JIT: 'X <= uint.MaxValue' to 'X >> 32'

[EgorBo]

bool Foo(ulong x) => x <= uint.MaxValue; // same for >

Proposed diff for RyuJIT (what it currently emits vs expected)

; Method T:Foo(ulong):bool:this
-      mov      eax, 0xFFFFFFFF
-      cmp      rdx, rax
-      setbe    al
+      shr      rdi, 32
+      sete     al
       movzx    rax, al
       ret      

grep.app: https://grep.app/search?q=%28%3C%3D%7C%3E%29%20uint.MaxValue&regexp=true

category:implementation
theme:expression-opts
skill-level:beginner
cost:small
impact:small

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

Issue Details

---

bool Foo(ulong x) => x <= uint.MaxValue;

Proposed diff for RyuJIT (what it currently emits vs expected)

; Method T:Foo3(ulong):bool:this
-      mov      eax, 0xFFFFFFFF
-      cmp      rdx, rax
-      setbe    al
+      shr      rdi, 32
+      sete     al
       movzx    rax, al
       ret      

grep: https://grep.app/search?q=%3C%3D%20uint.MaxValue&words=true

Author:	EgorBo
Assignees:	-
Labels:	area-CodeGen-coreclr, untriaged
Milestone:	-

[En3Tho]

I can take this one on :)

Where is the best place to put this opt?
Can it work with long + int in similar way (shifting 31 bits instead of 32)? (I guess no?)
Also, similar opt can be done with ushort/byte, uint/ushort, uint/byte etc, short/byte etc. Do we have to consider them too? Byte.MaxValue seems quite a common comparand
https://grep.app/search?q=%28%3C%3D%7C%3E%29%20byte.MaxValue&regexp=true

[EgorBo]

I can take this one on :)

Where is the best place to put this opt? Can it work with long + int in similar way (shifting 31 bits instead of 32)? (I guess no?) Also, similar opt can be done with ushort/byte, uint/ushort, uint/byte etc, short/byte etc. Do we have to consider them too? Byte.MaxValue seems quite a common comparand https://grep.app/search?q=%28%3C%3D%7C%3E%29%20byte.MaxValue&regexp=true

Great! Assigning to you 🙂
For the rest of the questions I'd suggest playing with godbolt, e.g. https://godbolt.org/z/r3oxfYjo4

Where is the best place to put this opt?

Up to you or whatever simpler/produces more diffs 🙂 shr sets flags so it doesn't matter for it whether it feeds a condition or just a local.

[jakobbotsch]

I would like to avoid us introducing more uses of GTF_SET_FLAGS (c.f. #74867). So I think it should be done either via containment in lowering or as a peephole during codegen.

[EgorBo]

or as a peephole during codegen.

I think it's better to be done before lsra so it knows that we don't need a temp reg

[jakobbotsch]

I think it's better to be done before lsra so it knows that we don't need a temp reg

I agree, but it's more complicated.

[jakobbotsch]

We might consider transforming x <= uint.MaxValue into (x >> 32) == 0 as a separate transformation, it probably has some improvements even without avoiding the test. Introducing support for consuming CPU flags directly from arithmetic nodes via containment can be applied in general for many operations on xarch, so it should maybe be a separate step.

[En3Tho]

Here is a small reference table of comparsions: https://godbolt.org/z/49j88sh4x

In fact only (ulong)x > uint.MaxValue is optimized via shr