BBVectorize: Don't vectorize vector-manipulation chains

Don't choose a vectorization plan containing only shuffles and
vector inserts/extracts. Due to inperfections in the cost model,
these can lead to infinite recusion.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@167811 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Transforms/Vectorize/BBVectorize.cpp b/lib/Transforms/Vectorize/BBVectorize.cpp
index 32a37bad7fe309becea1219e432cc316b1a98dc0..a2c3ad7d46bdee451b5e78870e14c7068c74920d 100644 (file)
--- a/lib/Transforms/Vectorize/BBVectorize.cpp
+++ b/lib/Transforms/Vectorize/BBVectorize.cpp
@@ -1703,10 +1703,20 @@ assert(n < 10 && "hrmm, really?");
         // The set of pairs that have already contributed to the total cost.
         DenseSet<ValuePair> IncomingPairs;
 
+        // If the cost model were perfect, this might not be necessary; but we
+        // need to make sure that we don't get stuck vectorizing our own
+        // shuffle chains.
+        bool HasNontrivialInsts = false;
+
         // The node weights represent the cost savings associated with
         // fusing the pair of instructions.
         for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
              E = PrunedTree.end(); S != E; ++S) {
+          if (!isa<ShuffleVectorInst>(S->first) &&
+              !isa<InsertElementInst>(S->first) &&
+              !isa<ExtractElementInst>(S->first))
+            HasNontrivialInsts = true;
+
           bool FlipOrder = false;
 
           if (getDepthFactor(S->first)) {
@@ -1943,6 +1953,13 @@ assert(n < 10 && "hrmm, really?");
             }
           }
         }
+
+        if (!HasNontrivialInsts) {
+          DEBUG(if (DebugPairSelection) dbgs() <<
+                "\tNo non-trivial instructions in tree;"
+                " override to zero effective size\n");
+          EffSize = 0;
+        }
       } else {
         for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
              E = PrunedTree.end(); S != E; ++S)

diff --git a/test/Transforms/BBVectorize/X86/sh-rec3.ll b/test/Transforms/BBVectorize/X86/sh-rec3.ll
new file mode 100644 (file)
index 0000000..fd2cc8b
--- /dev/null
+++ b/test/Transforms/BBVectorize/X86/sh-rec3.ll
@@ -0,0 +1,170 @@
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+; RUN: opt < %s -basicaa -mtriple=x86_64-unknown-linux-gnu -mcpu=corei7 -bb-vectorize -S | FileCheck %s
+
+%struct.gsm_state.2.8.39.44.45.55.56.57.58.59.62.63.64.65.74.75.76.77.80.87.92.93.94.95.96.97.110.111.112.113.114.128.130.135.136.137.138.139.140.141.142.143.144.145.148.149.150.151.152.169.170.177.178.179.184.185.186.187.188.201.208.209.219.220.221.223.224.225.230.231.232.233.235.236.237.238.245.246.248.249.272.274.279.280.281.282.283.286.293.298.299.314.315.316.317.318.319.320.321.322.323.324.325.326.327.328.329.330.331.332.333.334.335.336.337.338.339.340.341.342.343.344.345.346.347.348.349.350.351.352.353.565 = type { [280 x i16], i16, i64, i32, [8 x i16], [2 x [8 x i16]], i16, i16, [9 x i16], i16, i8, i8 }
+
+define void @gsm_encode(%struct.gsm_state.2.8.39.44.45.55.56.57.58.59.62.63.64.65.74.75.76.77.80.87.92.93.94.95.96.97.110.111.112.113.114.128.130.135.136.137.138.139.140.141.142.143.144.145.148.149.150.151.152.169.170.177.178.179.184.185.186.187.188.201.208.209.219.220.221.223.224.225.230.231.232.233.235.236.237.238.245.246.248.249.272.274.279.280.281.282.283.286.293.298.299.314.315.316.317.318.319.320.321.322.323.324.325.326.327.328.329.330.331.332.333.334.335.336.337.338.339.340.341.342.343.344.345.346.347.348.349.350.351.352.353.565* %s, i16* %source, i8* %c) nounwind uwtable {
+entry:
+  %LARc28 = alloca [2 x i64], align 16
+  %LARc28.sub = getelementptr inbounds [2 x i64]* %LARc28, i64 0, i64 0
+  %tmpcast = bitcast [2 x i64]* %LARc28 to [8 x i16]*
+  %Nc = alloca [4 x i16], align 2
+  %Mc = alloca [4 x i16], align 2
+  %bc = alloca [4 x i16], align 2
+  %xmc = alloca [52 x i16], align 16
+  %arraydecay = bitcast [2 x i64]* %LARc28 to i16*
+  %arraydecay1 = getelementptr inbounds [4 x i16]* %Nc, i64 0, i64 0
+  %arraydecay2 = getelementptr inbounds [4 x i16]* %bc, i64 0, i64 0
+  %arraydecay3 = getelementptr inbounds [4 x i16]* %Mc, i64 0, i64 0
+  %arraydecay5 = getelementptr inbounds [52 x i16]* %xmc, i64 0, i64 0
+  call void @Gsm_Coder(%struct.gsm_state.2.8.39.44.45.55.56.57.58.59.62.63.64.65.74.75.76.77.80.87.92.93.94.95.96.97.110.111.112.113.114.128.130.135.136.137.138.139.140.141.142.143.144.145.148.149.150.151.152.169.170.177.178.179.184.185.186.187.188.201.208.209.219.220.221.223.224.225.230.231.232.233.235.236.237.238.245.246.248.249.272.274.279.280.281.282.283.286.293.298.299.314.315.316.317.318.319.320.321.322.323.324.325.326.327.328.329.330.331.332.333.334.335.336.337.338.339.340.341.342.343.344.345.346.347.348.349.350.351.352.353.565* %s, i16* %source, i16* %arraydecay, i16* %arraydecay1, i16* %arraydecay2, i16* %arraydecay3, i16* undef, i16* %arraydecay5) nounwind
+  %0 = load i64* %LARc28.sub, align 16
+  %1 = trunc i64 %0 to i32
+  %conv1 = lshr i32 %1, 2
+  %and = and i32 %conv1, 15
+  %or = or i32 %and, 208
+  %conv6 = trunc i32 %or to i8
+  %incdec.ptr = getelementptr inbounds i8* %c, i64 1
+  store i8 %conv6, i8* %c, align 1
+  %conv84 = trunc i64 %0 to i8
+  %and9 = shl i8 %conv84, 6
+  %incdec.ptr15 = getelementptr inbounds i8* %c, i64 2
+  store i8 %and9, i8* %incdec.ptr, align 1
+  %2 = lshr i64 %0, 50
+  %shr226.tr = trunc i64 %2 to i8
+  %conv25 = and i8 %shr226.tr, 7
+  %incdec.ptr26 = getelementptr inbounds i8* %c, i64 3
+  store i8 %conv25, i8* %incdec.ptr15, align 1
+  %incdec.ptr42 = getelementptr inbounds i8* %c, i64 4
+  store i8 0, i8* %incdec.ptr26, align 1
+  %arrayidx52 = getelementptr inbounds [8 x i16]* %tmpcast, i64 0, i64 7
+  %3 = load i16* %arrayidx52, align 2
+  %conv537 = trunc i16 %3 to i8
+  %and54 = and i8 %conv537, 7
+  %incdec.ptr57 = getelementptr inbounds i8* %c, i64 5
+  store i8 %and54, i8* %incdec.ptr42, align 1
+  %incdec.ptr68 = getelementptr inbounds i8* %c, i64 6
+  store i8 0, i8* %incdec.ptr57, align 1
+  %4 = load i16* %arraydecay3, align 2
+  %conv748 = trunc i16 %4 to i8
+  %and75 = shl i8 %conv748, 5
+  %shl76 = and i8 %and75, 96
+  %incdec.ptr84 = getelementptr inbounds i8* %c, i64 7
+  store i8 %shl76, i8* %incdec.ptr68, align 1
+  %arrayidx94 = getelementptr inbounds [52 x i16]* %xmc, i64 0, i64 1
+  %5 = load i16* %arrayidx94, align 2
+  %conv959 = trunc i16 %5 to i8
+  %and96 = shl i8 %conv959, 1
+  %shl97 = and i8 %and96, 14
+  %or103 = or i8 %shl97, 1
+  %incdec.ptr105 = getelementptr inbounds i8* %c, i64 8
+  store i8 %or103, i8* %incdec.ptr84, align 1
+  %arrayidx115 = getelementptr inbounds [52 x i16]* %xmc, i64 0, i64 4
+  %6 = bitcast i16* %arrayidx115 to i32*
+  %7 = load i32* %6, align 8
+  %conv11610 = trunc i32 %7 to i8
+  %and117 = and i8 %conv11610, 7
+  %incdec.ptr120 = getelementptr inbounds i8* %c, i64 9
+  store i8 %and117, i8* %incdec.ptr105, align 1
+  %8 = lshr i32 %7, 16
+  %and12330 = shl nuw nsw i32 %8, 5
+  %and123 = trunc i32 %and12330 to i8
+  %incdec.ptr136 = getelementptr inbounds i8* %c, i64 10
+  store i8 %and123, i8* %incdec.ptr120, align 1
+  %incdec.ptr157 = getelementptr inbounds i8* %c, i64 11
+  store i8 0, i8* %incdec.ptr136, align 1
+  %incdec.ptr172 = getelementptr inbounds i8* %c, i64 12
+  store i8 0, i8* %incdec.ptr157, align 1
+  %arrayidx173 = getelementptr inbounds [4 x i16]* %Nc, i64 0, i64 1
+  %9 = load i16* %arrayidx173, align 2
+  %conv17412 = zext i16 %9 to i32
+  %and175 = shl nuw nsw i32 %conv17412, 1
+  %arrayidx177 = getelementptr inbounds [4 x i16]* %bc, i64 0, i64 1
+  %10 = load i16* %arrayidx177, align 2
+  %conv17826 = zext i16 %10 to i32
+  %shr17913 = lshr i32 %conv17826, 1
+  %and180 = and i32 %shr17913, 1
+  %or181 = or i32 %and175, %and180
+  %conv182 = trunc i32 %or181 to i8
+  %incdec.ptr183 = getelementptr inbounds i8* %c, i64 13
+  store i8 %conv182, i8* %incdec.ptr172, align 1
+  %arrayidx188 = getelementptr inbounds [4 x i16]* %Mc, i64 0, i64 1
+  %11 = load i16* %arrayidx188, align 2
+  %conv18914 = trunc i16 %11 to i8
+  %and190 = shl i8 %conv18914, 5
+  %shl191 = and i8 %and190, 96
+  %incdec.ptr199 = getelementptr inbounds i8* %c, i64 14
+  store i8 %shl191, i8* %incdec.ptr183, align 1
+  %arrayidx209 = getelementptr inbounds [52 x i16]* %xmc, i64 0, i64 14
+  %12 = load i16* %arrayidx209, align 4
+  %conv21015 = trunc i16 %12 to i8
+  %and211 = shl i8 %conv21015, 1
+  %shl212 = and i8 %and211, 14
+  %or218 = or i8 %shl212, 1
+  %incdec.ptr220 = getelementptr inbounds i8* %c, i64 15
+  store i8 %or218, i8* %incdec.ptr199, align 1
+  %arrayidx225 = getelementptr inbounds [52 x i16]* %xmc, i64 0, i64 16
+  %13 = bitcast i16* %arrayidx225 to i64*
+  %14 = load i64* %13, align 16
+  %conv22616 = trunc i64 %14 to i8
+  %and227 = shl i8 %conv22616, 3
+  %shl228 = and i8 %and227, 56
+  %incdec.ptr235 = getelementptr inbounds i8* %c, i64 16
+  store i8 %shl228, i8* %incdec.ptr220, align 1
+  %15 = lshr i64 %14, 32
+  %and23832 = shl nuw nsw i64 %15, 5
+  %and238 = trunc i64 %and23832 to i8
+  %incdec.ptr251 = getelementptr inbounds i8* %c, i64 17
+  store i8 %and238, i8* %incdec.ptr235, align 1
+  %arrayidx266 = getelementptr inbounds [52 x i16]* %xmc, i64 0, i64 23
+  %incdec.ptr272 = getelementptr inbounds i8* %c, i64 18
+  store i8 0, i8* %incdec.ptr251, align 1
+  %16 = load i16* %arrayidx266, align 2
+  %conv27418 = trunc i16 %16 to i8
+  %and275 = shl i8 %conv27418, 6
+  %incdec.ptr287 = getelementptr inbounds i8* %c, i64 19
+  store i8 %and275, i8* %incdec.ptr272, align 1
+  %arrayidx288 = getelementptr inbounds [4 x i16]* %Nc, i64 0, i64 2
+  %17 = load i16* %arrayidx288, align 2
+  %conv28919 = zext i16 %17 to i32
+  %and290 = shl nuw nsw i32 %conv28919, 1
+  %arrayidx292 = getelementptr inbounds [4 x i16]* %bc, i64 0, i64 2
+  %18 = load i16* %arrayidx292, align 2
+  %conv29327 = zext i16 %18 to i32
+  %shr29420 = lshr i32 %conv29327, 1
+  %and295 = and i32 %shr29420, 1
+  %or296 = or i32 %and290, %and295
+  %conv297 = trunc i32 %or296 to i8
+  %incdec.ptr298 = getelementptr inbounds i8* %c, i64 20
+  store i8 %conv297, i8* %incdec.ptr287, align 1
+  %conv30021 = trunc i16 %18 to i8
+  %and301 = shl i8 %conv30021, 7
+  %incdec.ptr314 = getelementptr inbounds i8* %c, i64 21
+  store i8 %and301, i8* %incdec.ptr298, align 1
+  %incdec.ptr335 = getelementptr inbounds i8* %c, i64 22
+  store i8 0, i8* %incdec.ptr314, align 1
+  %arrayidx340 = getelementptr inbounds [52 x i16]* %xmc, i64 0, i64 29
+  %19 = load i16* %arrayidx340, align 2
+  %conv34122 = trunc i16 %19 to i8
+  %and342 = shl i8 %conv34122, 3
+  %shl343 = and i8 %and342, 56
+  %incdec.ptr350 = getelementptr inbounds i8* %c, i64 23
+  store i8 %shl343, i8* %incdec.ptr335, align 1
+  %arrayidx355 = getelementptr inbounds [52 x i16]* %xmc, i64 0, i64 32
+  %20 = bitcast i16* %arrayidx355 to i32*
+  %21 = load i32* %20, align 16
+  %conv35623 = shl i32 %21, 2
+  %shl358 = and i32 %conv35623, 28
+  %22 = lshr i32 %21, 17
+  %and363 = and i32 %22, 3
+  %or364 = or i32 %shl358, %and363
+  %conv365 = trunc i32 %or364 to i8
+  store i8 %conv365, i8* %incdec.ptr350, align 1
+  unreachable
+; CHECK: @gsm_encode
+}
+
+declare void @Gsm_Coder(%struct.gsm_state.2.8.39.44.45.55.56.57.58.59.62.63.64.65.74.75.76.77.80.87.92.93.94.95.96.97.110.111.112.113.114.128.130.135.136.137.138.139.140.141.142.143.144.145.148.149.150.151.152.169.170.177.178.179.184.185.186.187.188.201.208.209.219.220.221.223.224.225.230.231.232.233.235.236.237.238.245.246.248.249.272.274.279.280.281.282.283.286.293.298.299.314.315.316.317.318.319.320.321.322.323.324.325.326.327.328.329.330.331.332.333.334.335.336.337.338.339.340.341.342.343.344.345.346.347.348.349.350.351.352.353.565*, i16*, i16*, i16*, i16*, i16*, i16*, i16*)
+
+declare void @llvm.trap() noreturn nounwind