search

epiqc / ScaffCC

Compilation, analysis and optimization framework for the Scaffold quantum programming language.
BSD 2-Clause "Simplified" License
192 stars 53 forks source link

[PASS] O1 optimizations wrongly use -loop-simplify #11

Open xiangzhai opened 7 years ago

xiangzhai commented 7 years ago

Hi ScaffCC developers,

-loop-simplify will produce the disable unroll meta data (!llvm.loop !2), for example:

$ cat cat_state.n044.ll
; ModuleID = 'cat_state.n045a.ll'
source_filename = "cat_state.n04_merged.scaffold"
target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"

; Function Attrs: noinline nounwind uwtable
define void @catN(i16* %bit, i32 %n) local_unnamed_addr #0 {
entry:
   %0 = load i16, i16* %bit, align 2
   tail call void @llvm.H.i16(i16 %0)
   %cmp1 = icmp sgt i32 %n, 1
   br i1 %cmp1, label %for.body.lr.ph, label %for.end

for.body.lr.ph:                                   ; preds = %entry
   %1 = add i32 %n, -1
   %2 = add i32 %n, -2
   %xtraiter = and i32 %1, 1
   %3 = icmp ult i32 %2, 1
   br i1 %3, label %for.cond.for.end_crit_edge.unr-lcssa, label 
%for.body.lr.ph.new

for.body.lr.ph.new:                               ; preds = %for.body.lr.ph
   %unroll_iter = sub i32 %1, %xtraiter
   br label %for.body

for.body:                                         ; preds = %for.body, 
%for.body.lr.ph.new
   %inc3 = phi i32 [ 1, %for.body.lr.ph.new ], [ %inc.1, %for.body ]
   %niter = phi i32 [ %unroll_iter, %for.body.lr.ph.new ], [ 
%niter.nsub.1, %for.body ]
   %idxprom = sext i32 %inc3 to i64
   %arrayidx1 = getelementptr inbounds i16, i16* %bit, i64 %idxprom
   %4 = load i16, i16* %arrayidx1, align 2
   %sub = add nsw i32 %inc3, -1
   %idxprom2 = sext i32 %sub to i64
   %arrayidx3 = getelementptr inbounds i16, i16* %bit, i64 %idxprom2
   %5 = load i16, i16* %arrayidx3, align 2
   tail call void @llvm.CNOT.i16.i16(i16 %4, i16 %5)
   %inc = add nsw i32 %inc3, 1
   %niter.nsub = sub i32 %niter, 1
   %idxprom.1 = sext i32 %inc to i64
   %arrayidx1.1 = getelementptr inbounds i16, i16* %bit, i64 %idxprom.1
   %6 = load i16, i16* %arrayidx1.1, align 2
   %idxprom2.1 = sext i32 %inc3 to i64
   %arrayidx3.1 = getelementptr inbounds i16, i16* %bit, i64 %idxprom2.1
   %7 = load i16, i16* %arrayidx3.1, align 2
   tail call void @llvm.CNOT.i16.i16(i16 %6, i16 %7)
   %inc.1 = add nsw i32 %inc, 1
   %niter.nsub.1 = sub i32 %niter.nsub, 1
   %niter.ncmp.1 = icmp ne i32 %niter.nsub.1, 0
   br i1 %niter.ncmp.1, label %for.body, label 
%for.cond.for.end_crit_edge.unr-lcssa

for.cond.for.end_crit_edge.unr-lcssa:             ; preds = %for.body, 
%for.body.lr.ph
   %inc3.unr = phi i32 [ 1, %for.body.lr.ph ], [ %inc.1, %for.body ]
   %lcmp.mod = icmp ne i32 %xtraiter, 0
   br i1 %lcmp.mod, label %for.body.epil, label %for.end

for.body.epil:                                    ; preds = 
%for.cond.for.end_crit_edge.unr-lcssa
   %inc3.epil = phi i32 [ %inc3.unr, %for.cond.for.end_crit_edge.unr-lcssa ]
   %idxprom.epil = sext i32 %inc3.epil to i64
   %arrayidx1.epil = getelementptr inbounds i16, i16* %bit, i64 
%idxprom.epil
   %8 = load i16, i16* %arrayidx1.epil, align 2
   %sub.epil = add nsw i32 %inc3.epil, -1
   %idxprom2.epil = sext i32 %sub.epil to i64
   %arrayidx3.epil = getelementptr inbounds i16, i16* %bit, i64 
%idxprom2.epil
   %9 = load i16, i16* %arrayidx3.epil, align 2
   tail call void @llvm.CNOT.i16.i16(i16 %8, i16 %9)
   %inc.epil = add nsw i32 %inc3.epil, 1
   %cmp.epil = icmp slt i32 %inc.epil, %n
   br label %for.end

for.end:                                          ; preds = 
%for.body.epil, %for.cond.for.end_crit_edge.unr-lcssa, %entry
   ret void
}

; Function Attrs: nounwind
declare void @llvm.H.i16(i16) #1

; Function Attrs: nounwind
declare void @llvm.CNOT.i16.i16(i16, i16) #1

; Function Attrs: noinline nounwind uwtable
define void @unCatN(i16* %bit, i32 %n) local_unnamed_addr #0 {
entry:
   %storemerge1 = add nsw i32 %n, -1
   %cmp2 = icmp sgt i32 %n, 1
   br i1 %cmp2, label %for.body.peel, label %for.end

for.body.peel:                                    ; preds = %entry
   %idxprom.peel = sext i32 %storemerge1 to i64
   %arrayidx.peel = getelementptr inbounds i16, i16* %bit, i64 %idxprom.peel
   %0 = load i16, i16* %arrayidx.peel, align 2
   %sub1.peel = add nsw i32 %n, -2
   %idxprom2.peel = sext i32 %sub1.peel to i64
   %arrayidx3.peel = getelementptr inbounds i16, i16* %bit, i64 
%idxprom2.peel
   %1 = load i16, i16* %arrayidx3.peel, align 2
   tail call void @llvm.CNOT.i16.i16(i16 %0, i16 %1)
   %storemerge.peel = add nsw i32 %storemerge1, -1
   %cmp.peel = icmp sgt i32 %storemerge1, 1
   br i1 %cmp.peel, label %for.body.lr.ph.peel.newph, label %for.end

for.body.lr.ph.peel.newph:                        ; preds = %for.body.peel
   br label %for.body

for.body:                                         ; preds = %for.body, 
%for.body.lr.ph.peel.newph
   %storemerge5 = phi i32 [ %storemerge.peel, %for.body.lr.ph.peel.newph 
], [ %storemerge, %for.body ]
   %storemerge.in3 = phi i32 [ %storemerge1, %for.body.lr.ph.peel.newph 
], [ %storemerge5, %for.body ]
   %idxprom = sext i32 %storemerge5 to i64
   %arrayidx = getelementptr inbounds i16, i16* %bit, i64 %idxprom
   %2 = load i16, i16* %arrayidx, align 2
   %sub1 = add nsw i32 %storemerge.in3, -2
   %idxprom2 = sext i32 %sub1 to i64
   %arrayidx3 = getelementptr inbounds i16, i16* %bit, i64 %idxprom2
   %3 = load i16, i16* %arrayidx3, align 2
   tail call void @llvm.CNOT.i16.i16(i16 %2, i16 %3)
   %storemerge = add nsw i32 %storemerge5, -1
   %cmp = icmp sgt i32 %storemerge5, 1
   br i1 %cmp, label %for.body, label %for.end, !llvm.loop !2

for.end:                                          ; preds = %for.body, 
%for.body.peel, %entry
   %.lcssa = phi i16* [ %bit, %entry ], [ %bit, %for.body.peel ], [ 
%bit, %for.body ]
   %4 = load i16, i16* %.lcssa, align 2
   tail call void @llvm.H.i16(i16 %4)
   ret void
}

; Function Attrs: noinline nounwind uwtable
define i32 @main() local_unnamed_addr #0 {
entry:
   %bits = alloca [4 x i16], align 2
   %arraydecay = getelementptr inbounds [4 x i16], [4 x i16]* %bits, i64 
0, i64 0
   call void @catN_IP4_IPx_IPx_IPx_DPx_DPx_DPx_DPx(i16* %arraydecay, i32 
undef)
   ret i32 0
}

define void @catN_IP4_IPx_IPx_IPx_DPx_DPx_DPx_DPx(i16* %bit, i32 %n) {
entry.:
   %0 = load i16, i16* %bit, align 2
   tail call void @llvm.H.i16(i16 %0)
   %arrayidx1. = getelementptr inbounds i16, i16* %bit, i64 1
   %1 = load i16, i16* %arrayidx1., align 2
   %2 = load i16, i16* %bit, align 2
   tail call void @llvm.CNOT.i16.i16(i16 %1, i16 %2)
   %arrayidx1.1. = getelementptr inbounds i16, i16* %bit, i64 2
   %3 = load i16, i16* %arrayidx1.1., align 2
   %arrayidx3.1. = getelementptr inbounds i16, i16* %bit, i64 1
   %4 = load i16, i16* %arrayidx3.1., align 2
   tail call void @llvm.CNOT.i16.i16(i16 %3, i16 %4)
   %arrayidx1.epil. = getelementptr inbounds i16, i16* %bit, i64 3
   %5 = load i16, i16* %arrayidx1.epil., align 2
   %arrayidx3.epil. = getelementptr inbounds i16, i16* %bit, i64 2
   %6 = load i16, i16* %arrayidx3.epil., align 2
   tail call void @llvm.CNOT.i16.i16(i16 %5, i16 %6)
   ret void
}

attributes #0 = { noinline nounwind uwtable 
"correctly-rounded-divide-sqrt-fp-math"="false" 
"disable-tail-calls"="false" "less-precise-fpmad"="false" 
"no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf" 
"no-infs-fp-math"="false" "no-jump-tables"="false" 
"no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" 
"no-trapping-math"="false" "stack-protector-buffer-size"="8" 
"target-cpu"="x86-64" "target-features"="+fxsr,+mmx,+sse,+sse2,+x87" 
"unsafe-fp-math"="false" "use-soft-float"="false" }
attributes #1 = { nounwind }

!llvm.module.flags = !{!0}
!llvm.ident = !{!1}

!0 = !{i32 1, !"wchar_size", i32 4}
!1 = !{!"clang version 6.0.0 (git at github.com:llvm-mirror/clang.git 
0aed123216ad4a38a9c2b16f1783895fd5cb1a04) 
(git at github.com:llvm-mirror/llvm.git 
d209b37aec1e392dabbf9b5324ea4a60c36fbc55)"}
!2 = distinct !{!2, !3}
!3 = !{!"llvm.loop.unroll.disable"}

Then it failed to Unroll:

$(OPT) -S cat_state.n044.ll -mem2reg -loops -loop-simplify -loop-rotate 
-lcssa -loop-unroll -unroll-threshold=100000000 -sccp -simplifycfg -o 
cat_state.n045.ll

There are still for-loops in the cat_state.n045.ll, although it is able to workaround use -internalize -globaldce to remove unCatN and catN DeadFunction which including for-loops :)

Regards, Leslie Zhai

xiangzhai commented 7 years ago

Fixed in https://github.com/ScaffCC/ScaffCC/commit/7cf9c3cef0901b83dfb41f07f9dea083fd52b15d#diff-4449e30adec5beea4c33c58954ce4d0d