rtl: remove unnecessary basic rtl instructions

include/rtl/c_op.h

@@ -8,20 +8,21 @@
 #define c_xor(a, b) ((a) ^ (b))
 #define c_shl(a, b) ((a) << (b))
 #define c_shr(a, b) ((a) >> (b))
-#define c_shr64(a, b) ((uint32_t)(a) >> (b))
-#define c_sar(a, b) ((int32_t)(a) >> (b))
-#define c_sar64(a, b)  ((int64_t)(a) >> (b))
+#define c_sar(a, b) ((sword_t)(a) >> (b))
+
 #define c_mul_lo(a, b) ((a) * (b))
-#define c_mul_hi(a, b) (((uint64_t)(a) * (uint64_t)(b)) >> 32)
-#define c_imul_lo(a, b) ((int32_t)(a) * (int32_t)(b))
-#define c_imul_hi(a, b) (((int64_t)(int32_t)(a) * (int64_t)(int32_t)(b)) >> 32)
+#define c_imul_lo(a, b) ((sword_t)(a) * (sword_t)(b))
+#ifdef ISA64
+# define c_mul_hi(a, b) (((__uint128_t)(a) * (__uint128_t)(b)) >> 64)
+# define c_imul_hi(a, b) (((__int128_t)(sword_t)(a) * (__int128_t)(sword_t)(b)) >> 64)
+#else
+# define c_mul_hi(a, b) (((uint64_t)(a) * (uint64_t)(b)) >> 32)
+# define c_imul_hi(a, b) (((int64_t)(sword_t)(a) * (int64_t)(sword_t)(b)) >> 32)
+#endif
+
 #define c_div_q(a, b) ((a) / (b))
-#define c_div_q64(a, b) (((a) & 0xffffffff) / ((b) & 0xffffffff))
 #define c_div_r(a, b)  ((a) % (b))
-#define c_div_r64(a, b) (((a) & 0xffffffff) % ((b) & 0xffffffff))
-#define c_idiv_q(a, b) ((int32_t)(a) / (int32_t)(b))
-#define c_idiv_q64(a, b)  ((int32_t)((uint32_t)((a) & 0xffffffff)) / (int32_t)((uint32_t)((b) & 0xffffffff)))
-#define c_idiv_r(a, b)  ((int32_t)(a) % (int32_t)(b))
-#define c_idiv_r64(a, b)  ((int32_t)((uint32_t)((a) & 0xffffffff)) % (int32_t)((uint32_t)((b) & 0xffffffff)))
+#define c_idiv_q(a, b) ((sword_t)(a) / (sword_t)(b))
+#define c_idiv_r(a, b)  ((sword_t)(a) % (sword_t)(b))
 
 #endif

include/rtl/rtl-wrapper.h

@@ -14,21 +14,15 @@
 #define rtl_xor       concat(RTL_PREFIX, _rtl_xor     )
 #define rtl_shl       concat(RTL_PREFIX, _rtl_shl     )
 #define rtl_shr       concat(RTL_PREFIX, _rtl_shr     )
-#define rtl_shr64     concat(RTL_PREFIX, _rtl_shr64   )
 #define rtl_sar       concat(RTL_PREFIX, _rtl_sar     )
-#define rtl_sar64     concat(RTL_PREFIX, _rtl_sar64   )
 #define rtl_mul_lo    concat(RTL_PREFIX, _rtl_mul_lo  )
 #define rtl_mul_hi    concat(RTL_PREFIX, _rtl_mul_hi  )
 #define rtl_imul_lo   concat(RTL_PREFIX, _rtl_imul_lo )
 #define rtl_imul_hi   concat(RTL_PREFIX, _rtl_imul_hi )
 #define rtl_div_q     concat(RTL_PREFIX, _rtl_div_q   )
-#define rtl_div_q64   concat(RTL_PREFIX, _rtl_div_q64 )
 #define rtl_div_r     concat(RTL_PREFIX, _rtl_div_r   )
-#define rtl_div_r64   concat(RTL_PREFIX, _rtl_div_r64 )
 #define rtl_idiv_q    concat(RTL_PREFIX, _rtl_idiv_q  )
-#define rtl_idiv_q64  concat(RTL_PREFIX, _rtl_idiv_q64)
 #define rtl_idiv_r    concat(RTL_PREFIX, _rtl_idiv_r  )
-#define rtl_idiv_r64  concat(RTL_PREFIX, _rtl_idiv_r64)
 #define rtl_div64_q   concat(RTL_PREFIX, _rtl_div64_q )
 #define rtl_div64_r   concat(RTL_PREFIX, _rtl_div64_r )
 #define rtl_idiv64_q  concat(RTL_PREFIX, _rtl_idiv64_q)

include/rtl/rtl.h

@@ -49,15 +49,6 @@ make_rtl_arith_logic(div_r)
 make_rtl_arith_logic(idiv_q)
 make_rtl_arith_logic(idiv_r)
 
-#ifdef ISA64
-make_rtl_arith_logic(shr64)
-make_rtl_arith_logic(sar64)
-make_rtl_arith_logic(div_q64)
-make_rtl_arith_logic(div_r64)
-make_rtl_arith_logic(idiv_q64)
-make_rtl_arith_logic(idiv_r64)
-#endif
-
 static inline void interpret_rtl_div64_q(rtlreg_t* dest,
     const rtlreg_t* src1_hi, const rtlreg_t* src1_lo, const rtlreg_t* src2) {
   uint64_t dividend = ((uint64_t)(*src1_hi) << 32) | (*src1_lo);
@@ -155,7 +146,7 @@ static inline void rtl_sext(rtlreg_t* dest, const rtlreg_t* src1, int width) {
   } else {
     assert(width == 1 || width == 2 || width == 4);
     rtl_shli(dest, src1, (8 - width) * 8);
-    rtl_sar64i(dest, dest, (8 - width) * 8);
+    rtl_sari(dest, dest, (8 - width) * 8);
   }
 #else
   if (width == 4) {

src/isa/riscv64/exec/compute.c

@@ -117,6 +117,7 @@ make_EHelper(sllw) {
 }
 
 make_EHelper(srlw) {
+  rtl_andi(&id_src2->val, &id_src2->val, 0x1f);
   if (decinfo.isa.instr.funct7 == 32) {
     // sraw
     rtl_sar(&s0, &id_src->val, &id_src2->val);
@@ -125,9 +126,8 @@ make_EHelper(srlw) {
   }
   else {
     // srlw
-    rtl_andi(&id_src2->val, &id_src2->val, 0x1f);
-
-    rtl_shr64(&s0, &id_src->val, &id_src2->val);
+    rtl_andi(&id_src->val, &id_src->val, 0xffffffffu);
+    rtl_shr(&s0, &id_src->val, &id_src2->val);
     rtl_sext(&s0, &s0, 4);
     print_asm_template3(srlw);
   }

src/isa/riscv64/exec/muldiv.c

 #include "cpu/exec.h"
 
 make_EHelper(mul) {
-#ifdef ISA64
   rtl_mul_lo(&s0, &id_src->val, &id_src2->val);
-#else
-  rtl_imul_lo(&s0, &id_src->val, &id_src2->val);
-#endif
   rtl_sr(id_dest->reg, &s0, 4);
 
   print_asm_template3(mul);
@@ -54,7 +50,7 @@ make_EHelper(remu) {
 }
 
 make_EHelper(mulw) {
-  rtl_imul_lo(&s0, &id_src->val, &id_src2->val);
+  rtl_mul_lo(&s0, &id_src->val, &id_src2->val);
   rtl_sext(&s0, &s0, 4);
   rtl_sr(id_dest->reg, &s0, 4);
 
@@ -62,15 +58,19 @@ make_EHelper(mulw) {
 }
 
 make_EHelper(divw) {
-  rtl_idiv_q64(&s0, &id_src->val, &id_src2->val);
+  rtl_sext(&s0, &id_src->val, 4);
+  rtl_sext(&s1, &id_src2->val, 4);
+  rtl_idiv_q(&s0, &s0, &s1);
   rtl_sext(&s0, &s0, 4);
   rtl_sr(id_dest->reg, &s0, 4);
 
-  print_asm_template3(div);
+  print_asm_template3(divw);
 }
 
 make_EHelper(remw) {
-  rtl_idiv_r64(&s0, &id_src->val, &id_src2->val);
+  rtl_sext(&s0, &id_src->val, 4);
+  rtl_sext(&s1, &id_src2->val, 4);
+  rtl_idiv_r(&s0, &s0, &s1);
   rtl_sext(&s0, &s0, 4);
   rtl_sr(id_dest->reg, &s0, 4);
 
@@ -78,7 +78,9 @@ make_EHelper(remw) {
 }
 
 make_EHelper(divuw) {
-  rtl_div_q64(&s0, &id_src->val, &id_src2->val);
+  rtl_andi(&s0, &id_src->val, 0xffffffffu);
+  rtl_andi(&s1, &id_src2->val, 0xffffffffu);
+  rtl_div_q(&s0, &s0, &s1);
   rtl_sext(&s0, &s0, 4);
   rtl_sr(id_dest->reg, &s0, 4);
 
@@ -86,9 +88,11 @@ make_EHelper(divuw) {
 }
 
 make_EHelper(remuw) {
-  rtl_idiv_r64(&s0, &id_src->val, &id_src2->val);
+  rtl_andi(&s0, &id_src->val, 0xffffffffu);
+  rtl_andi(&s1, &id_src2->val, 0xffffffffu);
+  rtl_div_r(&s0, &s0, &s1);
   rtl_sext(&s0, &s0, 4);
   rtl_sr(id_dest->reg, &s0, 4);
 
   print_asm_template3(remuw);
-}
\ No newline at end of file
+}