llvm dev - Jan 2020 - Disabling select instructions

Hi,

I would like to know if there's a way to avoid select instructions during
the IR generation. What are the optimization passes that can result in a
select instruction?
i.e. I want to preserve branches in my code without disabling any other
optimizations applicable.

For example,
void foo(int* x, int* y){
    if(*x > 0){
        *y = *x + 10;
    }
    else{
        *y = *x + 20;
    }
}

with O1 I get,

define void @foo(i32* nocapture readonly, i32* nocapture)
local_unnamed_addr #0 {
  %3 = load i32, i32* %0, align 4, !tbaa !2
  %4 = icmp sgt i32 %3, 0
  %5 = select i1 %4, i32 10, i32 20
  %6 = add nsw i32 %5, %3
  store i32 %6, i32* %1, align 4, !tbaa !2
  ret void
}

But with O0,

define void @foo(i32*, i32*) #0 {
  %3 = alloca i32*, align 8
  %4 = alloca i32*, align 8
  store i32* %0, i32** %3, align 8
  store i32* %1, i32** %4, align 8
  %5 = load i32*, i32** %3, align 8
  %6 = load i32, i32* %5, align 4
  %7 = icmp sgt i32 %6, 0
  br i1 %7, label %8, label %13

; <label>:8:                                      ; preds = %2
  %9 = load i32*, i32** %3, align 8
  %10 = load i32, i32* %9, align 4
  %11 = add nsw i32 %10, 10
  %12 = load i32*, i32** %4, align 8
  store i32 %11, i32* %12, align 4
  br label %18

; <label>:13:                                     ; preds = %2
  %14 = load i32*, i32** %3, align 8
  %15 = load i32, i32* %14, align 4
  %16 = add nsw i32 %15, 20
  %17 = load i32*, i32** %4, align 8
  store i32 %16, i32* %17, align 4
  br label %18

; <label>:18:                                     ; preds = %13, %8
  ret void
}

Thanks,
Charitha
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20200130/fc78bdc5/attachment.html>

Several different passes introduce select instructions, such as 
InstCombine, SimplifyCFG, and SROA. Searching for calls to 
"CreateSelect" in the source tree should give you a fairly complete
list.

It may be better to let the optimization pipeline proceed undisturbed, 
but then run a custom pass at the end that turns selects back into 
control flow. This would be an easy pass to write.

John


On 1/30/20 3:31 PM, Charitha Saumya via llvm-dev wrote:
> Hi,
> 
> I would like to know if there's a way to avoid select instructions 
> during the IR generation. What are the optimization passes that can 
> result in a select instruction?
> i.e. I want to preserve branches in my code without disabling any other 
> optimizations applicable.
> 
> For example,
> void foo(int* x, int* y){
>      if(*x > 0){
>          *y = *x + 10;
>      }
>      else{
>          *y = *x + 20;
>      }
> }
> 
> with O1 I get,
> 
> define void @foo(i32* nocapture readonly, i32* nocapture) 
> local_unnamed_addr #0 {
>    %3 = load i32, i32* %0, align 4, !tbaa !2
>    %4 = icmp sgt i32 %3, 0
>    %5 = select i1 %4, i32 10, i32 20
>    %6 = add nsw i32 %5, %3
>    store i32 %6, i32* %1, align 4, !tbaa !2
>    ret void
> }
> 
> But with O0,
> 
> define void @foo(i32*, i32*) #0 {
>    %3 = alloca i32*, align 8
>    %4 = alloca i32*, align 8
>    store i32* %0, i32** %3, align 8
>    store i32* %1, i32** %4, align 8
>    %5 = load i32*, i32** %3, align 8
>    %6 = load i32, i32* %5, align 4
>    %7 = icmp sgt i32 %6, 0
>    br i1 %7, label %8, label %13
> 
> ; <label>:8:                                      ; preds = %2
>    %9 = load i32*, i32** %3, align 8
>    %10 = load i32, i32* %9, align 4
>    %11 = add nsw i32 %10, 10
>    %12 = load i32*, i32** %4, align 8
>    store i32 %11, i32* %12, align 4
>    br label %18
> 
> ; <label>:13:                                     ; preds = %2
>    %14 = load i32*, i32** %3, align 8
>    %15 = load i32, i32* %14, align 4
>    %16 = add nsw i32 %15, 20
>    %17 = load i32*, i32** %4, align 8
>    store i32 %16, i32* %17, align 4
>    br label %18
> 
> ; <label>:18:                                     ; preds = %13, %8
>    ret void
> }
> 
> Thanks,
> Charitha
> 
> 
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>

I agree with John; also, if you decide to go this route, you can reuse the
code from CodeGenPrepare::optimizeSelectInst:

https://github.com/llvm/llvm-project/blob/master/llvm/lib/CodeGen/CodeGenPrepare.cpp#L6065

Alexey

On Thu, Jan 30, 2020 at 9:00 PM John Regehr via llvm-dev <
llvm-dev at lists.llvm.org> wrote:
> Several different passes introduce select instructions, such as
> InstCombine, SimplifyCFG, and SROA. Searching for calls to
> "CreateSelect" in the source tree should give you a fairly
complete list.
>
> It may be better to let the optimization pipeline proceed undisturbed,
> but then run a custom pass at the end that turns selects back into
> control flow. This would be an easy pass to write.
>
> John
>
>
> On 1/30/20 3:31 PM, Charitha Saumya via llvm-dev wrote:
> > Hi,
> >
> > I would like to know if there's a way to avoid select instructions
> > during the IR generation. What are the optimization passes that can
> > result in a select instruction?
> > i.e. I want to preserve branches in my code without disabling any
other
> > optimizations applicable.
> >
> > For example,
> > void foo(int* x, int* y){
> >      if(*x > 0){
> >          *y = *x + 10;
> >      }
> >      else{
> >          *y = *x + 20;
> >      }
> > }
> >
> > with O1 I get,
> >
> > define void @foo(i32* nocapture readonly, i32* nocapture)
> > local_unnamed_addr #0 {
> >    %3 = load i32, i32* %0, align 4, !tbaa !2
> >    %4 = icmp sgt i32 %3, 0
> >    %5 = select i1 %4, i32 10, i32 20
> >    %6 = add nsw i32 %5, %3
> >    store i32 %6, i32* %1, align 4, !tbaa !2
> >    ret void
> > }
> >
> > But with O0,
> >
> > define void @foo(i32*, i32*) #0 {
> >    %3 = alloca i32*, align 8
> >    %4 = alloca i32*, align 8
> >    store i32* %0, i32** %3, align 8
> >    store i32* %1, i32** %4, align 8
> >    %5 = load i32*, i32** %3, align 8
> >    %6 = load i32, i32* %5, align 4
> >    %7 = icmp sgt i32 %6, 0
> >    br i1 %7, label %8, label %13
> >
> > ; <label>:8:                                      ; preds = %2
> >    %9 = load i32*, i32** %3, align 8
> >    %10 = load i32, i32* %9, align 4
> >    %11 = add nsw i32 %10, 10
> >    %12 = load i32*, i32** %4, align 8
> >    store i32 %11, i32* %12, align 4
> >    br label %18
> >
> > ; <label>:13:                                     ; preds = %2
> >    %14 = load i32*, i32** %3, align 8
> >    %15 = load i32, i32* %14, align 4
> >    %16 = add nsw i32 %15, 20
> >    %17 = load i32*, i32** %4, align 8
> >    store i32 %16, i32* %17, align 4
> >    br label %18
> >
> > ; <label>:18:                                     ; preds = %13,
%8
> >    ret void
> > }
> >
> > Thanks,
> > Charitha
> >
> >
> > _______________________________________________
> > LLVM Developers mailing list
> > llvm-dev at lists.llvm.org
> > https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
> >
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20200131/86769fc1/attachment.html>