From: Sanjay Patel <spatel@rotateright.com>
Date: Thu, 23 May 2019 18:46:03 +0000 (+0000)
Subject: [InstCombine] be more careful when transforming a shuffle mask
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=a7df0fa47f78cb30e251416ca370c35ab52014e4;p=llvm

[InstCombine] be more careful when transforming a shuffle mask

This is reduced from a fuzzer test:
https://bugs.chromium.org/p/oss-fuzz/issues/detail?id=14890

Usually, demanded elements should be able to simplify shuffle
mask elements that are pointing to undef elements of its source
operands, but that doesn't happen in the test case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@361533 91177308-0d34-0410-b5e6-96231b3b80d8
---

diff --git a/lib/Transforms/InstCombine/InstCombineVectorOps.cpp b/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
index ecc4df179ef..c2ea0733a48 100644
--- a/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
+++ b/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
@@ -1635,8 +1635,8 @@ static Instruction *foldIdentityPaddedShuffles(ShuffleVectorInst &Shuf) {
   // We limit this transform to power-of-2 types because we expect that the
   // backend can convert the simplified IR patterns to identical nodes as the
   // original IR.
-  // TODO: If we can verify that behavior for arbitrary types, the power-of-2
-  // checks can be removed.
+  // TODO: If we can verify the same behavior for arbitrary types, the
+  //       power-of-2 checks can be removed.
   Value *X = Shuffle0->getOperand(0);
   Value *Y = Shuffle1->getOperand(0);
   if (X->getType() != Y->getType() ||
@@ -1663,10 +1663,27 @@ static Instruction *foldIdentityPaddedShuffles(ShuffleVectorInst &Shuf) {
   for (int i = 0, e = Mask.size(); i != e; ++i) {
     if (Mask[i] == -1)
       continue;
-    if (Mask[i] < WideElts)
+
+    // If this shuffle is choosing an undef element from 1 of the sources, that
+    // element is undef.
+    if (Mask[i] < WideElts) {
+      if (Shuffle0->getMaskValue(Mask[i]) == -1)
+        continue;
+    } else {
+      if (Shuffle1->getMaskValue(Mask[i] - WideElts) == -1)
+        continue;
+    }
+
+    // If this shuffle is choosing from the 1st narrow op, the mask element is
+    // the same. If this shuffle is choosing from the 2nd narrow op, the mask
+    // element is offset down to adjust for the narrow vector widths.
+    if (Mask[i] < WideElts) {
+      assert(Mask[i] < NarrowElts && "Unexpected shuffle mask");
       NewMask[i] = ConstantInt::get(I32Ty, Mask[i]);
-    else
+    } else {
+      assert(Mask[i] < (WideElts + NarrowElts) && "Unexpected shuffle mask");
       NewMask[i] = ConstantInt::get(I32Ty, Mask[i] - (WideElts - NarrowElts));
+    }
   }
   return new ShuffleVectorInst(X, Y, ConstantVector::get(NewMask));
 }
diff --git a/test/Transforms/InstCombine/vec_shuffle.ll b/test/Transforms/InstCombine/vec_shuffle.ll
index 3dcd7d7bf33..65af43ea2a1 100644
--- a/test/Transforms/InstCombine/vec_shuffle.ll
+++ b/test/Transforms/InstCombine/vec_shuffle.ll
@@ -1235,3 +1235,20 @@ define <4 x double> @not_insert_subvector_shuffles_with_same_size(<2 x double> %
   %s3 = shufflevector <4 x double> %s2, <4 x double> %s1, <4 x i32> <i32 0, i32 5, i32 undef, i32 undef>
   ret <4 x double> %s3
 }
+
+; Demanded vector elements may not be able to simplify a shuffle mask
+; before we try to narrow it. This used to crash.
+
+define <4 x float> @insert_subvector_crash_invalid_mask_elt(<2 x float> %x, <4 x float>* %p) {
+; CHECK-LABEL: @insert_subvector_crash_invalid_mask_elt(
+; CHECK-NEXT:    [[WIDEN:%.*]] = shufflevector <2 x float> [[X:%.*]], <2 x float> undef, <4 x i32> <i32 0, i32 1, i32 undef, i32 undef>
+; CHECK-NEXT:    [[I:%.*]] = shufflevector <2 x float> [[X]], <2 x float> undef, <4 x i32> <i32 0, i32 1, i32 undef, i32 undef>
+; CHECK-NEXT:    store <4 x float> [[I]], <4 x float>* [[P:%.*]], align 16
+; CHECK-NEXT:    ret <4 x float> [[WIDEN]]
+;
+  %widen = shufflevector <2 x float> %x, <2 x float> undef, <4 x i32> <i32 0, i32 1, i32 undef, i32 undef>
+  %ext2 = extractelement <2 x float> %x, i32 0
+  %I = insertelement <4 x float> %widen, float %ext2, i16 0
+  store <4 x float> %I, <4 x float>* %p
+  ret <4 x float> %widen
+}