[dhartglassMSFT]

Method:
bool M1(object o) => o != null && (o is int or uint or long);

Intro block has code with an extra compare like cmp+cset+cbz (null check, write result of null check to local, compare and branch on local). The first cmp is needed because the local which receives the null check is live in later blocks.

To fix, JIT will look for shapes like this during forwardsub, replace the branch with the null check itself, and store a constant to the null-check-local along the path its used.

TPDiff reports +0.01%.

Diffs from motivating example:

x64

        test     rcx, rcx
-       setne    al
-       movzx    rax, al
-       test     eax, eax
-       je       RETURN
+       jne      TYPE_CHECKS
+       xor      eax, eax
+       jmp      RETURN

 TYPE_CHECKS:
        rest_of_function

ARM64

-       cmp      x0, #0
-       cset     x1, ne
-       cbz      w1, RETURN
+       cbnz     x0, TYPE_CHECKS
+       mov      w0, wzr
+       b        RETURN

 TYPE_CHECKS:
       rest_of_function
...
 RETURN:
-       mov      w0, w1
        ldp      fp, lr, [sp], #0x10
        ret      lr

Full arch diff results. I'm still looking through regressions. Some are cases of loop cloning being enabled, or changes in lsra.

Target	Contexts with diffs	Smaller	Larger	Same size	Net code-size delta
Windows x64	279	248	25	6	−3,621 bytes
Linux x64	299	266	28	5	−3,707 bytes
Windows x86	248	205	37	6	−2,776 bytes
Windows ARM64	285	224	7	54	−2,196 bytes
Linux ARM64	317	242	15	60	−2,448 bytes
macOS ARM64	26	22	1	3	−244 bytes
Linux ARM	235	222	7	6	−4,156 bytes
Browser WASM	1	1	0	0	−2 bytes

[azure-pipelines]

Azure Pipelines:
Successfully started running 5 pipeline(s).
11 pipeline(s) were filtered out due to trigger conditions.
There may be pipelines that require an authorized user to comment /azp run to run.

[dotnet-policy-service]

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

[dhartglassMSFT]

@dotnet/jit-contrib @EgorBo
If there's a better place than forward sub for this let me know. I wanted an early phase that had liveness information

[EgorBo]

Not sure I understand the limitation on varDsc->lvIsStructField and genActualType(varDsc->TypeGet()) != TYP_INT). Also, IsLiveInOutOfHandler feels redudant - Liveness already adds what exception handlers need to a block's live-in set

[EgorBo]

am I missing something - why do we reverse only EQ here?

[EgorBo]

This feels like it's doing what optRedundantRelop supposed to do? Do you mind if I leave this to @AndyAyersMS or @jakobbotsch to take a look, perhaps we can generalize it into some strength reduction / code sink phase or something, because current diffs looks quite small.