From d233d8afa59675f6f4fc4cbf699e5da49772b53c Mon Sep 17 00:00:00 2001
From: Denis Drakhnya <numas13@gmail.com>
Date: Fri, 15 Jan 2021 22:49:11 +0200
Subject: [PATCH] e2k: Add packed add and sub using saturation.

---
 target/e2k/helper.h        | 12 ++++++++++++
 target/e2k/helper_vec.c    | 15 +++++++++++++++
 target/e2k/translate/alc.c | 16 ++++++++--------
 3 files changed, 35 insertions(+), 8 deletions(-)

diff --git a/target/e2k/helper.h b/target/e2k/helper.h
index 0e4f4baf30..56af5c390d 100644
--- a/target/e2k/helper.h
+++ b/target/e2k/helper.h
@@ -58,6 +58,18 @@ DEF_HELPER_2(pcmpgth, i64, i64, i64)
 DEF_HELPER_2(pcmpgtw, i64, i64, i64)
 DEF_HELPER_2(pcmpgtd, i64, i64, i64)
 
+/* Packed Add using saturation */
+DEF_HELPER_2(paddsb, i64, i64, i64)
+DEF_HELPER_2(paddsh, i64, i64, i64)
+DEF_HELPER_2(paddusb, i64, i64, i64)
+DEF_HELPER_2(paddush, i64, i64, i64)
+
+/* Packed Sub using saturation */
+DEF_HELPER_2(psubsb, i64, i64, i64)
+DEF_HELPER_2(psubsh, i64, i64, i64)
+DEF_HELPER_2(psubusb, i64, i64, i64)
+DEF_HELPER_2(psubush, i64, i64, i64)
+
 /* Float 32/64 Ops */
 #define DEF_HELPER_3_32_64(name) \
     DEF_HELPER_3(name##s, i32, env, i32, i32) \
diff --git a/target/e2k/helper_vec.c b/target/e2k/helper_vec.c
index edca9d44cf..7498bcca52 100644
--- a/target/e2k/helper_vec.c
+++ b/target/e2k/helper_vec.c
@@ -148,6 +148,21 @@ GEN_HELPER_PACKED_CMP(pcmpgth, sh, >)
 GEN_HELPER_PACKED_CMP(pcmpgtw, sw, >)
 GEN_HELPER_PACKED_CMP(pcmpgtd, sd, >)
 
+#define GEN_HELPER_PACKED_BINOP_SATURATE(name, type, op, t, min, max) \
+    GEN_HELPER_PACKED(name, type, { \
+        dst.type[i] = MIN(MAX(s1.type[i] op s2.type[i], min), max); \
+    })
+
+GEN_HELPER_PACKED_BINOP_SATURATE(paddsb, sb, +, int, -128, 127)
+GEN_HELPER_PACKED_BINOP_SATURATE(paddsh, sh, +, int, -32768, 32767)
+GEN_HELPER_PACKED_BINOP_SATURATE(paddusb, ub, +, int, 0, 255)
+GEN_HELPER_PACKED_BINOP_SATURATE(paddush, uh, +, int, 0, 65535)
+
+GEN_HELPER_PACKED_BINOP_SATURATE(psubsb, sb, -, int, -128, 127)
+GEN_HELPER_PACKED_BINOP_SATURATE(psubsh, sh, -, int, -32768, 32767)
+GEN_HELPER_PACKED_BINOP_SATURATE(psubusb, ub, -, int, 0, 255)
+GEN_HELPER_PACKED_BINOP_SATURATE(psubush, uh, -, int, 0, 65535)
+
 uint64_t HELPER(pmovmskb)(uint64_t src1, uint64_t src2)
 {
     unsigned int i;
diff --git a/target/e2k/translate/alc.c b/target/e2k/translate/alc.c
index 444af1e238..24cebea104 100644
--- a/target/e2k/translate/alc.c
+++ b/target/e2k/translate/alc.c
@@ -2958,10 +2958,18 @@ static void gen_op(DisasContext *ctx, Instr *instr)
     case OP_PADDH: gen_alopf1_ddd(instr, tcg_gen_vec_add16_i64); break;
     case OP_PADDW: gen_alopf1_ddd(instr, tcg_gen_vec_add32_i64); break;
     case OP_PADDD: gen_alopf1_ddd(instr, tcg_gen_add_i64); break;
+    case OP_PADDSB: gen_alopf1_ddd(instr, gen_helper_paddsb); break;
+    case OP_PADDSH: gen_alopf1_ddd(instr, gen_helper_paddsh); break;
+    case OP_PADDUSB: gen_alopf1_ddd(instr, gen_helper_paddusb); break;
+    case OP_PADDUSH: gen_alopf1_ddd(instr, gen_helper_paddush); break;
     case OP_PSUBB: gen_alopf1_ddd(instr, tcg_gen_vec_sub8_i64); break;
     case OP_PSUBH: gen_alopf1_ddd(instr, tcg_gen_vec_sub16_i64); break;
     case OP_PSUBW: gen_alopf1_ddd(instr, tcg_gen_vec_sub32_i64); break;
     case OP_PSUBD: gen_alopf1_ddd(instr, tcg_gen_sub_i64); break;
+    case OP_PSUBSB: gen_alopf1_ddd(instr, gen_helper_psubsb); break;
+    case OP_PSUBSH: gen_alopf1_ddd(instr, gen_helper_psubsh); break;
+    case OP_PSUBUSB: gen_alopf1_ddd(instr, gen_helper_psubusb); break;
+    case OP_PSUBUSH: gen_alopf1_ddd(instr, gen_helper_psubush); break;
     case OP_GETTAGS: gen_gettag_i32(instr); break;
     case OP_GETTAGD: gen_gettag_i64(instr); break;
     case OP_PUTTAGS: gen_puttag_i32(instr); break;
@@ -3116,14 +3124,6 @@ static void gen_op(DisasContext *ctx, Instr *instr)
     case OP_FSQRTTD:
     case OP_PFMULS:
     case OP_PFMULD:
-    case OP_PADDSB:
-    case OP_PADDSH:
-    case OP_PADDUSB:
-    case OP_PADDUSH:
-    case OP_PSUBSB:
-    case OP_PSUBSH:
-    case OP_PSUBUSB:
-    case OP_PSUBUSH:
     case OP_PSADBW:
     case OP_PMULHUH:
     case OP_PMULHH: