From: Craig Topper <craig.topper@gmail.com>
Date: Sat, 4 Feb 2017 23:26:46 +0000 (+0000)
Subject: [X86] Add support for folding (insert_subvector vec1, (extract_subvector vec2, idx1... 
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=1cfb6c6e52e0f9e3ebfba92731f87e5f7304afe1;p=llvm

[X86] Add support for folding (insert_subvector vec1, (extract_subvector vec2, idx1), idx1) -> (blendi vec2, vec1).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@294112 91177308-0d34-0410-b5e6-96231b3b80d8
---

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index f51920f5e6e..2090d32d08c 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -34121,8 +34121,7 @@ static SDValue combineInsertSubvector(SDNode *N, SelectionDAG &DAG,
   // We are still creating an INSERT_SUBVECTOR below with an undef node to
   // extend the subvector to the size of the result vector. Make sure that
   // we are not recursing on that node by checking for undef here.
-  if (IdxVal == 0 && OpVT.is256BitVector() && SubVecVT.is128BitVector() &&
-      !Vec.isUndef()) {
+  if (IdxVal == 0 && OpVT.is256BitVector() && !Vec.isUndef()) {
     SDValue Vec256 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT,
                                  DAG.getUNDEF(OpVT), SubVec, N->getOperand(2));
 
@@ -34144,6 +34143,30 @@ static SDValue combineInsertSubvector(SDNode *N, SelectionDAG &DAG,
     return DAG.getBitcast(OpVT, Vec256);
   }
 
+  // If we're inserting into the upper half of a 256-bit vector with a vector
+  // that was extracted from the upper half of a 256-bit vector, we should
+  // use a blend instead.
+  if (SubVec.getOpcode() == ISD::EXTRACT_SUBVECTOR && OpVT.is256BitVector() &&
+      SubVec.getOperand(0).getSimpleValueType() == OpVT &&
+      Idx == SubVec.getOperand(1) && IdxVal == OpVT.getVectorNumElements()/2) {
+
+    // Integers must be cast to 32-bit because there is only vpblendd;
+    // vpblendw can't be used for this because it has a handicapped mask.
+    // If we don't have AVX2, then cast to float. Using a wrong domain blend
+    // is still more efficient than using the wrong domain vinsertf128 that
+    // will be created by InsertSubVector().
+    MVT CastVT = OpVT;
+    if (OpVT.isInteger())
+      CastVT = Subtarget.hasAVX2() ? MVT::v8i32 : MVT::v8f32;
+
+    // The blend instruction, and therefore its mask, depend on the data type.
+    unsigned MaskVal = CastVT.getScalarSizeInBits() == 64 ? 0x0c : 0xf0;
+    SDValue Mask = DAG.getConstant(MaskVal, dl, MVT::i8);
+    Vec = DAG.getNode(X86ISD::BLENDI, dl, CastVT, DAG.getBitcast(CastVT, Vec),
+                      DAG.getBitcast(CastVT, SubVec.getOperand(0)), Mask);
+    return DAG.getBitcast(OpVT, Vec);
+  }
+
   // Fold two 16-byte or 32-byte subvector loads into one 32-byte or 64-byte
   // load:
   // (insert_subvector (insert_subvector undef, (load16 addr), 0),
diff --git a/test/CodeGen/X86/avx-vperm2x128.ll b/test/CodeGen/X86/avx-vperm2x128.ll
index f7f54a01d7f..280b218b8b5 100644
--- a/test/CodeGen/X86/avx-vperm2x128.ll
+++ b/test/CodeGen/X86/avx-vperm2x128.ll
@@ -466,8 +466,7 @@ define <4 x i64> @shuffle_v4i64_67zz(<4 x i64> %a, <4 x i64> %b) {
 ; AVX1:       ## BB#0:
 ; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],zero,zero
 ; AVX1-NEXT:    vpaddq %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX1-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: shuffle_v4i64_67zz:
diff --git a/test/CodeGen/X86/vector-shuffle-256-v16.ll b/test/CodeGen/X86/vector-shuffle-256-v16.ll
index 682a34d3cdb..eb036353eaf 100644
--- a/test/CodeGen/X86/vector-shuffle-256-v16.ll
+++ b/test/CodeGen/X86/vector-shuffle-256-v16.ll
@@ -2653,8 +2653,7 @@ define <16 x i16> @shuffle_v16i16_01_zz_02_zz_04_uu_06_07_08_09_10_11_12_13_14_1
 ; AVX1-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm0[1,1,2,3,4,5,6,7]
 ; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
 ; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0],xmm2[1],xmm1[2],xmm2[3],xmm1[4,5,6,7]
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
 ; AVX1-NEXT:    retq
 ;
 ; AVX2OR512VL-LABEL: shuffle_v16i16_01_zz_02_zz_04_uu_06_07_08_09_10_11_12_13_14_15:
diff --git a/test/CodeGen/X86/vector-shuffle-256-v32.ll b/test/CodeGen/X86/vector-shuffle-256-v32.ll
index 8b0e25ce43f..31563e82d37 100644
--- a/test/CodeGen/X86/vector-shuffle-256-v32.ll
+++ b/test/CodeGen/X86/vector-shuffle-256-v32.ll
@@ -1089,8 +1089,7 @@ define <32 x i8> @shuffle_v32i8_01_zz_02_zz_04_uu_06_07_08_09_10_11_12_13_14_15_
 ; AVX1-LABEL: shuffle_v32i8_01_zz_02_zz_04_uu_06_07_08_09_10_11_12_13_14_15_u6_17_18_19_20_21_22_23_24_25_26_27_28_29_30_31:
 ; AVX1:       # BB#0:
 ; AVX1-NEXT:    vpshufb {{.*#+}} xmm1 = xmm0[1],zero,xmm0[2],zero,xmm0[4,u,6,7,8,9,10,11,12,13,14,15]
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
 ; AVX1-NEXT:    retq
 ;
 ; AVX2OR512VL-LABEL: shuffle_v32i8_01_zz_02_zz_04_uu_06_07_08_09_10_11_12_13_14_15_u6_17_18_19_20_21_22_23_24_25_26_27_28_29_30_31: