[RV64_DYNAREC] Minor fixes and improvements to CMP opcodes (#2353)

* [RV64_DYNAREC] Minor fixes and improvements to CMP opcodes

* review

6 files changed, 26 insertions, 21 deletions

diff --git a/src/dynarec/rv64/dynarec_rv64_00_0.c b/src/dynarec/rv64/dynarec_rv64_00_0.c
index 1a0290d6..4399c7c1 100644
--- a/src/dynarec/rv64/dynarec_rv64_00_0.c
+++ b/src/dynarec/rv64/dynarec_rv64_00_0.c
@@ -584,10 +584,10 @@ uintptr_t dynarec64_00_0(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t ip, int
             SETFLAGS(X_ALL, SF_SET_PENDING, NAT_FLAGS_FUSION);
             i64 = F32S;
             if (i64) {
-                MOV64xw(x2, i64);
+                MOV64x(x2, i64);
                 emit_cmp32(dyn, ninst, rex, xRAX, x2, x3, x4, x5, x6);
             } else
-                emit_cmp32_0(dyn, ninst, rex, xRAX, x3, x4);
+                emit_cmp32_0(dyn, ninst, rex, nextop, xRAX, x3, x4, x5);
             break;
 
         default:
diff --git a/src/dynarec/rv64/dynarec_rv64_00_2.c b/src/dynarec/rv64/dynarec_rv64_00_2.c
index 75d03ad7..e77956ff 100644
--- a/src/dynarec/rv64/dynarec_rv64_00_2.c
+++ b/src/dynarec/rv64/dynarec_rv64_00_2.c
@@ -253,15 +253,10 @@ uintptr_t dynarec64_00_2(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t ip, int
                     else
                         i64 = F8S;
                     if (i64) {
-                        MOV64xw(x2, i64);
+                        MOV64x(x2, i64);
                         emit_cmp32(dyn, ninst, rex, ed, x2, x3, x4, x5, x6);
-                    } else {
-                        if (!rex.w && MODREG) {
-                            ZEXTW2(x1, ed);
-                            ed = x1;
-                        }
-                        emit_cmp32_0(dyn, ninst, rex, ed, x3, x4);
-                    }
+                    } else
+                        emit_cmp32_0(dyn, ninst, rex, nextop, ed, x3, x4, x5);
                     break;
             }
             break;
diff --git a/src/dynarec/rv64/dynarec_rv64_64.c b/src/dynarec/rv64/dynarec_rv64_64.c
index 8d1e2d7f..6447c7f2 100644
--- a/src/dynarec/rv64/dynarec_rv64_64.c
+++ b/src/dynarec/rv64/dynarec_rv64_64.c
@@ -372,10 +372,10 @@ uintptr_t dynarec64_64(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
                     else
                         i64 = F8S;
                     if (i64) {
-                        MOV64xw(x2, i64);
+                        MOV64x(x2, i64);
                         emit_cmp32(dyn, ninst, rex, ed, x2, x3, x4, x5, x6);
                     } else
-                        emit_cmp32_0(dyn, ninst, rex, ed, x3, x4);
+                        emit_cmp32_0(dyn, ninst, rex, nextop, ed, x3, x4, x5);
                     break;
             }
             break;
diff --git a/src/dynarec/rv64/dynarec_rv64_67.c b/src/dynarec/rv64/dynarec_rv64_67.c
index f079ae49..a42cc19e 100644
--- a/src/dynarec/rv64/dynarec_rv64_67.c
+++ b/src/dynarec/rv64/dynarec_rv64_67.c
@@ -458,10 +458,10 @@ uintptr_t dynarec64_67(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
             SETFLAGS(X_ALL, SF_SET_PENDING, NAT_FLAGS_FUSION);
             i64 = F32S;
             if (i64) {
-                MOV64xw(x2, i64);
+                MOV64x(x2, i64);
                 emit_cmp32(dyn, ninst, rex, xRAX, x2, x3, x4, x5, x6);
             } else
-                emit_cmp32_0(dyn, ninst, rex, xRAX, x3, x4);
+                emit_cmp32_0(dyn, ninst, rex, nextop, xRAX, x3, x4, x5);
             break;
         case 0x63:
             INST_NAME("MOVSXD Gd, Ed");
@@ -638,10 +638,10 @@ uintptr_t dynarec64_67(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
                     else
                         i64 = F8S;
                     if (i64) {
-                        MOV64xw(x2, i64);
+                        MOV64x(x2, i64);
                         emit_cmp32(dyn, ninst, rex, ed, x2, x3, x4, x5, x6);
                     } else
-                        emit_cmp32_0(dyn, ninst, rex, ed, x3, x4);
+                        emit_cmp32_0(dyn, ninst, rex, nextop, ed, x3, x4, x5);
                     break;
             }
             break;
diff --git a/src/dynarec/rv64/dynarec_rv64_emit_tests.c b/src/dynarec/rv64/dynarec_rv64_emit_tests.c
index 1fa0b8ab..bc9e79ce 100644
--- a/src/dynarec/rv64/dynarec_rv64_emit_tests.c
+++ b/src/dynarec/rv64/dynarec_rv64_emit_tests.c
@@ -255,7 +255,7 @@ void emit_cmp32(dynarec_rv64_t* dyn, int ninst, rex_t rex, int s1, int s2, int s
 }
 
 // emit CMP32 instruction, from cmp s1, 0, using s3 and s4 as scratch
-void emit_cmp32_0(dynarec_rv64_t* dyn, int ninst, rex_t rex, int s1, int s3, int s4)
+void emit_cmp32_0(dynarec_rv64_t* dyn, int ninst, rex_t rex, uint8_t nextop, int s1, int s3, int s4, int s5)
 {
     CLEAR_FLAGS();
     IFX_PENDOR0 {
@@ -275,11 +275,18 @@ void emit_cmp32_0(dynarec_rv64_t* dyn, int ninst, rex_t rex, int s1, int s3, int
             SET_FLAGS_NEZ(s3, F_SF, s4);
         }
     }
+    int res = s1;
+    IFX (X_ZF | X_PF) {
+        if (!rex.w && MODREG) {
+            ZEXTW2(s5, s1);
+            res = s5;
+        }
+    }
     IFX (X_ZF) {
-        SET_FLAGS_EQZ(s1, F_ZF, s3);
+        SET_FLAGS_EQZ(res, F_ZF, s3);
     }
     IFX (X_PF) {
-        emit_pf(dyn, ninst, s1, s3, s4);
+        emit_pf(dyn, ninst, res, s3, s4);
     }
     NAT_FLAGS_ENABLE_CARRY();
     NAT_FLAGS_ENABLE_SIGN();
@@ -289,10 +296,13 @@ void emit_cmp32_0(dynarec_rv64_t* dyn, int ninst, rex_t rex, int s1, int s3, int
         else {
             if (dyn->insts[ninst].nat_flags_needsign) {
                 SEXT_W(s3, s1);
+                NAT_FLAGS_OPS(s3, xZR);
+            } else if (res == s5) { // zero-up'd case
+                NAT_FLAGS_OPS(s5, xZR);
             } else {
                 ZEXTW2(s3, s1);
+                NAT_FLAGS_OPS(s3, xZR);
             }
-            NAT_FLAGS_OPS(s3, xZR);
         }
     }
 }
diff --git a/src/dynarec/rv64/dynarec_rv64_helper.h b/src/dynarec/rv64/dynarec_rv64_helper.h
index ca1e40d0..d1ecd7f7 100644
--- a/src/dynarec/rv64/dynarec_rv64_helper.h
+++ b/src/dynarec/rv64/dynarec_rv64_helper.h
@@ -1410,7 +1410,7 @@ void emit_cmp16(dynarec_rv64_t* dyn, int ninst, int s1, int s2, int s3, int s4,
 void emit_cmp32(dynarec_rv64_t* dyn, int ninst, rex_t rex, int s1, int s2, int s3, int s4, int s5, int s6);
 void emit_cmp8_0(dynarec_rv64_t* dyn, int ninst, int s1, int s3, int s4);
 void emit_cmp16_0(dynarec_rv64_t* dyn, int ninst, int s1, int s3, int s4);
-void emit_cmp32_0(dynarec_rv64_t* dyn, int ninst, rex_t rex, int s1, int s3, int s4);
+void emit_cmp32_0(dynarec_rv64_t* dyn, int ninst, rex_t rex, uint8_t nextop, int s1, int s3, int s4, int s5);
 void emit_test8(dynarec_rv64_t* dyn, int ninst, int s1, int s2, int s3, int s4, int s5);
 void emit_test8c(dynarec_rv64_t* dyn, int ninst, int s1, uint8_t c, int s3, int s4, int s5);
 void emit_test16(dynarec_rv64_t* dyn, int ninst, int s1, int s2, int s3, int s4, int s5);