thr3ads.net - llvm dev - [llvm-dev] simplifycfg not happening? [Sep 2015]

If this information is useful, please help other people find it:
Share via:

Russell Wallace via llvm-dev

2015-Sep-20 04:03 UTC

[llvm-dev] simplifycfg not happening?

The following function compiles with -O3 into the following IR.
http://llvm.org/docs/Passes.html#simplifycfg-simplify-the-cfg says

   - Eliminates a basic block that only contains an unconditional branch.

but the first and third blocks in the compiled function only contain an
unconditional branch; I would have expected them to be eliminated. What am
I missing?

double f(double *a) {
  for (int i = 0; i < 1000; i++)
    a[i] *= 2;
  for (int i = 0; i < 1000; i++)
    a[i] *= 2;
  return a[0] + a[1];
}

; Function Attrs: nounwind uwtable
define double @"\01?f@@YANPEAN at Z"(double* nocapture %a) #1 {
overflow.checked:
  br label %vector.body, !dbg !18

vector.body:                                      ; preds = %vector.body,
%overflow.checked
  %index = phi i64 [ 0, %overflow.checked ], [ %index.next.1, %vector.body
], !dbg !18
  %0 = getelementptr inbounds double, double* %a, i64 %index, !dbg !19
  %1 = bitcast double* %0 to <2 x double>*, !dbg !20
  %wide.load = load <2 x double>, <2 x double>* %1, align 8, !dbg
!20
  %2 = getelementptr double, double* %0, i64 2, !dbg !20
  %3 = bitcast double* %2 to <2 x double>*, !dbg !20
  %wide.load8 = load <2 x double>, <2 x double>* %3, align 8, !dbg
!20
  %4 = fmul <2 x double> %wide.load, <double 2.000000e+00, double
2.000000e+00>, !dbg !20
  %5 = fmul <2 x double> %wide.load8, <double 2.000000e+00, double
2.000000e+00>, !dbg !20
  %6 = bitcast double* %0 to <2 x double>*, !dbg !20
  store <2 x double> %4, <2 x double>* %6, align 8, !dbg !20
  %7 = bitcast double* %2 to <2 x double>*, !dbg !20
  store <2 x double> %5, <2 x double>* %7, align 8, !dbg !20
  %index.next = or i64 %index, 4, !dbg !18
  %8 = getelementptr inbounds double, double* %a, i64 %index.next, !dbg !19
  %9 = bitcast double* %8 to <2 x double>*, !dbg !20
  %wide.load.1 = load <2 x double>, <2 x double>* %9, align 8, !dbg
!20
  %10 = getelementptr double, double* %8, i64 2, !dbg !20
  %11 = bitcast double* %10 to <2 x double>*, !dbg !20
  %wide.load8.1 = load <2 x double>, <2 x double>* %11, align 8,
!dbg !20
  %12 = fmul <2 x double> %wide.load.1, <double 2.000000e+00, double
2.000000e+00>, !dbg !20
  %13 = fmul <2 x double> %wide.load8.1, <double 2.000000e+00, double
2.000000e+00>, !dbg !20
  %14 = bitcast double* %8 to <2 x double>*, !dbg !20
  store <2 x double> %12, <2 x double>* %14, align 8, !dbg !20
  %15 = bitcast double* %10 to <2 x double>*, !dbg !20
  store <2 x double> %13, <2 x double>* %15, align 8, !dbg !20
  %index.next.1 = add nsw i64 %index, 8, !dbg !18
  %16 = icmp eq i64 %index.next.1, 1000, !dbg !18
  br i1 %16, label %vector.body10.preheader, label %vector.body, !dbg !18,
!llvm.loop !21

vector.body10.preheader:                          ; preds = %vector.body
  br label %vector.body10, !dbg !24

vector.body10:                                    ; preds = %vector.body10,
%vector.body10.preheader
  %index13 = phi i64 [ 0, %vector.body10.preheader ], [ %index.next21.1,
%vector.body10 ], !dbg !25
  %17 = getelementptr inbounds double, double* %a, i64 %index13, !dbg !24
  %18 = bitcast double* %17 to <2 x double>*, !dbg !26
  %wide.load26 = load <2 x double>, <2 x double>* %18, align 8, !dbg
!26
  %19 = getelementptr double, double* %17, i64 2, !dbg !26
  %20 = bitcast double* %19 to <2 x double>*, !dbg !26
  %wide.load27 = load <2 x double>, <2 x double>* %20, align 8, !dbg
!26
  %21 = fmul <2 x double> %wide.load26, <double 2.000000e+00, double
2.000000e+00>, !dbg !26
  %22 = fmul <2 x double> %wide.load27, <double 2.000000e+00, double
2.000000e+00>, !dbg !26
  %23 = bitcast double* %17 to <2 x double>*, !dbg !26
  store <2 x double> %21, <2 x double>* %23, align 8, !dbg !26
  %24 = bitcast double* %19 to <2 x double>*, !dbg !26
  store <2 x double> %22, <2 x double>* %24, align 8, !dbg !26
  %index.next21 = or i64 %index13, 4, !dbg !25
  %25 = getelementptr inbounds double, double* %a, i64 %index.next21, !dbg
!24
  %26 = bitcast double* %25 to <2 x double>*, !dbg !26
  %wide.load26.1 = load <2 x double>, <2 x double>* %26, align 8,
!dbg !26
  %27 = getelementptr double, double* %25, i64 2, !dbg !26
  %28 = bitcast double* %27 to <2 x double>*, !dbg !26
  %wide.load27.1 = load <2 x double>, <2 x double>* %28, align 8,
!dbg !26
  %29 = fmul <2 x double> %wide.load26.1, <double 2.000000e+00, double
2.000000e+00>, !dbg !26
  %30 = fmul <2 x double> %wide.load27.1, <double 2.000000e+00, double
2.000000e+00>, !dbg !26
  %31 = bitcast double* %25 to <2 x double>*, !dbg !26
  store <2 x double> %29, <2 x double>* %31, align 8, !dbg !26
  %32 = bitcast double* %27 to <2 x double>*, !dbg !26
  store <2 x double> %30, <2 x double>* %32, align 8, !dbg !26
  %index.next21.1 = add nsw i64 %index13, 8, !dbg !25
  %33 = icmp eq i64 %index.next21.1, 1000, !dbg !25
  br i1 %33, label %middle.block11, label %vector.body10, !dbg !25,
!llvm.loop !27

middle.block11:                                   ; preds = %vector.body10
  %34 = load double, double* %a, align 8, !dbg !28
  %35 = getelementptr inbounds double, double* %a, i64 1, !dbg !29
  %36 = load double, double* %35, align 8, !dbg !29
  %37 = fadd double %34, %36, !dbg !30
  ret double %37, !dbg !31
}
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20150920/4ada174f/attachment.html>

Xiangyang Guo via llvm-dev

2015-Sep-20 04:27 UTC

head link

[llvm-dev] simplifycfg not happening?

Hi,

if you use opt -simplifycfg, the third BB can be eliminated.

2015-09-20 0:03 GMT-04:00 Russell Wallace via llvm-dev <
llvm-dev at lists.llvm.org>:
> The following function compiles with -O3 into the following IR.
> http://llvm.org/docs/Passes.html#simplifycfg-simplify-the-cfg says
>
>    - Eliminates a basic block that only contains an unconditional branch.
>
> but the first and third blocks in the compiled function only contain an
> unconditional branch; I would have expected them to be eliminated. What am
> I missing?
>
> double f(double *a) {
>   for (int i = 0; i < 1000; i++)
>     a[i] *= 2;
>   for (int i = 0; i < 1000; i++)
>     a[i] *= 2;
>   return a[0] + a[1];
> }
>
> ; Function Attrs: nounwind uwtable
> define double @"\01?f@@YANPEAN at Z"(double* nocapture %a) #1 {
> overflow.checked:
>   br label %vector.body, !dbg !18
>
> vector.body:                                      ; preds = %vector.body,
> %overflow.checked
>   %index = phi i64 [ 0, %overflow.checked ], [ %index.next.1, %vector.body
> ], !dbg !18
>   %0 = getelementptr inbounds double, double* %a, i64 %index, !dbg !19
>   %1 = bitcast double* %0 to <2 x double>*, !dbg !20
>   %wide.load = load <2 x double>, <2 x double>* %1, align 8,
!dbg !20
>   %2 = getelementptr double, double* %0, i64 2, !dbg !20
>   %3 = bitcast double* %2 to <2 x double>*, !dbg !20
>   %wide.load8 = load <2 x double>, <2 x double>* %3, align 8,
!dbg !20
>   %4 = fmul <2 x double> %wide.load, <double 2.000000e+00, double
> 2.000000e+00>, !dbg !20
>   %5 = fmul <2 x double> %wide.load8, <double 2.000000e+00, double
> 2.000000e+00>, !dbg !20
>   %6 = bitcast double* %0 to <2 x double>*, !dbg !20
>   store <2 x double> %4, <2 x double>* %6, align 8, !dbg !20
>   %7 = bitcast double* %2 to <2 x double>*, !dbg !20
>   store <2 x double> %5, <2 x double>* %7, align 8, !dbg !20
>   %index.next = or i64 %index, 4, !dbg !18
>   %8 = getelementptr inbounds double, double* %a, i64 %index.next, !dbg !19
>   %9 = bitcast double* %8 to <2 x double>*, !dbg !20
>   %wide.load.1 = load <2 x double>, <2 x double>* %9, align 8,
!dbg !20
>   %10 = getelementptr double, double* %8, i64 2, !dbg !20
>   %11 = bitcast double* %10 to <2 x double>*, !dbg !20
>   %wide.load8.1 = load <2 x double>, <2 x double>* %11, align
8, !dbg !20
>   %12 = fmul <2 x double> %wide.load.1, <double 2.000000e+00,
double
> 2.000000e+00>, !dbg !20
>   %13 = fmul <2 x double> %wide.load8.1, <double 2.000000e+00,
double
> 2.000000e+00>, !dbg !20
>   %14 = bitcast double* %8 to <2 x double>*, !dbg !20
>   store <2 x double> %12, <2 x double>* %14, align 8, !dbg !20
>   %15 = bitcast double* %10 to <2 x double>*, !dbg !20
>   store <2 x double> %13, <2 x double>* %15, align 8, !dbg !20
>   %index.next.1 = add nsw i64 %index, 8, !dbg !18
>   %16 = icmp eq i64 %index.next.1, 1000, !dbg !18
>   br i1 %16, label %vector.body10.preheader, label %vector.body, !dbg !18,
> !llvm.loop !21
>
> vector.body10.preheader:                          ; preds = %vector.body
>   br label %vector.body10, !dbg !24
>
> vector.body10:                                    ; preds >
%vector.body10, %vector.body10.preheader
>   %index13 = phi i64 [ 0, %vector.body10.preheader ], [ %index.next21.1,
> %vector.body10 ], !dbg !25
>   %17 = getelementptr inbounds double, double* %a, i64 %index13, !dbg !24
>   %18 = bitcast double* %17 to <2 x double>*, !dbg !26
>   %wide.load26 = load <2 x double>, <2 x double>* %18, align 8,
!dbg !26
>   %19 = getelementptr double, double* %17, i64 2, !dbg !26
>   %20 = bitcast double* %19 to <2 x double>*, !dbg !26
>   %wide.load27 = load <2 x double>, <2 x double>* %20, align 8,
!dbg !26
>   %21 = fmul <2 x double> %wide.load26, <double 2.000000e+00,
double
> 2.000000e+00>, !dbg !26
>   %22 = fmul <2 x double> %wide.load27, <double 2.000000e+00,
double
> 2.000000e+00>, !dbg !26
>   %23 = bitcast double* %17 to <2 x double>*, !dbg !26
>   store <2 x double> %21, <2 x double>* %23, align 8, !dbg !26
>   %24 = bitcast double* %19 to <2 x double>*, !dbg !26
>   store <2 x double> %22, <2 x double>* %24, align 8, !dbg !26
>   %index.next21 = or i64 %index13, 4, !dbg !25
>   %25 = getelementptr inbounds double, double* %a, i64 %index.next21, !dbg
> !24
>   %26 = bitcast double* %25 to <2 x double>*, !dbg !26
>   %wide.load26.1 = load <2 x double>, <2 x double>* %26, align
8, !dbg !26
>   %27 = getelementptr double, double* %25, i64 2, !dbg !26
>   %28 = bitcast double* %27 to <2 x double>*, !dbg !26
>   %wide.load27.1 = load <2 x double>, <2 x double>* %28, align
8, !dbg !26
>   %29 = fmul <2 x double> %wide.load26.1, <double 2.000000e+00,
double
> 2.000000e+00>, !dbg !26
>   %30 = fmul <2 x double> %wide.load27.1, <double 2.000000e+00,
double
> 2.000000e+00>, !dbg !26
>   %31 = bitcast double* %25 to <2 x double>*, !dbg !26
>   store <2 x double> %29, <2 x double>* %31, align 8, !dbg !26
>   %32 = bitcast double* %27 to <2 x double>*, !dbg !26
>   store <2 x double> %30, <2 x double>* %32, align 8, !dbg !26
>   %index.next21.1 = add nsw i64 %index13, 8, !dbg !25
>   %33 = icmp eq i64 %index.next21.1, 1000, !dbg !25
>   br i1 %33, label %middle.block11, label %vector.body10, !dbg !25,
> !llvm.loop !27
>
> middle.block11:                                   ; preds = %vector.body10
>   %34 = load double, double* %a, align 8, !dbg !28
>   %35 = getelementptr inbounds double, double* %a, i64 1, !dbg !29
>   %36 = load double, double* %35, align 8, !dbg !29
>   %37 = fadd double %34, %36, !dbg !30
>   ret double %37, !dbg !31
> }
>
>
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://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/20150920/c1fada28/attachment.html>

Russell Wallace via llvm-dev

2015-Sep-20 04:35 UTC

head link

[llvm-dev] simplifycfg not happening?

You're right, it can indeed.

Is there a reason -O3 doesn't do this? I had been expecting -O3 to perform
full optimization.

The first block still remains in any case. Is the first block needed for
some purpose I'm not taking into account?

On Sun, Sep 20, 2015 at 5:27 AM, Xiangyang Guo <eceguo at gmail.com>
wrote:
> Hi,
>
> if you use opt -simplifycfg, the third BB can be eliminated.
>
> 2015-09-20 0:03 GMT-04:00 Russell Wallace via llvm-dev <
> llvm-dev at lists.llvm.org>:
>
>> The following function compiles with -O3 into the following IR.
>> http://llvm.org/docs/Passes.html#simplifycfg-simplify-the-cfg says
>>
>>    - Eliminates a basic block that only contains an unconditional
branch.
>>
>> but the first and third blocks in the compiled function only contain an
>> unconditional branch; I would have expected them to be eliminated. What
am
>> I missing?
>>
>> double f(double *a) {
>>   for (int i = 0; i < 1000; i++)
>>     a[i] *= 2;
>>   for (int i = 0; i < 1000; i++)
>>     a[i] *= 2;
>>   return a[0] + a[1];
>> }
>>
>> ; Function Attrs: nounwind uwtable
>> define double @"\01?f@@YANPEAN at Z"(double* nocapture %a) #1
{
>> overflow.checked:
>>   br label %vector.body, !dbg !18
>>
>> vector.body:                                      ; preds =
%vector.body,
>> %overflow.checked
>>   %index = phi i64 [ 0, %overflow.checked ], [ %index.next.1,
>> %vector.body ], !dbg !18
>>   %0 = getelementptr inbounds double, double* %a, i64 %index, !dbg !19
>>   %1 = bitcast double* %0 to <2 x double>*, !dbg !20
>>   %wide.load = load <2 x double>, <2 x double>* %1, align
8, !dbg !20
>>   %2 = getelementptr double, double* %0, i64 2, !dbg !20
>>   %3 = bitcast double* %2 to <2 x double>*, !dbg !20
>>   %wide.load8 = load <2 x double>, <2 x double>* %3, align
8, !dbg !20
>>   %4 = fmul <2 x double> %wide.load, <double 2.000000e+00,
double
>> 2.000000e+00>, !dbg !20
>>   %5 = fmul <2 x double> %wide.load8, <double 2.000000e+00,
double
>> 2.000000e+00>, !dbg !20
>>   %6 = bitcast double* %0 to <2 x double>*, !dbg !20
>>   store <2 x double> %4, <2 x double>* %6, align 8, !dbg
!20
>>   %7 = bitcast double* %2 to <2 x double>*, !dbg !20
>>   store <2 x double> %5, <2 x double>* %7, align 8, !dbg
!20
>>   %index.next = or i64 %index, 4, !dbg !18
>>   %8 = getelementptr inbounds double, double* %a, i64 %index.next, !dbg
>> !19
>>   %9 = bitcast double* %8 to <2 x double>*, !dbg !20
>>   %wide.load.1 = load <2 x double>, <2 x double>* %9, align
8, !dbg !20
>>   %10 = getelementptr double, double* %8, i64 2, !dbg !20
>>   %11 = bitcast double* %10 to <2 x double>*, !dbg !20
>>   %wide.load8.1 = load <2 x double>, <2 x double>* %11,
align 8, !dbg !20
>>   %12 = fmul <2 x double> %wide.load.1, <double 2.000000e+00,
double
>> 2.000000e+00>, !dbg !20
>>   %13 = fmul <2 x double> %wide.load8.1, <double 2.000000e+00,
double
>> 2.000000e+00>, !dbg !20
>>   %14 = bitcast double* %8 to <2 x double>*, !dbg !20
>>   store <2 x double> %12, <2 x double>* %14, align 8, !dbg
!20
>>   %15 = bitcast double* %10 to <2 x double>*, !dbg !20
>>   store <2 x double> %13, <2 x double>* %15, align 8, !dbg
!20
>>   %index.next.1 = add nsw i64 %index, 8, !dbg !18
>>   %16 = icmp eq i64 %index.next.1, 1000, !dbg !18
>>   br i1 %16, label %vector.body10.preheader, label %vector.body, !dbg
>> !18, !llvm.loop !21
>>
>> vector.body10.preheader:                          ; preds =
%vector.body
>>   br label %vector.body10, !dbg !24
>>
>> vector.body10:                                    ; preds >>
%vector.body10, %vector.body10.preheader
>>   %index13 = phi i64 [ 0, %vector.body10.preheader ], [
%index.next21.1,
>> %vector.body10 ], !dbg !25
>>   %17 = getelementptr inbounds double, double* %a, i64 %index13, !dbg
!24
>>   %18 = bitcast double* %17 to <2 x double>*, !dbg !26
>>   %wide.load26 = load <2 x double>, <2 x double>* %18,
align 8, !dbg !26
>>   %19 = getelementptr double, double* %17, i64 2, !dbg !26
>>   %20 = bitcast double* %19 to <2 x double>*, !dbg !26
>>   %wide.load27 = load <2 x double>, <2 x double>* %20,
align 8, !dbg !26
>>   %21 = fmul <2 x double> %wide.load26, <double 2.000000e+00,
double
>> 2.000000e+00>, !dbg !26
>>   %22 = fmul <2 x double> %wide.load27, <double 2.000000e+00,
double
>> 2.000000e+00>, !dbg !26
>>   %23 = bitcast double* %17 to <2 x double>*, !dbg !26
>>   store <2 x double> %21, <2 x double>* %23, align 8, !dbg
!26
>>   %24 = bitcast double* %19 to <2 x double>*, !dbg !26
>>   store <2 x double> %22, <2 x double>* %24, align 8, !dbg
!26
>>   %index.next21 = or i64 %index13, 4, !dbg !25
>>   %25 = getelementptr inbounds double, double* %a, i64 %index.next21,
>> !dbg !24
>>   %26 = bitcast double* %25 to <2 x double>*, !dbg !26
>>   %wide.load26.1 = load <2 x double>, <2 x double>* %26,
align 8, !dbg !26
>>   %27 = getelementptr double, double* %25, i64 2, !dbg !26
>>   %28 = bitcast double* %27 to <2 x double>*, !dbg !26
>>   %wide.load27.1 = load <2 x double>, <2 x double>* %28,
align 8, !dbg !26
>>   %29 = fmul <2 x double> %wide.load26.1, <double
2.000000e+00, double
>> 2.000000e+00>, !dbg !26
>>   %30 = fmul <2 x double> %wide.load27.1, <double
2.000000e+00, double
>> 2.000000e+00>, !dbg !26
>>   %31 = bitcast double* %25 to <2 x double>*, !dbg !26
>>   store <2 x double> %29, <2 x double>* %31, align 8, !dbg
!26
>>   %32 = bitcast double* %27 to <2 x double>*, !dbg !26
>>   store <2 x double> %30, <2 x double>* %32, align 8, !dbg
!26
>>   %index.next21.1 = add nsw i64 %index13, 8, !dbg !25
>>   %33 = icmp eq i64 %index.next21.1, 1000, !dbg !25
>>   br i1 %33, label %middle.block11, label %vector.body10, !dbg !25,
>> !llvm.loop !27
>>
>> middle.block11:                                   ; preds =
%vector.body10
>>   %34 = load double, double* %a, align 8, !dbg !28
>>   %35 = getelementptr inbounds double, double* %a, i64 1, !dbg !29
>>   %36 = load double, double* %35, align 8, !dbg !29
>>   %37 = fadd double %34, %36, !dbg !30
>>   ret double %37, !dbg !31
>> }
>>
>>
>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
>> http://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/20150920/94971409/attachment.html>

Apparently Analagous Threads

Search for more possibly parallel threads

llvm dev - Sep 2015 - simplifycfg not happening?

[llvm-dev] simplifycfg not happening?

[llvm-dev] simplifycfg not happening?

[llvm-dev] simplifycfg not happening?

Apparently Analagous Threads