Added more insertps optimizations

[filcab]

Summary

When inserting an element that's coming from a vector load or a broadcast
of a vector (or scalar) load, combine the load into the insertps
instruction.

Added test cases.

[filcab]

Ping.

[filcab]

Pinging the patch, but also added a CHECK-NOT to a test.

Changed CHECK to CHECK-NOT instead of removing.

[andreadb]

Hi Filipe,

I think there is a better way to fix this bug.

For example, you can simply add the following ISel patterns to X86InstrSSE.td instead of adding new target combine rules in X86ISelLowering.cpp.

let Predicates = [UseSSE41] in {
  def : Pat<(v4f32 (X86insertps (v4f32 VR128:$src1), (loadv4f32 addr:$src2),
                imm:$src3)),
            (INSERTPSrm VR128:$src1, addr:$src2, imm:$src3)>;
  def : Pat<(v4f32 (X86insertps (v4f32 VR128:$src1), (X86PShufd (v4f32 
                 (scalar_to_vector (loadf32 addr:$src2))), (i8 0)), imm:$src3)),
            (INSERTPSrm VR128:$src1, addr:$src2, imm:$src3)>;
}

let Predicates = [UseAVX] in {
  def : Pat<(v4f32 (X86insertps (v4f32 VR128:$src1), (loadv4f32 addr:$src2),
                imm:$src3)),
            (VINSERTPSrm VR128:$src1, addr:$src2, imm:$src3)>;
  def : Pat<(v4f32 (X86insertps (v4f32 VR128:$src1),
                (X86VBroadcast (loadf32 addr:$src2)), imm:$src3)),
            (VINSERTPSrm VR128:$src1, addr:$src2, imm:$src3)>;
}

I am pretty sure that the four patterns above would cover all the interesting cases.
I personally prefer to have tablegen patterns instead of complicated target combine rules that would only trigger on the already legalized and immediately before instruction selection.

Also, you should add tests to verify the AVX codegen as well. Currently your new tests only verifies that we do the correct thing with SSE4.1. However, your patch also affects AVX (in fact, you specifically added combine rules for the case where the inserted element comes from a vbroadcast).

-Andrea

[filcab]

Hi Andrea,

Wouldn't the tablegen patterns be problematic when we use the load result
several times?
If we use it several times, then we shouldn't generate additional loads. Or
do you think it doesn't matter since they should be close together, since
we're generating for one BB only?

That was the reasoning behind doing it with code and guarding them with
MayFoldLoad, which includes hasOneUse().

I'll add the AVX tests today.

Filipe

[andreadb]

Hi Andrea,

Wouldn't the tablegen patterns be problematic when we use the load result
several times?
If we use it several times, then we shouldn't generate additional loads. Or
do you think it doesn't matter since they should be close together, since
we're generating for one BB only?

The ISel pattern matcher will only reduce a dag sequence if the chain is "profitable to fold". A chain is never profitable, for example, if it contains nodes with more than one use.

In the case of your new test function 'insertps_from_vector_load', if I add another use for variable %1, then the following rule will no longer trigger.

def : Pat<(v4f32 (X86insertps (v4f32 VR128:$src1), (loadv4f32 addr:$src2),
                imm:$src3)),
            (INSERTPSrm VR128:$src1, addr:$src2, imm:$src3)>;

That is because eventually, method 'X86DAGtoDAGISel::IsProfitableToFold' will return false if the 'loadv4f32' has more than one use (see X86ISelDAGToDAG.cpp - around line 303). Basically, the matcher will try to verify that all the intermediate nodes between the X86Insertps (excluded) and the loadv4f32 (included), only have a single use.

Given how the ISel Matcher works, in case of a load result used several times, we will not generate additional loads.

That was the reasoning behind doing it with code and guarding them with
MayFoldLoad, which includes hasOneUse().

That reasoning is correct. However, the pattern Matcher will do the same and will never try to fold the loadv4f32 if it has more than one use. (P.s.: you can see it by yourself if you some experiments adding more uses for the load instruction and then you debug ISel passing flag -debug to llc).

I hope this helps.

[filcab]

Thanks Andrea. I will check that out and resubmit the patch.

Filipe

[filcab]

Added the patterns that Andrea mentioned.

We still need to have PerformSELECTCombine for the case when we do a
vector load and extract one element of that vector. In this case we want
to narrow it to an element load.

The other cases are simplified by just dealing with them with patterns in X86InstrSSE.td

Added test cases for SSE4.1, AVX, and AVX2.

[filcab]

Fixed the diff to not contain extra stuff that was in some trunk commits.

[filcab]

Actually add AVX tests.
Sorry about the revisions. I'll wait for comments, now.

[andreadb]

Hi Filipe,

I tested yout patch and it works for me.
If you address the (minor) comments below, then the patch looks good to me!

lib/Target/X86/X86ISelLowering.cpp

20287

It might be useful to have a comment here explaining why you need a shift.
When the source is a memory operand, the Count_S bits of the immediate operand are not used to select the floating point element from the source memory location.
That's why we have to extract the 'Count_S' bits from the immediate operand and use them as 'index' for a new load instruction.

lib/Target/X86/X86InstrSSE.td

6553

You forgot to enclose both patterns between curly braces.
It still works fine because we never produce an X86insertps dag node if we don't have SSE4.1 :-)

6564

Same here, you should enclose the following two patterns between curly braces.

test/CodeGen/X86/avx.ll

6–28 ↗

(On Diff #9504)

These three tests are not part of this patch.
I think you should add those in a separate commit.

[filcab]

Closed by commit rL209156 (authored by @filcab).

[filcab]

Thanks Andrea. Committed as r209156.

lib/Target/X86/X86InstrSSE.td

6553

The lack of {} is what you get when you add a second pattern and forget to add the {}.

lib/Target/X86/X86ISelLowering.cpp

SDValue getMOVLP(SDValue &Op, SDLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
 
   assert(VT != MVT::v4i32 && "unsupported shuffle type");
 
   // Invert the operand order and use SHUFPS to match it.
   return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V2, V1,
                               getShuffleSHUFImmediate(SVOp), DAG);
 }
 
+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(), Addr.getSimpleValueType()));
+
+  SDValue NewLoad =
+      DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
+                  DAG.getMachineFunction().getMachineMemOperand(
+                      Load->getMemOperand(), 0, EVT.getStoreSize()));
+  return NewLoad;
+}
+
 // It is only safe to call this function if isINSERTPSMask is true for
 // this shufflevector mask.
 static SDValue getINSERTPS(ShuffleVectorSDNode *SVOp, SDLoc &dl,
                            SelectionDAG &DAG) {
   // Generate an insertps instruction when inserting an f32 from memory onto a
   // v4f32 or when copying a member from one v4f32 to another.
   // We also use it for transferring i32 from one register to another,
   // since it simply copies the same bits.
   // If we're transferring an i32 from memory to a specific element in a
   // register, we output a generic DAG that will match the PINSRD
   // instruction.
-  // TODO: Optimize for AVX cases too (VINSERTPS)
   MVT VT = SVOp->getSimpleValueType(0);
   MVT EVT = VT.getVectorElementType();
   SDValue V1 = SVOp->getOperand(0);
   SDValue V2 = SVOp->getOperand(1);
   auto Mask = SVOp->getMask();
   assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
          "unsupported vector type for insertps/pinsrd");
 
DestIndex = std::find_if(Mask.begin(), Mask.end(),
                              [](const int &i) { return i >= 4; }) -
                 Mask.begin();
   }
 
   if (MayFoldLoad(From)) {
     // Trivial case, when From comes from a load and is only used by the
     // shuffle. Make it use insertps from the vector that we need from that
     // load.
-    SDValue Addr = From.getOperand(1);
-    SDValue NewAddr =
-        DAG.getNode(ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
-                    DAG.getConstant(DestIndex * EVT.getStoreSize(),
-                                    Addr.getSimpleValueType()));
-
-    LoadSDNode *Load = cast<LoadSDNode>(From);
     SDValue NewLoad =
-        DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
-                    DAG.getMachineFunction().getMachineMemOperand(
-                        Load->getMemOperand(), 0, EVT.getStoreSize()));
+        NarrowVectorLoadToElement(cast<LoadSDNode>(From), DestIndex, DAG);
+    if (!NewLoad.getNode())
+      return SDValue();
 
     if (EVT == MVT::f32) {
       // Create this as a scalar to vector to match the instruction pattern.
       SDValue LoadScalarToVector =
           DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, NewLoad);
       SDValue InsertpsMask = DAG.getIntPtrConstant(DestIndex << 4);
       return DAG.getNode(X86ISD::INSERTPS, dl, VT, To, LoadScalarToVector,
                          InsertpsMask);
if (IsSEXT1 && IsVZero0) {
         return DAG.getNOT(DL, RHS.getOperand(0), VT);
       return RHS.getOperand(0);
     }
   }
 
   return SDValue();
 }
 
+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)) {
+    unsigned DestIndex =
+        cast<ConstantSDNode>(N->getOperand(2))->getZExtValue() >> 6;
+    Ld = NarrowVectorLoadToElement(cast<LoadSDNode>(Ld), DestIndex, DAG);
+  } else
+    return SDValue();
+
+  // 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));
+}
+
 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
 // as "sbb reg,reg", since it can be extended without zext and produces
 // an all-ones bit which is more useful than 0/1 in some cases.
 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
                                MVT VT) {
   if (VT == MVT::i8)
     return DAG.getNode(ISD::AND, DL, VT,
                        DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
   case X86ISD::MOVSS:
   case X86ISD::MOVSD:
   case X86ISD::VPERMILP:
   case X86ISD::VPERM2X128:
   case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
   case ISD::FMA:            return PerformFMACombine(N, DAG, Subtarget);
   case ISD::INTRINSIC_WO_CHAIN:
     return PerformINTRINSIC_WO_CHAINCombine(N, DAG, Subtarget);
+  case X86ISD::INSERTPS:
+    return PerformINSERTPSCombine(N, DAG, Subtarget);
   }
 
   return SDValue();
 }
 
 /// isTypeDesirableForOp - Return true if the target has native support for
 /// the specified value type and it is 'desirable' to use the type for the
 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16

lib/Target/X86/X86InstrSSE.td

 
 let ExeDomain = SSEPackedSingle in {
   let Predicates = [UseAVX] in
     defm VINSERTPS : SS41I_insertf32<0x21, "vinsertps", 0>, VEX_4V;
   let Constraints = "$src1 = $dst" in
     defm INSERTPS : SS41I_insertf32<0x21, "insertps", 1, SSE_INSERT_ITINS>;
 }
 
+let Predicates = [UseSSE41] in
+  // If we're inserting an element from a load or a null pshuf of a load,
+  // fold the load into the insertps instruction.
+  def : Pat<(v4f32 (X86insertps (v4f32 VR128:$src1), (X86PShufd (v4f32
+                       (scalar_to_vector (loadf32 addr:$src2))), (i8 0)),
+                   imm:$src3)),
+            (INSERTPSrm VR128:$src1, addr:$src2, imm:$src3)>;
+  def : Pat<(v4f32 (X86insertps (v4f32 VR128:$src1), (X86PShufd
+                      (loadv4f32 addr:$src2), (i8 0)), imm:$src3)),
+            (INSERTPSrm VR128:$src1, addr:$src2, imm:$src3)>;
+
+let Predicates = [UseAVX] in
+  // If we're inserting an element from a vbroadcast of a load, fold the
+  // load into the X86insertps instruction.
+  def : Pat<(v4f32 (X86insertps (v4f32 VR128:$src1),
+                (X86VBroadcast (loadf32 addr:$src2)), imm:$src3)),
+            (VINSERTPSrm VR128:$src1, addr:$src2, imm:$src3)>;
+  def : Pat<(v4f32 (X86insertps (v4f32 VR128:$src1),
+                (X86VBroadcast (loadv4f32 addr:$src2)), imm:$src3)),
+            (VINSERTPSrm VR128:$src1, addr:$src2, imm:$src3)>;
+
 //===----------------------------------------------------------------------===//
 // SSE4.1 - Round Instructions
 //===----------------------------------------------------------------------===//
 
 multiclass sse41_fp_unop_rm<bits<8> opcps, bits<8> opcpd, string OpcodeStr,
                             X86MemOperand x86memop, RegisterClass RC,
                             PatFrag mem_frag32, PatFrag mem_frag64,
                             Intrinsic V4F32Int, Intrinsic V2F64Int> {

test/CodeGen/X86/avx2.ll

-; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=core-avx2 -mattr=+avx2 | FileCheck %s
+; RUN: llc < %s -mtriple=i686-apple-darwin -mcpu=core-avx2 | FileCheck %s -check-prefix=X32 --check-prefix=CHECK
+; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=core-avx2 | FileCheck %s -check-prefix=X64 --check-prefix=CHECK
+
+declare <4 x float> @llvm.x86.sse41.insertps(<4 x float>, <4 x float>, i32) nounwind readnone
 
 define <4 x i32> @blendvb_fallback_v4i32(<4 x i1> %mask, <4 x i32> %x, <4 x i32> %y) {
 ; CHECK-LABEL: @blendvb_fallback_v4i32
 ; CHECK: vblendvps
 ; CHECK: ret
   %ret = select <4 x i1> %mask, <4 x i32> %x, <4 x i32> %y
   ret <4 x i32> %ret
 }
 
 define <8 x i32> @blendvb_fallback_v8i32(<8 x i1> %mask, <8 x i32> %x, <8 x i32> %y) {
 ; CHECK-LABEL: @blendvb_fallback_v8i32
 ; CHECK: vblendvps
 ; CHECK: ret
   %ret = select <8 x i1> %mask, <8 x i32> %x, <8 x i32> %y
   ret <8 x i32> %ret
 }
 
 define <8 x float> @blendvb_fallback_v8f32(<8 x i1> %mask, <8 x float> %x, <8 x float> %y) {
 ; CHECK-LABEL: @blendvb_fallback_v8f32
 ; CHECK: vblendvps
 ; CHECK: ret
   %ret = select <8 x i1> %mask, <8 x float> %x, <8 x float> %y
   ret <8 x float> %ret
 }
+
+define <4 x float> @insertps_from_vector_load(<4 x float> %a, <4 x float>* nocapture readonly %pb) {
+; CHECK-LABEL: insertps_from_vector_load:
+; On X32, account for the argument's move to registers
+; X32: movl    4(%esp), %eax
+; CHECK-NOT: mov
+; CHECK: insertps    $48
+; CHECK-NEXT: ret
+  %1 = load <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)
+  ret <4 x float> %2
+}
+
+;; Use a non-zero CountS for insertps
+define <4 x float> @insertps_from_vector_load_offset(<4 x float> %a, <4 x float>* nocapture readonly %pb) {
+; CHECK-LABEL: insertps_from_vector_load_offset:
+; On X32, account for the argument's move to registers
+; 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-NEXT: ret
+  %1 = load <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)
+  ret <4 x float> %2
+}
+
+define <4 x float> @insertps_from_vector_load_offset_2(<4 x float> %a, <4 x float>* nocapture readonly %pb, i64 %index) {
+; CHECK-LABEL: insertps_from_vector_load_offset_2:
+; On X32, account for the argument's move to registers
+; X32: movl    4(%esp), %eax
+; 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-NEXT: ret
+  %1 = getelementptr inbounds <4 x float>* %pb, i64 %index
+  %2 = load <4 x float>* %1, align 16
+  %3 = tail call <4 x float> @llvm.x86.sse41.insertps(<4 x float> %a, <4 x float> %2, i32 192)
+  ret <4 x float> %3
+}
+
+define <4 x float> @insertps_from_broadcast_loadf32(<4 x float> %a, float* nocapture readonly %fb, i64 %index) {
+; CHECK-LABEL: insertps_from_broadcast_loadf32:
+; On X32, account for the arguments' move to registers
+; X32: movl    8(%esp), %eax
+; X32: movl    4(%esp), %ecx
+; CHECK-NOT: mov
+; CHECK: insertps    $48
+; CHECK-NEXT: ret
+  %1 = getelementptr inbounds float* %fb, i64 %index
+  %2 = load float* %1, align 4
+  %3 = insertelement <4 x float> undef, float %2, i32 0
+  %4 = insertelement <4 x float> %3, float %2, i32 1
+  %5 = insertelement <4 x float> %4, float %2, i32 2
+  %6 = insertelement <4 x float> %5, float %2, i32 3
+  %7 = tail call <4 x float> @llvm.x86.sse41.insertps(<4 x float> %a, <4 x float> %6, i32 48)
+  ret <4 x float> %7
+}
+
+define <4 x float> @insertps_from_broadcast_loadv4f32(<4 x float> %a, <4 x float>* nocapture readonly %b) {
+; CHECK-LABEL: insertps_from_broadcast_loadv4f32:
+; On X32, account for the arguments' move to registers
+; X32: movl    4(%esp), %{{...}}
+; CHECK-NOT: mov
+; CHECK: insertps    $48
+; CHECK-NEXT: ret
+  %1 = load <4 x float>* %b, align 4
+  %2 = extractelement <4 x float> %1, i32 0
+  %3 = insertelement <4 x float> undef, float %2, i32 0
+  %4 = insertelement <4 x float> %3, float %2, i32 1
+  %5 = insertelement <4 x float> %4, float %2, i32 2
+  %6 = insertelement <4 x float> %5, float %2, i32 3
+  %7 = tail call <4 x float> @llvm.x86.sse41.insertps(<4 x float> %a, <4 x float> %6, i32 48)
+  ret <4 x float> %7
+}
+
+;; FIXME: We're emitting an extraneous pshufd/vbroadcast.
+define <4 x float> @insertps_from_broadcast_multiple_use(<4 x float> %a, <4 x float> %b, <4 x float> %c, <4 x float> %d, float* nocapture readonly %fb, i64 %index) {
+; CHECK-LABEL: insertps_from_broadcast_multiple_use:
+; On X32, account for the arguments' move to registers
+; X32: movl    8(%esp), %eax
+; X32: movl    4(%esp), %ecx
+; CHECK: vbroadcastss
+; CHECK-NOT: mov
+; CHECK: insertps    $48
+; CHECK: insertps    $48
+; CHECK: insertps    $48
+; CHECK: insertps    $48
+; CHECK: vaddps
+; CHECK: vaddps
+; CHECK: vaddps
+; CHECK-NEXT: ret
+  %1 = getelementptr inbounds float* %fb, i64 %index
+  %2 = load float* %1, align 4
+  %3 = insertelement <4 x float> undef, float %2, i32 0
+  %4 = insertelement <4 x float> %3, float %2, i32 1
+  %5 = insertelement <4 x float> %4, float %2, i32 2
+  %6 = insertelement <4 x float> %5, float %2, i32 3
+  %7 = tail call <4 x float> @llvm.x86.sse41.insertps(<4 x float> %a, <4 x float> %6, i32 48)
+  %8 = tail call <4 x float> @llvm.x86.sse41.insertps(<4 x float> %b, <4 x float> %6, i32 48)
+  %9 = tail call <4 x float> @llvm.x86.sse41.insertps(<4 x float> %c, <4 x float> %6, i32 48)
+  %10 = tail call <4 x float> @llvm.x86.sse41.insertps(<4 x float> %d, <4 x float> %6, i32 48)
+  %11 = fadd <4 x float> %7, %8
+  %12 = fadd <4 x float> %9, %10
+  %13 = fadd <4 x float> %11, %12
+  ret <4 x float> %13
+}

test/CodeGen/X86/fold-load-vec.ll

 ; RUN: llc < %s -march=x86-64 -mcpu=corei7 -mattr=+sse4.1 | FileCheck %s
 
 ; rdar://12721174
 ; We should not fold movss into pshufd since pshufd expects m128 while movss
 ; loads from m32.
 define void @sample_test(<4 x float>* %source, <2 x float>* %dest) nounwind {
 ; CHECK: sample_test
-; CHECK: movaps
+; CHECK-NOT: movaps
 ; CHECK: insertps
 entry:
   %source.addr = alloca <4 x float>*, align 8
   %dest.addr = alloca <2 x float>*, align 8
   %tmp = alloca <2 x float>, align 8
   store <4 x float>* %source, <4 x float>** %source.addr, align 8
   store <2 x float>* %dest, <2 x float>** %dest.addr, align 8
   store <2 x float> zeroinitializer, <2 x float>* %tmp, align 8

test/CodeGen/X86/sse41.ll

 
 define <8 x i16> @blendvb_fallback(<8 x i1> %mask, <8 x i16> %x, <8 x i16> %y) {
 ; CHECK-LABEL: blendvb_fallback
 ; CHECK: blendvb
 ; CHECK: ret
   %ret = select <8 x i1> %mask, <8 x i16> %x, <8 x i16> %y
   ret <8 x i16> %ret
 }
+
+define <4 x float> @insertps_from_vector_load(<4 x float> %a, <4 x float>* nocapture readonly %pb) {
+; CHECK-LABEL: insertps_from_vector_load:
+; On X32, account for the argument's move to registers
+; X32: movl    4(%esp), %eax
+; CHECK-NOT: mov
+; CHECK: insertps    $48
+; CHECK-NEXT: ret
+  %1 = load <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)
+  ret <4 x float> %2
+}
+
+;; Use a non-zero CountS for insertps
+define <4 x float> @insertps_from_vector_load_offset(<4 x float> %a, <4 x float>* nocapture readonly %pb) {
+; CHECK-LABEL: insertps_from_vector_load_offset:
+; On X32, account for the argument's move to registers
+; 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-NEXT: ret
+  %1 = load <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)
+  ret <4 x float> %2
+}
+
+define <4 x float> @insertps_from_vector_load_offset_2(<4 x float> %a, <4 x float>* nocapture readonly %pb, i64 %index) {
+; CHECK-LABEL: insertps_from_vector_load_offset_2:
+; On X32, account for the argument's move to registers
+; X32: movl    4(%esp), %eax
+; 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: insertps    $192, 12(%{{...}},%{{...}}), %
+; CHECK-NEXT: ret
+  %1 = getelementptr inbounds <4 x float>* %pb, i64 %index
+  %2 = load <4 x float>* %1, align 16
+  %3 = tail call <4 x float> @llvm.x86.sse41.insertps(<4 x float> %a, <4 x float> %2, i32 192)
+  ret <4 x float> %3
+}
+
+define <4 x float> @insertps_from_broadcast_loadf32(<4 x float> %a, float* nocapture readonly %fb, i64 %index) {
+; CHECK-LABEL: insertps_from_broadcast_loadf32:
+; On X32, account for the arguments' move to registers
+; X32: movl    8(%esp), %eax
+; X32: movl    4(%esp), %ecx
+; CHECK-NOT: mov
+; CHECK: insertps    $48
+; CHECK-NEXT: ret
+  %1 = getelementptr inbounds float* %fb, i64 %index
+  %2 = load float* %1, align 4
+  %3 = insertelement <4 x float> undef, float %2, i32 0
+  %4 = insertelement <4 x float> %3, float %2, i32 1
+  %5 = insertelement <4 x float> %4, float %2, i32 2
+  %6 = insertelement <4 x float> %5, float %2, i32 3
+  %7 = tail call <4 x float> @llvm.x86.sse41.insertps(<4 x float> %a, <4 x float> %6, i32 48)
+  ret <4 x float> %7
+}
+
+define <4 x float> @insertps_from_broadcast_loadv4f32(<4 x float> %a, <4 x float>* nocapture readonly %b) {
+; CHECK-LABEL: insertps_from_broadcast_loadv4f32:
+; On X32, account for the arguments' move to registers
+; X32: movl    4(%esp), %{{...}}
+; CHECK-NOT: mov
+; CHECK: insertps    $48
+; CHECK-NEXT: ret
+  %1 = load <4 x float>* %b, align 4
+  %2 = extractelement <4 x float> %1, i32 0
+  %3 = insertelement <4 x float> undef, float %2, i32 0
+  %4 = insertelement <4 x float> %3, float %2, i32 1
+  %5 = insertelement <4 x float> %4, float %2, i32 2
+  %6 = insertelement <4 x float> %5, float %2, i32 3
+  %7 = tail call <4 x float> @llvm.x86.sse41.insertps(<4 x float> %a, <4 x float> %6, i32 48)
+  ret <4 x float> %7
+}
+
+;; FIXME: We're emitting an extraneous pshufd/vbroadcast.
+define <4 x float> @insertps_from_broadcast_multiple_use(<4 x float> %a, <4 x float> %b, <4 x float> %c, <4 x float> %d, float* nocapture readonly %fb, i64 %index) {
+; CHECK-LABEL: insertps_from_broadcast_multiple_use:
+; On X32, account for the arguments' move to registers
+; X32: movl    8(%esp), %eax
+; X32: movl    4(%esp), %ecx
+; CHECK: movss
+; CHECK-NOT: mov
+; CHECK: insertps    $48
+; CHECK: insertps    $48
+; CHECK: insertps    $48
+; CHECK: insertps    $48
+; CHECK: addps
+; CHECK: addps
+; CHECK: addps
+; CHECK-NEXT: ret
+  %1 = getelementptr inbounds float* %fb, i64 %index
+  %2 = load float* %1, align 4
+  %3 = insertelement <4 x float> undef, float %2, i32 0
+  %4 = insertelement <4 x float> %3, float %2, i32 1
+  %5 = insertelement <4 x float> %4, float %2, i32 2
+  %6 = insertelement <4 x float> %5, float %2, i32 3
+  %7 = tail call <4 x float> @llvm.x86.sse41.insertps(<4 x float> %a, <4 x float> %6, i32 48)
+  %8 = tail call <4 x float> @llvm.x86.sse41.insertps(<4 x float> %b, <4 x float> %6, i32 48)
+  %9 = tail call <4 x float> @llvm.x86.sse41.insertps(<4 x float> %c, <4 x float> %6, i32 48)
+  %10 = tail call <4 x float> @llvm.x86.sse41.insertps(<4 x float> %d, <4 x float> %6, i32 48)
+  %11 = fadd <4 x float> %7, %8
+  %12 = fadd <4 x float> %9, %10
+  %13 = fadd <4 x float> %11, %12
+  ret <4 x float> %13
+}