[llvm] r181842 - ARM ISel: Don't create illegal types during LowerMUL

Renato Golin renato.golin at linaro.org
Wed May 15 01:21:52 PDT 2013


Hi Bill, this is the last patch to fix all 3.3 rc1 problems, which should
give us a clean rc2 on ARM.

Cheers,
Renato
On 15 May 2013 01:31, "Arnold Schwaighofer" <aschwaighofer at apple.com> wrote:

Author: arnolds
Date: Tue May 14 17:33:24 2013
New Revision: 181842

URL: http://llvm.org/viewvc/llvm-project?rev=181842&view=rev
Log:
ARM ISel: Don't create illegal types during LowerMUL

The transformation happening here is that we want to turn a
"mul(ext(X), ext(X))" into a "vmull(X, X)", stripping off the extension. We
have
to make sure that X still has a valid vector type - possibly recreate an
extension to a smaller type. In case of a extload of a memory type smaller
than
64 bit we used create a ext(load()). The problem with doing this - instead
of
recreating an extload - is that an illegal type is exposed.

This patch fixes this by creating extloads instead of ext(load()) sequences.

Fixes PR15970.

radar://13871383

Modified:
    llvm/trunk/lib/Target/ARM/ARMISelLowering.cpp
    llvm/trunk/test/CodeGen/ARM/vmul.ll

Modified: llvm/trunk/lib/Target/ARM/ARMISelLowering.cpp
URL:
http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/ARM/ARMISelLowering.cpp?rev=181842&r1=181841&r2=181842&view=diff
==============================================================================
--- llvm/trunk/lib/Target/ARM/ARMISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/ARM/ARMISelLowering.cpp Tue May 14 17:33:24 2013
@@ -5257,6 +5257,23 @@ static bool isZeroExtended(SDNode *N, Se
   return false;
 }

+static EVT getExtensionTo64Bits(const EVT &OrigVT) {
+  if (OrigVT.getSizeInBits() >= 64)
+    return OrigVT;
+
+  assert(OrigVT.isSimple() && "Expecting a simple value type");
+
+  MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
+  switch (OrigSimpleTy) {
+  default: llvm_unreachable("Unexpected Vector Type");
+  case MVT::v2i8:
+  case MVT::v2i16:
+     return MVT::v2i32;
+  case MVT::v4i8:
+    return  MVT::v4i16;
+  }
+}
+
 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the
total
 /// value size to 64 bits. We need a 64-bit D register as an operand to
VMULL.
 /// We insert the required extension here to get the vector to fill a D
register.
@@ -5272,18 +5289,8 @@ static SDValue AddRequiredExtensionForVM
     return N;

   // Must extend size to at least 64 bits to be used as an operand for
VMULL.
-  MVT::SimpleValueType OrigSimpleTy = OrigTy.getSimpleVT().SimpleTy;
-  EVT NewVT;
-  switch (OrigSimpleTy) {
-  default: llvm_unreachable("Unexpected Orig Vector Type");
-  case MVT::v2i8:
-  case MVT::v2i16:
-    NewVT = MVT::v2i32;
-    break;
-  case MVT::v4i8:
-    NewVT = MVT::v4i16;
-    break;
-  }
+  EVT NewVT = getExtensionTo64Bits(OrigTy);
+
   return DAG.getNode(ExtOpcode, N->getDebugLoc(), NewVT, N);
 }

@@ -5293,22 +5300,22 @@ static SDValue AddRequiredExtensionForVM
 /// reach a total size of 64 bits. We have to add the extension separately
 /// because ARM does not have a sign/zero extending load for vectors.
 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG&
DAG) {
-  SDValue NonExtendingLoad =
-    DAG.getLoad(LD->getMemoryVT(), LD->getDebugLoc(), LD->getChain(),
+  EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());
+
+  // The load already has the right type.
+  if (ExtendedTy == LD->getMemoryVT())
+    return DAG.getLoad(LD->getMemoryVT(), LD->getDebugLoc(),
LD->getChain(),
                 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
                 LD->isNonTemporal(), LD->isInvariant(),
                 LD->getAlignment());
-  unsigned ExtOp = 0;
-  switch (LD->getExtensionType()) {
-  default: llvm_unreachable("Unexpected LoadExtType");
-  case ISD::EXTLOAD:
-  case ISD::SEXTLOAD: ExtOp = ISD::SIGN_EXTEND; break;
-  case ISD::ZEXTLOAD: ExtOp = ISD::ZERO_EXTEND; break;
-  }
-  MVT::SimpleValueType MemType = LD->getMemoryVT().getSimpleVT().SimpleTy;
-  MVT::SimpleValueType ExtType =
LD->getValueType(0).getSimpleVT().SimpleTy;
-  return AddRequiredExtensionForVMULL(NonExtendingLoad, DAG,
-                                      MemType, ExtType, ExtOp);
+
+  // We need to create a zextload/sextload. We cannot just create a load
+  // followed by a zext/zext node because LowerMUL is also run during
normal
+  // operation legalization where we can't create illegal types.
+  return DAG.getExtLoad(LD->getExtensionType(), LD->getDebugLoc(),
ExtendedTy,
+                        LD->getChain(), LD->getBasePtr(),
LD->getPointerInfo(),
+                        LD->getMemoryVT(), LD->isVolatile(),
+                        LD->isNonTemporal(), LD->getAlignment());
 }

 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,

Modified: llvm/trunk/test/CodeGen/ARM/vmul.ll
URL:
http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/ARM/vmul.ll?rev=181842&r1=181841&r2=181842&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/ARM/vmul.ll (original)
+++ llvm/trunk/test/CodeGen/ARM/vmul.ll Tue May 14 17:33:24 2013
@@ -599,3 +599,27 @@ for.end179:
 declare <8 x i16> @llvm.arm.neon.vrshiftu.v8i16(<8 x i16>, <8 x i16>)
nounwind readnone
 declare <8 x i16> @llvm.arm.neon.vqsubu.v8i16(<8 x i16>, <8 x i16>)
nounwind readnone
 declare <8 x i8> @llvm.arm.neon.vqmovnu.v8i8(<8 x i16>) nounwind readnone
+
+; vmull lowering would create a zext(v4i8 load()) instead of a
zextload(v4i8),
+; creating an illegal type during legalization and causing an assert.
+; PR15970
+define void @no_illegal_types_vmull_sext(<4 x i32> %a) {
+entry:
+  %wide.load283.i = load <4 x i8>* undef, align 1
+  %0 = sext <4 x i8> %wide.load283.i to <4 x i32>
+  %1 = sub nsw <4 x i32> %0, %a
+  %2 = mul nsw <4 x i32> %1, %1
+  %predphi290.v.i = select <4 x i1> undef, <4 x i32> undef, <4 x i32> %2
+  store <4 x i32> %predphi290.v.i, <4 x i32>* undef, align 4
+  ret void
+}
+define void @no_illegal_types_vmull_zext(<4 x i32> %a) {
+entry:
+  %wide.load283.i = load <4 x i8>* undef, align 1
+  %0 = zext <4 x i8> %wide.load283.i to <4 x i32>
+  %1 = sub nsw <4 x i32> %0, %a
+  %2 = mul nsw <4 x i32> %1, %1
+  %predphi290.v.i = select <4 x i1> undef, <4 x i32> undef, <4 x i32> %2
+  store <4 x i32> %predphi290.v.i, <4 x i32>* undef, align 4
+  ret void
+}


_______________________________________________
llvm-commits mailing list
llvm-commits at cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20130515/88297515/attachment.html>


More information about the llvm-commits mailing list