DAG combine should not increase alignment of loads / stores with alignment less

than ABI alignment. These are loads / stores from / to "packed" data structures.
Their alignments are intentionally under-specified.

rdar://10301431


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145273 91177308-0d34-0410-b5e6-96231b3b80d8

3 files changed, 66 insertions, 13 deletions

diff --git a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index d8208a4433..2c4886a3c8 100644
--- a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -6206,13 +6206,20 @@ SDValue DAGCombiner::visitLOAD(SDNode *N) {
 
   // Try to infer better alignment information than the load already has.
   if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
-    if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
-      if (Align > LD->getAlignment())
-        return DAG.getExtLoad(LD->getExtensionType(), N->getDebugLoc(),
-                              LD->getValueType(0),
-                              Chain, Ptr, LD->getPointerInfo(),
-                              LD->getMemoryVT(),
-                              LD->isVolatile(), LD->isNonTemporal(), Align);
+    unsigned ABIAlign = TLI.getTargetData()->
+      getABITypeAlignment(LD->getMemoryVT().getTypeForEVT(*DAG.getContext()));
+    unsigned LDAlign = LD->getAlignment();
+    // Do not touch loads with explicit alignments that are smaller than ABI
+    // alignment to avoid breaking loads from "packed" types.
+    if (!LDAlign || LDAlign >= ABIAlign) {
+      if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
+        if (Align > LDAlign)
+          return DAG.getExtLoad(LD->getExtensionType(), N->getDebugLoc(),
+                                LD->getValueType(0),
+                                Chain, Ptr, LD->getPointerInfo(),
+                                LD->getMemoryVT(),
+                                LD->isVolatile(), LD->isNonTemporal(), Align);
+      }
     }
   }
 
@@ -6669,11 +6676,18 @@ SDValue DAGCombiner::visitSTORE(SDNode *N) {
 
   // Try to infer better alignment information than the store already has.
   if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
-    if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
-      if (Align > ST->getAlignment())
-        return DAG.getTruncStore(Chain, N->getDebugLoc(), Value,
-                                 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
-                                 ST->isVolatile(), ST->isNonTemporal(), Align);
+    unsigned ABIAlign = TLI.getTargetData()->
+      getABITypeAlignment(ST->getMemoryVT().getTypeForEVT(*DAG.getContext()));
+    unsigned STAlign = ST->getAlignment();
+    // Do not touch stores with explicit alignments that are smaller than ABI
+    // alignment to avoid breaking stores from "packed" types.
+    if (!STAlign || STAlign >= ABIAlign) {
+      if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
+        if (Align > STAlign)
+          return DAG.getTruncStore(Chain, N->getDebugLoc(), Value,
+                                   Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
+                                   ST->isVolatile(), ST->isNonTemporal(),Align);
+      }
     }
   }
 
diff --git a/test/CodeGen/ARM/2011-11-28-DAGCombineBug.ll b/test/CodeGen/ARM/2011-11-28-DAGCombineBug.ll
new file mode 100644
index 0000000000..5409f8c608
--- /dev/null
+++ b/test/CodeGen/ARM/2011-11-28-DAGCombineBug.ll
@@ -0,0 +1,38 @@
+; RUN: llc < %s -mtriple=thumbv7-apple-ios5.0.0 | FileCheck %s
+; rdar://10464621
+
+; DAG combine increases loads from packed types. ARM load / store optimizer then
+; combined them into a ldm which causes runtime exception.
+
+%struct.InformationBlock = type <{ i32, %struct.FlagBits, %struct.FlagBits }>
+%struct.FlagBits = type <{ [4 x i32] }>
+
+@infoBlock = external global %struct.InformationBlock
+
+define hidden void @foo() {
+; CHECK: foo:
+; CHECK: ldr.w
+; CHECK: ldr.w
+; CHECK-NOT: ldm
+entry:
+  %tmp13 = load i32* getelementptr inbounds (%struct.InformationBlock* @infoBlock, i32 0, i32 1, i32 0, i32 0), align 1
+  %tmp15 = load i32* getelementptr inbounds (%struct.InformationBlock* @infoBlock, i32 0, i32 1, i32 0, i32 1), align 1
+  %tmp17 = load i32* getelementptr inbounds (%struct.InformationBlock* @infoBlock, i32 0, i32 1, i32 0, i32 2), align 1
+  %tmp19 = load i32* getelementptr inbounds (%struct.InformationBlock* @infoBlock, i32 0, i32 1, i32 0, i32 3), align 1
+  %tmp = load i32* getelementptr inbounds (%struct.InformationBlock* @infoBlock, i32 0, i32 2, i32 0, i32 0), align 1
+  %tmp3 = load i32* getelementptr inbounds (%struct.InformationBlock* @infoBlock, i32 0, i32 2, i32 0, i32 1), align 1
+  %tmp4 = load i32* getelementptr inbounds (%struct.InformationBlock* @infoBlock, i32 0, i32 2, i32 0, i32 2), align 1
+  %tmp5 = load i32* getelementptr inbounds (%struct.InformationBlock* @infoBlock, i32 0, i32 2, i32 0, i32 3), align 1
+  %insert21 = insertvalue [4 x i32] undef, i32 %tmp13, 0
+  %insert23 = insertvalue [4 x i32] %insert21, i32 %tmp15, 1
+  %insert25 = insertvalue [4 x i32] %insert23, i32 %tmp17, 2
+  %insert27 = insertvalue [4 x i32] %insert25, i32 %tmp19, 3
+  %insert = insertvalue [4 x i32] undef, i32 %tmp, 0
+  %insert7 = insertvalue [4 x i32] %insert, i32 %tmp3, 1
+  %insert9 = insertvalue [4 x i32] %insert7, i32 %tmp4, 2
+  %insert11 = insertvalue [4 x i32] %insert9, i32 %tmp5, 3
+  tail call void @bar([4 x i32] %insert27, [4 x i32] %insert11)
+  ret void
+}
+
+declare void @bar([4 x i32], [4 x i32])
diff --git a/test/CodeGen/X86/widen_load-1.ll b/test/CodeGen/X86/widen_load-1.ll
index 639617f177..a9fd190418 100644
--- a/test/CodeGen/X86/widen_load-1.ll
+++ b/test/CodeGen/X86/widen_load-1.ll
@@ -1,5 +1,6 @@
 ; RUN: llc %s -o - -march=x86-64 -mtriple=x86_64-unknown-linux-gnu | FileCheck %s
 ; PR4891
+; PR5626
 
 ; This load should be before the call, not after.
 
@@ -20,7 +21,7 @@ define void @reset(<2 x float>* noalias %garbage1) {
   store i32 0, i32* %changed, align 4
   %r2 = getelementptr float* bitcast ([20 x i64]* @compl to float*), i64 32 ; <float*> [#uses=1]
   %r3 = bitcast float* %r2 to <2 x float>*        ; <<2 x float>*> [#uses=1]
-  %r4 = load <2 x float>* %r3, align 4            ; <<2 x float>> [#uses=1]
+  %r4 = load <2 x float>* %r3, align 8            ; <<2 x float>> [#uses=1]
   call void @killcommon(i32* %changed)
   br label %"file complex.c, line 34, bb4"