cmove is not emitted for a simple expression

[EgorBo]

int foo(bool cond, int a, int b) => 1 + (cond ? a : b); // or e.g. "cond ? 1 : 2"

Current codegen:

; Method Program:foo(bool,int,int):int:this (FullOpts)
 test dl, dl
 jne SHORT G_M62184_IG04
 mov eax, r9d
 jmp SHORT G_M62184_IG05
G_M62184_IG04: ;; offset=0x0009
 mov eax, r8d
G_M62184_IG05: ;; offset=0x000C
 inc eax
 ret 
; Total bytes of code: 15

Expected: branchless code (cmove)

cc @jakobbotsch @dotnet/jit-contrib

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

Issue Details

---

int foo(bool cond, int a, int b) => 1 + (cond ? a : b); // or e.g. "cond ? 1 : 2"

Current codegen:

; Method Program:foo(bool,int,int):int:this (FullOpts)
 test dl, dl
 jne SHORT G_M62184_IG04
 mov eax, r9d
 jmp SHORT G_M62184_IG05
G_M62184_IG04: ;; offset=0x0009
 mov eax, r8d
G_M62184_IG05: ;; offset=0x000C
 inc eax
 ret 
; Total bytes of code: 15

Expected: branchless code (cmove)

cc @jakobbotsch @dotnet/jit-contrib

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

[EgorBo]

Works if I move + 1 to the right side:

 int foo(bool cond, int a, int b) => (cond ? a : b) + 1;

; Method Program:foo(bool,int,int):int:this (FullOpts)
 test dl, dl
 cmove r8d, r9d
 mov eax, r8d
 inc eax
 ret 
; Total bytes of code: 12

[stephentoub]

Likely a separate issue, but as we discussed offline, it'd also be nice if:

int foo(bool cond, int a, int b) => (cond ? 1 : 2) + 1;

could become the equivalent of:

G_M000_IG02:
 mov eax, 2
 mov edx, 3
 test cl, cl
 cmove eax, edx

G_M000_IG03:
 ret 

instead of:

G_M000_IG02:
 mov eax, 1
 mov edx, 2
 test cl, cl
 cmove eax, edx
 inc eax

G_M000_IG03:
 ret

[jakobbotsch]

This is presumably the same underlying issue as #87072, where Roslyn leaves things on the IL stack across control flow. I'm not at my PC, but I would guess that in this case that introduces additional stores in each basic block that disqualifies it from if conversion.

If that's right then we can likely teach if conversion about some of these cases, but it might fit better as an additional opportunity handled by tail merging.

[jakobbotsch]

Likely a separate issue, but as we discussed offline, it'd also be nice if:

int foo(bool cond, int a, int b) => (cond ? 1 : 2) + 1;

could become the equivalent of:

G_M000_IG02:
 mov eax, 2
 mov edx, 3
 test cl, cl
 cmove eax, edx

G_M000_IG03:
 ret 

instead of:

G_M000_IG02:
 mov eax, 1
 mov edx, 2
 test cl, cl
 cmove eax, edx
 inc eax

G_M000_IG03:
 ret

It might be worth it to invoke forward sub on the new statements created by if-conversion, and then run a folding pass over potentially resulting combined trees. For example, it would optimize ADD(SELECT(cond, cns1, cns2), cns3) into SELECT(cond, cns1 + cns3, cns2 + cns3). That would be an easy way to handle some of the simple cases.

It would require a bit of work to make forward sub work with the node threading we have at the point of if-conversion.