[PATCH] D46591: [AArch64] Fix performPostLD1Combine to check for constant lane index.

[Geoff Berry via Phabricator via llvm-commits]

gberry created this revision.
gberry added a reviewer: t.p.northover.
Herald added subscribers: kristof.beyls, rengolin.
Herald added a reviewer: javed.absar.

performPostLD1Combine in AArch64ISelLowering looks for vector
insert_vector_elt of a loaded value which it can optimize into a single
LD1LANE instruction.  The code checking for the pattern was not checking
if the lane index was a constant which could cause two problems:

- an assert when lowering the LD1LANE ISD node since it assumes an constant operand

- an assert in isel if the lane index value depends on the post-incremented base register

Both of these issues are avoided by simply checking that the lane index
is a constant.

Fixes bug 35822.


Repository:
  rL LLVM

https://reviews.llvm.org/D46591

Files:
  lib/Target/AArch64/AArch64ISelLowering.cpp
  test/CodeGen/AArch64/arm64-indexed-vector-ldst-2.ll


Index: test/CodeGen/AArch64/arm64-indexed-vector-ldst-2.ll
===================================================================
--- test/CodeGen/AArch64/arm64-indexed-vector-ldst-2.ll
+++ test/CodeGen/AArch64/arm64-indexed-vector-ldst-2.ll
@@ -28,6 +28,28 @@
   ret void
 }
 
+; Avoid an assert/bad codegen in LD1LANEPOST lowering by not forming
+; LD1LANEPOST ISD nodes with a non-constant lane index.
+define <4 x i32> @f2(i32 *%p, <4 x i1> %m, <4 x i32> %v1, <4 x i32> %v2, i32 %idx) {
+  %L0 = load i32, i32* %p
+  %p1 = getelementptr i32, i32* %p, i64 1
+  %L1 = load i32, i32* %p1
+  %v = select <4 x i1> %m, <4 x i32> %v1, <4 x i32> %v2
+  %vret = insertelement <4 x i32> %v, i32 %L0, i32 %idx
+  store i32 %L1, i32 *%p
+  ret <4 x i32> %vret
+}
+
+; Check that a cycle is avoided during isel between the LD1LANEPOST instruction and the load of %L1.
+define <4 x i32> @f3(i32 *%p, <4 x i1> %m, <4 x i32> %v1, <4 x i32> %v2) {
+  %L0 = load i32, i32* %p
+  %p1 = getelementptr i32, i32* %p, i64 1
+  %L1 = load i32, i32* %p1
+  %v = select <4 x i1> %m, <4 x i32> %v1, <4 x i32> %v2
+  %vret = insertelement <4 x i32> %v, i32 %L0, i32 %L1
+  ret <4 x i32> %vret
+}
+
 ; Function Attrs: nounwind readnone
 declare i64 @llvm.objectsize.i64.p0i8(i8*, i1) #1
 
Index: lib/Target/AArch64/AArch64ISelLowering.cpp
===================================================================
--- lib/Target/AArch64/AArch64ISelLowering.cpp
+++ lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -9931,6 +9931,14 @@
   if (LD->getOpcode() != ISD::LOAD)
     return SDValue();
 
+  // The vector lane must be a constant in the LD1LANE opcode.
+  SDValue Lane;
+  if (IsLaneOp) {
+    Lane = N->getOperand(2);
+    if (!isa<ConstantSDNode>(Lane))
+      return SDValue();
+  }
+
   LoadSDNode *LoadSDN = cast<LoadSDNode>(LD);
   EVT MemVT = LoadSDN->getMemoryVT();
   // Check if memory operand is the same type as the vector element.
@@ -9987,7 +9995,7 @@
     Ops.push_back(LD->getOperand(0));  // Chain
     if (IsLaneOp) {
       Ops.push_back(Vector);           // The vector to be inserted
-      Ops.push_back(N->getOperand(2)); // The lane to be inserted in the vector
+      Ops.push_back(Lane);             // The lane to be inserted in the vector
     }
     Ops.push_back(Addr);
     Ops.push_back(Inc);


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D46591.145730.patch
Type: text/x-patch
Size: 2294 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20180508/2e122ba8/attachment.bin>