X86: Add a couple of target-specific dag combines that turn VSELECTS into psubus...

We match the pattern "x >= y ? x-y : 0" into "subus x, y" and two special cases
if y is a constant. DAGCombiner canonicalizes those so we first have to undo the
canonicalization for those cases. The pattern occurs in gzip when the loop
vectorizer is enabled. Part of PR14613.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@170273 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index d4ee985ca6593c1490f8e69a2c6f16e1a5ff194f..70089fcc65887efa4907a0b0fd68ab45e7858a1b 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -10097,6 +10097,14 @@ static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
     return DAG.getNode(X86ISD::PMULUDQ, dl, Op.getValueType(),
                        Op.getOperand(1), Op.getOperand(2));
 
+  // SSE2/AVX2 sub with unsigned saturation intrinsics
+  case Intrinsic::x86_sse2_psubus_b:
+  case Intrinsic::x86_sse2_psubus_w:
+  case Intrinsic::x86_avx2_psubus_b:
+  case Intrinsic::x86_avx2_psubus_w:
+    return DAG.getNode(X86ISD::SUBUS, dl, Op.getValueType(),
+                       Op.getOperand(1), Op.getOperand(2));
+
   // SSE3/AVX horizontal add/sub intrinsics
   case Intrinsic::x86_sse3_hadd_ps:
   case Intrinsic::x86_sse3_hadd_pd:
@@ -11961,6 +11969,7 @@ const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
   case X86ISD::PSIGN:              return "X86ISD::PSIGN";
   case X86ISD::BLENDV:             return "X86ISD::BLENDV";
   case X86ISD::BLENDI:             return "X86ISD::BLENDI";
+  case X86ISD::SUBUS:              return "X86ISD::SUBUS";
   case X86ISD::HADD:               return "X86ISD::HADD";
   case X86ISD::HSUB:               return "X86ISD::HSUB";
   case X86ISD::FHADD:              return "X86ISD::FHADD";
@@ -14913,6 +14922,65 @@ static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
     }
   }
 
+  // Match VSELECTs into subs with unsigned saturation.
+  if (!DCI.isBeforeLegalize() &&
+      N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
+      // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
+      ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
+       (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
+    ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
+
+    // Check if one of the arms of the VSELECT is a zero vector. If it's on the
+    // left side invert the predicate to simplify logic below.
+    SDValue Other;
+    if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
+      Other = RHS;
+      CC = ISD::getSetCCInverse(CC, true);
+    } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
+      Other = LHS;
+    }
+
+    if (Other.getNode() && Other->getNumOperands() == 2 &&
+        DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
+      SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
+      SDValue CondRHS = Cond->getOperand(1);
+
+      // Look for a general sub with unsigned saturation first.
+      // x >= y ? x-y : 0 --> subus x, y
+      // x >  y ? x-y : 0 --> subus x, y
+      if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
+          Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
+        return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
+
+      // If the RHS is a constant we have to reverse the const canonicalization.
+      // x > C-1 ? x+-C : 0 --> subus x, C
+      if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
+          isSplatVector(CondRHS.getNode()) && isSplatVector(OpRHS.getNode())) {
+        APInt A = cast<ConstantSDNode>(OpRHS.getOperand(0))->getAPIntValue();
+        if (CondRHS.getConstantOperandVal(0) == -A-1) {
+          SmallVector<SDValue, 32> V(VT.getVectorNumElements(),
+                                     DAG.getConstant(-A, VT.getScalarType()));
+          return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS,
+                             DAG.getNode(ISD::BUILD_VECTOR, DL, VT,
+                                         V.data(), V.size()));
+        }
+      }
+
+      // Another special case: If C was a sign bit, the sub has been
+      // canonicalized into a xor.
+      // FIXME: Would it be better to use ComputeMaskedBits to determine whether
+      //        it's safe to decanonicalize the xor?
+      // x s< 0 ? x^C : 0 --> subus x, C
+      if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
+          ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
+          isSplatVector(OpRHS.getNode())) {
+        APInt A = cast<ConstantSDNode>(OpRHS.getOperand(0))->getAPIntValue();
+        if (A.isSignBit())
+          return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
+      }
+    }
+  }
+
   // If we know that this node is legal then we know that it is going to be
   // matched by one of the SSE/AVX BLEND instructions. These instructions only
   // depend on the highest bit in each word. Try to use SimplifyDemandedBits

diff --git a/lib/Target/X86/X86ISelLowering.h b/lib/Target/X86/X86ISelLowering.h
index cbe83a7c244dcae3fc6682fa04005dd297467f0c..60f34099dde883a2012fca447ccbd1d1ae762571 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.h
+++ b/lib/Target/X86/X86ISelLowering.h
@@ -182,6 +182,9 @@ namespace llvm {
       /// BLENDI - Blend where the selector is an immediate.
       BLENDI,
 
+      // SUBUS - Integer sub with unsigned saturation.
+      SUBUS,
+
       /// HADD - Integer horizontal add.
       HADD,
 

diff --git a/lib/Target/X86/X86InstrFragmentsSIMD.td b/lib/Target/X86/X86InstrFragmentsSIMD.td
index 09ab995166981a8def6d597ced2a34a1f156a350..7d16d2741d98e5a09d5bc037c359d1c9156b1edc 100644 (file)
--- a/lib/Target/X86/X86InstrFragmentsSIMD.td
+++ b/lib/Target/X86/X86InstrFragmentsSIMD.td
@@ -128,6 +128,7 @@ def X86vsrai   : SDNode<"X86ISD::VSRAI", SDTIntShiftOp>;
 def SDTX86CmpPTest : SDTypeProfile<1, 2, [SDTCisVT<0, i32>,
                                           SDTCisVec<1>,
                                           SDTCisSameAs<2, 1>]>;
+def X86subus   : SDNode<"X86ISD::SUBUS", SDTIntBinOp>;
 def X86ptest   : SDNode<"X86ISD::PTEST", SDTX86CmpPTest>;
 def X86testp   : SDNode<"X86ISD::TESTP", SDTX86CmpPTest>;
 

diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
index 1912a936cecc0884780f493ec04eb621d0378d66..54032fe97f63ec6783ee9dfee84766ec91054121 100644 (file)
--- a/lib/Target/X86/X86InstrSSE.td
+++ b/lib/Target/X86/X86InstrSSE.td
@@ -3724,6 +3724,12 @@ defm VPSUBD : PDI_binop_rm<0xFA, "vpsubd", sub, v4i32, VR128, memopv2i64,
                             i128mem, SSE_INTALU_ITINS_P, 0, 0>, VEX_4V;
 defm VPSUBQ : PDI_binop_rm<0xFB, "vpsubq", sub, v2i64, VR128, memopv2i64,
                             i128mem, SSE_INTALUQ_ITINS_P, 0, 0>, VEX_4V;
+defm VPSUBUSB : PDI_binop_rm<0xD8, "vpsubusb", X86subus, v16i8, VR128,
+                             memopv2i64, i128mem, SSE_INTALU_ITINS_P, 0, 0>,
+                             VEX_4V;
+defm VPSUBUSW : PDI_binop_rm<0xD9, "vpsubusw", X86subus, v8i16, VR128,
+                             memopv2i64, i128mem, SSE_INTALU_ITINS_P, 0, 0>,
+                             VEX_4V;
 defm VPMULUDQ : PDI_binop_rm2<0xF4, "vpmuludq", X86pmuludq, v2i64, v4i32, VR128,
                               memopv2i64, i128mem, SSE_INTMUL_ITINS_P, 1, 0>,
                               VEX_4V;
@@ -3735,12 +3741,6 @@ defm VPSUBSB  : PDI_binop_rm_int<0xE8, "vpsubsb" , int_x86_sse2_psubs_b,
 defm VPSUBSW  : PDI_binop_rm_int<0xE9, "vpsubsw" , int_x86_sse2_psubs_w,
                                  VR128, memopv2i64, i128mem,
                                  SSE_INTALU_ITINS_P, 0, 0>, VEX_4V;
-defm VPSUBUSB : PDI_binop_rm_int<0xD8, "vpsubusb", int_x86_sse2_psubus_b,
-                                 VR128, memopv2i64, i128mem,
-                                 SSE_INTALU_ITINS_P, 0, 0>, VEX_4V;
-defm VPSUBUSW : PDI_binop_rm_int<0xD9, "vpsubusw", int_x86_sse2_psubus_w,
-                                 VR128, memopv2i64, i128mem,
-                                 SSE_INTALU_ITINS_P, 0, 0>, VEX_4V;
 defm VPADDSB  : PDI_binop_rm_int<0xEC, "vpaddsb" , int_x86_sse2_padds_b,
                                  VR128, memopv2i64, i128mem,
                                  SSE_INTALU_ITINS_P, 1, 0>, VEX_4V;
@@ -3804,6 +3804,12 @@ defm VPSUBDY  : PDI_binop_rm<0xFA, "vpsubd", sub, v8i32, VR256, memopv4i64,
                              i256mem, SSE_INTALU_ITINS_P, 0, 0>, VEX_4V, VEX_L;
 defm VPSUBQY  : PDI_binop_rm<0xFB, "vpsubq", sub, v4i64, VR256, memopv4i64,
                              i256mem, SSE_INTALUQ_ITINS_P, 0, 0>, VEX_4V, VEX_L;
+defm VPSUBUSBY : PDI_binop_rm<0xD8, "vpsubusb", X86subus, v32i8, VR256,
+                              memopv4i64, i256mem, SSE_INTALU_ITINS_P, 0, 0>,
+                              VEX_4V, VEX_L;
+defm VPSUBUSWY : PDI_binop_rm<0xD9, "vpsubusw", X86subus, v16i16, VR256,
+                              memopv4i64, i256mem, SSE_INTALU_ITINS_P, 0, 0>,
+                              VEX_4V, VEX_L;
 defm VPMULUDQY : PDI_binop_rm2<0xF4, "vpmuludq", X86pmuludq, v4i64, v8i32,
                                VR256, memopv4i64, i256mem,
                                SSE_INTMUL_ITINS_P, 1, 0>, VEX_4V, VEX_L;
@@ -3815,12 +3821,6 @@ defm VPSUBSBY  : PDI_binop_rm_int<0xE8, "vpsubsb" , int_x86_avx2_psubs_b,
 defm VPSUBSWY  : PDI_binop_rm_int<0xE9, "vpsubsw" , int_x86_avx2_psubs_w,
                                   VR256, memopv4i64, i256mem,
                                   SSE_INTALU_ITINS_P, 0, 0>, VEX_4V, VEX_L;
-defm VPSUBUSBY : PDI_binop_rm_int<0xD8, "vpsubusb", int_x86_avx2_psubus_b,
-                                  VR256, memopv4i64, i256mem,
-                                  SSE_INTALU_ITINS_P, 0, 0>, VEX_4V, VEX_L;
-defm VPSUBUSWY : PDI_binop_rm_int<0xD9, "vpsubusw", int_x86_avx2_psubus_w,
-                                  VR256, memopv4i64, i256mem,
-                                  SSE_INTALU_ITINS_P, 0, 0>, VEX_4V, VEX_L;
 defm VPADDSBY  : PDI_binop_rm_int<0xEC, "vpaddsb" , int_x86_avx2_padds_b,
                                   VR256, memopv4i64, i256mem,
                                   SSE_INTALU_ITINS_P, 1, 0>, VEX_4V, VEX_L;
@@ -3884,6 +3884,10 @@ defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32, VR128, memopv2i64,
                           i128mem, SSE_INTALU_ITINS_P>;
 defm PSUBQ : PDI_binop_rm<0xFB, "psubq", sub, v2i64, VR128, memopv2i64,
                           i128mem, SSE_INTALUQ_ITINS_P>;
+defm PSUBUSB : PDI_binop_rm<0xD8, "psubusb", X86subus, v16i8, VR128, memopv2i64,
+                            i128mem, SSE_INTALU_ITINS_P>;
+defm PSUBUSW : PDI_binop_rm<0xD9, "psubusw", X86subus, v8i16, VR128, memopv2i64,
+                            i128mem, SSE_INTALU_ITINS_P>;
 defm PMULUDQ : PDI_binop_rm2<0xF4, "pmuludq", X86pmuludq, v2i64, v4i32, VR128,
                              memopv2i64, i128mem, SSE_INTMUL_ITINS_P, 1>;
 
@@ -3894,12 +3898,6 @@ defm PSUBSB  : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b,
 defm PSUBSW  : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w,
                                 VR128, memopv2i64, i128mem,
                                 SSE_INTALU_ITINS_P>;
-defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b,
-                                VR128, memopv2i64, i128mem,
-                                SSE_INTALU_ITINS_P>;
-defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w,
-                                VR128, memopv2i64, i128mem,
-                                SSE_INTALU_ITINS_P>;
 defm PADDSB  : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b,
                                 VR128, memopv2i64, i128mem,
                                 SSE_INTALU_ITINS_P, 1>;

diff --git a/test/CodeGen/X86/psubus.ll b/test/CodeGen/X86/psubus.ll
new file mode 100644 (file)
index 0000000..aff4afb
--- /dev/null
+++ b/test/CodeGen/X86/psubus.ll
@@ -0,0 +1,340 @@
+; RUN: llc -mcpu=core2 < %s | FileCheck %s -check-prefix=SSE2
+; RUN: llc -mcpu=corei7-avx < %s | FileCheck %s -check-prefix=AVX1
+; RUN: llc -mcpu=core-avx2 < %s | FileCheck %s -check-prefix=AVX2
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-macosx10.8.0"
+
+define void @test1(i16* nocapture %head) nounwind {
+vector.ph:
+  br label %vector.body
+
+vector.body:                                      ; preds = %vector.body, %vector.ph
+  %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
+  %0 = getelementptr inbounds i16* %head, i64 %index
+  %1 = bitcast i16* %0 to <8 x i16>*
+  %2 = load <8 x i16>* %1, align 2
+  %3 = icmp slt <8 x i16> %2, zeroinitializer
+  %4 = xor <8 x i16> %2, <i16 -32768, i16 -32768, i16 -32768, i16 -32768, i16 -32768, i16 -32768, i16 -32768, i16 -32768>
+  %5 = select <8 x i1> %3, <8 x i16> %4, <8 x i16> zeroinitializer
+  store <8 x i16> %5, <8 x i16>* %1, align 2
+  %index.next = add i64 %index, 8
+  %6 = icmp eq i64 %index.next, 16384
+  br i1 %6, label %for.end, label %vector.body
+
+for.end:                                          ; preds = %vector.body
+  ret void
+
+; SSE2: @test1
+; SSE2: psubusw LCPI0_0(%rip), %xmm0
+
+; AVX1: @test1
+; AVX1: vpsubusw LCPI0_0(%rip), %xmm0, %xmm0
+
+; AVX2: @test1
+; AVX2: vpsubusw LCPI0_0(%rip), %xmm0, %xmm0
+}
+
+define void @test2(i16* nocapture %head) nounwind {
+vector.ph:
+  br label %vector.body
+
+vector.body:                                      ; preds = %vector.body, %vector.ph
+  %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
+  %0 = getelementptr inbounds i16* %head, i64 %index
+  %1 = bitcast i16* %0 to <8 x i16>*
+  %2 = load <8 x i16>* %1, align 2
+  %3 = icmp ugt <8 x i16> %2, <i16 32766, i16 32766, i16 32766, i16 32766, i16 32766, i16 32766, i16 32766, i16 32766>
+  %4 = add <8 x i16> %2, <i16 -32767, i16 -32767, i16 -32767, i16 -32767, i16 -32767, i16 -32767, i16 -32767, i16 -32767>
+  %5 = select <8 x i1> %3, <8 x i16> %4, <8 x i16> zeroinitializer
+  store <8 x i16> %5, <8 x i16>* %1, align 2
+  %index.next = add i64 %index, 8
+  %6 = icmp eq i64 %index.next, 16384
+  br i1 %6, label %for.end, label %vector.body
+
+for.end:                                          ; preds = %vector.body
+  ret void
+
+; SSE2: @test2
+; SSE2: psubusw LCPI1_0(%rip), %xmm0
+
+; AVX1: @test2
+; AVX1: vpsubusw LCPI1_0(%rip), %xmm0, %xmm0
+
+; AVX2: @test2
+; AVX2: vpsubusw LCPI1_0(%rip), %xmm0, %xmm0
+}
+
+define void @test3(i16* nocapture %head, i16 zeroext %w) nounwind {
+vector.ph:
+  %0 = insertelement <8 x i16> undef, i16 %w, i32 0
+  %broadcast15 = shufflevector <8 x i16> %0, <8 x i16> undef, <8 x i32> zeroinitializer
+  br label %vector.body
+
+vector.body:                                      ; preds = %vector.body, %vector.ph
+  %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
+  %1 = getelementptr inbounds i16* %head, i64 %index
+  %2 = bitcast i16* %1 to <8 x i16>*
+  %3 = load <8 x i16>* %2, align 2
+  %4 = icmp ult <8 x i16> %3, %broadcast15
+  %5 = sub <8 x i16> %3, %broadcast15
+  %6 = select <8 x i1> %4, <8 x i16> zeroinitializer, <8 x i16> %5
+  store <8 x i16> %6, <8 x i16>* %2, align 2
+  %index.next = add i64 %index, 8
+  %7 = icmp eq i64 %index.next, 16384
+  br i1 %7, label %for.end, label %vector.body
+
+for.end:                                          ; preds = %vector.body
+  ret void
+
+; SSE2: @test3
+; SSE2: psubusw %xmm0, %xmm1
+
+; AVX1: @test3
+; AVX1: vpsubusw %xmm0, %xmm1, %xmm1
+
+; AVX2: @test3
+; AVX2: vpsubusw %xmm0, %xmm1, %xmm1
+}
+
+define void @test4(i8* nocapture %head) nounwind {
+vector.ph:
+  br label %vector.body
+
+vector.body:                                      ; preds = %vector.body, %vector.ph
+  %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
+  %0 = getelementptr inbounds i8* %head, i64 %index
+  %1 = bitcast i8* %0 to <16 x i8>*
+  %2 = load <16 x i8>* %1, align 1
+  %3 = icmp slt <16 x i8> %2, zeroinitializer
+  %4 = xor <16 x i8> %2, <i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128>
+  %5 = select <16 x i1> %3, <16 x i8> %4, <16 x i8> zeroinitializer
+  store <16 x i8> %5, <16 x i8>* %1, align 1
+  %index.next = add i64 %index, 16
+  %6 = icmp eq i64 %index.next, 16384
+  br i1 %6, label %for.end, label %vector.body
+
+for.end:                                          ; preds = %vector.body
+  ret void
+
+; SSE2: @test4
+; SSE2: psubusb LCPI3_0(%rip), %xmm0
+
+; AVX1: @test4
+; AVX1: vpsubusb LCPI3_0(%rip), %xmm0, %xmm0
+
+; AVX2: @test4
+; AVX2: vpsubusb LCPI3_0(%rip), %xmm0, %xmm0
+}
+
+define void @test5(i8* nocapture %head) nounwind {
+vector.ph:
+  br label %vector.body
+
+vector.body:                                      ; preds = %vector.body, %vector.ph
+  %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
+  %0 = getelementptr inbounds i8* %head, i64 %index
+  %1 = bitcast i8* %0 to <16 x i8>*
+  %2 = load <16 x i8>* %1, align 1
+  %3 = icmp ugt <16 x i8> %2, <i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126>
+  %4 = add <16 x i8> %2, <i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127>
+  %5 = select <16 x i1> %3, <16 x i8> %4, <16 x i8> zeroinitializer
+  store <16 x i8> %5, <16 x i8>* %1, align 1
+  %index.next = add i64 %index, 16
+  %6 = icmp eq i64 %index.next, 16384
+  br i1 %6, label %for.end, label %vector.body
+
+for.end:                                          ; preds = %vector.body
+  ret void
+
+; SSE2: @test5
+; SSE2: psubusb LCPI4_0(%rip), %xmm0
+
+; AVX1: @test5
+; AVX1: vpsubusb LCPI4_0(%rip), %xmm0, %xmm0
+
+; AVX2: @test5
+; AVX2: vpsubusb LCPI4_0(%rip), %xmm0, %xmm0
+}
+
+define void @test6(i8* nocapture %head, i8 zeroext %w) nounwind {
+vector.ph:
+  %0 = insertelement <16 x i8> undef, i8 %w, i32 0
+  %broadcast15 = shufflevector <16 x i8> %0, <16 x i8> undef, <16 x i32> zeroinitializer
+  br label %vector.body
+
+vector.body:                                      ; preds = %vector.body, %vector.ph
+  %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
+  %1 = getelementptr inbounds i8* %head, i64 %index
+  %2 = bitcast i8* %1 to <16 x i8>*
+  %3 = load <16 x i8>* %2, align 1
+  %4 = icmp ult <16 x i8> %3, %broadcast15
+  %5 = sub <16 x i8> %3, %broadcast15
+  %6 = select <16 x i1> %4, <16 x i8> zeroinitializer, <16 x i8> %5
+  store <16 x i8> %6, <16 x i8>* %2, align 1
+  %index.next = add i64 %index, 16
+  %7 = icmp eq i64 %index.next, 16384
+  br i1 %7, label %for.end, label %vector.body
+
+for.end:                                          ; preds = %vector.body
+  ret void
+
+; SSE2: @test6
+; SSE2: psubusb %xmm0, %xmm1
+
+; AVX1: @test6
+; AVX1: vpsubusb %xmm0, %xmm1, %xmm1
+
+; AVX2: @test6
+; AVX2: vpsubusb %xmm0, %xmm1, %xmm1
+}
+
+define void @test7(i16* nocapture %head) nounwind {
+vector.ph:
+  br label %vector.body
+
+vector.body:                                      ; preds = %vector.body, %vector.ph
+  %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
+  %0 = getelementptr inbounds i16* %head, i64 %index
+  %1 = bitcast i16* %0 to <16 x i16>*
+  %2 = load <16 x i16>* %1, align 2
+  %3 = icmp slt <16 x i16> %2, zeroinitializer
+  %4 = xor <16 x i16> %2, <i16 -32768, i16 -32768, i16 -32768, i16 -32768, i16 -32768, i16 -32768, i16 -32768, i16 -32768, i16 -32768, i16 -32768, i16 -32768, i16 -32768, i16 -32768, i16 -32768, i16 -32768, i16 -32768>
+  %5 = select <16 x i1> %3, <16 x i16> %4, <16 x i16> zeroinitializer
+  store <16 x i16> %5, <16 x i16>* %1, align 2
+  %index.next = add i64 %index, 8
+  %6 = icmp eq i64 %index.next, 16384
+  br i1 %6, label %for.end, label %vector.body
+
+for.end:                                          ; preds = %vector.body
+  ret void
+
+; AVX2: @test7
+; AVX2: vpsubusw LCPI6_0(%rip), %ymm0, %ymm0
+}
+
+define void @test8(i16* nocapture %head) nounwind {
+vector.ph:
+  br label %vector.body
+
+vector.body:                                      ; preds = %vector.body, %vector.ph
+  %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
+  %0 = getelementptr inbounds i16* %head, i64 %index
+  %1 = bitcast i16* %0 to <16 x i16>*
+  %2 = load <16 x i16>* %1, align 2
+  %3 = icmp ugt <16 x i16> %2, <i16 32766, i16 32766, i16 32766, i16 32766, i16 32766, i16 32766, i16 32766, i16 32766, i16 32766, i16 32766, i16 32766, i16 32766, i16 32766, i16 32766, i16 32766, i16 32766>
+  %4 = add <16 x i16> %2, <i16 -32767, i16 -32767, i16 -32767, i16 -32767, i16 -32767, i16 -32767, i16 -32767, i16 -32767, i16 -32767, i16 -32767, i16 -32767, i16 -32767, i16 -32767, i16 -32767, i16 -32767, i16 -32767>
+  %5 = select <16 x i1> %3, <16 x i16> %4, <16 x i16> zeroinitializer
+  store <16 x i16> %5, <16 x i16>* %1, align 2
+  %index.next = add i64 %index, 8
+  %6 = icmp eq i64 %index.next, 16384
+  br i1 %6, label %for.end, label %vector.body
+
+for.end:                                          ; preds = %vector.body
+  ret void
+
+; AVX2: @test8
+; AVX2: vpsubusw LCPI7_0(%rip), %ymm0, %ymm0
+}
+
+define void @test9(i16* nocapture %head, i16 zeroext %w) nounwind {
+vector.ph:
+  %0 = insertelement <16 x i16> undef, i16 %w, i32 0
+  %broadcast15 = shufflevector <16 x i16> %0, <16 x i16> undef, <16 x i32> zeroinitializer
+  br label %vector.body
+
+vector.body:                                      ; preds = %vector.body, %vector.ph
+  %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
+  %1 = getelementptr inbounds i16* %head, i64 %index
+  %2 = bitcast i16* %1 to <16 x i16>*
+  %3 = load <16 x i16>* %2, align 2
+  %4 = icmp ult <16 x i16> %3, %broadcast15
+  %5 = sub <16 x i16> %3, %broadcast15
+  %6 = select <16 x i1> %4, <16 x i16> zeroinitializer, <16 x i16> %5
+  store <16 x i16> %6, <16 x i16>* %2, align 2
+  %index.next = add i64 %index, 8
+  %7 = icmp eq i64 %index.next, 16384
+  br i1 %7, label %for.end, label %vector.body
+
+for.end:                                          ; preds = %vector.body
+  ret void
+
+
+; AVX2: @test9
+; AVX2: vpsubusw %ymm0, %ymm1, %ymm1
+}
+
+define void @test10(i8* nocapture %head) nounwind {
+vector.ph:
+  br label %vector.body
+
+vector.body:                                      ; preds = %vector.body, %vector.ph
+  %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
+  %0 = getelementptr inbounds i8* %head, i64 %index
+  %1 = bitcast i8* %0 to <32 x i8>*
+  %2 = load <32 x i8>* %1, align 1
+  %3 = icmp slt <32 x i8> %2, zeroinitializer
+  %4 = xor <32 x i8> %2, <i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128, i8 -128>
+  %5 = select <32 x i1> %3, <32 x i8> %4, <32 x i8> zeroinitializer
+  store <32 x i8> %5, <32 x i8>* %1, align 1
+  %index.next = add i64 %index, 16
+  %6 = icmp eq i64 %index.next, 16384
+  br i1 %6, label %for.end, label %vector.body
+
+for.end:                                          ; preds = %vector.body
+  ret void
+
+
+; AVX2: @test10
+; AVX2: vpsubusb LCPI9_0(%rip), %ymm0, %ymm0
+}
+
+define void @test11(i8* nocapture %head) nounwind {
+vector.ph:
+  br label %vector.body
+
+vector.body:                                      ; preds = %vector.body, %vector.ph
+  %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
+  %0 = getelementptr inbounds i8* %head, i64 %index
+  %1 = bitcast i8* %0 to <32 x i8>*
+  %2 = load <32 x i8>* %1, align 1
+  %3 = icmp ugt <32 x i8> %2, <i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126, i8 126>
+  %4 = add <32 x i8> %2, <i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127, i8 -127>
+  %5 = select <32 x i1> %3, <32 x i8> %4, <32 x i8> zeroinitializer
+  store <32 x i8> %5, <32 x i8>* %1, align 1
+  %index.next = add i64 %index, 16
+  %6 = icmp eq i64 %index.next, 16384
+  br i1 %6, label %for.end, label %vector.body
+
+for.end:                                          ; preds = %vector.body
+  ret void
+
+; AVX2: @test11
+; AVX2: vpsubusb LCPI10_0(%rip), %ymm0, %ymm0
+}
+
+define void @test12(i8* nocapture %head, i8 zeroext %w) nounwind {
+vector.ph:
+  %0 = insertelement <32 x i8> undef, i8 %w, i32 0
+  %broadcast15 = shufflevector <32 x i8> %0, <32 x i8> undef, <32 x i32> zeroinitializer
+  br label %vector.body
+
+vector.body:                                      ; preds = %vector.body, %vector.ph
+  %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
+  %1 = getelementptr inbounds i8* %head, i64 %index
+  %2 = bitcast i8* %1 to <32 x i8>*
+  %3 = load <32 x i8>* %2, align 1
+  %4 = icmp ult <32 x i8> %3, %broadcast15
+  %5 = sub <32 x i8> %3, %broadcast15
+  %6 = select <32 x i1> %4, <32 x i8> zeroinitializer, <32 x i8> %5
+  store <32 x i8> %6, <32 x i8>* %2, align 1
+  %index.next = add i64 %index, 16
+  %7 = icmp eq i64 %index.next, 16384
+  br i1 %7, label %for.end, label %vector.body
+
+for.end:                                          ; preds = %vector.body
+  ret void
+
+; AVX2: @test12
+; AVX2: vpsubusb %ymm0, %ymm1, %ymm1
+}