From 06e6c385cbf550ab73ad70bd3e02d9ad2fd3089c Mon Sep 17 00:00:00 2001
From: Pete Cooper <peter_cooper@apple.com>
Date: Sat, 16 Jun 2012 01:43:26 +0000
Subject: [PATCH] Fix crash from r158529 on Bullet.

Dynamic GEPs created by SROA needed to insert extra "i32 0"
operands to index through structs and arrays to get to the
vector being indexed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@158590 91177308-0d34-0410-b5e6-96231b3b80d8
---
 .../Scalar/ScalarReplAggregates.cpp           | 11 +++-
 .../ScalarRepl/dynamic-vector-gep.ll          | 55 +++++++++++++++++++
 2 files changed, 65 insertions(+), 1 deletion(-)

diff --git a/lib/Transforms/Scalar/ScalarReplAggregates.cpp b/lib/Transforms/Scalar/ScalarReplAggregates.cpp
index 9fc4164cc0e..bc428802e89 100644
--- a/lib/Transforms/Scalar/ScalarReplAggregates.cpp
+++ b/lib/Transforms/Scalar/ScalarReplAggregates.cpp
@@ -2011,8 +2011,17 @@ void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
     uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy);
     NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx));
   }
-  if (NonConstantIdx)
+  if (NonConstantIdx) {
+    Type* GepTy = T;
+    // This GEP has a dynamic index.  We need to add "i32 0" to index through
+    // any structs or arrays in the original type until we get to the vector
+    // to index.
+    while (!isa<VectorType>(GepTy)) {
+      NewArgs.push_back(Constant::getNullValue(i32Ty));
+      GepTy = cast<CompositeType>(GepTy)->getTypeAtIndex(0U);
+    }
     NewArgs.push_back(NonConstantIdx);
+  }
   Instruction *Val = NewElts[Idx];
   if (NewArgs.size() > 1) {
     Val = GetElementPtrInst::CreateInBounds(Val, NewArgs, "", GEPI);
diff --git a/test/Transforms/ScalarRepl/dynamic-vector-gep.ll b/test/Transforms/ScalarRepl/dynamic-vector-gep.ll
index 48a0da9dabd..af7f4398d10 100644
--- a/test/Transforms/ScalarRepl/dynamic-vector-gep.ll
+++ b/test/Transforms/ScalarRepl/dynamic-vector-gep.ll
@@ -80,4 +80,59 @@ entry:
   ret float %ret
 }
 
+; CHECK: test5
+; CHECK: %[[alloc0:[\.a-z0-9]*]] = alloca <4 x float>
+; CHECK: %[[alloc1:[\.a-z0-9]*]] = alloca <4 x float>
+; CHECK: store <4 x float> zeroinitializer, <4 x float>* %[[alloc0]]
+; CHECK: store <4 x float> zeroinitializer, <4 x float>* %[[alloc1]]
+; CHECK: %ptr1 = getelementptr inbounds <4 x float>* %[[alloc0]], i32 0, i32 %idx1
+; CHECK: store float 1.000000e+00, float* %ptr1
+; CHECK: %ptr2 = getelementptr inbounds <4 x float>* %[[alloc1]], i32 0, i32 %idx2
+; CHECK: %ret = load float* %ptr2
+
+%vector.pair = type { %vector.anon, %vector.anon }
+%vector.anon = type { %vector }
+%vector = type { <4 x float> }
+
+; Dynamic GEPs on vectors were crashing when the vector was inside a struct
+; as the new GEP for the new alloca might not include all the indices from
+; the original GEP, just the indices it needs to get to the correct offset of
+; some type, not necessarily the dynamic vector.
+; This test makes sure we don't have this crash.
+define float @test5(i32 %idx1, i32 %idx2) {
+entry:
+  %0 = alloca %vector.pair
+  store %vector.pair zeroinitializer, %vector.pair* %0
+  %ptr1 = getelementptr %vector.pair* %0, i32 0, i32 0, i32 0, i32 0, i32 %idx1
+  store float 1.0, float* %ptr1
+  %ptr2 = getelementptr %vector.pair* %0, i32 0, i32 1, i32 0, i32 0, i32 %idx2
+  %ret = load float* %ptr2
+  ret float %ret
+}
+
+; CHECK: test6
+; CHECK: %[[alloc0:[\.a-z0-9]*]] = alloca <4 x float>
+; CHECK: %[[alloc1:[\.a-z0-9]*]] = alloca <4 x float>
+; CHECK: store <4 x float> zeroinitializer, <4 x float>* %[[alloc0]]
+; CHECK: store <4 x float> zeroinitializer, <4 x float>* %[[alloc1]]
+; CHECK: %ptr1 = getelementptr inbounds <4 x float>* %[[alloc0]], i32 0, i32 %idx1
+; CHECK: store float 1.000000e+00, float* %ptr1
+; CHECK: %ptr2 = getelementptr inbounds <4 x float>* %[[alloc1]], i32 0, i32 %idx2
+; CHECK: %ret = load float* %ptr2
+
+%array.pair = type { [2 x %array.anon], %array.anon }
+%array.anon = type { [2 x %vector] }
+
+; This is the same as test5 and tests the same crash, but on arrays.
+define float @test6(i32 %idx1, i32 %idx2) {
+entry:
+  %0 = alloca %array.pair
+  store %array.pair zeroinitializer, %array.pair* %0
+  %ptr1 = getelementptr %array.pair* %0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 %idx1
+  store float 1.0, float* %ptr1
+  %ptr2 = getelementptr %array.pair* %0, i32 0, i32 1, i32 0, i32 0, i32 0, i32 %idx2
+  %ret = load float* %ptr2
+  ret float %ret
+}
+
 declare void @llvm.memset.p0i8.i32(i8*, i8, i32, i32, i1)