From cd6c57917db22a3913a2cdbadfa79fed3547bdec Mon Sep 17 00:00:00 2001
From: Akira Hatanaka <ahatanaka@mips.com>
Date: Tue, 30 Apr 2013 22:37:26 +0000
Subject: [PATCH] [mips] Instruction selection patterns for DSP-ASE vector
 select and compare instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@180820 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/Mips/MipsDSPInstrInfo.td           |  73 ++
 lib/Target/Mips/MipsISelLowering.cpp          |   4 +-
 lib/Target/Mips/MipsISelLowering.h            |   4 +
 lib/Target/Mips/MipsRegisterInfo.td           |   4 +
 lib/Target/Mips/MipsSEISelLowering.cpp        |  63 +-
 test/CodeGen/Mips/dsp-patterns-cmp-vselect.ll | 641 ++++++++++++++++++
 6 files changed, 786 insertions(+), 3 deletions(-)
 create mode 100644 test/CodeGen/Mips/dsp-patterns-cmp-vselect.ll

diff --git a/lib/Target/Mips/MipsDSPInstrInfo.td b/lib/Target/Mips/MipsDSPInstrInfo.td
index 730bca7d529..6719cd54a9b 100644
--- a/lib/Target/Mips/MipsDSPInstrInfo.td
+++ b/lib/Target/Mips/MipsDSPInstrInfo.td
@@ -79,6 +79,8 @@ def MipsMSUBU_DSP : MipsDSPBase<"MSUBU_DSP", SDT_MipsDPA>;
 def MipsSHLL_DSP : MipsDSPBase<"SHLL_DSP", SDT_MipsSHIFT_DSP>;
 def MipsSHRA_DSP : MipsDSPBase<"SHRA_DSP", SDT_MipsSHIFT_DSP>;
 def MipsSHRL_DSP : MipsDSPBase<"SHRL_DSP", SDT_MipsSHIFT_DSP>;
+def MipsSETCC_DSP : MipsDSPBase<"SETCC_DSP", SDTSetCC>;
+def MipsSELECT_CC_DSP : MipsDSPBase<"SELECT_CC_DSP", SDTSelectCC>;
 
 // Flags.
 class UseAC {
@@ -1237,6 +1239,26 @@ let isPseudo = 1 in {
 
 def COPY_AC_DSP : PseudoSE<(outs ACRegsDSP:$dst), (ins ACRegsDSP:$src), []>;
 
+// Pseudo CMP and PICK instructions.
+class PseudoCMP<Instruction RealInst> :
+  PseudoDSP<(outs DSPCC:$cmp), (ins DSPRegs:$rs, DSPRegs:$rt), []>,
+  PseudoInstExpansion<(RealInst DSPRegs:$rs, DSPRegs:$rt)>, NeverHasSideEffects;
+
+class PseudoPICK<Instruction RealInst> :
+  PseudoDSP<(outs DSPRegs:$rd), (ins DSPCC:$cmp, DSPRegs:$rs, DSPRegs:$rt), []>,
+  PseudoInstExpansion<(RealInst DSPRegs:$rd, DSPRegs:$rs, DSPRegs:$rt)>,
+  NeverHasSideEffects;
+
+def PseudoCMP_EQ_PH : PseudoCMP<CMP_EQ_PH>;
+def PseudoCMP_LT_PH : PseudoCMP<CMP_LT_PH>;
+def PseudoCMP_LE_PH : PseudoCMP<CMP_LE_PH>;
+def PseudoCMPU_EQ_QB : PseudoCMP<CMPU_EQ_QB>;
+def PseudoCMPU_LT_QB : PseudoCMP<CMPU_LT_QB>;
+def PseudoCMPU_LE_QB : PseudoCMP<CMPU_LE_QB>;
+
+def PseudoPICK_PH : PseudoPICK<PICK_PH>;
+def PseudoPICK_QB : PseudoPICK<PICK_QB>;
+
 // Patterns.
 class DSPPat<dag pattern, dag result, Predicate pred = HasDSP> :
   Pat<pattern, result>, Requires<[pred]>;
@@ -1298,6 +1320,57 @@ def : DSPShiftPat<SHLL_QB, v4i8, int_mips_shll_qb, immZExt3>;
 def : DSPShiftPat<SHRA_QB, v4i8, int_mips_shra_qb, immZExt3, HasDSPR2>;
 def : DSPShiftPat<SHRL_QB, v4i8, int_mips_shrl_qb, immZExt3>;
 
+// SETCC/SELECT_CC patterns.
+class DSPSetCCPat<Instruction Cmp, Instruction Pick, ValueType ValTy,
+                  CondCode CC> :
+  DSPPat<(ValTy (MipsSETCC_DSP ValTy:$a, ValTy:$b, CC)),
+         (ValTy (Pick (ValTy (Cmp ValTy:$a, ValTy:$b)),
+                      (ValTy (COPY_TO_REGCLASS (ADDiu ZERO, -1), DSPRegs)),
+                      (ValTy ZERO)))>;
+
+class DSPSetCCPatInv<Instruction Cmp, Instruction Pick, ValueType ValTy,
+                     CondCode CC> :
+  DSPPat<(ValTy (MipsSETCC_DSP ValTy:$a, ValTy:$b, CC)),
+         (ValTy (Pick (ValTy (Cmp ValTy:$a, ValTy:$b)),
+                      (ValTy ZERO),
+                      (ValTy (COPY_TO_REGCLASS (ADDiu ZERO, -1), DSPRegs))))>;
+
+class DSPSelectCCPat<Instruction Cmp, Instruction Pick, ValueType ValTy,
+                     CondCode CC> :
+  DSPPat<(ValTy (MipsSELECT_CC_DSP ValTy:$a, ValTy:$b, ValTy:$c, ValTy:$d, CC)),
+         (ValTy (Pick (ValTy (Cmp ValTy:$a, ValTy:$b)), $c, $d))>;
+
+class DSPSelectCCPatInv<Instruction Cmp, Instruction Pick, ValueType ValTy,
+                        CondCode CC> :
+  DSPPat<(ValTy (MipsSELECT_CC_DSP ValTy:$a, ValTy:$b, ValTy:$c, ValTy:$d, CC)),
+         (ValTy (Pick (ValTy (Cmp ValTy:$a, ValTy:$b)), $d, $c))>;
+
+def : DSPSetCCPat<PseudoCMP_EQ_PH, PseudoPICK_PH, v2i16, SETEQ>;
+def : DSPSetCCPat<PseudoCMP_LT_PH, PseudoPICK_PH, v2i16, SETLT>;
+def : DSPSetCCPat<PseudoCMP_LE_PH, PseudoPICK_PH, v2i16, SETLE>;
+def : DSPSetCCPatInv<PseudoCMP_EQ_PH, PseudoPICK_PH, v2i16, SETNE>;
+def : DSPSetCCPatInv<PseudoCMP_LT_PH, PseudoPICK_PH, v2i16, SETGE>;
+def : DSPSetCCPatInv<PseudoCMP_LE_PH, PseudoPICK_PH, v2i16, SETGT>;
+def : DSPSetCCPat<PseudoCMPU_EQ_QB, PseudoPICK_QB, v4i8, SETEQ>;
+def : DSPSetCCPat<PseudoCMPU_LT_QB, PseudoPICK_QB, v4i8, SETULT>;
+def : DSPSetCCPat<PseudoCMPU_LE_QB, PseudoPICK_QB, v4i8, SETULE>;
+def : DSPSetCCPatInv<PseudoCMPU_EQ_QB, PseudoPICK_QB, v4i8, SETNE>;
+def : DSPSetCCPatInv<PseudoCMPU_LT_QB, PseudoPICK_QB, v4i8, SETUGE>;
+def : DSPSetCCPatInv<PseudoCMPU_LE_QB, PseudoPICK_QB, v4i8, SETUGT>;
+
+def : DSPSelectCCPat<PseudoCMP_EQ_PH, PseudoPICK_PH, v2i16, SETEQ>;
+def : DSPSelectCCPat<PseudoCMP_LT_PH, PseudoPICK_PH, v2i16, SETLT>;
+def : DSPSelectCCPat<PseudoCMP_LE_PH, PseudoPICK_PH, v2i16, SETLE>;
+def : DSPSelectCCPatInv<PseudoCMP_EQ_PH, PseudoPICK_PH, v2i16, SETNE>;
+def : DSPSelectCCPatInv<PseudoCMP_LT_PH, PseudoPICK_PH, v2i16, SETGE>;
+def : DSPSelectCCPatInv<PseudoCMP_LE_PH, PseudoPICK_PH, v2i16, SETGT>;
+def : DSPSelectCCPat<PseudoCMPU_EQ_QB, PseudoPICK_QB, v4i8, SETEQ>;
+def : DSPSelectCCPat<PseudoCMPU_LT_QB, PseudoPICK_QB, v4i8, SETULT>;
+def : DSPSelectCCPat<PseudoCMPU_LE_QB, PseudoPICK_QB, v4i8, SETULE>;
+def : DSPSelectCCPatInv<PseudoCMPU_EQ_QB, PseudoPICK_QB, v4i8, SETNE>;
+def : DSPSelectCCPatInv<PseudoCMPU_LT_QB, PseudoPICK_QB, v4i8, SETUGE>;
+def : DSPSelectCCPatInv<PseudoCMPU_LE_QB, PseudoPICK_QB, v4i8, SETUGT>;
+
 // Extr patterns.
 class EXTR_W_TY1_R2_Pat<SDPatternOperator OpNode, Instruction Instr> :
   DSPPat<(i32 (OpNode CPURegs:$rs, ACRegsDSP:$ac)),
diff --git a/lib/Target/Mips/MipsISelLowering.cpp b/lib/Target/Mips/MipsISelLowering.cpp
index 8a0673feb64..611d48326e6 100644
--- a/lib/Target/Mips/MipsISelLowering.cpp
+++ b/lib/Target/Mips/MipsISelLowering.cpp
@@ -200,6 +200,8 @@ const char *MipsTargetLowering::getTargetNodeName(unsigned Opcode) const {
   case MipsISD::SHLL_DSP:          return "MipsISD::SHLL_DSP";
   case MipsISD::SHRA_DSP:          return "MipsISD::SHRA_DSP";
   case MipsISD::SHRL_DSP:          return "MipsISD::SHRL_DSP";
+  case MipsISD::SETCC_DSP:         return "MipsISD::SETCC_DSP";
+  case MipsISD::SELECT_CC_DSP:     return "MipsISD::SELECT_CC_DSP";
   default:                         return NULL;
   }
 }
@@ -213,7 +215,7 @@ MipsTargetLowering(MipsTargetMachine &TM)
   // Mips does not have i1 type, so use i32 for
   // setcc operations results (slt, sgt, ...).
   setBooleanContents(ZeroOrOneBooleanContent);
-  setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct?
+  setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
 
   // Load extented operations for i1 types must be promoted
   setLoadExtAction(ISD::EXTLOAD,  MVT::i1,  Promote);
diff --git a/lib/Target/Mips/MipsISelLowering.h b/lib/Target/Mips/MipsISelLowering.h
index c1574fac578..5587e8f5814 100644
--- a/lib/Target/Mips/MipsISelLowering.h
+++ b/lib/Target/Mips/MipsISelLowering.h
@@ -148,6 +148,10 @@ namespace llvm {
       SHRA_DSP,
       SHRL_DSP,
 
+      // DSP setcc and select_cc nodes.
+      SETCC_DSP,
+      SELECT_CC_DSP,
+
       // Load/Store Left/Right nodes.
       LWL = ISD::FIRST_TARGET_MEMORY_OPCODE,
       LWR,
diff --git a/lib/Target/Mips/MipsRegisterInfo.td b/lib/Target/Mips/MipsRegisterInfo.td
index 30699ac4210..e314e946a01 100644
--- a/lib/Target/Mips/MipsRegisterInfo.td
+++ b/lib/Target/Mips/MipsRegisterInfo.td
@@ -265,6 +265,7 @@ let Namespace = "Mips" in {
   def AC0_64 : ACC<0, "ac0", [LO64, HI64]>;
 
   def DSPCtrl : Register<"dspctrl">;
+  def DSPCCond : Register<"">;
 }
 
 //===----------------------------------------------------------------------===//
@@ -362,6 +363,9 @@ def ACRegsDSP : RegisterClass<"Mips", [untyped], 64, (sequence "AC%u", 0, 3)> {
   let Size = 64;
 }
 
+def DSPCC : RegisterClass<"Mips", [v4i8, v2i16], 32, (add DSPCCond)>;
+
+// Register Operands.
 def CPURegsAsmOperand : AsmOperandClass {
   let Name = "CPURegsAsm";
   let ParserMethod = "parseCPURegs";
diff --git a/lib/Target/Mips/MipsSEISelLowering.cpp b/lib/Target/Mips/MipsSEISelLowering.cpp
index 7e103bbef74..8544bb89107 100644
--- a/lib/Target/Mips/MipsSEISelLowering.cpp
+++ b/lib/Target/Mips/MipsSEISelLowering.cpp
@@ -56,6 +56,8 @@ MipsSETargetLowering::MipsSETargetLowering(MipsTargetMachine &TM)
     setTargetDAGCombine(ISD::SHL);
     setTargetDAGCombine(ISD::SRA);
     setTargetDAGCombine(ISD::SRL);
+    setTargetDAGCombine(ISD::SETCC);
+    setTargetDAGCombine(ISD::VSELECT);
   }
 
   if (Subtarget->hasDSPR2())
@@ -373,9 +375,57 @@ static SDValue performSRLCombine(SDNode *N, SelectionDAG &DAG,
   return performDSPShiftCombine(MipsISD::SHRL_DSP, N, Ty, DAG, Subtarget);
 }
 
+static bool isLegalDSPCondCode(EVT Ty, ISD::CondCode CC) {
+  bool IsV216 = (Ty == MVT::v2i16);
+
+  switch (CC) {
+  case ISD::SETEQ:
+  case ISD::SETNE:  return true;
+  case ISD::SETLT:
+  case ISD::SETLE:
+  case ISD::SETGT:
+  case ISD::SETGE:  return IsV216;
+  case ISD::SETULT:
+  case ISD::SETULE:
+  case ISD::SETUGT:
+  case ISD::SETUGE: return !IsV216;
+  default:          return false;
+  }
+}
+
+static SDValue performSETCCCombine(SDNode *N, SelectionDAG &DAG) {
+  EVT Ty = N->getValueType(0);
+
+  if ((Ty != MVT::v2i16) && (Ty != MVT::v4i8))
+    return SDValue();
+
+  if (!isLegalDSPCondCode(Ty, cast<CondCodeSDNode>(N->getOperand(2))->get()))
+    return SDValue();
+
+  return DAG.getNode(MipsISD::SETCC_DSP, N->getDebugLoc(), Ty, N->getOperand(0),
+                     N->getOperand(1), N->getOperand(2));
+}
+
+static SDValue performVSELECTCombine(SDNode *N, SelectionDAG &DAG) {
+  EVT Ty = N->getValueType(0);
+
+  if ((Ty != MVT::v2i16) && (Ty != MVT::v4i8))
+    return SDValue();
+
+  SDValue SetCC = N->getOperand(0);
+
+  if (SetCC.getOpcode() != MipsISD::SETCC_DSP)
+    return SDValue();
+
+  return DAG.getNode(MipsISD::SELECT_CC_DSP, N->getDebugLoc(), Ty,
+                     SetCC.getOperand(0), SetCC.getOperand(1), N->getOperand(1),
+                     N->getOperand(2), SetCC.getOperand(2));
+}
+
 SDValue
 MipsSETargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
   SelectionDAG &DAG = DCI.DAG;
+  SDValue Val;
 
   switch (N->getOpcode()) {
   case ISD::ADDE:
@@ -388,9 +438,18 @@ MipsSETargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
     return performSRACombine(N, DAG, DCI, Subtarget);
   case ISD::SRL:
     return performSRLCombine(N, DAG, DCI, Subtarget);
-  default:
-    return MipsTargetLowering::PerformDAGCombine(N, DCI);
+  case ISD::VSELECT:
+    return performVSELECTCombine(N, DAG);
+  case ISD::SETCC: {
+    Val = performSETCCCombine(N, DAG);
+    break;
   }
+  }
+
+  if (Val.getNode())
+    return Val;
+
+  return MipsTargetLowering::PerformDAGCombine(N, DCI);
 }
 
 MachineBasicBlock *
diff --git a/test/CodeGen/Mips/dsp-patterns-cmp-vselect.ll b/test/CodeGen/Mips/dsp-patterns-cmp-vselect.ll
new file mode 100644
index 00000000000..9f2f0661f99
--- /dev/null
+++ b/test/CodeGen/Mips/dsp-patterns-cmp-vselect.ll
@@ -0,0 +1,641 @@
+; RUN: llc -march=mips -mattr=dsp < %s | FileCheck %s
+
+; CHECK: select_v2q15_eq_:
+; CHECK: cmp.eq.ph ${{[0-9]+}}, ${{[0-9]+}}
+; CHECK: pick.ph ${{[0-9]+}}, $6, $7
+
+define { i32 } @select_v2q15_eq_(i32 %a0.coerce, i32 %a1.coerce, i32 %a2.coerce, i32 %a3.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <2 x i16>
+  %1 = bitcast i32 %a1.coerce to <2 x i16>
+  %2 = bitcast i32 %a2.coerce to <2 x i16>
+  %3 = bitcast i32 %a3.coerce to <2 x i16>
+  %cmp = icmp eq <2 x i16> %0, %1
+  %or = select <2 x i1> %cmp, <2 x i16> %2, <2 x i16> %3
+  %4 = bitcast <2 x i16> %or to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %4, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: select_v2q15_lt_:
+; CHECK: cmp.lt.ph $4, $5
+; CHECK: pick.ph ${{[0-9]+}}, $6, $7
+
+define { i32 } @select_v2q15_lt_(i32 %a0.coerce, i32 %a1.coerce, i32 %a2.coerce, i32 %a3.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <2 x i16>
+  %1 = bitcast i32 %a1.coerce to <2 x i16>
+  %2 = bitcast i32 %a2.coerce to <2 x i16>
+  %3 = bitcast i32 %a3.coerce to <2 x i16>
+  %cmp = icmp slt <2 x i16> %0, %1
+  %or = select <2 x i1> %cmp, <2 x i16> %2, <2 x i16> %3
+  %4 = bitcast <2 x i16> %or to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %4, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: select_v2q15_le_:
+; CHECK: cmp.le.ph $4, $5
+; CHECK: pick.ph ${{[0-9]+}}, $6, $7
+
+define { i32 } @select_v2q15_le_(i32 %a0.coerce, i32 %a1.coerce, i32 %a2.coerce, i32 %a3.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <2 x i16>
+  %1 = bitcast i32 %a1.coerce to <2 x i16>
+  %2 = bitcast i32 %a2.coerce to <2 x i16>
+  %3 = bitcast i32 %a3.coerce to <2 x i16>
+  %cmp = icmp sle <2 x i16> %0, %1
+  %or = select <2 x i1> %cmp, <2 x i16> %2, <2 x i16> %3
+  %4 = bitcast <2 x i16> %or to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %4, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: select_v2q15_ne_:
+; CHECK: cmp.eq.ph ${{[0-9]+}}, ${{[0-9]+}}
+; CHECK: pick.ph ${{[0-9]+}}, $7, $6
+
+define { i32 } @select_v2q15_ne_(i32 %a0.coerce, i32 %a1.coerce, i32 %a2.coerce, i32 %a3.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <2 x i16>
+  %1 = bitcast i32 %a1.coerce to <2 x i16>
+  %2 = bitcast i32 %a2.coerce to <2 x i16>
+  %3 = bitcast i32 %a3.coerce to <2 x i16>
+  %cmp = icmp ne <2 x i16> %0, %1
+  %or = select <2 x i1> %cmp, <2 x i16> %2, <2 x i16> %3
+  %4 = bitcast <2 x i16> %or to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %4, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: select_v2q15_gt_:
+; CHECK: cmp.le.ph $4, $5
+; CHECK: pick.ph ${{[0-9]+}}, $7, $6
+
+define { i32 } @select_v2q15_gt_(i32 %a0.coerce, i32 %a1.coerce, i32 %a2.coerce, i32 %a3.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <2 x i16>
+  %1 = bitcast i32 %a1.coerce to <2 x i16>
+  %2 = bitcast i32 %a2.coerce to <2 x i16>
+  %3 = bitcast i32 %a3.coerce to <2 x i16>
+  %cmp = icmp sgt <2 x i16> %0, %1
+  %or = select <2 x i1> %cmp, <2 x i16> %2, <2 x i16> %3
+  %4 = bitcast <2 x i16> %or to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %4, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: select_v2q15_ge_:
+; CHECK: cmp.lt.ph $4, $5
+; CHECK: pick.ph ${{[0-9]+}}, $7, $6
+
+define { i32 } @select_v2q15_ge_(i32 %a0.coerce, i32 %a1.coerce, i32 %a2.coerce, i32 %a3.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <2 x i16>
+  %1 = bitcast i32 %a1.coerce to <2 x i16>
+  %2 = bitcast i32 %a2.coerce to <2 x i16>
+  %3 = bitcast i32 %a3.coerce to <2 x i16>
+  %cmp = icmp sge <2 x i16> %0, %1
+  %or = select <2 x i1> %cmp, <2 x i16> %2, <2 x i16> %3
+  %4 = bitcast <2 x i16> %or to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %4, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: select_v4ui8_eq_:
+; CHECK: cmpu.eq.qb ${{[0-9]+}}, ${{[0-9]+}}
+; CHECK: pick.qb ${{[0-9]+}}, $6, $7
+
+define { i32 } @select_v4ui8_eq_(i32 %a0.coerce, i32 %a1.coerce, i32 %a2.coerce, i32 %a3.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <4 x i8>
+  %1 = bitcast i32 %a1.coerce to <4 x i8>
+  %2 = bitcast i32 %a2.coerce to <4 x i8>
+  %3 = bitcast i32 %a3.coerce to <4 x i8>
+  %cmp = icmp eq <4 x i8> %0, %1
+  %or = select <4 x i1> %cmp, <4 x i8> %2, <4 x i8> %3
+  %4 = bitcast <4 x i8> %or to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %4, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: select_v4ui8_lt_:
+; CHECK: cmpu.lt.qb $4, $5
+; CHECK: pick.qb ${{[0-9]+}}, $6, $7
+
+define { i32 } @select_v4ui8_lt_(i32 %a0.coerce, i32 %a1.coerce, i32 %a2.coerce, i32 %a3.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <4 x i8>
+  %1 = bitcast i32 %a1.coerce to <4 x i8>
+  %2 = bitcast i32 %a2.coerce to <4 x i8>
+  %3 = bitcast i32 %a3.coerce to <4 x i8>
+  %cmp = icmp ult <4 x i8> %0, %1
+  %or = select <4 x i1> %cmp, <4 x i8> %2, <4 x i8> %3
+  %4 = bitcast <4 x i8> %or to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %4, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: select_v4ui8_le_:
+; CHECK: cmpu.le.qb $4, $5
+; CHECK: pick.qb ${{[0-9]+}}, $6, $7
+
+define { i32 } @select_v4ui8_le_(i32 %a0.coerce, i32 %a1.coerce, i32 %a2.coerce, i32 %a3.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <4 x i8>
+  %1 = bitcast i32 %a1.coerce to <4 x i8>
+  %2 = bitcast i32 %a2.coerce to <4 x i8>
+  %3 = bitcast i32 %a3.coerce to <4 x i8>
+  %cmp = icmp ule <4 x i8> %0, %1
+  %or = select <4 x i1> %cmp, <4 x i8> %2, <4 x i8> %3
+  %4 = bitcast <4 x i8> %or to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %4, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: select_v4ui8_ne_:
+; CHECK: cmpu.eq.qb ${{[0-9]+}}, ${{[0-9]+}}
+; CHECK: pick.qb ${{[0-9]+}}, $7, $6
+
+define { i32 } @select_v4ui8_ne_(i32 %a0.coerce, i32 %a1.coerce, i32 %a2.coerce, i32 %a3.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <4 x i8>
+  %1 = bitcast i32 %a1.coerce to <4 x i8>
+  %2 = bitcast i32 %a2.coerce to <4 x i8>
+  %3 = bitcast i32 %a3.coerce to <4 x i8>
+  %cmp = icmp ne <4 x i8> %0, %1
+  %or = select <4 x i1> %cmp, <4 x i8> %2, <4 x i8> %3
+  %4 = bitcast <4 x i8> %or to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %4, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: select_v4ui8_gt_:
+; CHECK: cmpu.le.qb $4, $5
+; CHECK: pick.qb ${{[0-9]+}}, $7, $6
+
+define { i32 } @select_v4ui8_gt_(i32 %a0.coerce, i32 %a1.coerce, i32 %a2.coerce, i32 %a3.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <4 x i8>
+  %1 = bitcast i32 %a1.coerce to <4 x i8>
+  %2 = bitcast i32 %a2.coerce to <4 x i8>
+  %3 = bitcast i32 %a3.coerce to <4 x i8>
+  %cmp = icmp ugt <4 x i8> %0, %1
+  %or = select <4 x i1> %cmp, <4 x i8> %2, <4 x i8> %3
+  %4 = bitcast <4 x i8> %or to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %4, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: select_v4ui8_ge_:
+; CHECK: cmpu.lt.qb $4, $5
+; CHECK: pick.qb ${{[0-9]+}}, $7, $6
+
+define { i32 } @select_v4ui8_ge_(i32 %a0.coerce, i32 %a1.coerce, i32 %a2.coerce, i32 %a3.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <4 x i8>
+  %1 = bitcast i32 %a1.coerce to <4 x i8>
+  %2 = bitcast i32 %a2.coerce to <4 x i8>
+  %3 = bitcast i32 %a3.coerce to <4 x i8>
+  %cmp = icmp uge <4 x i8> %0, %1
+  %or = select <4 x i1> %cmp, <4 x i8> %2, <4 x i8> %3
+  %4 = bitcast <4 x i8> %or to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %4, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: select_v2ui16_lt_:
+; CHECK-NOT: cmp
+; CHECK-NOT: pick
+
+define { i32 } @select_v2ui16_lt_(i32 %a0.coerce, i32 %a1.coerce, i32 %a2.coerce, i32 %a3.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <2 x i16>
+  %1 = bitcast i32 %a1.coerce to <2 x i16>
+  %2 = bitcast i32 %a2.coerce to <2 x i16>
+  %3 = bitcast i32 %a3.coerce to <2 x i16>
+  %cmp = icmp ult <2 x i16> %0, %1
+  %or = select <2 x i1> %cmp, <2 x i16> %2, <2 x i16> %3
+  %4 = bitcast <2 x i16> %or to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %4, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: select_v2ui16_le_:
+; CHECK-NOT: cmp
+; CHECK-NOT: pick
+
+define { i32 } @select_v2ui16_le_(i32 %a0.coerce, i32 %a1.coerce, i32 %a2.coerce, i32 %a3.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <2 x i16>
+  %1 = bitcast i32 %a1.coerce to <2 x i16>
+  %2 = bitcast i32 %a2.coerce to <2 x i16>
+  %3 = bitcast i32 %a3.coerce to <2 x i16>
+  %cmp = icmp ule <2 x i16> %0, %1
+  %or = select <2 x i1> %cmp, <2 x i16> %2, <2 x i16> %3
+  %4 = bitcast <2 x i16> %or to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %4, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: select_v2ui16_gt_:
+; CHECK-NOT: cmp
+; CHECK-NOT: pick
+
+define { i32 } @select_v2ui16_gt_(i32 %a0.coerce, i32 %a1.coerce, i32 %a2.coerce, i32 %a3.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <2 x i16>
+  %1 = bitcast i32 %a1.coerce to <2 x i16>
+  %2 = bitcast i32 %a2.coerce to <2 x i16>
+  %3 = bitcast i32 %a3.coerce to <2 x i16>
+  %cmp = icmp ugt <2 x i16> %0, %1
+  %or = select <2 x i1> %cmp, <2 x i16> %2, <2 x i16> %3
+  %4 = bitcast <2 x i16> %or to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %4, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: select_v2ui16_ge_:
+; CHECK-NOT: cmp
+; CHECK-NOT: pick
+
+define { i32 } @select_v2ui16_ge_(i32 %a0.coerce, i32 %a1.coerce, i32 %a2.coerce, i32 %a3.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <2 x i16>
+  %1 = bitcast i32 %a1.coerce to <2 x i16>
+  %2 = bitcast i32 %a2.coerce to <2 x i16>
+  %3 = bitcast i32 %a3.coerce to <2 x i16>
+  %cmp = icmp uge <2 x i16> %0, %1
+  %or = select <2 x i1> %cmp, <2 x i16> %2, <2 x i16> %3
+  %4 = bitcast <2 x i16> %or to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %4, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: select_v4i8_lt_:
+; CHECK-NOT: cmp
+; CHECK-NOT: pick
+
+define { i32 } @select_v4i8_lt_(i32 %a0.coerce, i32 %a1.coerce, i32 %a2.coerce, i32 %a3.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <4 x i8>
+  %1 = bitcast i32 %a1.coerce to <4 x i8>
+  %2 = bitcast i32 %a2.coerce to <4 x i8>
+  %3 = bitcast i32 %a3.coerce to <4 x i8>
+  %cmp = icmp slt <4 x i8> %0, %1
+  %or = select <4 x i1> %cmp, <4 x i8> %2, <4 x i8> %3
+  %4 = bitcast <4 x i8> %or to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %4, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: select_v4i8_le_:
+; CHECK-NOT: cmp
+; CHECK-NOT: pick
+
+define { i32 } @select_v4i8_le_(i32 %a0.coerce, i32 %a1.coerce, i32 %a2.coerce, i32 %a3.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <4 x i8>
+  %1 = bitcast i32 %a1.coerce to <4 x i8>
+  %2 = bitcast i32 %a2.coerce to <4 x i8>
+  %3 = bitcast i32 %a3.coerce to <4 x i8>
+  %cmp = icmp sle <4 x i8> %0, %1
+  %or = select <4 x i1> %cmp, <4 x i8> %2, <4 x i8> %3
+  %4 = bitcast <4 x i8> %or to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %4, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: select_v4i8_gt_:
+; CHECK-NOT: cmp
+; CHECK-NOT: pick
+
+define { i32 } @select_v4i8_gt_(i32 %a0.coerce, i32 %a1.coerce, i32 %a2.coerce, i32 %a3.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <4 x i8>
+  %1 = bitcast i32 %a1.coerce to <4 x i8>
+  %2 = bitcast i32 %a2.coerce to <4 x i8>
+  %3 = bitcast i32 %a3.coerce to <4 x i8>
+  %cmp = icmp sgt <4 x i8> %0, %1
+  %or = select <4 x i1> %cmp, <4 x i8> %2, <4 x i8> %3
+  %4 = bitcast <4 x i8> %or to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %4, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: select_v4i8_ge_:
+; CHECK-NOT: cmp
+; CHECK-NOT: pick
+
+define { i32 } @select_v4i8_ge_(i32 %a0.coerce, i32 %a1.coerce, i32 %a2.coerce, i32 %a3.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <4 x i8>
+  %1 = bitcast i32 %a1.coerce to <4 x i8>
+  %2 = bitcast i32 %a2.coerce to <4 x i8>
+  %3 = bitcast i32 %a3.coerce to <4 x i8>
+  %cmp = icmp sge <4 x i8> %0, %1
+  %or = select <4 x i1> %cmp, <4 x i8> %2, <4 x i8> %3
+  %4 = bitcast <4 x i8> %or to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %4, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: compare_v2q15_eq_:
+; CHECK: cmp.eq.ph ${{[0-9]+}}, ${{[0-9]+}}
+; CHECK: pick.ph ${{[0-9]+}}, ${{[a-z0-9]+}}, ${{[a-z0-9]+}}
+
+define { i32 } @compare_v2q15_eq_(i32 %a0.coerce, i32 %a1.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <2 x i16>
+  %1 = bitcast i32 %a1.coerce to <2 x i16>
+  %cmp = icmp eq <2 x i16> %0, %1
+  %sext = sext <2 x i1> %cmp to <2 x i16>
+  %2 = bitcast <2 x i16> %sext to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %2, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: compare_v2q15_lt_:
+; CHECK: cmp.lt.ph $4, $5
+; CHECK: pick.ph ${{[0-9]+}}, ${{[a-z0-9]+}}, ${{[a-z0-9]+}}
+
+define { i32 } @compare_v2q15_lt_(i32 %a0.coerce, i32 %a1.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <2 x i16>
+  %1 = bitcast i32 %a1.coerce to <2 x i16>
+  %cmp = icmp slt <2 x i16> %0, %1
+  %sext = sext <2 x i1> %cmp to <2 x i16>
+  %2 = bitcast <2 x i16> %sext to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %2, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: compare_v2q15_le_:
+; CHECK: cmp.le.ph $4, $5
+; CHECK: pick.ph ${{[0-9]+}}, ${{[a-z0-9]+}}, ${{[a-z0-9]+}}
+
+define { i32 } @compare_v2q15_le_(i32 %a0.coerce, i32 %a1.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <2 x i16>
+  %1 = bitcast i32 %a1.coerce to <2 x i16>
+  %cmp = icmp sle <2 x i16> %0, %1
+  %sext = sext <2 x i1> %cmp to <2 x i16>
+  %2 = bitcast <2 x i16> %sext to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %2, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: compare_v2q15_ne_:
+; CHECK: cmp.eq.ph ${{[0-9]+}}, ${{[0-9]+}}
+; CHECK: pick.ph ${{[0-9]+}}, ${{[a-z0-9]+}}, ${{[a-z0-9]+}}
+
+define { i32 } @compare_v2q15_ne_(i32 %a0.coerce, i32 %a1.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <2 x i16>
+  %1 = bitcast i32 %a1.coerce to <2 x i16>
+  %cmp = icmp ne <2 x i16> %0, %1
+  %sext = sext <2 x i1> %cmp to <2 x i16>
+  %2 = bitcast <2 x i16> %sext to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %2, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: compare_v2q15_gt_:
+; CHECK: cmp.le.ph $4, $5
+; CHECK: pick.ph ${{[0-9]+}}, ${{[a-z0-9]+}}, ${{[a-z0-9]+}}
+
+define { i32 } @compare_v2q15_gt_(i32 %a0.coerce, i32 %a1.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <2 x i16>
+  %1 = bitcast i32 %a1.coerce to <2 x i16>
+  %cmp = icmp sgt <2 x i16> %0, %1
+  %sext = sext <2 x i1> %cmp to <2 x i16>
+  %2 = bitcast <2 x i16> %sext to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %2, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: compare_v2q15_ge_:
+; CHECK: cmp.lt.ph $4, $5
+; CHECK: pick.ph ${{[0-9]+}}, ${{[a-z0-9]+}}, ${{[a-z0-9]+}}
+
+define { i32 } @compare_v2q15_ge_(i32 %a0.coerce, i32 %a1.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <2 x i16>
+  %1 = bitcast i32 %a1.coerce to <2 x i16>
+  %cmp = icmp sge <2 x i16> %0, %1
+  %sext = sext <2 x i1> %cmp to <2 x i16>
+  %2 = bitcast <2 x i16> %sext to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %2, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: compare_v4ui8_eq_:
+; CHECK: cmpu.eq.qb ${{[0-9]+}}, ${{[0-9]+}}
+; CHECK: pick.qb ${{[0-9]+}}, ${{[a-z0-9]+}}, ${{[a-z0-9]+}}
+
+define { i32 } @compare_v4ui8_eq_(i32 %a0.coerce, i32 %a1.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <4 x i8>
+  %1 = bitcast i32 %a1.coerce to <4 x i8>
+  %cmp = icmp eq <4 x i8> %0, %1
+  %sext = sext <4 x i1> %cmp to <4 x i8>
+  %2 = bitcast <4 x i8> %sext to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %2, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: compare_v4ui8_lt_:
+; CHECK: cmpu.lt.qb $4, $5
+; CHECK: pick.qb ${{[0-9]+}}, ${{[a-z0-9]+}}, ${{[a-z0-9]+}}
+
+define { i32 } @compare_v4ui8_lt_(i32 %a0.coerce, i32 %a1.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <4 x i8>
+  %1 = bitcast i32 %a1.coerce to <4 x i8>
+  %cmp = icmp ult <4 x i8> %0, %1
+  %sext = sext <4 x i1> %cmp to <4 x i8>
+  %2 = bitcast <4 x i8> %sext to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %2, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: compare_v4ui8_le_:
+; CHECK: cmpu.le.qb $4, $5
+; CHECK: pick.qb ${{[0-9]+}}, ${{[a-z0-9]+}}, ${{[a-z0-9]+}}
+
+define { i32 } @compare_v4ui8_le_(i32 %a0.coerce, i32 %a1.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <4 x i8>
+  %1 = bitcast i32 %a1.coerce to <4 x i8>
+  %cmp = icmp ule <4 x i8> %0, %1
+  %sext = sext <4 x i1> %cmp to <4 x i8>
+  %2 = bitcast <4 x i8> %sext to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %2, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: compare_v4ui8_ne_:
+; CHECK: cmpu.eq.qb ${{[0-9]+}}, ${{[0-9]+}}
+; CHECK: pick.qb ${{[0-9]+}}, ${{[a-z0-9]+}}, ${{[a-z0-9]+}}
+
+define { i32 } @compare_v4ui8_ne_(i32 %a0.coerce, i32 %a1.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <4 x i8>
+  %1 = bitcast i32 %a1.coerce to <4 x i8>
+  %cmp = icmp ne <4 x i8> %0, %1
+  %sext = sext <4 x i1> %cmp to <4 x i8>
+  %2 = bitcast <4 x i8> %sext to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %2, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: compare_v4ui8_gt_:
+; CHECK: cmpu.le.qb $4, $5
+; CHECK: pick.qb ${{[0-9]+}}, ${{[a-z0-9]+}}, ${{[a-z0-9]+}}
+
+define { i32 } @compare_v4ui8_gt_(i32 %a0.coerce, i32 %a1.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <4 x i8>
+  %1 = bitcast i32 %a1.coerce to <4 x i8>
+  %cmp = icmp ugt <4 x i8> %0, %1
+  %sext = sext <4 x i1> %cmp to <4 x i8>
+  %2 = bitcast <4 x i8> %sext to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %2, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: compare_v4ui8_ge_:
+; CHECK: cmpu.lt.qb $4, $5
+; CHECK: pick.qb ${{[0-9]+}}, ${{[a-z0-9]+}}, ${{[a-z0-9]+}}
+
+define { i32 } @compare_v4ui8_ge_(i32 %a0.coerce, i32 %a1.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <4 x i8>
+  %1 = bitcast i32 %a1.coerce to <4 x i8>
+  %cmp = icmp uge <4 x i8> %0, %1
+  %sext = sext <4 x i1> %cmp to <4 x i8>
+  %2 = bitcast <4 x i8> %sext to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %2, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: compare_v2ui16_lt_:
+; CHECK-NOT: cmp
+; CHECK-NOT: pick
+
+define { i32 } @compare_v2ui16_lt_(i32 %a0.coerce, i32 %a1.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <2 x i16>
+  %1 = bitcast i32 %a1.coerce to <2 x i16>
+  %cmp = icmp ult <2 x i16> %0, %1
+  %sext = sext <2 x i1> %cmp to <2 x i16>
+  %2 = bitcast <2 x i16> %sext to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %2, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: compare_v2ui16_le_:
+; CHECK-NOT: cmp
+; CHECK-NOT: pick
+
+define { i32 } @compare_v2ui16_le_(i32 %a0.coerce, i32 %a1.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <2 x i16>
+  %1 = bitcast i32 %a1.coerce to <2 x i16>
+  %cmp = icmp ule <2 x i16> %0, %1
+  %sext = sext <2 x i1> %cmp to <2 x i16>
+  %2 = bitcast <2 x i16> %sext to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %2, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: compare_v2ui16_gt_:
+; CHECK-NOT: cmp
+; CHECK-NOT: pick
+
+define { i32 } @compare_v2ui16_gt_(i32 %a0.coerce, i32 %a1.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <2 x i16>
+  %1 = bitcast i32 %a1.coerce to <2 x i16>
+  %cmp = icmp ugt <2 x i16> %0, %1
+  %sext = sext <2 x i1> %cmp to <2 x i16>
+  %2 = bitcast <2 x i16> %sext to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %2, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: compare_v2ui16_ge_:
+; CHECK-NOT: cmp
+; CHECK-NOT: pick
+
+define { i32 } @compare_v2ui16_ge_(i32 %a0.coerce, i32 %a1.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <2 x i16>
+  %1 = bitcast i32 %a1.coerce to <2 x i16>
+  %cmp = icmp uge <2 x i16> %0, %1
+  %sext = sext <2 x i1> %cmp to <2 x i16>
+  %2 = bitcast <2 x i16> %sext to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %2, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: compare_v4i8_lt_:
+; CHECK-NOT: cmp
+; CHECK-NOT: pick
+
+define { i32 } @compare_v4i8_lt_(i32 %a0.coerce, i32 %a1.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <4 x i8>
+  %1 = bitcast i32 %a1.coerce to <4 x i8>
+  %cmp = icmp slt <4 x i8> %0, %1
+  %sext = sext <4 x i1> %cmp to <4 x i8>
+  %2 = bitcast <4 x i8> %sext to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %2, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: compare_v4i8_le_:
+; CHECK-NOT: cmp
+; CHECK-NOT: pick
+
+define { i32 } @compare_v4i8_le_(i32 %a0.coerce, i32 %a1.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <4 x i8>
+  %1 = bitcast i32 %a1.coerce to <4 x i8>
+  %cmp = icmp sle <4 x i8> %0, %1
+  %sext = sext <4 x i1> %cmp to <4 x i8>
+  %2 = bitcast <4 x i8> %sext to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %2, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: compare_v4i8_gt_:
+; CHECK-NOT: cmp
+; CHECK-NOT: pick
+
+define { i32 } @compare_v4i8_gt_(i32 %a0.coerce, i32 %a1.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <4 x i8>
+  %1 = bitcast i32 %a1.coerce to <4 x i8>
+  %cmp = icmp sgt <4 x i8> %0, %1
+  %sext = sext <4 x i1> %cmp to <4 x i8>
+  %2 = bitcast <4 x i8> %sext to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %2, 0
+  ret { i32 } %.fca.0.insert
+}
+
+; CHECK: compare_v4i8_ge_:
+; CHECK-NOT: cmp
+; CHECK-NOT: pick
+
+define { i32 } @compare_v4i8_ge_(i32 %a0.coerce, i32 %a1.coerce) {
+entry:
+  %0 = bitcast i32 %a0.coerce to <4 x i8>
+  %1 = bitcast i32 %a1.coerce to <4 x i8>
+  %cmp = icmp sge <4 x i8> %0, %1
+  %sext = sext <4 x i1> %cmp to <4 x i8>
+  %2 = bitcast <4 x i8> %sext to i32
+  %.fca.0.insert = insertvalue { i32 } undef, i32 %2, 0
+  ret { i32 } %.fca.0.insert
+}
-- 
2.34.1