JIT: missing conditional branch merging across multiple BBs

[Sergio0694]

Description

RyuJit currently misses some optimization opportunities when there are multiple BBs with conditional branches that could be merged, and instead result in avoidable conditional branches in the final codegen. Consider this example:

static unsafe bool Test(int* a, int* b)
{
    return a < b || a == b;
}

This results in the following codegen:

L0000: cmp ecx, edx
L0002: jb short L000d
L0004: cmp ecx, edx
L0006: sete al
L0009: movzx eax, al
L000c: ret
L000d: mov eax, 1
L0012: ret

Here the two BBs with the LT and EQ tests could just be combined in a single LE one, which could result in eg. a single seta in the final codegen here, or in a single jle if the method was inlined and the caller was to do a conditional branch on its return.

Courtesy of @EgorBo for these trees, we basically need the following optimization:

*  JTRUE   
\--*  LT  
   +--*  LCL_VAR  arg0         
    \--*  LCL_VAR  arg1         

*  JTRUE or RETURN   
\--*  EQ           
   +--*  LCL_VAR  arg0         
    \--*  LCL_VAR  arg1

to

*  JTRUE   
\--*  LE  
   +--*  LCL_VAR  arg0         
    \--*  LCL_VAR  arg1

And the same for all possible conditional combinations as well (eg. GT+EQ to GE, etc.).

Configuration

Tested on .NET 5 stable, and also on .NET 6 (daily builds, with Disasmo).

Regression?

Nope.

category:correctness
theme:ir
skill-level:intermediate
cost:small
impact:small

[ghost]

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

Issue Details

---

Description

RyuJit currently misses some optimization opportunities when there are multiple BBs with conditional branches that could be merged, and instead result in avoidable conditional branches in the final codegen. Consider this example:

static unsafe bool Test(int* a, int* b)
{
    return a < b || a == b;
}

This results in the following codegen:

L0000: cmp ecx, edx
L0002: jb short L000d
L0004: cmp ecx, edx
L0006: sete al
L0009: movzx eax, al
L000c: ret
L000d: mov eax, 1
L0012: ret

Here the two BBs with the LT and EQ tests could just be combined in a single LE one, which could result in eg. a single seta in the final codegen here, or in a single jle if the method was inlined and the caller was to do a conditional branch on its return.

Courtesy of @EgorBo for these trees, we basically need the following optimization:

*  JTRUE   
\--*  LT  
   +--*  LCL_VAR  arg0         
    \--*  LCL_VAR  arg1         

*  JTRUE or RETURN   
\--*  EQ           
   +--*  LCL_VAR  arg0         
    \--*  LCL_VAR  arg1

to

*  JTRUE   
\--*  LE  
   +--*  LCL_VAR  arg0         
    \--*  LCL_VAR  arg1

And the same for all possible conditional combinations as well (eg. GT+EQ to GE, etc.).

Configuration

Tested on .NET 5 stable, and also on .NET 6 (daily builds, with Disasmo).

Regression?

Nope.

Author:	Sergio0694
Assignees:	-
Labels:	tenet-performance, area-CodeGen-coreclr, untriaged
Milestone:	-

[EgorBo]

From my understanding it can be done in optOptimizeBools without introducing additional loops over BBs.