From 6507ccc30f29948a81661048e8a0ac3ae8e9a436 Mon Sep 17 00:00:00 2001
From: "Yueh-Ting (eop) Chen" <yueh.ting.chen@gmail.com>
Date: Wed, 1 Dec 2021 03:37:05 +0800
Subject: [PATCH] Simplify mulhsu (#870)

---
 riscv/decode.h           | 47 ++++++++++++++++++++++++++++++++++++++++
 riscv/insns/vmulhsu_vv.h | 38 ++------------------------------
 riscv/insns/vmulhsu_vx.h | 38 ++------------------------------
 3 files changed, 51 insertions(+), 72 deletions(-)

diff --git a/riscv/decode.h b/riscv/decode.h
index 2eccce2d..47df4511 100644
--- a/riscv/decode.h
+++ b/riscv/decode.h
@@ -738,6 +738,16 @@ static inline bool is_aligned(const unsigned val, const unsigned pos)
   type_sew_t<x>::type &vd = P.VU.elt<type_sew_t<x>::type>(rd_num, i, true); \
   type_usew_t<x>::type vs2 = P.VU.elt<type_usew_t<x>::type>(rs2_num, RS1);
 
+#define VV_SU_PARAMS(x) \
+  type_sew_t<x>::type &vd = P.VU.elt<type_sew_t<x>::type>(rd_num, i, true); \
+  type_usew_t<x>::type vs1 = P.VU.elt<type_usew_t<x>::type>(rs1_num, i); \
+  type_sew_t<x>::type vs2 = P.VU.elt<type_sew_t<x>::type>(rs2_num, i);
+
+#define VX_SU_PARAMS(x) \
+  type_sew_t<x>::type &vd = P.VU.elt<type_sew_t<x>::type>(rd_num, i, true); \
+  type_usew_t<x>::type rs1 = (type_usew_t<x>::type)RS1; \
+  type_sew_t<x>::type vs2 = P.VU.elt<type_sew_t<x>::type>(rs2_num, i);
+
 #define VV_UCMP_PARAMS(x) \
   type_usew_t<x>::type vs1 = P.VU.elt<type_usew_t<x>::type>(rs1_num, i); \
   type_usew_t<x>::type vs2 = P.VU.elt<type_usew_t<x>::type>(rs2_num, i);
@@ -1112,6 +1122,43 @@ static inline bool is_aligned(const unsigned val, const unsigned pos)
   } \
   VI_LOOP_END 
 
+// signed unsigned operation loop (e.g. mulhsu)
+#define VI_VV_SU_LOOP(BODY) \
+  VI_CHECK_SSS(true) \
+  VI_LOOP_BASE \
+  if (sew == e8){ \
+    VV_SU_PARAMS(e8); \
+    BODY; \
+  }else if(sew == e16){ \
+    VV_SU_PARAMS(e16); \
+    BODY; \
+  }else if(sew == e32){ \
+    VV_SU_PARAMS(e32); \
+    BODY; \
+  }else if(sew == e64){ \
+    VV_SU_PARAMS(e64); \
+    BODY; \
+  } \
+  VI_LOOP_END 
+
+#define VI_VX_SU_LOOP(BODY) \
+  VI_CHECK_SSS(false) \
+  VI_LOOP_BASE \
+  if (sew == e8){ \
+    VX_SU_PARAMS(e8); \
+    BODY; \
+  }else if(sew == e16){ \
+    VX_SU_PARAMS(e16); \
+    BODY; \
+  }else if(sew == e32){ \
+    VX_SU_PARAMS(e32); \
+    BODY; \
+  }else if(sew == e64){ \
+    VX_SU_PARAMS(e64); \
+    BODY; \
+  } \
+  VI_LOOP_END 
+
 // narrow operation loop
 #define VI_VV_LOOP_NARROW(BODY) \
 VI_NARROW_CHECK_COMMON; \
diff --git a/riscv/insns/vmulhsu_vv.h b/riscv/insns/vmulhsu_vv.h
index f77a7d3f..e1c0ba60 100644
--- a/riscv/insns/vmulhsu_vv.h
+++ b/riscv/insns/vmulhsu_vv.h
@@ -1,38 +1,4 @@
 // vmulhsu.vv vd, vs2, vs1
-VI_CHECK_SSS(true);
-VI_LOOP_BASE
-switch(sew) {
-case e8: {
-  auto &vd = P.VU.elt<int8_t>(rd_num, i, true);
-  auto vs2 = P.VU.elt<int8_t>(rs2_num, i);
-  auto vs1 = P.VU.elt<uint8_t>(rs1_num, i);
-
-  vd = ((int16_t)vs2 * (uint16_t)vs1) >> sew;
-  break;
-}
-case e16: {
-  auto &vd = P.VU.elt<int16_t>(rd_num, i, true);
-  auto vs2 = P.VU.elt<int16_t>(rs2_num, i);
-  auto vs1 = P.VU.elt<uint16_t>(rs1_num, i);
-
-  vd = ((int32_t)vs2 * (uint32_t)vs1) >> sew;
-  break;
-}
-case e32: {
-  auto &vd = P.VU.elt<int32_t>(rd_num, i, true);
-  auto vs2 = P.VU.elt<int32_t>(rs2_num, i);
-  auto vs1 = P.VU.elt<uint32_t>(rs1_num, i);
-
-  vd = ((int64_t)vs2 * (uint64_t)vs1) >> sew;
-  break;
-}
-default: {
-  auto &vd = P.VU.elt<int64_t>(rd_num, i, true);
-  auto vs2 = P.VU.elt<int64_t>(rs2_num, i);
-  auto vs1 = P.VU.elt<uint64_t>(rs1_num, i);
-
+VI_VV_SU_LOOP({
   vd = ((int128_t)vs2 * (uint128_t)vs1) >> sew;
-  break;
-}
-}
-VI_LOOP_END
+})
diff --git a/riscv/insns/vmulhsu_vx.h b/riscv/insns/vmulhsu_vx.h
index b0699f6f..4619ea89 100644
--- a/riscv/insns/vmulhsu_vx.h
+++ b/riscv/insns/vmulhsu_vx.h
@@ -1,38 +1,4 @@
 // vmulhsu.vx vd, vs2, rs1
-VI_CHECK_SSS(false);
-VI_LOOP_BASE
-switch(sew) {
-case e8: {
-  auto &vd = P.VU.elt<int8_t>(rd_num, i, true);
-  auto vs2 = P.VU.elt<int8_t>(rs2_num, i);
-  uint8_t rs1 = RS1;
-
-  vd = ((int16_t)vs2 * (uint16_t)rs1) >> sew;
-  break;
-}
-case e16: {
-  auto &vd = P.VU.elt<int16_t>(rd_num, i, true);
-  auto vs2 = P.VU.elt<int16_t>(rs2_num, i);
-  uint16_t rs1 = RS1;
-
-  vd = ((int32_t)vs2 * (uint32_t)rs1) >> sew;
-  break;
-}
-case e32: {
-  auto &vd = P.VU.elt<int32_t>(rd_num, i, true);
-  auto vs2 = P.VU.elt<int32_t>(rs2_num, i);
-  uint32_t rs1 = RS1;
-
-  vd = ((int64_t)vs2 * (uint64_t)rs1) >> sew;
-  break;
-}
-default: {
-  auto &vd = P.VU.elt<int64_t>(rd_num, i, true);
-  auto vs2 = P.VU.elt<int64_t>(rs2_num, i);
-  uint64_t rs1 = RS1;
-
+VI_VX_SU_LOOP({
   vd = ((int128_t)vs2 * (uint128_t)rs1) >> sew;
-  break;
-}
-}
-VI_LOOP_END
+})