[X86][SSE] Add general memory folding for (V)INSERTPS instruction

This patch improves the memory folding of the inserted float element for the (V)INSERTPS instruction.

The existing implementation occurs in the DAGCombiner and relies on the narrowing of a whole vector load into a scalar load (and then converted into a vector) to (hopefully) allow folding to occur later on. Not only has this proven problematic for debug builds, it also prevents other memory folds (notably stack reloads) from happening.

This patch removes the old implementation and moves the folding code to the X86 foldMemoryOperand handler. A new private 'special case' function - foldMemoryOperandCustom - has been added to deal with memory folding of instructions that can't just use the lookup tables - (V)INSERTPS is the first of several that could be done.

It also tweaks the memory operand folding code with an additional pointer offset that allows existing memory addresses to be modified, in this case to convert the vector address to the explicit address of the scalar element that will be inserted.

Unlike the previous implementation we now set the insertion source index to zero, although this is ignored for the (V)INSERTPSrm version, anything that relied on shuffle decodes (such as unfolding of insertps loads) was incorrectly calculating the source address - I've added a test for this at insertps-unfold-load-bug.ll

Differential Revision: http://reviews.llvm.org/D13988

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252074 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 3434f8c7f0bf5bd5b308455f2191f57db4677365..685067a28c60de023e74ff418e7878ddabfd4738 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -26174,52 +26174,6 @@ static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
   return SDValue();
 }
 
-static SDValue NarrowVectorLoadToElement(LoadSDNode *Load, unsigned Index,
-                                         SelectionDAG &DAG) {
-  SDLoc dl(Load);
-  MVT VT = Load->getSimpleValueType(0);
-  MVT EVT = VT.getVectorElementType();
-  SDValue Addr = Load->getOperand(1);
-  SDValue NewAddr = DAG.getNode(
-      ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
-      DAG.getConstant(Index * EVT.getStoreSize(), dl,
-                      Addr.getSimpleValueType()));
-
-  SDValue NewLoad =
-      DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
-                  DAG.getMachineFunction().getMachineMemOperand(
-                      Load->getMemOperand(), 0, EVT.getStoreSize()));
-  return NewLoad;
-}
-
-static SDValue PerformINSERTPSCombine(SDNode *N, SelectionDAG &DAG,
-                                      const X86Subtarget *Subtarget) {
-  SDLoc dl(N);
-  MVT VT = N->getOperand(1)->getSimpleValueType(0);
-  assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
-         "X86insertps is only defined for v4x32");
-
-  SDValue Ld = N->getOperand(1);
-  if (MayFoldLoad(Ld)) {
-    // Extract the countS bits from the immediate so we can get the proper
-    // address when narrowing the vector load to a specific element.
-    // When the second source op is a memory address, insertps doesn't use
-    // countS and just gets an f32 from that address.
-    unsigned DestIndex =
-        cast<ConstantSDNode>(N->getOperand(2))->getZExtValue() >> 6;
-
-    Ld = NarrowVectorLoadToElement(cast<LoadSDNode>(Ld), DestIndex, DAG);
-
-    // Create this as a scalar to vector to match the instruction pattern.
-    SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld);
-    // countS bits are ignored when loading from memory on insertps, which
-    // means we don't need to explicitly set them to 0.
-    return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0),
-                       LoadScalarToVector, N->getOperand(2));
-  }
-  return SDValue();
-}
-
 static SDValue PerformBLENDICombine(SDNode *N, SelectionDAG &DAG) {
   SDValue V0 = N->getOperand(0);
   SDValue V1 = N->getOperand(1);
@@ -26685,11 +26639,6 @@ SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
   case X86ISD::VPERM2X128:
   case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
   case ISD::FMA:            return PerformFMACombine(N, DAG, Subtarget);
-  case X86ISD::INSERTPS: {
-    if (getTargetMachine().getOptLevel() > CodeGenOpt::None)
-      return PerformINSERTPSCombine(N, DAG, Subtarget);
-    break;
-  }
   case X86ISD::BLENDI:    return PerformBLENDICombine(N, DAG);
   }
 

diff --git a/lib/Target/X86/X86InstrInfo.cpp b/lib/Target/X86/X86InstrInfo.cpp
index 81c61d997c025d665c6d1f67431722922d899992..a5b3a12e9102c579146c69b2b1702d585d1dac1b 100644 (file)
--- a/lib/Target/X86/X86InstrInfo.cpp
+++ b/lib/Target/X86/X86InstrInfo.cpp
@@ -4876,12 +4876,35 @@ bool X86InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
   return false;
 }
 
-static void addOperands(MachineInstrBuilder &MIB, ArrayRef<MachineOperand> MOs) {
+static void addOperands(MachineInstrBuilder &MIB, ArrayRef<MachineOperand> MOs,
+                        int PtrOffset = 0) {
   unsigned NumAddrOps = MOs.size();
-  for (unsigned i = 0; i != NumAddrOps; ++i)
-    MIB.addOperand(MOs[i]);
-  if (NumAddrOps < 4) // FrameIndex only
-    addOffset(MIB, 0);
+
+  if (NumAddrOps < 4) {
+    // FrameIndex only - add an immediate offset (whether its zero or not).
+    for (unsigned i = 0; i != NumAddrOps; ++i)
+      MIB.addOperand(MOs[i]);
+    addOffset(MIB, PtrOffset);
+  } else {
+    // General Memory Addressing - we need to add any offset to an existing
+    // offset.
+    assert(MOs.size() == 5 && "Unexpected memory operand list length");
+    for (unsigned i = 0; i != NumAddrOps; ++i) {
+      const MachineOperand &MO = MOs[i];
+      if (i == 3 && PtrOffset != 0) {
+        assert((MO.isImm() || MO.isGlobal()) &&
+               "Unexpected memory operand type");
+        if (MO.isImm()) {
+          MIB.addImm(MO.getImm() + PtrOffset);
+        } else {
+          MIB.addGlobalAddress(MO.getGlobal(), MO.getOffset() + PtrOffset,
+                               MO.getTargetFlags());
+        }
+      } else {
+        MIB.addOperand(MO);
+      }
+    }
+  }
 }
 
 static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
@@ -4916,7 +4939,8 @@ static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
 static MachineInstr *FuseInst(MachineFunction &MF, unsigned Opcode,
                               unsigned OpNo, ArrayRef<MachineOperand> MOs,
                               MachineBasicBlock::iterator InsertPt,
-                              MachineInstr *MI, const TargetInstrInfo &TII) {
+                              MachineInstr *MI, const TargetInstrInfo &TII,
+                              int PtrOffset = 0) {
   // Omit the implicit operands, something BuildMI can't do.
   MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
                                               MI->getDebugLoc(), true);
@@ -4926,7 +4950,7 @@ static MachineInstr *FuseInst(MachineFunction &MF, unsigned Opcode,
     MachineOperand &MO = MI->getOperand(i);
     if (i == OpNo) {
       assert(MO.isReg() && "Expected to fold into reg operand!");
-      addOperands(MIB, MOs);
+      addOperands(MIB, MOs, PtrOffset);
     } else {
       MIB.addOperand(MO);
     }
@@ -4948,6 +4972,40 @@ static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
   return MIB.addImm(0);
 }
 
+MachineInstr *X86InstrInfo::foldMemoryOperandCustom(
+    MachineFunction &MF, MachineInstr *MI, unsigned OpNum,
+    ArrayRef<MachineOperand> MOs, MachineBasicBlock::iterator InsertPt,
+    unsigned Size, unsigned Align) const {
+  switch (MI->getOpcode()) {
+  case X86::INSERTPSrr:
+  case X86::VINSERTPSrr:
+    // Attempt to convert the load of inserted vector into a fold load
+    // of a single float.
+    if (OpNum == 2) {
+      unsigned Imm = MI->getOperand(MI->getNumOperands() - 1).getImm();
+      unsigned ZMask = Imm & 15;
+      unsigned DstIdx = (Imm >> 4) & 3;
+      unsigned SrcIdx = (Imm >> 6) & 3;
+
+      unsigned RCSize = getRegClass(MI->getDesc(), OpNum, &RI, MF)->getSize();
+      if (Size <= RCSize && 4 <= Align) {
+        int PtrOffset = SrcIdx * 4;
+        unsigned NewImm = (DstIdx << 4) | ZMask;
+        unsigned NewOpCode =
+            (MI->getOpcode() == X86::VINSERTPSrr ? X86::VINSERTPSrm
+                                                 : X86::INSERTPSrm);
+        MachineInstr *NewMI =
+            FuseInst(MF, NewOpCode, OpNum, MOs, InsertPt, MI, *this, PtrOffset);
+        NewMI->getOperand(NewMI->getNumOperands() - 1).setImm(NewImm);
+        return NewMI;
+      }
+    }
+    break;
+  };
+
+  return nullptr;
+}
+
 MachineInstr *X86InstrInfo::foldMemoryOperandImpl(
     MachineFunction &MF, MachineInstr *MI, unsigned OpNum,
     ArrayRef<MachineOperand> MOs, MachineBasicBlock::iterator InsertPt,
@@ -4977,6 +5035,12 @@ MachineInstr *X86InstrInfo::foldMemoryOperandImpl(
     return nullptr;
 
   MachineInstr *NewMI = nullptr;
+
+  // Attempt to fold any custom cases we have.
+  if (NewMI =
+          foldMemoryOperandCustom(MF, MI, OpNum, MOs, InsertPt, Size, Align))
+    return NewMI;
+
   // Folding a memory location into the two-address part of a two-address
   // instruction is different than folding it other places.  It requires
   // replacing the *two* registers with the memory location.

diff --git a/lib/Target/X86/X86InstrInfo.h b/lib/Target/X86/X86InstrInfo.h
index e83a10f3b81c5eb987a48dfdb89bcd8ba2947f18..3f6b030c442da8630438968266e0e00e3d94daf9 100644 (file)
--- a/lib/Target/X86/X86InstrInfo.h
+++ b/lib/Target/X86/X86InstrInfo.h
@@ -512,6 +512,14 @@ private:
                                               MachineBasicBlock::iterator &MBBI,
                                               LiveVariables *LV) const;
 
+  /// Handles memory folding for special case instructions, for instance those
+  /// requiring custom manipulation of the address.
+  MachineInstr *foldMemoryOperandCustom(MachineFunction &MF, MachineInstr *MI,
+                                        unsigned OpNum,
+                                        ArrayRef<MachineOperand> MOs,
+                                        MachineBasicBlock::iterator InsertPt,
+                                        unsigned Size, unsigned Align) const;
+
   /// isFrameOperand - Return true and the FrameIndex if the specified
   /// operand and follow operands form a reference to the stack frame.
   bool isFrameOperand(const MachineInstr *MI, unsigned int Op,

diff --git a/test/CodeGen/X86/avx.ll b/test/CodeGen/X86/avx.ll
index f71ec5c10e69aa6151c44e59fd0f4c3b6a808eb4..341dd867e4ff47d18343180a9423201aa293cedf 100644 (file)
--- a/test/CodeGen/X86/avx.ll
+++ b/test/CodeGen/X86/avx.ll
@@ -32,7 +32,7 @@ define <4 x float> @insertps_from_vector_load(<4 x float> %a, <4 x float>* nocap
 ; On X32, account for the argument's move to registers
 ; X32: movl    4(%esp), %eax
 ; CHECK-NOT: mov
-; CHECK: insertps    $48
+; CHECK: vinsertps    $48, (%{{...}}), {{.*#+}} xmm0 = xmm0[0,1,2],mem[0]
 ; CHECK-NEXT: ret
   %1 = load <4 x float>, <4 x float>* %pb, align 16
   %2 = tail call <4 x float> @llvm.x86.sse41.insertps(<4 x float> %a, <4 x float> %1, i32 48)
@@ -46,7 +46,7 @@ define <4 x float> @insertps_from_vector_load_offset(<4 x float> %a, <4 x float>
 ; X32: movl    4(%esp), %eax
 ; CHECK-NOT: mov
 ;; Try to match a bit more of the instr, since we need the load's offset.
-; CHECK: insertps    $96, 4(%{{...}}), %
+; CHECK: vinsertps    $32, 4(%{{...}}), {{.*#+}} xmm0 = xmm0[0,1],mem[0],xmm0[3]
 ; CHECK-NEXT: ret
   %1 = load <4 x float>, <4 x float>* %pb, align 16
   %2 = tail call <4 x float> @llvm.x86.sse41.insertps(<4 x float> %a, <4 x float> %1, i32 96)
@@ -60,7 +60,7 @@ define <4 x float> @insertps_from_vector_load_offset_2(<4 x float> %a, <4 x floa
 ; X32: movl    8(%esp), %ecx
 ; CHECK-NOT: mov
 ;; Try to match a bit more of the instr, since we need the load's offset.
-; CHECK: vinsertps    $192, 12(%{{...}},%{{...}}), %
+; CHECK: vinsertps    $0, 12(%{{...}},%{{...}}), {{.*#+}} xmm0 = mem[0],xmm0[1,2,3]
 ; CHECK-NEXT: ret
   %1 = getelementptr inbounds <4 x float>, <4 x float>* %pb, i64 %index
   %2 = load <4 x float>, <4 x float>* %1, align 16

diff --git a/test/CodeGen/X86/insertps-unfold-load-bug.ll b/test/CodeGen/X86/insertps-unfold-load-bug.ll
new file mode 100644 (file)
index 0000000..bf7c4bc
--- /dev/null
+++ b/test/CodeGen/X86/insertps-unfold-load-bug.ll
@@ -0,0 +1,33 @@
+; RUN: llc -mtriple=i686-unknown-unknown -mattr=+sse4.1 < %s | FileCheck %s -check-prefix=X32
+; RUN: llc -mtriple=x86_64-unknown-unknown -mattr=+sse4.1 < %s | FileCheck %s -check-prefix=X64
+
+; Test for case where insertps was folding the load of the insertion element, but a later optimization
+; was then manipulating the load.
+
+define <4 x float> @insertps_unfold(<4 x float>* %v0, <4 x float>* %v1) {
+; X32-LABEL: insertps_unfold:
+; X32:       # BB#0:
+; X32-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
+; X32-NEXT:    movaps (%eax), %xmm0
+; X32-NEXT:    insertps {{.*#+}} xmm0 = xmm0[0,1,2],mem[0]
+; X32-NEXT:    addps %xmm1, %xmm0
+; X32-NEXT:    retl
+;
+; X64-LABEL: insertps_unfold:
+; X64:       # BB#0:
+; X64-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
+; X64-NEXT:    movaps (%rdi), %xmm0
+; X64-NEXT:    insertps {{.*#+}} xmm0 = xmm0[0,1,2],mem[0]
+; X64-NEXT:    addps %xmm1, %xmm0
+; X64-NEXT:    retq
+  %a = getelementptr inbounds <4 x float>, <4 x float>* %v1, i64 0, i64 1
+  %b = load float, float* %a, align 4
+  %c = insertelement <4 x float> undef, float %b, i32 0
+  %d = load <4 x float>, <4 x float>* %v1, align 16
+  %e = load <4 x float>, <4 x float>* %v0, align 16
+  %f = shufflevector <4 x float> %e, <4 x float> %d, <4 x i32> <i32 0, i32 1, i32 2, i32 5>
+  %g = fadd <4 x float> %c, %f
+  ret <4 x float> %g
+}

diff --git a/test/CodeGen/X86/sse41.ll b/test/CodeGen/X86/sse41.ll
index d624c8dcbb470810f0f718d8a752cc3055502b02..0a83a9753b81a0939e482508d1341cfbb57f8614 100644 (file)
--- a/test/CodeGen/X86/sse41.ll
+++ b/test/CodeGen/X86/sse41.ll
@@ -794,12 +794,12 @@ define <4 x float> @insertps_from_vector_load(<4 x float> %a, <4 x float>* nocap
 ; X32-LABEL: insertps_from_vector_load:
 ; X32:       ## BB#0:
 ; X32-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X32-NEXT:    insertps {{.*#+}} xmm0 = xmm0[0,1,2],mem[0]
+; X32-NEXT:    insertps    $48, (%{{...}}), {{.*#+}} xmm0 = xmm0[0,1,2],mem[0]
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: insertps_from_vector_load:
 ; X64:       ## BB#0:
-; X64-NEXT:    insertps {{.*#+}} xmm0 = xmm0[0,1,2],mem[0]
+; X64-NEXT:    insertps    $48, (%{{...}}), {{.*#+}} xmm0 = xmm0[0,1,2],mem[0]
 ; X64-NEXT:    retq
   %1 = load <4 x float>, <4 x float>* %pb, align 16
   %2 = tail call <4 x float> @llvm.x86.sse41.insertps(<4 x float> %a, <4 x float> %1, i32 48)
@@ -812,12 +812,12 @@ define <4 x float> @insertps_from_vector_load_offset(<4 x float> %a, <4 x float>
 ; X32-LABEL: insertps_from_vector_load_offset:
 ; X32:       ## BB#0:
 ; X32-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X32-NEXT:    insertps {{.*#+}} xmm0 = xmm0[0,1],mem[1],xmm0[3]
+; X32-NEXT:    insertps    $32, 4(%{{...}}), {{.*#+}} xmm0 = xmm0[0,1],mem[0],xmm0[3]
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: insertps_from_vector_load_offset:
 ; X64:       ## BB#0:
-; X64-NEXT:    insertps {{.*#+}} xmm0 = xmm0[0,1],mem[1],xmm0[3]
+; X64-NEXT:    insertps    $32, 4(%{{...}}), {{.*#+}} xmm0 = xmm0[0,1],mem[0],xmm0[3]
 ; X64-NEXT:    retq
   %1 = load <4 x float>, <4 x float>* %pb, align 16
   %2 = tail call <4 x float> @llvm.x86.sse41.insertps(<4 x float> %a, <4 x float> %1, i32 96)
@@ -831,13 +831,13 @@ define <4 x float> @insertps_from_vector_load_offset_2(<4 x float> %a, <4 x floa
 ; X32-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X32-NEXT:    movl {{[0-9]+}}(%esp), %ecx
 ; X32-NEXT:    shll $4, %ecx
-; X32-NEXT:    insertps {{.*#+}} xmm0 = mem[3],xmm0[1,2,3]
+; X32-NEXT:    insertps    $0, 12(%{{...}},%{{...}}), {{.*#+}} xmm0 = mem[0],xmm0[1,2,3]
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: insertps_from_vector_load_offset_2:
 ; X64:       ## BB#0:
 ; X64-NEXT:    shlq $4, %rsi
-; X64-NEXT:    insertps {{.*#+}} xmm0 = mem[3],xmm0[1,2,3]
+; X64-NEXT:    insertps    $0, 12(%{{...}},%{{...}}), {{.*#+}} xmm0 = mem[0],xmm0[1,2,3]
 ; X64-NEXT:    retq
   %1 = getelementptr inbounds <4 x float>, <4 x float>* %pb, i64 %index
   %2 = load <4 x float>, <4 x float>* %1, align 16
@@ -968,12 +968,12 @@ define <4 x float> @pr20087(<4 x float> %a, <4 x float> *%ptr) {
 ; X32-LABEL: pr20087:
 ; X32:       ## BB#0:
 ; X32-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X32-NEXT:    insertps {{.*#+}} xmm0 = xmm0[0],zero,xmm0[2],mem[2]
+; X32-NEXT:    insertps {{.*#+}} xmm0 = xmm0[0],zero,xmm0[2],mem[0]
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: pr20087:
 ; X64:       ## BB#0:
-; X64-NEXT:    insertps {{.*#+}} xmm0 = xmm0[0],zero,xmm0[2],mem[2]
+; X64-NEXT:    insertps {{.*#+}} xmm0 = xmm0[0],zero,xmm0[2],mem[0]
 ; X64-NEXT:    retq
   %load = load <4 x float> , <4 x float> *%ptr
   %ret = shufflevector <4 x float> %load, <4 x float> %a, <4 x i32> <i32 4, i32 undef, i32 6, i32 2>

diff --git a/test/CodeGen/X86/stack-folding-fp-avx1.ll b/test/CodeGen/X86/stack-folding-fp-avx1.ll
index c7a2143b5b28022eb48138dd4bdef3898759f905..b86ec0ea22ff1677c1f939cfd7dcf29f008a1c9d 100644 (file)
--- a/test/CodeGen/X86/stack-folding-fp-avx1.ll
+++ b/test/CodeGen/X86/stack-folding-fp-avx1.ll
@@ -946,7 +946,15 @@ define <8 x float> @stack_fold_insertf128(<4 x float> %a0, <4 x float> %a1) {
   ret <8 x float> %2
 }
 
-; TODO stack_fold_insertps
+define <4 x float> @stack_fold_insertps(<4 x float> %a0, <4 x float> %a1) {
+  ;CHECK-LABEL: stack_fold_insertps
+  ;CHECK:       vinsertps $17, {{-?[0-9]*}}(%rsp), {{%xmm[0-9][0-9]*}}, {{%xmm[0-9][0-9]*}} {{.*#+}} 16-byte Folded Reload
+  ;CHECK-NEXT:                                                                              {{.*#+}} xmm0 = zero,mem[0],xmm0[2,3]
+  %1 = tail call <2 x i64> asm sideeffect "nop", "=x,~{xmm2},~{xmm3},~{xmm4},~{xmm5},~{xmm6},~{xmm7},~{xmm8},~{xmm9},~{xmm10},~{xmm11},~{xmm12},~{xmm13},~{xmm14},~{xmm15},~{flags}"()
+  %2 = call <4 x float> @llvm.x86.sse41.insertps(<4 x float> %a0, <4 x float> %a1, i8 209)
+  ret <4 x float> %2
+}
+declare <4 x float> @llvm.x86.sse41.insertps(<4 x float>, <4 x float>, i8) nounwind readnone
 
 define <2 x double> @stack_fold_maxpd(<2 x double> %a0, <2 x double> %a1) {
   ;CHECK-LABEL: stack_fold_maxpd

diff --git a/test/CodeGen/X86/stack-folding-fp-sse42.ll b/test/CodeGen/X86/stack-folding-fp-sse42.ll
index e90a89d36c09257fc47618d449a0fca8c10af6f9..105115bc7d25c0c64dc11c278de40aef69c84640 100644 (file)
--- a/test/CodeGen/X86/stack-folding-fp-sse42.ll
+++ b/test/CodeGen/X86/stack-folding-fp-sse42.ll
@@ -637,7 +637,15 @@ define <4 x float> @stack_fold_hsubps(<4 x float> %a0, <4 x float> %a1) {
 }
 declare <4 x float> @llvm.x86.sse3.hsub.ps(<4 x float>, <4 x float>) nounwind readnone
 
-; TODO stack_fold_insertps
+define <4 x float> @stack_fold_insertps(<4 x float> %a0, <4 x float> %a1) {
+  ;CHECK-LABEL: stack_fold_insertps
+  ;CHECK:       insertps $17, {{-?[0-9]*}}(%rsp), {{%xmm[0-9][0-9]*}} {{.*#+}} 16-byte Folded Reload
+  ;CHECK-NEXT:                                                        {{.*#+}} xmm0 = zero,mem[0],xmm0[2,3]
+  %1 = tail call <2 x i64> asm sideeffect "nop", "=x,~{xmm2},~{xmm3},~{xmm4},~{xmm5},~{xmm6},~{xmm7},~{xmm8},~{xmm9},~{xmm10},~{xmm11},~{xmm12},~{xmm13},~{xmm14},~{xmm15},~{flags}"()
+  %2 = call <4 x float> @llvm.x86.sse41.insertps(<4 x float> %a0, <4 x float> %a1, i8 209)
+  ret <4 x float> %2
+}
+declare <4 x float> @llvm.x86.sse41.insertps(<4 x float>, <4 x float>, i8) nounwind readnone
 
 define <2 x double> @stack_fold_maxpd(<2 x double> %a0, <2 x double> %a1) {
   ;CHECK-LABEL: stack_fold_maxpd