From cdba4e1bccb84a909db3f4183147f2e543f2ea10 Mon Sep 17 00:00:00 2001
From: Colin LeMahieu <colinl@codeaurora.org>
Date: Tue, 3 Feb 2015 18:01:45 +0000
Subject: [PATCH] [Hexagon] Adding vector intrinsics for alu32/alu and
 xtype/alu.

llvm-svn: 227993
---
 llvm/lib/Target/Hexagon/HexagonIntrinsics.td      | 268 +++++--------
 llvm/lib/Target/Hexagon/HexagonIntrinsicsV4.td    |  31 ++
 llvm/test/CodeGen/Hexagon/intrinsics/alu32_alu.ll |  66 +++
 llvm/test/CodeGen/Hexagon/intrinsics/xtype_alu.ll | 467 ++++++++++++++++++++++
 4 files changed, 657 insertions(+), 175 deletions(-)

diff --git a/llvm/lib/Target/Hexagon/HexagonIntrinsics.td b/llvm/lib/Target/Hexagon/HexagonIntrinsics.td
index e4d8c44..ec2b5db 100644
--- a/llvm/lib/Target/Hexagon/HexagonIntrinsics.td
+++ b/llvm/lib/Target/Hexagon/HexagonIntrinsics.td
@@ -629,6 +629,22 @@ def : Pat <(i32 (int_hexagon_C2_cmpltu (I32:$src1),
       (i32 (C2_cmpgtu (I32:$src2), (I32:$src1)))>;
 
 /********************************************************************
+*            ALU32/VH                                               *
+*********************************************************************/
+// Vector add, subtract, average halfwords
+def: T_RR_pat<A2_svaddh,   int_hexagon_A2_svaddh>;
+def: T_RR_pat<A2_svaddhs,  int_hexagon_A2_svaddhs>;
+def: T_RR_pat<A2_svadduhs, int_hexagon_A2_svadduhs>;
+
+def: T_RR_pat<A2_svsubh,   int_hexagon_A2_svsubh>;
+def: T_RR_pat<A2_svsubhs,  int_hexagon_A2_svsubhs>;
+def: T_RR_pat<A2_svsubuhs, int_hexagon_A2_svsubuhs>;
+
+def: T_RR_pat<A2_svavgh,   int_hexagon_A2_svavgh>;
+def: T_RR_pat<A2_svavghs,  int_hexagon_A2_svavghs>;
+def: T_RR_pat<A2_svnavgh,  int_hexagon_A2_svnavgh>;
+
+/********************************************************************
 *            ALU64/ALU                                              *
 *********************************************************************/
 def: T_RR_pat<A2_addsat,   int_hexagon_A2_addsat>;
@@ -647,6 +663,83 @@ def: T_PP_pat<C2_cmpgtup,  int_hexagon_C2_cmpgtup>;
 def: T_PP_pat<S2_parityp,  int_hexagon_S2_parityp>;
 def: T_RR_pat<S2_packhl,   int_hexagon_S2_packhl>;
 
+/********************************************************************
+*            ALU64/VB                                               *
+*********************************************************************/
+// ALU64 - Vector add
+def : T_PP_pat <A2_vaddub,   int_hexagon_A2_vaddub>;
+def : T_PP_pat <A2_vaddubs,  int_hexagon_A2_vaddubs>;
+def : T_PP_pat <A2_vaddh,    int_hexagon_A2_vaddh>;
+def : T_PP_pat <A2_vaddhs,   int_hexagon_A2_vaddhs>;
+def : T_PP_pat <A2_vadduhs,  int_hexagon_A2_vadduhs>;
+def : T_PP_pat <A2_vaddw,    int_hexagon_A2_vaddw>;
+def : T_PP_pat <A2_vaddws,   int_hexagon_A2_vaddws>;
+
+// ALU64 - Vector average
+def : T_PP_pat <A2_vavgub,   int_hexagon_A2_vavgub>;
+def : T_PP_pat <A2_vavgubr,  int_hexagon_A2_vavgubr>;
+def : T_PP_pat <A2_vavgh,    int_hexagon_A2_vavgh>;
+def : T_PP_pat <A2_vavghr,   int_hexagon_A2_vavghr>;
+def : T_PP_pat <A2_vavghcr,  int_hexagon_A2_vavghcr>;
+def : T_PP_pat <A2_vavguh,   int_hexagon_A2_vavguh>;
+def : T_PP_pat <A2_vavguhr,  int_hexagon_A2_vavguhr>;
+
+def : T_PP_pat <A2_vavgw,    int_hexagon_A2_vavgw>;
+def : T_PP_pat <A2_vavgwr,   int_hexagon_A2_vavgwr>;
+def : T_PP_pat <A2_vavgwcr,  int_hexagon_A2_vavgwcr>;
+def : T_PP_pat <A2_vavguw,   int_hexagon_A2_vavguw>;
+def : T_PP_pat <A2_vavguwr,  int_hexagon_A2_vavguwr>;
+
+// ALU64 - Vector negative average
+def : T_PP_pat <A2_vnavgh,   int_hexagon_A2_vnavgh>;
+def : T_PP_pat <A2_vnavghr,  int_hexagon_A2_vnavghr>;
+def : T_PP_pat <A2_vnavghcr, int_hexagon_A2_vnavghcr>;
+def : T_PP_pat <A2_vnavgw,   int_hexagon_A2_vnavgw>;
+def : T_PP_pat <A2_vnavgwr,  int_hexagon_A2_vnavgwr>;
+def : T_PP_pat <A2_vnavgwcr, int_hexagon_A2_vnavgwcr>;
+
+// ALU64 - Vector max
+def : T_PP_pat <A2_vmaxh,    int_hexagon_A2_vmaxh>;
+def : T_PP_pat <A2_vmaxw,    int_hexagon_A2_vmaxw>;
+def : T_PP_pat <A2_vmaxub,   int_hexagon_A2_vmaxub>;
+def : T_PP_pat <A2_vmaxuh,   int_hexagon_A2_vmaxuh>;
+def : T_PP_pat <A2_vmaxuw,   int_hexagon_A2_vmaxuw>;
+
+// ALU64 - Vector min
+def : T_PP_pat <A2_vminh,    int_hexagon_A2_vminh>;
+def : T_PP_pat <A2_vminw,    int_hexagon_A2_vminw>;
+def : T_PP_pat <A2_vminub,   int_hexagon_A2_vminub>;
+def : T_PP_pat <A2_vminuh,   int_hexagon_A2_vminuh>;
+def : T_PP_pat <A2_vminuw,   int_hexagon_A2_vminuw>;
+
+// ALU64 - Vector sub
+def : T_PP_pat <A2_vsubub,   int_hexagon_A2_vsubub>;
+def : T_PP_pat <A2_vsububs,  int_hexagon_A2_vsububs>;
+def : T_PP_pat <A2_vsubh,    int_hexagon_A2_vsubh>;
+def : T_PP_pat <A2_vsubhs,   int_hexagon_A2_vsubhs>;
+def : T_PP_pat <A2_vsubuhs,  int_hexagon_A2_vsubuhs>;
+def : T_PP_pat <A2_vsubw,    int_hexagon_A2_vsubw>;
+def : T_PP_pat <A2_vsubws,   int_hexagon_A2_vsubws>;
+
+// ALU64 - Vector compare bytes
+def : T_PP_pat <A2_vcmpbeq,  int_hexagon_A2_vcmpbeq>;
+def : T_PP_pat <A4_vcmpbgt,  int_hexagon_A4_vcmpbgt>;
+def : T_PP_pat <A2_vcmpbgtu, int_hexagon_A2_vcmpbgtu>;
+
+// ALU64 - Vector compare halfwords
+def : T_PP_pat <A2_vcmpheq,  int_hexagon_A2_vcmpheq>;
+def : T_PP_pat <A2_vcmphgt,  int_hexagon_A2_vcmphgt>;
+def : T_PP_pat <A2_vcmphgtu, int_hexagon_A2_vcmphgtu>;
+
+// ALU64 - Vector compare words
+def : T_PP_pat <A2_vcmpweq,  int_hexagon_A2_vcmpweq>;
+def : T_PP_pat <A2_vcmpwgt,  int_hexagon_A2_vcmpwgt>;
+def : T_PP_pat <A2_vcmpwgtu, int_hexagon_A2_vcmpwgtu>;
+
+// ALU64 / VB / Vector mux.
+def : Pat<(int_hexagon_C2_vmux PredRegs:$Pu, DoubleRegs:$Rs, DoubleRegs:$Rt),
+          (C2_vmux PredRegs:$Pu, DoubleRegs:$Rs, DoubleRegs:$Rt)>;
+
 // MPY - Multiply and use full result
 // Rdd = mpy[u](Rs, Rt)
 def : T_RR_pat <M2_dpmpyss_s0, int_hexagon_M2_dpmpyss_s0>;
@@ -1387,60 +1480,6 @@ class di_LDInstPI_diu4<string opc, Intrinsic IntID>
            [],
            "$src1 = $dst">;
 
-/********************************************************************
-*            ALU32/VH                                               *
-*********************************************************************/
-
-// ALU32 / VH / Vector add halfwords.
-// Rd32=vadd[u]h(Rs32,Rt32:sat]
-def HEXAGON_A2_svaddh:
-  si_ALU32_sisi                   <"vaddh",    int_hexagon_A2_svaddh>;
-def HEXAGON_A2_svaddhs:
-  si_ALU32_sisi_sat               <"vaddh",    int_hexagon_A2_svaddhs>;
-def HEXAGON_A2_svadduhs:
-  si_ALU32_sisi_sat               <"vadduh",   int_hexagon_A2_svadduhs>;
-
-// ALU32 / VH / Vector average halfwords.
-def HEXAGON_A2_svavgh:
-  si_ALU32_sisi                   <"vavgh",    int_hexagon_A2_svavgh>;
-def HEXAGON_A2_svavghs:
-  si_ALU32_sisi_rnd               <"vavgh",    int_hexagon_A2_svavghs>;
-def HEXAGON_A2_svnavgh:
-  si_ALU32_sisi                   <"vnavgh",   int_hexagon_A2_svnavgh>;
-
-// ALU32 / VH / Vector subtract halfwords.
-def HEXAGON_A2_svsubh:
-  si_ALU32_sisi                   <"vsubh",    int_hexagon_A2_svsubh>;
-def HEXAGON_A2_svsubhs:
-  si_ALU32_sisi_sat               <"vsubh",    int_hexagon_A2_svsubhs>;
-def HEXAGON_A2_svsubuhs:
-  si_ALU32_sisi_sat               <"vsubuh",   int_hexagon_A2_svsubuhs>;
-
-// ALU64 / ALU / Transfer register.
-def HEXAGON_A2_tfrp:
-  di_ALU64_di                     <"",         int_hexagon_A2_tfrp>;
-
-/********************************************************************
-*            ALU64/VB                                               *
-*********************************************************************/
-
-// ALU64 / VB / Vector add unsigned bytes.
-def HEXAGON_A2_vaddub:
-  di_ALU64_didi                   <"vaddub",   int_hexagon_A2_vaddub>;
-def HEXAGON_A2_vaddubs:
-  di_ALU64_didi_sat               <"vaddub",   int_hexagon_A2_vaddubs>;
-
-// ALU64 / VB / Vector average unsigned bytes.
-def HEXAGON_A2_vavgub:
-  di_ALU64_didi                   <"vavgub",   int_hexagon_A2_vavgub>;
-def HEXAGON_A2_vavgubr:
-  di_ALU64_didi_rnd               <"vavgub",   int_hexagon_A2_vavgubr>;
-
-// ALU64 / VB / Vector compare unsigned bytes.
-def HEXAGON_A2_vcmpbeq:
-  qi_ALU64_didi                   <"vcmpb.eq", int_hexagon_A2_vcmpbeq>;
-def HEXAGON_A2_vcmpbgtu:
-  qi_ALU64_didi                   <"vcmpb.gtu",int_hexagon_A2_vcmpbgtu>;
 
 // ALU64 / VB / Vector maximum/minimum unsigned bytes.
 def HEXAGON_A2_vmaxub:
@@ -1454,127 +1493,6 @@ def HEXAGON_A2_vsubub:
 def HEXAGON_A2_vsububs:
   di_ALU64_didi_sat               <"vsubub",   int_hexagon_A2_vsububs>;
 
-// ALU64 / VB / Vector mux.
-def HEXAGON_C2_vmux:
-  di_ALU64_qididi                 <"vmux",     int_hexagon_C2_vmux>;
-
-
-/********************************************************************
-*            ALU64/VH                                               *
-*********************************************************************/
-
-// ALU64 / VH / Vector add halfwords.
-// Rdd64=vadd[u]h(Rss64,Rtt64:sat]
-def HEXAGON_A2_vaddh:
-  di_ALU64_didi                   <"vaddh",    int_hexagon_A2_vaddh>;
-def HEXAGON_A2_vaddhs:
-  di_ALU64_didi_sat               <"vaddh",    int_hexagon_A2_vaddhs>;
-def HEXAGON_A2_vadduhs:
-  di_ALU64_didi_sat               <"vadduh",   int_hexagon_A2_vadduhs>;
-
-// ALU64 / VH / Vector average halfwords.
-// Rdd64=v[n]avg[u]h(Rss64,Rtt64:rnd/:crnd][:sat]
-def HEXAGON_A2_vavgh:
-  di_ALU64_didi                   <"vavgh",    int_hexagon_A2_vavgh>;
-def HEXAGON_A2_vavghcr:
-  di_ALU64_didi_crnd              <"vavgh",    int_hexagon_A2_vavghcr>;
-def HEXAGON_A2_vavghr:
-  di_ALU64_didi_rnd               <"vavgh",    int_hexagon_A2_vavghr>;
-def HEXAGON_A2_vavguh:
-  di_ALU64_didi                   <"vavguh",   int_hexagon_A2_vavguh>;
-def HEXAGON_A2_vavguhr:
-  di_ALU64_didi_rnd               <"vavguh",   int_hexagon_A2_vavguhr>;
-def HEXAGON_A2_vnavgh:
-  di_ALU64_didi                   <"vnavgh",   int_hexagon_A2_vnavgh>;
-def HEXAGON_A2_vnavghcr:
-  di_ALU64_didi_crnd_sat          <"vnavgh",   int_hexagon_A2_vnavghcr>;
-def HEXAGON_A2_vnavghr:
-  di_ALU64_didi_rnd_sat           <"vnavgh",   int_hexagon_A2_vnavghr>;
-
-// ALU64 / VH / Vector compare halfwords.
-def HEXAGON_A2_vcmpheq:
-  qi_ALU64_didi                   <"vcmph.eq", int_hexagon_A2_vcmpheq>;
-def HEXAGON_A2_vcmphgt:
-  qi_ALU64_didi                   <"vcmph.gt", int_hexagon_A2_vcmphgt>;
-def HEXAGON_A2_vcmphgtu:
-  qi_ALU64_didi                   <"vcmph.gtu",int_hexagon_A2_vcmphgtu>;
-
-// ALU64 / VH / Vector maximum halfwords.
-def HEXAGON_A2_vmaxh:
-  di_ALU64_didi                   <"vmaxh",    int_hexagon_A2_vmaxh>;
-def HEXAGON_A2_vmaxuh:
-  di_ALU64_didi                   <"vmaxuh",   int_hexagon_A2_vmaxuh>;
-
-// ALU64 / VH / Vector minimum halfwords.
-def HEXAGON_A2_vminh:
-  di_ALU64_didi                   <"vminh",    int_hexagon_A2_vminh>;
-def HEXAGON_A2_vminuh:
-  di_ALU64_didi                   <"vminuh",   int_hexagon_A2_vminuh>;
-
-// ALU64 / VH / Vector subtract halfwords.
-def HEXAGON_A2_vsubh:
-  di_ALU64_didi                   <"vsubh",    int_hexagon_A2_vsubh>;
-def HEXAGON_A2_vsubhs:
-  di_ALU64_didi_sat               <"vsubh",    int_hexagon_A2_vsubhs>;
-def HEXAGON_A2_vsubuhs:
-  di_ALU64_didi_sat               <"vsubuh",   int_hexagon_A2_vsubuhs>;
-
-
-/********************************************************************
-*            ALU64/VW                                               *
-*********************************************************************/
-
-// ALU64 / VW / Vector add words.
-// Rdd32=vaddw(Rss32,Rtt32)[:sat]
-def HEXAGON_A2_vaddw:
-  di_ALU64_didi                   <"vaddw",    int_hexagon_A2_vaddw>;
-def HEXAGON_A2_vaddws:
-  di_ALU64_didi_sat               <"vaddw",   int_hexagon_A2_vaddws>;
-
-// ALU64 / VW / Vector average words.
-def HEXAGON_A2_vavguw:
-  di_ALU64_didi                   <"vavguw",   int_hexagon_A2_vavguw>;
-def HEXAGON_A2_vavguwr:
-  di_ALU64_didi_rnd               <"vavguw",   int_hexagon_A2_vavguwr>;
-def HEXAGON_A2_vavgw:
-  di_ALU64_didi                   <"vavgw",    int_hexagon_A2_vavgw>;
-def HEXAGON_A2_vavgwcr:
-  di_ALU64_didi_crnd              <"vavgw",    int_hexagon_A2_vavgwcr>;
-def HEXAGON_A2_vavgwr:
-  di_ALU64_didi_rnd               <"vavgw",    int_hexagon_A2_vavgwr>;
-def HEXAGON_A2_vnavgw:
-  di_ALU64_didi                   <"vnavgw",   int_hexagon_A2_vnavgw>;
-def HEXAGON_A2_vnavgwcr:
-  di_ALU64_didi_crnd_sat          <"vnavgw",   int_hexagon_A2_vnavgwcr>;
-def HEXAGON_A2_vnavgwr:
-  di_ALU64_didi_rnd_sat           <"vnavgw",   int_hexagon_A2_vnavgwr>;
-
-// ALU64 / VW / Vector compare words.
-def HEXAGON_A2_vcmpweq:
-  qi_ALU64_didi                   <"vcmpw.eq", int_hexagon_A2_vcmpweq>;
-def HEXAGON_A2_vcmpwgt:
-  qi_ALU64_didi                   <"vcmpw.gt", int_hexagon_A2_vcmpwgt>;
-def HEXAGON_A2_vcmpwgtu:
-  qi_ALU64_didi                   <"vcmpw.gtu",int_hexagon_A2_vcmpwgtu>;
-
-// ALU64 / VW / Vector maximum words.
-def HEXAGON_A2_vmaxw:
-  di_ALU64_didi                   <"vmaxw",    int_hexagon_A2_vmaxw>;
-def HEXAGON_A2_vmaxuw:
-  di_ALU64_didi                   <"vmaxuw",   int_hexagon_A2_vmaxuw>;
-
-// ALU64 / VW / Vector minimum words.
-def HEXAGON_A2_vminw:
-  di_ALU64_didi                   <"vminw",    int_hexagon_A2_vminw>;
-def HEXAGON_A2_vminuw:
-  di_ALU64_didi                   <"vminuw",   int_hexagon_A2_vminuw>;
-
-// ALU64 / VW / Vector subtract words.
-def HEXAGON_A2_vsubw:
-  di_ALU64_didi                   <"vsubw",    int_hexagon_A2_vsubw>;
-def HEXAGON_A2_vsubws:
-  di_ALU64_didi_sat               <"vsubw",    int_hexagon_A2_vsubws>;
-
 /********************************************************************
 *            MTYPE/ALU                                              *
 *********************************************************************/
diff --git a/llvm/lib/Target/Hexagon/HexagonIntrinsicsV4.td b/llvm/lib/Target/Hexagon/HexagonIntrinsicsV4.td
index 58a702c..ae2d463 100644
--- a/llvm/lib/Target/Hexagon/HexagonIntrinsicsV4.td
+++ b/llvm/lib/Target/Hexagon/HexagonIntrinsicsV4.td
@@ -34,6 +34,10 @@ def : MType_R32_pat <int_hexagon_M2_hmmpyh_s1, M2_hmmpyh_s1>;
 def : MType_R32_pat <int_hexagon_M2_hmmpyl_s1, M2_hmmpyl_s1>;
 def : MType_R32_pat <int_hexagon_M2_mpy_up_s1_sat, M2_mpy_up_s1_sat>;
 
+// Vector reduce add unsigned halfwords
+def : Pat <(int_hexagon_M2_vraddh DoubleRegs:$src1, DoubleRegs:$src2),
+           (M2_vraddh DoubleRegs:$src1, DoubleRegs:$src2)>;
+
 def : T_P_pat <S2_brevp, int_hexagon_S2_brevp>;
 
 def: T_P_pat  <S2_ct0p,      int_hexagon_S2_ct0p>;
@@ -82,12 +86,39 @@ def : T_RP_pat  <S4_extract_rp, int_hexagon_S4_extract_rp>;
 def : T_PII_pat <S4_extractp, int_hexagon_S4_extractp>;
 def : T_RII_pat <S4_extract, int_hexagon_S4_extract>;
 
+// Vector conditional negate
+// Rdd=vcnegh(Rss,Rt)
+def : T_PR_pat <S2_vcnegh, int_hexagon_S2_vcnegh>;
+
 // Shift an immediate left by register amount
 def : T_IR_pat<S4_lsli, int_hexagon_S4_lsli>;
 
+// Vector reduce maximum halfwords
+def : T_PPR_pat <A4_vrmaxh, int_hexagon_A4_vrmaxh>;
+def : T_PPR_pat <A4_vrmaxuh, int_hexagon_A4_vrmaxuh>;
+
+// Vector reduce maximum words
+def : T_PPR_pat <A4_vrmaxw, int_hexagon_A4_vrmaxw>;
+def : T_PPR_pat <A4_vrmaxuw, int_hexagon_A4_vrmaxuw>;
+
+// Vector reduce minimum halfwords
+def : T_PPR_pat <A4_vrminh, int_hexagon_A4_vrminh>;
+def : T_PPR_pat <A4_vrminuh, int_hexagon_A4_vrminuh>;
+
+// Vector reduce minimum words
+def : T_PPR_pat <A4_vrminw, int_hexagon_A4_vrminw>;
+def : T_PPR_pat <A4_vrminuw, int_hexagon_A4_vrminuw>;
+
+// Vector conditional negate
+def : T_PPR_pat<S2_vrcnegh, int_hexagon_S2_vrcnegh>;
+
 // Logical xor with xor accumulation
 def : T_PPP_pat<M4_xor_xacc, int_hexagon_M4_xor_xacc>;
 
+// ALU64 - Vector min/max byte
+def : T_PP_pat <A2_vminb, int_hexagon_A2_vminb>;
+def : T_PP_pat <A2_vmaxb, int_hexagon_A2_vmaxb>;
+
 // Shift and add/sub/and/or
 def : T_IRI_pat <S4_andi_asl_ri, int_hexagon_S4_andi_asl_ri>;
 def : T_IRI_pat <S4_ori_asl_ri,  int_hexagon_S4_ori_asl_ri>;
diff --git a/llvm/test/CodeGen/Hexagon/intrinsics/alu32_alu.ll b/llvm/test/CodeGen/Hexagon/intrinsics/alu32_alu.ll
index 6ff9387..37f9f40 100644
--- a/llvm/test/CodeGen/Hexagon/intrinsics/alu32_alu.ll
+++ b/llvm/test/CodeGen/Hexagon/intrinsics/alu32_alu.ll
@@ -127,6 +127,72 @@ define i32 @A2_tfr(i32 %a) {
 }
 ; CHECK: r0 = r0
 
+; Vector add halfwords
+declare i32 @llvm.hexagon.A2.svaddh(i32, i32)
+define i32 @A2_svaddh(i32 %a, i32 %b) {
+  %z = call i32 @llvm.hexagon.A2.svaddh(i32 %a, i32 %b)
+  ret i32 %z
+}
+; CHECK: r0 = vaddh(r0, r1)
+
+declare i32 @llvm.hexagon.A2.svaddhs(i32, i32)
+define i32 @A2_svaddhs(i32 %a, i32 %b) {
+  %z = call i32 @llvm.hexagon.A2.svaddhs(i32 %a, i32 %b)
+  ret i32 %z
+}
+; CHECK: r0 = vaddh(r0, r1):sat
+
+declare i32 @llvm.hexagon.A2.svadduhs(i32, i32)
+define i32 @A2_svadduhs(i32 %a, i32 %b) {
+  %z = call i32 @llvm.hexagon.A2.svadduhs(i32 %a, i32 %b)
+  ret i32 %z
+}
+; CHECK: r0 = vadduh(r0, r1):sat
+
+; Vector average halfwords
+declare i32 @llvm.hexagon.A2.svavgh(i32, i32)
+define i32 @A2_svavgh(i32 %a, i32 %b) {
+  %z = call i32 @llvm.hexagon.A2.svavgh(i32 %a, i32 %b)
+  ret i32 %z
+}
+; CHECK: r0 = vavgh(r0, r1)
+
+declare i32 @llvm.hexagon.A2.svavghs(i32, i32)
+define i32 @A2_svavghs(i32 %a, i32 %b) {
+  %z = call i32 @llvm.hexagon.A2.svavghs(i32 %a, i32 %b)
+  ret i32 %z
+}
+; CHECK: r0 = vavgh(r0, r1):rnd
+
+declare i32 @llvm.hexagon.A2.svnavgh(i32, i32)
+define i32 @A2_svnavgh(i32 %a, i32 %b) {
+  %z = call i32 @llvm.hexagon.A2.svnavgh(i32 %a, i32 %b)
+  ret i32 %z
+}
+; CHECK: r0 = vnavgh(r0, r1)
+
+; Vector subtract halfwords
+declare i32 @llvm.hexagon.A2.svsubh(i32, i32)
+define i32 @A2_svsubh(i32 %a, i32 %b) {
+  %z = call i32 @llvm.hexagon.A2.svsubh(i32 %a, i32 %b)
+  ret i32 %z
+}
+; CHECK: r0 = vsubh(r0, r1)
+
+declare i32 @llvm.hexagon.A2.svsubhs(i32, i32)
+define i32 @A2_svsubhs(i32 %a, i32 %b) {
+  %z = call i32 @llvm.hexagon.A2.svsubhs(i32 %a, i32 %b)
+  ret i32 %z
+}
+; CHECK: r0 = vsubh(r0, r1):sat
+
+declare i32 @llvm.hexagon.A2.svsubuhs(i32, i32)
+define i32 @A2_svsubuhs(i32 %a, i32 %b) {
+  %z = call i32 @llvm.hexagon.A2.svsubuhs(i32 %a, i32 %b)
+  ret i32 %z
+}
+; CHECK: r0 = vsubuh(r0, r1):sat
+
 ; Zero extend
 declare i32 @llvm.hexagon.A2.zxth(i32)
 define i32 @A2_zxth(i32 %a) {
diff --git a/llvm/test/CodeGen/Hexagon/intrinsics/xtype_alu.ll b/llvm/test/CodeGen/Hexagon/intrinsics/xtype_alu.ll
index febf4a6..40d9b6b 100644
--- a/llvm/test/CodeGen/Hexagon/intrinsics/xtype_alu.ll
+++ b/llvm/test/CodeGen/Hexagon/intrinsics/xtype_alu.ll
@@ -551,3 +551,470 @@ define i64 @A2_sxtw(i32 %a) {
   ret i64 %z
 }
 ; CHECK:  = sxtw(r0)
+
+; Vector absolute value halfwords
+declare i64 @llvm.hexagon.A2.vabsh(i64)
+define i64 @A2_vabsh(i64 %a) {
+  %z = call i64 @llvm.hexagon.A2.vabsh(i64 %a)
+  ret i64 %z
+}
+; CHECK: r1:0 = vabsh(r1:0)
+
+declare i64 @llvm.hexagon.A2.vabshsat(i64)
+define i64 @A2_vabshsat(i64 %a) {
+  %z = call i64 @llvm.hexagon.A2.vabshsat(i64 %a)
+  ret i64 %z
+}
+; CHECK: r1:0 = vabsh(r1:0):sat
+
+; Vector absolute value words
+declare i64 @llvm.hexagon.A2.vabsw(i64)
+define i64 @A2_vabsw(i64 %a) {
+  %z = call i64 @llvm.hexagon.A2.vabsw(i64 %a)
+  ret i64 %z
+}
+; CHECK: r1:0 = vabsw(r1:0)
+
+declare i64 @llvm.hexagon.A2.vabswsat(i64)
+define i64 @A2_vabswsat(i64 %a) {
+  %z = call i64 @llvm.hexagon.A2.vabswsat(i64 %a)
+  ret i64 %z
+}
+; CHECK: r1:0 = vabsw(r1:0):sat
+
+; Vector absolute difference halfwords
+declare i64 @llvm.hexagon.M2.vabsdiffh(i64, i64)
+define i64 @M2_vabsdiffh(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.M2.vabsdiffh(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vabsdiffh(r1:0, r3:2)
+
+; Vector absolute difference words
+declare i64 @llvm.hexagon.M2.vabsdiffw(i64, i64)
+define i64 @M2_vabsdiffw(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.M2.vabsdiffw(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vabsdiffw(r1:0, r3:2)
+
+; Vector add halfwords
+declare i64 @llvm.hexagon.A2.vaddh(i64, i64)
+define i64 @A2_vaddh(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vaddh(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vaddh(r1:0, r3:2)
+
+declare i64 @llvm.hexagon.A2.vaddhs(i64, i64)
+define i64 @A2_vaddhs(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vaddhs(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vaddh(r1:0, r3:2):sat
+
+declare i64 @llvm.hexagon.A2.vadduhs(i64, i64)
+define i64 @A2_vadduhs(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vadduhs(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vadduh(r1:0, r3:2):sat
+
+; Vector add halfwords with saturate and pack to unsigned bytes
+declare i32 @llvm.hexagon.A5.vaddhubs(i64, i64)
+define i32 @A5_vaddhubs(i64 %a, i64 %b) {
+  %z = call i32 @llvm.hexagon.A5.vaddhubs(i64 %a, i64 %b)
+  ret i32 %z
+}
+; CHECK: r0 = vaddhub(r1:0, r3:2):sat
+
+; Vector reduce add unsigned bytes
+declare i64 @llvm.hexagon.A2.vraddub(i64, i64)
+define i64 @A2_vraddub(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vraddub(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vraddub(r1:0, r3:2)
+
+declare i64 @llvm.hexagon.A2.vraddub.acc(i64, i64, i64)
+define i64 @A2_vraddub_acc(i64 %a, i64 %b, i64 %c) {
+  %z = call i64 @llvm.hexagon.A2.vraddub.acc(i64 %a, i64 %b, i64 %c)
+  ret i64 %z
+}
+; CHECK: r1:0 += vraddub(r3:2, r5:4)
+
+; Vector reduce add halfwords
+declare i32 @llvm.hexagon.M2.vradduh(i64, i64)
+define i32 @M2_vradduh(i64 %a, i64 %b) {
+  %z = call i32 @llvm.hexagon.M2.vradduh(i64 %a, i64 %b)
+  ret i32 %z
+}
+; CHECK: r0 = vradduh(r1:0, r3:2)
+
+declare i32 @llvm.hexagon.M2.vraddh(i64, i64)
+define i32 @M2_vraddh(i64 %a, i64 %b) {
+  %z = call i32 @llvm.hexagon.M2.vraddh(i64 %a, i64 %b)
+  ret i32 %z
+}
+; CHECK: r0 = vraddh(r1:0, r3:2)
+
+; Vector add bytes
+declare i64 @llvm.hexagon.A2.vaddub(i64, i64)
+define i64 @A2_vaddub(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vaddub(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vaddub(r1:0, r3:2)
+
+declare i64 @llvm.hexagon.A2.vaddubs(i64, i64)
+define i64 @A2_vaddubs(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vaddubs(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vaddub(r1:0, r3:2):sat
+
+; Vector add words
+declare i64 @llvm.hexagon.A2.vaddw(i64, i64)
+define i64 @A2_vaddw(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vaddw(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vaddw(r1:0, r3:2)
+
+declare i64 @llvm.hexagon.A2.vaddws(i64, i64)
+define i64 @A2_vaddws(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vaddws(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vaddw(r1:0, r3:2):sat
+
+; Vector average halfwords
+declare i64 @llvm.hexagon.A2.vavgh(i64, i64)
+define i64 @A2_vavgh(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vavgh(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vavgh(r1:0, r3:2)
+
+declare i64 @llvm.hexagon.A2.vavghr(i64, i64)
+define i64 @A2_vavghr(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vavghr(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vavgh(r1:0, r3:2):rnd
+
+declare i64 @llvm.hexagon.A2.vavghcr(i64, i64)
+define i64 @A2_vavghcr(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vavghcr(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vavgh(r1:0, r3:2):crnd
+
+declare i64 @llvm.hexagon.A2.vavguh(i64, i64)
+define i64 @A2_vavguh(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vavguh(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vavguh(r1:0, r3:2)
+
+declare i64 @llvm.hexagon.A2.vavguhr(i64, i64)
+define i64 @A2_vavguhr(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vavguhr(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vavguh(r1:0, r3:2):rnd
+
+declare i64 @llvm.hexagon.A2.vnavgh(i64, i64)
+define i64 @A2_vnavgh(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vnavgh(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vnavgh(r1:0, r3:2)
+
+declare i64 @llvm.hexagon.A2.vnavghr(i64, i64)
+define i64 @A2_vnavghr(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vnavghr(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vnavgh(r1:0, r3:2):rnd
+
+declare i64 @llvm.hexagon.A2.vnavghcr(i64, i64)
+define i64 @A2_vnavghcr(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vnavghcr(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vnavgh(r1:0, r3:2):crnd
+
+; Vector average unsigned bytes
+declare i64 @llvm.hexagon.A2.vavgub(i64, i64)
+define i64 @A2_vavgub(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vavgub(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vavgub(r1:0, r3:2)
+
+declare i64 @llvm.hexagon.A2.vavgubr(i64, i64)
+define i64 @A2_vavgubr(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vavgubr(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vavgub(r1:0, r3:2):rnd
+
+; Vector average words
+declare i64 @llvm.hexagon.A2.vavgw(i64, i64)
+define i64 @A2_vavgw(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vavgw(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vavgw(r1:0, r3:2)
+
+declare i64 @llvm.hexagon.A2.vavgwr(i64, i64)
+define i64 @A2_vavgwr(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vavgwr(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vavgw(r1:0, r3:2):rnd
+
+declare i64 @llvm.hexagon.A2.vavgwcr(i64, i64)
+define i64 @A2_vavgwcr(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vavgwcr(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vavgw(r1:0, r3:2):crnd
+
+declare i64 @llvm.hexagon.A2.vavguw(i64, i64)
+define i64 @A2_vavguw(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vavguw(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vavguw(r1:0, r3:2)
+
+declare i64 @llvm.hexagon.A2.vavguwr(i64, i64)
+define i64 @A2_vavguwr(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vavguwr(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vavguw(r1:0, r3:2):rnd
+
+declare i64 @llvm.hexagon.A2.vnavgw(i64, i64)
+define i64 @A2_vnavgw(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vnavgw(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vnavgw(r1:0, r3:2)
+
+declare i64 @llvm.hexagon.A2.vnavgwr(i64, i64)
+define i64 @A2_vnavgwr(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vnavgwr(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vnavgw(r1:0, r3:2):rnd
+
+declare i64 @llvm.hexagon.A2.vnavgwcr(i64, i64)
+define i64 @A2_vnavgwcr(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vnavgwcr(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vnavgw(r1:0, r3:2):crnd
+
+; Vector conditional negate
+declare i64 @llvm.hexagon.S2.vcnegh(i64, i32)
+define i64 @S2_vcnegh(i64 %a, i32 %b) {
+  %z = call i64 @llvm.hexagon.S2.vcnegh(i64 %a, i32 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vcnegh(r1:0, r2)
+
+declare i64 @llvm.hexagon.S2.vrcnegh(i64, i64, i32)
+define i64 @S2_vrcnegh(i64 %a, i64 %b, i32 %c) {
+  %z = call i64 @llvm.hexagon.S2.vrcnegh(i64 %a, i64 %b, i32 %c)
+  ret i64 %z
+}
+; CHECK: r1:0 += vrcnegh(r3:2, r4)
+
+; Vector maximum bytes
+declare i64 @llvm.hexagon.A2.vmaxub(i64, i64)
+define i64 @A2_vmaxub(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vmaxub(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vmaxub(r1:0, r3:2)
+
+declare i64 @llvm.hexagon.A2.vmaxb(i64, i64)
+define i64 @A2_vmaxb(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vmaxb(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vmaxb(r1:0, r3:2)
+
+; Vector maximum halfwords
+declare i64 @llvm.hexagon.A2.vmaxh(i64, i64)
+define i64 @A2_vmaxh(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vmaxh(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vmaxh(r1:0, r3:2)
+
+declare i64 @llvm.hexagon.A2.vmaxuh(i64, i64)
+define i64 @A2_vmaxuh(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vmaxuh(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vmaxuh(r1:0, r3:2)
+
+; Vector reduce maximum halfwords
+declare i64 @llvm.hexagon.A4.vrmaxh(i64, i64, i32)
+define i64 @A4_vrmaxh(i64 %a, i64 %b, i32 %c) {
+  %z = call i64 @llvm.hexagon.A4.vrmaxh(i64 %a, i64 %b, i32 %c)
+  ret i64 %z
+}
+; CHECK: r1:0 = vrmaxh(r3:2, r4)
+
+declare i64 @llvm.hexagon.A4.vrmaxuh(i64, i64, i32)
+define i64 @A4_vrmaxuh(i64 %a, i64 %b, i32 %c) {
+  %z = call i64 @llvm.hexagon.A4.vrmaxuh(i64 %a, i64 %b, i32 %c)
+  ret i64 %z
+}
+; CHECK: r1:0 = vrmaxuh(r3:2, r4)
+
+; Vector reduce maximum words
+declare i64 @llvm.hexagon.A4.vrmaxw(i64, i64, i32)
+define i64 @A4_vrmaxw(i64 %a, i64 %b, i32 %c) {
+  %z = call i64 @llvm.hexagon.A4.vrmaxw(i64 %a, i64 %b, i32 %c)
+  ret i64 %z
+}
+; CHECK: r1:0 = vrmaxw(r3:2, r4)
+
+declare i64 @llvm.hexagon.A4.vrmaxuw(i64, i64, i32)
+define i64 @A4_vrmaxuw(i64 %a, i64 %b, i32 %c) {
+  %z = call i64 @llvm.hexagon.A4.vrmaxuw(i64 %a, i64 %b, i32 %c)
+  ret i64 %z
+}
+; CHECK: r1:0 = vrmaxuw(r3:2, r4)
+
+; Vector minimum bytes
+declare i64 @llvm.hexagon.A2.vminub(i64, i64)
+define i64 @A2_vminub(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vminub(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vminub(r1:0, r3:2)
+
+declare i64 @llvm.hexagon.A2.vminb(i64, i64)
+define i64 @A2_vminb(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vminb(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vminb(r1:0, r3:2)
+
+; Vector minimum halfwords
+declare i64 @llvm.hexagon.A2.vminh(i64, i64)
+define i64 @A2_vminh(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vminh(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vminh(r1:0, r3:2)
+
+declare i64 @llvm.hexagon.A2.vminuh(i64, i64)
+define i64 @A2_vminuh(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vminuh(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vminuh(r1:0, r3:2)
+
+; Vector reduce minimum halfwords
+declare i64 @llvm.hexagon.A4.vrminh(i64, i64, i32)
+define i64 @A4_vrminh(i64 %a, i64 %b, i32 %c) {
+  %z = call i64 @llvm.hexagon.A4.vrminh(i64 %a, i64 %b, i32 %c)
+  ret i64 %z
+}
+; CHECK: r1:0 = vrminh(r3:2, r4)
+
+declare i64 @llvm.hexagon.A4.vrminuh(i64, i64, i32)
+define i64 @A4_vrminuh(i64 %a, i64 %b, i32 %c) {
+  %z = call i64 @llvm.hexagon.A4.vrminuh(i64 %a, i64 %b, i32 %c)
+  ret i64 %z
+}
+; CHECK: r1:0 = vrminuh(r3:2, r4)
+
+; Vector reduce minimum words
+declare i64 @llvm.hexagon.A4.vrminw(i64, i64, i32)
+define i64 @A4_vrminw(i64 %a, i64 %b, i32 %c) {
+  %z = call i64 @llvm.hexagon.A4.vrminw(i64 %a, i64 %b, i32 %c)
+  ret i64 %z
+}
+; CHECK: r1:0 = vrminw(r3:2, r4)
+
+declare i64 @llvm.hexagon.A4.vrminuw(i64, i64, i32)
+define i64 @A4_vrminuw(i64 %a, i64 %b, i32 %c) {
+  %z = call i64 @llvm.hexagon.A4.vrminuw(i64 %a, i64 %b, i32 %c)
+  ret i64 %z
+}
+; CHECK: r1:0 = vrminuw(r3:2, r4)
+
+; Vector sum of absolute differences unsigned bytes
+declare i64 @llvm.hexagon.A2.vrsadub(i64, i64)
+define i64 @A2_vrsadub(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vrsadub(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vrsadub(r1:0, r3:2)
+
+declare i64 @llvm.hexagon.A2.vrsadub.acc(i64, i64, i64)
+define i64 @A2_vrsadub_acc(i64 %a, i64 %b, i64 %c) {
+  %z = call i64 @llvm.hexagon.A2.vrsadub.acc(i64 %a, i64 %b, i64 %c)
+  ret i64 %z
+}
+; CHECK: r1:0 += vrsadub(r3:2, r5:4)
+
+; Vector subtract halfwords
+declare i64 @llvm.hexagon.A2.vsubh(i64, i64)
+define i64 @A2_vsubh(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vsubh(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vsubh(r1:0, r3:2)
+
+declare i64 @llvm.hexagon.A2.vsubhs(i64, i64)
+define i64 @A2_vsubhs(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vsubhs(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vsubh(r1:0, r3:2):sat
+
+declare i64 @llvm.hexagon.A2.vsubuhs(i64, i64)
+define i64 @A2_vsubuhs(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vsubuhs(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vsubuh(r1:0, r3:2):sat
+
+; Vector subtract bytes
+declare i64 @llvm.hexagon.A2.vsubub(i64, i64)
+define i64 @A2_vsubub(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vsubub(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vsubub(r1:0, r3:2)
+
+declare i64 @llvm.hexagon.A2.vsububs(i64, i64)
+define i64 @A2_vsububs(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vsububs(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vsubub(r1:0, r3:2):sat
+
+; Vector subtract words
+declare i64 @llvm.hexagon.A2.vsubw(i64, i64)
+define i64 @A2_vsubw(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vsubw(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vsubw(r1:0, r3:2)
+
+declare i64 @llvm.hexagon.A2.vsubws(i64, i64)
+define i64 @A2_vsubws(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.A2.vsubws(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = vsubw(r1:0, r3:2):sat
-- 
2.7.4