From 39b846ce0f72a9d460725385c8774dba4f707852 Mon Sep 17 00:00:00 2001
From: Colin LeMahieu <colinl@codeaurora.org>
Date: Wed, 28 Jan 2015 18:06:23 +0000
Subject: [PATCH] [Hexagon] Converting XTYPE/BIT intrinsic patterns and adding
 tests.

llvm-svn: 227335
---
 llvm/lib/Target/Hexagon/HexagonIntrinsics.td      | 150 +++++++---
 llvm/lib/Target/Hexagon/HexagonIntrinsicsV4.td    |  27 ++
 llvm/test/CodeGen/Hexagon/intrinsics/xtype_bit.ll | 329 ++++++++++++++++++++++
 3 files changed, 472 insertions(+), 34 deletions(-)
 create mode 100644 llvm/test/CodeGen/Hexagon/intrinsics/xtype_bit.ll

diff --git a/llvm/lib/Target/Hexagon/HexagonIntrinsics.td b/llvm/lib/Target/Hexagon/HexagonIntrinsics.td
index f11fc7d..ba4cfae 100644
--- a/llvm/lib/Target/Hexagon/HexagonIntrinsics.td
+++ b/llvm/lib/Target/Hexagon/HexagonIntrinsics.td
@@ -21,6 +21,10 @@ class T_R_pat <InstHexagon MI, Intrinsic IntID>
   : Pat <(IntID I32:$Rs),
          (MI I32:$Rs)>;
 
+class T_P_pat <InstHexagon MI, Intrinsic IntID>
+  : Pat <(IntID I64:$Rs),
+         (MI DoubleRegs:$Rs)>;
+
 class T_II_pat <InstHexagon MI, Intrinsic IntID, PatFrag Imm1, PatFrag Imm2>
   : Pat<(IntID Imm1:$Is, Imm2:$It),
         (MI Imm1:$Is, Imm2:$It)>;
@@ -37,6 +41,10 @@ class T_PI_pat <InstHexagon MI, Intrinsic IntID>
   : Pat<(IntID I64:$Rs, imm:$It),
         (MI DoubleRegs:$Rs, imm:$It)>;
 
+class T_RP_pat <InstHexagon MI, Intrinsic IntID>
+  : Pat<(IntID I32:$Rs, I64:$Rt),
+        (MI I32:$Rs, DoubleRegs:$Rt)>;
+
 class T_RR_pat <InstHexagon MI, Intrinsic IntID>
   : Pat <(IntID I32:$Rs, I32:$Rt),
          (MI I32:$Rs, I32:$Rt)>;
@@ -61,6 +69,10 @@ class T_RRI_pat <InstHexagon MI, Intrinsic IntID>
   : Pat <(IntID I32:$Rs, I32:$Rt, imm:$Iu),
          (MI I32:$Rs, I32:$Rt, imm:$Iu)>;
 
+class T_RII_pat <InstHexagon MI, Intrinsic IntID>
+  : Pat <(IntID I32:$Rs, imm:$It, imm:$Iu),
+         (MI I32:$Rs, imm:$It, imm:$Iu)>;
+
 class T_IRI_pat <InstHexagon MI, Intrinsic IntID>
   : Pat <(IntID imm:$It, I32:$Rs, imm:$Iu),
          (MI imm:$It, I32:$Rs, imm:$Iu)>;
@@ -73,6 +85,10 @@ class T_PPI_pat <InstHexagon MI, Intrinsic IntID>
   : Pat <(IntID I64:$Rs, I64:$Rt, imm:$Iu),
          (MI DoubleRegs:$Rs, DoubleRegs:$Rt, imm:$Iu)>;
 
+class T_PII_pat <InstHexagon MI, Intrinsic IntID>
+  : Pat <(IntID I64:$Rs, imm:$It, imm:$Iu),
+         (MI DoubleRegs:$Rs, imm:$It, imm:$Iu)>;
+
 class T_PPR_pat <InstHexagon MI, Intrinsic IntID>
   : Pat <(IntID I64:$Rs, I64:$Rt, I32:$Ru),
          (MI DoubleRegs:$Rs, DoubleRegs:$Rt, I32:$Ru)>;
@@ -549,6 +565,17 @@ def: T_PP_pat<A2_xorp,     int_hexagon_A2_xorp>;
 def: T_PP_pat<S2_parityp,  int_hexagon_S2_parityp>;
 def: T_RR_pat<S2_packhl,   int_hexagon_S2_packhl>;
 
+// MPY - Multiply and use full result
+// Rdd = mpy[u](Rs, Rt)
+def : T_RR_pat <M2_dpmpyss_s0, int_hexagon_M2_dpmpyss_s0>;
+def : T_RR_pat <M2_dpmpyuu_s0, int_hexagon_M2_dpmpyuu_s0>;
+
+// Rxx[+-]= mpy[u](Rs,Rt)
+def : T_PRR_pat <M2_dpmpyss_acc_s0, int_hexagon_M2_dpmpyss_acc_s0>;
+def : T_PRR_pat <M2_dpmpyss_nac_s0, int_hexagon_M2_dpmpyss_nac_s0>;
+def : T_PRR_pat <M2_dpmpyuu_acc_s0, int_hexagon_M2_dpmpyuu_acc_s0>;
+def : T_PRR_pat <M2_dpmpyuu_nac_s0, int_hexagon_M2_dpmpyuu_nac_s0>;
+
 // Multiply 32x32 and use lower result
 def : T_RRI_pat <M2_macsip, int_hexagon_M2_macsip>;
 def : T_RRI_pat <M2_macsin, int_hexagon_M2_macsin>;
@@ -566,9 +593,98 @@ def : T_RRI_pat <M2_naccii, int_hexagon_M2_naccii>;
 // XOR and XOR with destination
 def : T_RRR_pat <M2_xor_xacc, int_hexagon_M2_xor_xacc>;
 
+class MType_R32_pat <Intrinsic IntID, InstHexagon OutputInst> :
+      Pat <(IntID IntRegs:$src1, IntRegs:$src2),
+           (OutputInst IntRegs:$src1, IntRegs:$src2)>;
+
+// Multiply and use lower result
+def : MType_R32_pat <int_hexagon_M2_mpyi, M2_mpyi>;
+def : T_RI_pat<M2_mpysmi, int_hexagon_M2_mpysmi>;
+
+// Assembler mapped from Rd32=mpyui(Rs32,Rt32) to Rd32=mpyi(Rs32,Rt32)
+def : MType_R32_pat <int_hexagon_M2_mpyui, M2_mpyi>;
+
+// Multiply and use upper result
+def : MType_R32_pat <int_hexagon_M2_mpy_up, M2_mpy_up>;
+def : MType_R32_pat <int_hexagon_M2_mpyu_up, M2_mpyu_up>;
+def : MType_R32_pat <int_hexagon_M2_hmmpyh_rs1, M2_hmmpyh_rs1>;
+def : MType_R32_pat <int_hexagon_M2_hmmpyl_rs1, M2_hmmpyl_rs1>;
+def : MType_R32_pat <int_hexagon_M2_dpmpyss_rnd_s0, M2_dpmpyss_rnd_s0>;
+
+/********************************************************************
+*            STYPE/ALU                                              *
+*********************************************************************/
+def : T_P_pat <A2_absp, int_hexagon_A2_absp>;
+def : T_P_pat <A2_negp, int_hexagon_A2_negp>;
+def : T_P_pat <A2_notp, int_hexagon_A2_notp>;
+
+/********************************************************************
+*            STYPE/BIT                                              *
+*********************************************************************/
+
+// Count leading/trailing
+def: T_R_pat<S2_cl0,     int_hexagon_S2_cl0>;
+def: T_P_pat<S2_cl0p,    int_hexagon_S2_cl0p>;
+def: T_R_pat<S2_cl1,     int_hexagon_S2_cl1>;
+def: T_P_pat<S2_cl1p,    int_hexagon_S2_cl1p>;
+def: T_R_pat<S2_clb,     int_hexagon_S2_clb>;
+def: T_P_pat<S2_clbp,    int_hexagon_S2_clbp>;
+def: T_R_pat<S2_clbnorm, int_hexagon_S2_clbnorm>;
+def: T_R_pat<S2_ct0,     int_hexagon_S2_ct0>;
+def: T_R_pat<S2_ct1,     int_hexagon_S2_ct1>;
+
+// Compare bit mask
+def: T_RR_pat<C2_bitsclr,  int_hexagon_C2_bitsclr>;
+def: T_RI_pat<C2_bitsclri, int_hexagon_C2_bitsclri>;
+def: T_RR_pat<C2_bitsset,  int_hexagon_C2_bitsset>;
+
 // Shift by immediate and add
 def : T_RRI_pat<S2_addasl_rrri, int_hexagon_S2_addasl_rrri>;
 
+// Extract bitfield
+def : T_PII_pat<S2_extractup,    int_hexagon_S2_extractup>;
+def : T_RII_pat<S2_extractu,     int_hexagon_S2_extractu>;
+def : T_RP_pat <S2_extractu_rp,  int_hexagon_S2_extractu_rp>;
+def : T_PP_pat <S2_extractup_rp, int_hexagon_S2_extractup_rp>;
+
+// Insert bitfield
+def : Pat <(int_hexagon_S2_insert_rp IntRegs:$src1, IntRegs:$src2,
+                                     DoubleRegs:$src3),
+           (S2_insert_rp IntRegs:$src1, IntRegs:$src2, DoubleRegs:$src3)>;
+
+def : Pat<(i64 (int_hexagon_S2_insertp_rp (I64:$src1),
+                 (I64:$src2), (I64:$src3))),
+          (i64 (S2_insertp_rp (I64:$src1), (I64:$src2),
+                              (I64:$src3)))>;
+
+def : Pat<(int_hexagon_S2_insert IntRegs:$src1, IntRegs:$src2,
+                                 u5ImmPred:$src3, u5ImmPred:$src4),
+          (S2_insert IntRegs:$src1, IntRegs:$src2,
+                     u5ImmPred:$src3, u5ImmPred:$src4)>;
+
+def : Pat<(i64 (int_hexagon_S2_insertp (I64:$src1),
+                 (I64:$src2), u6ImmPred:$src3, u6ImmPred:$src4)),
+          (i64 (S2_insertp (I64:$src1), (I64:$src2),
+                           u6ImmPred:$src3, u6ImmPred:$src4))>;
+
+
+// Innterleave/deinterleave
+def : T_P_pat <S2_interleave, int_hexagon_S2_interleave>;
+def : T_P_pat <S2_deinterleave, int_hexagon_S2_deinterleave>;
+
+// Set/Clear/Toggle Bit
+def: T_RI_pat<S2_setbit_i,    int_hexagon_S2_setbit_i>;
+def: T_RI_pat<S2_clrbit_i,    int_hexagon_S2_clrbit_i>;
+def: T_RI_pat<S2_togglebit_i, int_hexagon_S2_togglebit_i>;
+
+def: T_RR_pat<S2_setbit_r,    int_hexagon_S2_setbit_r>;
+def: T_RR_pat<S2_clrbit_r,    int_hexagon_S2_clrbit_r>;
+def: T_RR_pat<S2_togglebit_r, int_hexagon_S2_togglebit_r>;
+
+// Test Bit
+def: T_RI_pat<S2_tstbit_i,    int_hexagon_S2_tstbit_i>;
+def: T_RR_pat<S2_tstbit_r,    int_hexagon_S2_tstbit_r>;
+
 /********************************************************************
 *            STYPE/SHIFT                                            *
 *********************************************************************/
@@ -2908,8 +3024,6 @@ def Hexagon_M2_mpysmi:
   si_MInst_sis9                   <"mpyi",     int_hexagon_M2_mpysmi>;
 def HEXAGON_M2_mpyi:
   si_MInst_sisi                   <"mpyi",     int_hexagon_M2_mpyi>;
-def HEXAGON_M2_mpyui:
-  si_MInst_sisi                   <"mpyui",    int_hexagon_M2_mpyui>;
 def HEXAGON_M2_macsip:
   si_MInst_sisiu8_acc             <"mpyi",     int_hexagon_M2_macsip>;
 def HEXAGON_M2_maci:
@@ -2989,32 +3103,6 @@ def HEXAGON_M2_mmacuhs_rs0:
 def HEXAGON_M2_mmacuhs_s0:
   di_MInst_dididi_acc_sat         <"vmpywouh", int_hexagon_M2_mmacuhs_s0>;
 
-// MTYPE / MPYH / Multiply and use upper result.
-def HEXAGON_M2_hmmpyh_rs1:
-  si_MInst_sisi_h_s1_rnd_sat      <"mpy",      int_hexagon_M2_hmmpyh_rs1>;
-def HEXAGON_M2_hmmpyl_rs1:
-  si_MInst_sisi_l_s1_rnd_sat      <"mpy",      int_hexagon_M2_hmmpyl_rs1>;
-def HEXAGON_M2_mpy_up:
-  si_MInst_sisi                   <"mpy",      int_hexagon_M2_mpy_up>;
-def HEXAGON_M2_dpmpyss_rnd_s0:
-  si_MInst_sisi_rnd               <"mpy",      int_hexagon_M2_dpmpyss_rnd_s0>;
-def HEXAGON_M2_mpyu_up:
-  si_MInst_sisi                   <"mpyu",     int_hexagon_M2_mpyu_up>;
-
-// MTYPE / MPYH / Multiply and use full result.
-def HEXAGON_M2_dpmpyuu_s0:
-  di_MInst_sisi                   <"mpyu",     int_hexagon_M2_dpmpyuu_s0>;
-def HEXAGON_M2_dpmpyuu_acc_s0:
-  di_MInst_disisi_acc             <"mpyu",     int_hexagon_M2_dpmpyuu_acc_s0>;
-def HEXAGON_M2_dpmpyuu_nac_s0:
-  di_MInst_disisi_nac             <"mpyu",     int_hexagon_M2_dpmpyuu_nac_s0>;
-def HEXAGON_M2_dpmpyss_s0:
-  di_MInst_sisi                   <"mpy",      int_hexagon_M2_dpmpyss_s0>;
-def HEXAGON_M2_dpmpyss_acc_s0:
-  di_MInst_disisi_acc             <"mpy",      int_hexagon_M2_dpmpyss_acc_s0>;
-def HEXAGON_M2_dpmpyss_nac_s0:
-  di_MInst_disisi_nac             <"mpy",      int_hexagon_M2_dpmpyss_nac_s0>;
-
 /********************************************************************
 *            MTYPE/VB                                               *
 *********************************************************************/
@@ -3096,15 +3184,9 @@ def HEXAGON_M2_vrmac_s0:
 // STYPE / ALU / Absolute value.
 def HEXAGON_A2_abs:
   si_SInst_si                     <"abs",     int_hexagon_A2_abs>;
-def HEXAGON_A2_absp:
-  di_SInst_di                     <"abs",     int_hexagon_A2_absp>;
 def HEXAGON_A2_abssat:
   si_SInst_si_sat                 <"abs",     int_hexagon_A2_abssat>;
 
-// STYPE / ALU / Logical Not.
-def HEXAGON_A2_notp:
-  di_SInst_di                     <"not",     int_hexagon_A2_notp>;
-
 // STYPE / ALU / Sign extend word to doubleword.
 def HEXAGON_A2_sxtw:
   di_SInst_si                     <"sxtw",     int_hexagon_A2_sxtw>;
diff --git a/llvm/lib/Target/Hexagon/HexagonIntrinsicsV4.td b/llvm/lib/Target/Hexagon/HexagonIntrinsicsV4.td
index 51ddb1b..894ead69 100644
--- a/llvm/lib/Target/Hexagon/HexagonIntrinsicsV4.td
+++ b/llvm/lib/Target/Hexagon/HexagonIntrinsicsV4.td
@@ -12,6 +12,20 @@
 // 80-V9418-12 Rev. A
 // June 15, 2010
 
+def : T_P_pat <S2_brevp, int_hexagon_S2_brevp>;
+
+def: T_P_pat  <S2_ct0p,      int_hexagon_S2_ct0p>;
+def: T_P_pat  <S2_ct1p,      int_hexagon_S2_ct1p>;
+def: T_RR_pat<C4_nbitsset,  int_hexagon_C4_nbitsset>;
+def: T_RR_pat<C4_nbitsclr,  int_hexagon_C4_nbitsclr>;
+def: T_RI_pat<C4_nbitsclri, int_hexagon_C4_nbitsclri>;
+
+// Extract bitfield
+def : T_PP_pat  <S4_extractp_rp, int_hexagon_S4_extractp_rp>;
+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>;
+
 // Shift an immediate left by register amount
 def : T_IR_pat<S4_lsli, int_hexagon_S4_lsli>;
 
@@ -25,6 +39,19 @@ def : T_IRI_pat <S4_ori_lsr_ri,  int_hexagon_S4_ori_lsr_ri>;
 def : T_IRI_pat <S4_addi_lsr_ri, int_hexagon_S4_addi_lsr_ri>;
 def : T_IRI_pat <S4_subi_lsr_ri, int_hexagon_S4_subi_lsr_ri>;
 
+// Split bitfield
+def : T_RI_pat <A4_bitspliti, int_hexagon_A4_bitspliti>;
+def : T_RR_pat <A4_bitsplit, int_hexagon_A4_bitsplit>;
+
+def: T_RR_pat<S4_parity,   int_hexagon_S4_parity>;
+
+def: T_RI_pat<S4_ntstbit_i,  int_hexagon_S4_ntstbit_i>;
+def: T_RR_pat<S4_ntstbit_r,  int_hexagon_S4_ntstbit_r>;
+
+def: T_RI_pat<S4_clbaddi,  int_hexagon_S4_clbaddi>;
+def: T_PI_pat<S4_clbpaddi, int_hexagon_S4_clbpaddi>;
+def: T_P_pat <S4_clbpnorm, int_hexagon_S4_clbpnorm>;
+
 //
 // ALU 32 types.
 //
diff --git a/llvm/test/CodeGen/Hexagon/intrinsics/xtype_bit.ll b/llvm/test/CodeGen/Hexagon/intrinsics/xtype_bit.ll
new file mode 100644
index 0000000..a7fc754
--- /dev/null
+++ b/llvm/test/CodeGen/Hexagon/intrinsics/xtype_bit.ll
@@ -0,0 +1,329 @@
+; RUN: llc -march=hexagon -O0 < %s | FileCheck %s
+; Hexagon Programmer's Reference Manual 11.10.2 XTYPE/BIT
+
+; Count leading
+declare i32 @llvm.hexagon.S2.clbp(i64)
+define i32 @S2_clbp(i64 %a) {
+  %z = call i32 @llvm.hexagon.S2.clbp(i64 %a)
+  ret i32 %z
+}
+; CHECK: r0 = clb(r1:0)
+
+declare i32 @llvm.hexagon.S2.cl0p(i64)
+define i32 @S2_cl0p(i64 %a) {
+  %z = call i32 @llvm.hexagon.S2.cl0p(i64 %a)
+  ret i32 %z
+}
+; CHECK: r0 = cl0(r1:0)
+
+declare i32 @llvm.hexagon.S2.cl1p(i64)
+define i32 @S2_cl1p(i64 %a) {
+  %z = call i32 @llvm.hexagon.S2.cl1p(i64 %a)
+  ret i32 %z
+}
+; CHECK: r0 = cl1(r1:0)
+
+declare i32 @llvm.hexagon.S4.clbpnorm(i64)
+define i32 @S4_clbpnorm(i64 %a) {
+  %z = call i32 @llvm.hexagon.S4.clbpnorm(i64 %a)
+  ret i32 %z
+}
+; CHECK: r0 = normamt(r1:0)
+
+declare i32 @llvm.hexagon.S4.clbpaddi(i64, i32)
+define i32 @S4_clbpaddi(i64 %a) {
+  %z = call i32 @llvm.hexagon.S4.clbpaddi(i64 %a, i32 0)
+  ret i32 %z
+}
+; CHECK: r0 = add(clb(r1:0), #0)
+
+declare i32 @llvm.hexagon.S4.clbaddi(i32, i32)
+define i32 @S4_clbaddi(i32 %a) {
+  %z = call i32 @llvm.hexagon.S4.clbaddi(i32 %a, i32 0)
+  ret i32 %z
+}
+; CHECK: r0 = add(clb(r0), #0)
+
+declare i32 @llvm.hexagon.S2.cl0(i32)
+define i32 @S2_cl0(i32 %a) {
+  %z = call i32 @llvm.hexagon.S2.cl0(i32 %a)
+  ret i32 %z
+}
+; CHECK: r0 = cl0(r0)
+
+declare i32 @llvm.hexagon.S2.cl1(i32)
+define i32 @S2_cl1(i32 %a) {
+  %z = call i32 @llvm.hexagon.S2.cl1(i32 %a)
+  ret i32 %z
+}
+; CHECK: r0 = cl1(r0)
+
+declare i32 @llvm.hexagon.S2.clbnorm(i32)
+define i32 @S4_clbnorm(i32 %a) {
+  %z = call i32 @llvm.hexagon.S2.clbnorm(i32 %a)
+  ret i32 %z
+}
+; CHECK: r0 = normamt(r0)
+
+; Count population
+declare i32 @llvm.hexagon.S5.popcountp(i64)
+define i32 @S5_popcountp(i64 %a) {
+  %z = call i32 @llvm.hexagon.S5.popcountp(i64 %a)
+  ret i32 %z
+}
+; CHECK: r0 = popcount(r1:0)
+
+; Count trailing
+declare i32 @llvm.hexagon.S2.ct0p(i64)
+define i32 @S2_ct0p(i64 %a) {
+  %z = call i32 @llvm.hexagon.S2.ct0p(i64 %a)
+  ret i32 %z
+}
+; CHECK: r0 = ct0(r1:0)
+
+declare i32 @llvm.hexagon.S2.ct1p(i64)
+define i32 @S2_ct1p(i64 %a) {
+  %z = call i32 @llvm.hexagon.S2.ct1p(i64 %a)
+  ret i32 %z
+}
+; CHECK: r0 = ct1(r1:0)
+
+declare i32 @llvm.hexagon.S2.ct0(i32)
+define i32 @S2_ct0(i32 %a) {
+  %z = call i32 @llvm.hexagon.S2.ct0(i32 %a)
+  ret i32 %z
+}
+; CHECK: r0 = ct0(r0)
+
+declare i32 @llvm.hexagon.S2.ct1(i32)
+define i32 @S2_ct1(i32 %a) {
+  %z = call i32 @llvm.hexagon.S2.ct1(i32 %a)
+  ret i32 %z
+}
+; CHECK: r0 = ct1(r0)
+
+; Extract bitfield
+declare i64 @llvm.hexagon.S2.extractup(i64, i32, i32)
+define i64 @S2_extractup(i64 %a) {
+  %z = call i64 @llvm.hexagon.S2.extractup(i64 %a, i32 0, i32 0)
+  ret i64 %z
+}
+; CHECK: r1:0 = extractu(r1:0, #0, #0)
+
+declare i64 @llvm.hexagon.S4.extractp(i64, i32, i32)
+define i64 @S2_extractp(i64 %a) {
+  %z = call i64 @llvm.hexagon.S4.extractp(i64 %a, i32 0, i32 0)
+  ret i64 %z
+}
+; CHECK: r1:0 = extract(r1:0, #0, #0)
+
+declare i32 @llvm.hexagon.S2.extractu(i32, i32, i32)
+define i32 @S2_extractu(i32 %a) {
+  %z = call i32 @llvm.hexagon.S2.extractu(i32 %a, i32 0, i32 0)
+  ret i32 %z
+}
+; CHECK: r0 = extractu(r0, #0, #0)
+
+declare i32 @llvm.hexagon.S4.extract(i32, i32, i32)
+define i32 @S2_extract(i32 %a) {
+  %z = call i32 @llvm.hexagon.S4.extract(i32 %a, i32 0, i32 0)
+  ret i32 %z
+}
+; CHECK: r0 = extract(r0, #0, #0)
+
+declare i64 @llvm.hexagon.S2.extractup.rp(i64, i64)
+define i64 @S2_extractup_rp(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.S2.extractup.rp(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = extractu(r1:0, r3:2)
+
+declare i64 @llvm.hexagon.S4.extractp.rp(i64, i64)
+define i64 @S4_extractp_rp(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.S4.extractp.rp(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = extract(r1:0, r3:2)
+
+declare i32 @llvm.hexagon.S2.extractu.rp(i32, i64)
+define i32 @S2_extractu_rp(i32 %a, i64 %b) {
+  %z = call i32 @llvm.hexagon.S2.extractu.rp(i32 %a, i64 %b)
+  ret i32 %z
+}
+; CHECK: r0 = extractu(r0, r3:2)
+
+declare i32 @llvm.hexagon.S4.extract.rp(i32, i64)
+define i32 @S4_extract_rp(i32 %a, i64 %b) {
+  %z = call i32 @llvm.hexagon.S4.extract.rp(i32 %a, i64 %b)
+  ret i32 %z
+}
+; CHECK: r0 = extract(r0, r3:2)
+
+; Insert bitfield
+declare i64 @llvm.hexagon.S2.insertp(i64, i64, i32, i32)
+define i64 @S2_insertp(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.S2.insertp(i64 %a, i64 %b, i32 0, i32 0)
+  ret i64 %z
+}
+; CHECK: r1:0 = insert(r3:2, #0, #0)
+
+declare i32 @llvm.hexagon.S2.insert(i32, i32, i32, i32)
+define i32 @S2_insert(i32 %a, i32 %b) {
+  %z = call i32 @llvm.hexagon.S2.insert(i32 %a, i32 %b, i32 0, i32 0)
+  ret i32 %z
+}
+; CHECK: r0 = insert(r1, #0, #0)
+
+declare i32 @llvm.hexagon.S2.insert.rp(i32, i32, i64)
+define i32 @S2_insert_rp(i32 %a, i32 %b, i64 %c) {
+  %z = call i32 @llvm.hexagon.S2.insert.rp(i32 %a, i32 %b, i64 %c)
+  ret i32 %z
+}
+; CHECK: r0 = insert(r1, r3:2)
+
+declare i64 @llvm.hexagon.S2.insertp.rp(i64, i64, i64)
+define i64 @S2_insertp_rp(i64 %a, i64 %b, i64 %c) {
+  %z = call i64 @llvm.hexagon.S2.insertp.rp(i64 %a, i64 %b, i64 %c)
+  ret i64 %z
+}
+; CHECK: r1:0 = insert(r3:2, r5:4)
+
+; Interleave/deinterleave
+declare i64 @llvm.hexagon.S2.deinterleave(i64)
+define i64 @S2_deinterleave(i64 %a) {
+  %z = call i64 @llvm.hexagon.S2.deinterleave(i64 %a)
+  ret i64 %z
+}
+; CHECK: r1:0 = deinterleave(r1:0)
+
+declare i64 @llvm.hexagon.S2.interleave(i64)
+define i64 @S2_interleave(i64 %a) {
+  %z = call i64 @llvm.hexagon.S2.interleave(i64 %a)
+  ret i64 %z
+}
+; CHECK: r1:0 = interleave(r1:0)
+
+; Linear feedback-shift operation
+declare i64 @llvm.hexagon.S2.lfsp(i64, i64)
+define i64 @S2_lfsp(i64 %a, i64 %b) {
+  %z = call i64 @llvm.hexagon.S2.lfsp(i64 %a, i64 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = lfs(r1:0, r3:2)
+
+; Masked parity
+declare i32 @llvm.hexagon.S2.parityp(i64, i64)
+define i32 @S2_parityp(i64 %a, i64 %b) {
+  %z = call i32 @llvm.hexagon.S2.parityp(i64 %a, i64 %b)
+  ret i32 %z
+}
+; CHECK: r0 = parity(r1:0, r3:2)
+
+declare i32 @llvm.hexagon.S4.parity(i32, i32)
+define i32 @S4_parity(i32 %a, i32 %b) {
+  %z = call i32 @llvm.hexagon.S4.parity(i32 %a, i32 %b)
+  ret i32 %z
+}
+; CHECK: r0 = parity(r0, r1)
+
+; Bit reverse
+declare i64 @llvm.hexagon.S2.brevp(i64)
+define i64 @S2_brevp(i64 %a) {
+  %z = call i64 @llvm.hexagon.S2.brevp(i64 %a)
+  ret i64 %z
+}
+; CHECK: r1:0 = brev(r1:0)
+
+declare i32 @llvm.hexagon.S2.brev(i32)
+define i32 @S2_brev(i32 %a) {
+  %z = call i32 @llvm.hexagon.S2.brev(i32 %a)
+  ret i32 %z
+}
+; CHECK: r0 = brev(r0)
+
+; Set/clear/toggle bit
+declare i32 @llvm.hexagon.S2.setbit.i(i32, i32)
+define i32 @S2_setbit_i(i32 %a) {
+  %z = call i32 @llvm.hexagon.S2.setbit.i(i32 %a, i32 0)
+  ret i32 %z
+}
+; CHECK: r0 = setbit(r0, #0)
+
+declare i32 @llvm.hexagon.S2.clrbit.i(i32, i32)
+define i32 @S2_clrbit_i(i32 %a) {
+  %z = call i32 @llvm.hexagon.S2.clrbit.i(i32 %a, i32 0)
+  ret i32 %z
+}
+; CHECK: r0 = clrbit(r0, #0)
+
+declare i32 @llvm.hexagon.S2.togglebit.i(i32, i32)
+define i32 @S2_togglebit_i(i32 %a) {
+  %z = call i32 @llvm.hexagon.S2.togglebit.i(i32 %a, i32 0)
+  ret i32 %z
+}
+; CHECK: r0 = togglebit(r0, #0)
+
+declare i32 @llvm.hexagon.S2.setbit.r(i32, i32)
+define i32 @S2_setbit_r(i32 %a, i32 %b) {
+  %z = call i32 @llvm.hexagon.S2.setbit.r(i32 %a, i32 %b)
+  ret i32 %z
+}
+; CHECK: r0 = setbit(r0, r1)
+
+declare i32 @llvm.hexagon.S2.clrbit.r(i32, i32)
+define i32 @S2_clrbit_r(i32 %a, i32 %b) {
+  %z = call i32 @llvm.hexagon.S2.clrbit.r(i32 %a, i32 %b)
+  ret i32 %z
+}
+; CHECK: r0 = clrbit(r0, r1)
+
+declare i32 @llvm.hexagon.S2.togglebit.r(i32, i32)
+define i32 @S2_togglebit_r(i32 %a, i32 %b) {
+  %z = call i32 @llvm.hexagon.S2.togglebit.r(i32 %a, i32 %b)
+  ret i32 %z
+}
+; CHECK: r0 = togglebit(r0, r1)
+
+; Split bitfield
+declare i64 @llvm.hexagon.A4.bitspliti(i32, i32)
+define i64 @A4_bitspliti(i32 %a) {
+  %z = call i64 @llvm.hexagon.A4.bitspliti(i32 %a, i32 0)
+  ret i64 %z
+}
+; CHECK:  = bitsplit(r0, #0)
+
+declare i64 @llvm.hexagon.A4.bitsplit(i32, i32)
+define i64 @A4_bitsplit(i32 %a, i32 %b) {
+  %z = call i64 @llvm.hexagon.A4.bitsplit(i32 %a, i32 %b)
+  ret i64 %z
+}
+; CHECK: r1:0 = bitsplit(r0, r1)
+
+; Table index
+declare i32 @llvm.hexagon.S2.tableidxb.goodsyntax(i32, i32, i32, i32)
+define i32 @S2_tableidxb_goodsyntax(i32 %a, i32 %b) {
+  %z = call i32 @llvm.hexagon.S2.tableidxb.goodsyntax(i32 %a, i32 %b, i32 0, i32 0)
+  ret i32 %z
+}
+; CHECK: r0 = tableidxb(r1, #0, #0)
+
+declare i32 @llvm.hexagon.S2.tableidxh.goodsyntax(i32, i32, i32, i32)
+define i32 @S2_tableidxh_goodsyntax(i32 %a, i32 %b) {
+  %z = call i32 @llvm.hexagon.S2.tableidxh.goodsyntax(i32 %a, i32 %b, i32 0, i32 0)
+  ret i32 %z
+}
+; CHECK: r0 = tableidxh(r1, #0, #0)
+
+declare i32 @llvm.hexagon.S2.tableidxw.goodsyntax(i32, i32, i32, i32)
+define i32 @S2_tableidxw_goodsyntax(i32 %a, i32 %b) {
+  %z = call i32 @llvm.hexagon.S2.tableidxw.goodsyntax(i32 %a, i32 %b, i32 0, i32 0)
+  ret i32 %z
+}
+; CHECK: r0 = tableidxw(r1, #0, #0)
+
+declare i32 @llvm.hexagon.S2.tableidxd.goodsyntax(i32, i32, i32, i32)
+define i32 @S2_tableidxd_goodsyntax(i32 %a, i32 %b) {
+  %z = call i32 @llvm.hexagon.S2.tableidxd.goodsyntax(i32 %a, i32 %b, i32 0, i32 0)
+  ret i32 %z
+}
+; CHECK: r0 = tableidxd(r1, #0, #0)
-- 
2.7.4