[X86] Don't split v8i32 all ones values if only AVX1 is available. Keep it intact...

This allows us to remove the check in ANDN combining that had to look through the extraction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@292881 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 418cda63d2bacbac430e14d8451f415defc060bf..9c00749bacff344b01edead73e56d38be806c51d 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -5090,8 +5090,7 @@ static SDValue concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
 }
 
 /// Returns a vector of specified type with all bits set.
-/// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
-/// no AVX2 support, use two <4 x i32> inserted in a <8 x i32> appropriately.
+/// Always build ones vectors as <4 x i32>, <8 x i32> or <16 x i32>.
 /// Then bitcast to their original type, ensuring they get CSE'd.
 static SDValue getOnesVector(EVT VT, const X86Subtarget &Subtarget,
                              SelectionDAG &DAG, const SDLoc &dl) {
@@ -5100,13 +5099,7 @@ static SDValue getOnesVector(EVT VT, const X86Subtarget &Subtarget,
 
   APInt Ones = APInt::getAllOnesValue(32);
   unsigned NumElts = VT.getSizeInBits() / 32;
-  SDValue Vec;
-  if (!Subtarget.hasInt256() && NumElts == 8) {
-    Vec = DAG.getConstant(Ones, dl, MVT::v4i32);
-    Vec = concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
-  } else {
-    Vec = DAG.getConstant(Ones, dl, MVT::getVectorVT(MVT::i32, NumElts));
-  }
+  SDValue Vec = DAG.getConstant(Ones, dl, MVT::getVectorVT(MVT::i32, NumElts));
   return DAG.getBitcast(VT, Vec);
 }
 
@@ -30633,20 +30626,9 @@ static SDValue combineANDXORWithAllOnesIntoANDNP(SDNode *N, SelectionDAG &DAG) {
 
   N01 = peekThroughBitcasts(N01);
 
-  // Either match a direct AllOnes for 128, 256, and 512-bit vectors, or an
-  // insert_subvector building a 256-bit AllOnes vector.
-  if (!ISD::isBuildVectorAllOnes(N01.getNode())) {
-    if (!VT.is256BitVector() || N01->getOpcode() != ISD::INSERT_SUBVECTOR)
-      return SDValue();
+  if (!ISD::isBuildVectorAllOnes(N01.getNode()))
+    return SDValue();
 
-    SDValue V1 = N01->getOperand(0);
-    SDValue V2 = N01->getOperand(1);
-    if (V1.getOpcode() != ISD::INSERT_SUBVECTOR ||
-        !V1.getOperand(0).isUndef() ||
-        !ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) ||
-        !ISD::isBuildVectorAllOnes(V2.getNode()))
-      return SDValue();
-  }
   return DAG.getNode(X86ISD::ANDNP, DL, VT, N00, N1);
 }
 

diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
index 81a61e7975c94984bdc64d666ed4cd7885b15d8a..922fa69343b2aff5e83a725a2dac0de82be3b61b 100644 (file)
--- a/lib/Target/X86/X86InstrSSE.td
+++ b/lib/Target/X86/X86InstrSSE.td
@@ -491,7 +491,6 @@ let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
                           [(set VR256:$dst, (v8i32 immAllOnesV))]>;
 }
 
-
 //===----------------------------------------------------------------------===//
 // SSE 1 & 2 - Move FP Scalar Instructions
 //
@@ -7862,6 +7861,15 @@ def VINSERTF128rm : AVXAIi8<0x18, MRMSrcMem, (outs VR256:$dst),
           []>, Sched<[WriteFShuffleLd, ReadAfterLd]>, VEX_4V, VEX_L;
 }
 
+
+// Without AVX2 we need to concat two v4i32 V_SETALLONES to create a 256-bit
+// all ones value.
+let Predicates = [HasAVX1Only] in
+def : Pat<(v8i32 immAllOnesV),
+          (VINSERTF128rr
+           (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)), (V_SETALLONES), sub_xmm),
+           (V_SETALLONES), 1)>;
+
 multiclass vinsert_lowering<string InstrStr, ValueType From, ValueType To,
                             PatFrag memop_frag> {
   def : Pat<(vinsert128_insert:$ins (To VR256:$src1), (From VR128:$src2),