This is an archive of the discontinued LLVM Phabricator instance.

Differential D34078

Enable support for floating-point division reductions
ClosedPublic

Authored by yuyichao on Jun 10 2017, 5:52 AM.

Details

Reviewers
mkuper
jmolloy
Commits
rG4bc88a0e9a2e: Enable support for floating-point division reductions
Summary

Similar to fsub, fdiv can also be vectorized using fmul.

Also http://llvm.org/viewvc/llvm-project?view=revision&revision=215200

Diff Detail

Event Timeline

yuyichao created this revision.Jun 10 2017, 5:52 AM
yuyichao added a comment.Jun 10 2017, 5:55 AM

Oh and this fixes https://bugs.llvm.org/show_bug.cgi?id=33345

yuyichao added a comment.Jun 15 2017, 11:36 AM

Ping?

mkuper edited reviewers, added: jmolloy; removed: llvm-commits.Jun 19 2017, 5:38 PM
mkuper added a subscriber: llvm-commits.
delena added a subscriber: delena.Jun 22 2017, 12:24 AM

I suggest to add one more test that shows that vectorization does not happen when the reduction variable is the second argument of fdiv.

vchuravy added a subscriber: vchuravy.Mar 25 2018, 1:29 PM
vchuravy added a project: Restricted Project.Jan 14 2020, 1:29 AM
vtjnash updated this revision to Diff 303518.Nov 6 2020, 11:58 AM
vtjnash edited the summary of this revision. (Show Details)
vtjnash added a subscriber: vtjnash.

add null test as requested, rebase

Herald added a project: Restricted Project. · View Herald TranscriptNov 6 2020, 11:58 AM
Herald added a subscriber: hiraditya. · View Herald Transcript
Harbormaster completed remote builds in B77921: Diff 303518.Nov 6 2020, 12:44 PM
This revision was not accepted when it landed; it landed in state Needs Review.Nov 23 2020, 5:06 PM
This revision was landed with ongoing or failed builds.
Closed by commit rG4bc88a0e9a2e: Enable support for floating-point division reductions (authored by yuyichao, committed by vchuravy). · Explain Why
This revision was automatically updated to reflect the committed changes.
vchuravy added a commit: rG4bc88a0e9a2e: Enable support for floating-point division reductions.

Revision Contents

PathSize
llvm/
lib/
Analysis/
1 line
test/
Transforms/
LoopVectorize/
44 lines

Diff 307219

llvm/lib/Analysis/IVDescriptors.cpp

Show First 20 Lines • Show All 572 Lines • ▼ Show 20 LinesRecurrenceDescriptor::isRecurrenceInstr(Instruction *I, RecurrenceKind Kind,
case Instruction::Mul: case Instruction::Mul:
return InstDesc(Kind == RK_IntegerMult, I); return InstDesc(Kind == RK_IntegerMult, I);
case Instruction::And: case Instruction::And:
return InstDesc(Kind == RK_IntegerAnd, I); return InstDesc(Kind == RK_IntegerAnd, I);
case Instruction::Or: case Instruction::Or:
return InstDesc(Kind == RK_IntegerOr, I); return InstDesc(Kind == RK_IntegerOr, I);
case Instruction::Xor: case Instruction::Xor:
return InstDesc(Kind == RK_IntegerXor, I); return InstDesc(Kind == RK_IntegerXor, I);
case Instruction::FDiv:
case Instruction::FMul: case Instruction::FMul:
return InstDesc(Kind == RK_FloatMult, I, UAI); return InstDesc(Kind == RK_FloatMult, I, UAI);
case Instruction::FSub: case Instruction::FSub:
case Instruction::FAdd: case Instruction::FAdd:
return InstDesc(Kind == RK_FloatAdd, I, UAI); return InstDesc(Kind == RK_FloatAdd, I, UAI);
case Instruction::Select: case Instruction::Select:
if (Kind == RK_FloatAdd || Kind == RK_FloatMult) if (Kind == RK_FloatAdd || Kind == RK_FloatMult)
return isConditionalRdxPattern(Kind, I); return isConditionalRdxPattern(Kind, I);
▲ Show 20 Lines • Show All 629 Lines • Show Last 20 Lines

llvm/test/Transforms/LoopVectorize/float-reduction.ll

Show All 38 Linesfor.body: ; preds = %for.body, %entry
%indvars.iv.next = add i64 %indvars.iv, 1 %indvars.iv.next = add i64 %indvars.iv, 1
%lftr.wideiv = trunc i64 %indvars.iv.next to i32 %lftr.wideiv = trunc i64 %indvars.iv.next to i32
%exitcond = icmp eq i32 %lftr.wideiv, 200 %exitcond = icmp eq i32 %lftr.wideiv, 200
br i1 %exitcond, label %for.end, label %for.body br i1 %exitcond, label %for.end, label %for.body
for.end: ; preds = %for.body for.end: ; preds = %for.body
ret float %sub ret float %sub
} }
;CHECK-LABEL: @foodiv(
;CHECK: fdiv fast <4 x float>
;CHECK: ret
define float @foodiv(float* nocapture %A, i32* nocapture %n) nounwind uwtable readonly ssp {
entry:
br label %for.body
for.body: ; preds = %for.body, %entry
%indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
%sum.04 = phi float [ 1.000000e+00, %entry ], [ %sub, %for.body ]
%arrayidx = getelementptr inbounds float, float* %A, i64 %indvars.iv
%0 = load float, float* %arrayidx, align 4
%sub = fdiv fast float %sum.04, %0
%indvars.iv.next = add i64 %indvars.iv, 1
%lftr.wideiv = trunc i64 %indvars.iv.next to i32
%exitcond = icmp eq i32 %lftr.wideiv, 200
br i1 %exitcond, label %for.end, label %for.body
for.end: ; preds = %for.body
ret float %sub
}
;CHECK-LABEL: @foonodiv(
;CHECK-NOT: fdiv fast <4 x float>
;CHECK: ret
define float @foonodiv(float* nocapture %A, i32* nocapture %n) nounwind uwtable readonly ssp {
entry:
br label %for.body
for.body: ; preds = %for.body, %entry
%indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
%sum.04 = phi float [ 1.000000e+00, %entry ], [ %sub, %for.body ]
%arrayidx = getelementptr inbounds float, float* %A, i64 %indvars.iv
%0 = load float, float* %arrayidx, align 4
%sub = fdiv fast float %0, %sum.04
%indvars.iv.next = add i64 %indvars.iv, 1
%lftr.wideiv = trunc i64 %indvars.iv.next to i32
%exitcond = icmp eq i32 %lftr.wideiv, 200
br i1 %exitcond, label %for.end, label %for.body
for.end: ; preds = %for.body
ret float %sub
}