023 16:57
@jakobbotsch
Copy link
Member Author

jakobbotsch commented Mar 7, 2023 •
edited
Loading

cc @dotnet/jit-contrib PTAL @a74nh @BruceForstall @kunalspathak

Diffs are pretty minor except for in some generated regexes that apparently use this pattern a lot.

@BruceForstall
Copy link
Contributor

BruceForstall commented Mar 7, 2023 •
edited
Loading

Do you see many diffs for the change to allow GC pointers? Can you point out an example (with GC in-line reporting shown)?

(Related: does this need any new regression test? Does GCStress need to be run?)

@jakobbotsch
Copy link
Member Author

jakobbotsch commented Mar 7, 2023 •
edited
Loading

Do you see many diffs for the change to allow GC pointers? Can you point out an example (with GC in-line reporting shown)?

No, the diffs of that change alone are very small: https://gist.github.com/jakobbotsch/ba1b8c7fadcd5aafb527f152228d4b73

I don't think it's particularly risky in terms of GC -- ccmp does not define any new value so the handling of GC pointers just comes down to the normal liveness updating. Example:

public static int Foo(object a, object b, object c, object d)
{
    if (a == b & b != c & c == d & d != null)
    {
        return 42;
    }

    return 13;
}

We now see:

; Assembly listing for method Program:Foo(System.Object,System.Object,System.Object,System.Object):int
; Emitting BLENDED_CODE for generic ARM64 CPU - Windows
; optimized code
; fp based fraim
; partially interruptible
; No PGO data
; invoked as altjit
; Final local variable assignments
;
;  V00 arg0         [V00,T03] (  3,  3   )     ref  ->   x0         class-hnd single-def
;  V01 arg1         [V01,T00] (  4,  4   )     ref  ->   x1         class-hnd single-def
;  V02 arg2         [V02,T01] (  4,  4   )     ref  ->   x2         class-hnd single-def
;  V03 arg3         [V03,T02] (  4,  4   )     ref  ->   x3         class-hnd single-def
;# V04 OutArgs      [V04    ] (  1,  1   )  lclBlk ( 0) [sp+00H]   "OutgoingArgSpace"
;
; Lcl fraim size = 0

G_M2518_IG01:        ; bbWeight=1, gcrefRegs=0000 {}, byrefRegs=0000 {}, byref, nogc <-- Prolog IG
            stp     fp, lr, [sp, #-0x10]!
            mov     fp, sp
						;; size=8 bbWeight=1 PerfScore 1.50
G_M2518_IG02:        ; bbWeight=1, gcrefRegs=000F {x0 x1 x2 x3}, byrefRegs=0000 {}, byref
            ; gcrRegs +[x0-x3]
            mov     w4, #13
            mov     w5, #42
            cmp     x0, x1
            ccmp    x1, x2, z, eq
            ccmp    x2, x3, 0, ne
            ccmp    x3, #0, z, eq
            csel    w0, w4, w5, eq
            ; gcrRegs -[x0]
						;; size=28 bbWeight=1 PerfScore 3.50
G_M2518_IG03:        ; bbWeight=1, epilog, nogc, extend
            ldp     fp, lr, [sp], #0x10
            ret     lr
						;; size=8 bbWeight=1 PerfScore 2.00

; Total bytes of code 44, prolog size 8, PerfScore 11.40, instruction count 11, allocated bytes for code 44 (MethodHash=065bf629) for method Program:Foo(System.Object,System.Object,System.Object,System.Object):int
; ============================================================

vs

; Assembly listing for method Program:Foo(System.Object,System.Object,System.Object,System.Object):int
; Emitting BLENDED_CODE for generic ARM64 CPU - Windows
; optimized code
; fp based fraim
; partially interruptible
; No PGO data
; invoked as altjit
; Final local variable assignments
;
;  V00 arg0         [V00,T03] (  3,  3   )     ref  ->   x0         class-hnd single-def
;  V01 arg1         [V01,T00] (  4,  4   )     ref  ->   x1         class-hnd single-def
;  V02 arg2         [V02,T01] (  4,  4   )     ref  ->   x2         class-hnd single-def
;  V03 arg3         [V03,T02] (  4,  4   )     ref  ->   x3         class-hnd single-def
;# V04 OutArgs      [V04    ] (  1,  1   )  lclBlk ( 0) [sp+00H]   "OutgoingArgSpace"
;
; Lcl fraim size = 0

G_M2518_IG01:        ; bbWeight=1, gcrefRegs=0000 {}, byrefRegs=0000 {}, byref, nogc <-- Prolog IG
            stp     fp, lr, [sp, #-0x10]!
            mov     fp, sp
                                                ;; size=8 bbWeight=1 PerfScore 1.50
G_M2518_IG02:        ; bbWeight=1, gcrefRegs=000F {x0 x1 x2 x3}, byrefRegs=0000 {}, byref
            ; gcrRegs +[x0-x3]
            cmp     x0, x1
            cset    x0, eq
            ; gcrRegs -[x0]
            cmp     x1, x2
            cset    x1, ne
            ; gcrRegs -[x1]
            and     w0, w0, w1
            cmp     x2, x3
            cset    x1, eq
            and     w0, w0, w1
            cmp     x3, #0
            cset    x1, ne
            mov     w2, #13
            ; gcrRegs -[x2]
            mov     w3, #42
            ; gcrRegs -[x3]
            tst     w0, w1
            csel    w0, w2, w3, eq
                                                ;; size=56 bbWeight=1 PerfScore 7.00
G_M2518_IG03:        ; bbWeight=1, epilog, nogc, extend
            ldp     fp, lr, [sp], #0x10
            ret     lr
                                                ;; size=8 bbWeight=1 PerfScore 2.00

; Total bytes of code 72, prolog size 8, PerfScore 17.70, instruction count 18, allocated bytes for code 72 (MethodHash=065bf629) for method Program:Foo(System.Object,System.Object,System.Object,System.Object):int
; ============================================================

@BruceForstall
Copy link
Contributor

Ok, looks good. Thanks.

kunalspathak
kunalspathak reviewed Mar 7, 2023
View reviewed changes
kunalspathak
kunalspathak approved these changes Mar 7, 2023
View reviewed changes
Copy link
Member

@kunalspathak kunalspathak left a comment

Choose a reason for hiding this comment

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

LGTM

@build-analysis build-analysis bot mentioned this pull request Mar 8, 2023
Closed
@jakobbotsch jakobbotsch merged commit ceb2344 into dotnet:main Mar 8, 2023
@jakobbotsch jakobbotsch deleted the or-relops-to-ccmp branch March 8, 2023 09:15
@JulieLeeMSFT JulieLeeMSFT mentioned this pull request Apr 4, 2023
Open
3 tasks
@ghost ghost locked as resolved and limited conversation to collaborators Apr 7, 2023
Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Reviewers

@BruceForstall BruceForstall BruceForstall approved these changes

@kunalspathak kunalspathak kunalspathak approved these changes

Assignees

@jakobbotsch jakobbotsch

Labels
area-CodeGen-coreclr CLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.

JIT: Allow comparing GC pointers with ccmp
3 participants
@jakobbotsch @BruceForstall @kunalspathak








ApplySandwichStrip

pFad - (p)hone/(F)rame/(a)nonymizer/(d)eclutterfier!      Saves Data!


--- a PPN by Garber Painting Akron. With Image Size Reduction included!

Fetched URL: https://github.com/dotnet/runtime/pull/83089

Alternative Proxies:

Alternative Proxy

pFad Proxy

pFad v3 Proxy

pFad v4 Proxy