[RV64_DYNAREC] Enabled native flags optimization for SETcc opcodes (#2640)

4 files changed, 87 insertions, 30 deletions

diff --git a/src/dynarec/rv64/dynarec_rv64_0f.c b/src/dynarec/rv64/dynarec_rv64_0f.c
index 5627707b..ae66e1b2 100644
--- a/src/dynarec/rv64/dynarec_rv64_0f.c
+++ b/src/dynarec/rv64/dynarec_rv64_0f.c
@@ -1772,35 +1772,37 @@ uintptr_t dynarec64_0F(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
             GOCOND(0x80, "J", "Id");
 #undef GO
 
-#define GO(GETFLAGS, NO, YES, NATNO, NATYES, F)                                              \
-    READFLAGS(F);                                                                            \
-    tmp1 = x1;                                                                               \
-    tmp3 = x3;                                                                               \
-    GETFLAGS;                                                                                \
-    nextop = F8;                                                                             \
-    S##YES(x3, x1);                                                                          \
-    if (MODREG) {                                                                            \
-        if (rex.rex) {                                                                       \
-            eb1 = TO_NAT((nextop & 7) + (rex.b << 3));                                       \
-            eb2 = 0;                                                                         \
-        } else {                                                                             \
-            ed = (nextop & 7);                                                               \
-            eb2 = (ed >> 2) * 8;                                                             \
-            eb1 = TO_NAT(ed & 3);                                                            \
-        }                                                                                    \
-        if (eb2) {                                                                           \
-            LUI(x1, 0xffff0);                                                                \
-            ORI(x1, x1, 0xff);                                                               \
-            AND(eb1, eb1, x1);                                                               \
-            SLLI(x3, x3, 8);                                                                 \
-        } else {                                                                             \
-            ANDI(eb1, eb1, 0xf00);                                                           \
-        }                                                                                    \
-        OR(eb1, eb1, x3);                                                                    \
-    } else {                                                                                 \
-        addr = geted(dyn, addr, ninst, nextop, &ed, x2, x1, &fixedaddress, rex, NULL, 1, 0); \
-        SB(x3, ed, fixedaddress);                                                            \
-        SMWRITE();                                                                           \
+#define GO(GETFLAGS, NO, YES, NATNO, NATYES, F)                                                  \
+    READFLAGS_FUSION(F, x1, x2, x3, x4, x5);                                                     \
+    if (!dyn->insts[ninst].nat_flags_fusion) { GETFLAGS; }                                       \
+    nextop = F8;                                                                                 \
+    if (dyn->insts[ninst].nat_flags_fusion) {                                                    \
+        NATIVESET(NATYES, tmp3);                                                                 \
+    } else {                                                                                     \
+        S##YES(tmp3, tmp1);                                                                      \
+    }                                                                                            \
+    if (MODREG) {                                                                                \
+        if (rex.rex) {                                                                           \
+            eb1 = TO_NAT((nextop & 7) + (rex.b << 3));                                           \
+            eb2 = 0;                                                                             \
+        } else {                                                                                 \
+            ed = (nextop & 7);                                                                   \
+            eb2 = (ed >> 2) * 8;                                                                 \
+            eb1 = TO_NAT(ed & 3);                                                                \
+        }                                                                                        \
+        if (eb2) {                                                                               \
+            LUI(tmp1, 0xffff0);                                                                  \
+            ORI(tmp1, tmp1, 0xff);                                                               \
+            AND(eb1, eb1, tmp1);                                                                 \
+            SLLI(tmp3, tmp3, 8);                                                                 \
+        } else {                                                                                 \
+            ANDI(eb1, eb1, 0xf00);                                                               \
+        }                                                                                        \
+        OR(eb1, eb1, tmp3);                                                                      \
+    } else {                                                                                     \
+        addr = geted(dyn, addr, ninst, nextop, &ed, tmp2, tmp1, &fixedaddress, rex, NULL, 1, 0); \
+        SB(tmp3, ed, fixedaddress);                                                              \
+        SMWRITE();                                                                               \
     }
             GOCOND(0x90, "SET", "Eb");
 #undef GO
diff --git a/src/dynarec/rv64/dynarec_rv64_helper.h b/src/dynarec/rv64/dynarec_rv64_helper.h
index 900245fb..5eb8ac69 100644
--- a/src/dynarec/rv64/dynarec_rv64_helper.h
+++ b/src/dynarec/rv64/dynarec_rv64_helper.h
@@ -1814,6 +1814,7 @@ uintptr_t dynarec64_AVX_F3_0F(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t ip,
 // Dummy macros
 #define B__safe(a, b, c) XOR(xZR, xZR, xZR)
 #define B_(a, b, c)      XOR(xZR, xZR, xZR)
+#define S_(a, b, c)      XOR(xZR, xZR, xZR)
 
 #define NATIVEJUMP_safe(COND, val) \
     B##COND##_safe(dyn->insts[ninst].nat_flags_op1, dyn->insts[ninst].nat_flags_op2, val);
@@ -1821,6 +1822,9 @@ uintptr_t dynarec64_AVX_F3_0F(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t ip,
 #define NATIVEJUMP(COND, val) \
     B##COND(dyn->insts[ninst].nat_flags_op1, dyn->insts[ninst].nat_flags_op2, val);
 
+#define NATIVESET(COND, rd) \
+    S##COND(rd, dyn->insts[ninst].nat_flags_op1, dyn->insts[ninst].nat_flags_op2);
+
 #define NOTEST(s1)                                     \
     if (BOX64ENV(dynarec_test)) {                      \
         SW(xZR, xEmu, offsetof(x64emu_t, test.test));  \
diff --git a/src/dynarec/rv64/dynarec_rv64_pass0.h b/src/dynarec/rv64/dynarec_rv64_pass0.h
index ed213560..416e8bb0 100644
--- a/src/dynarec/rv64/dynarec_rv64_pass0.h
+++ b/src/dynarec/rv64/dynarec_rv64_pass0.h
@@ -15,7 +15,7 @@
     dyn->f.pending = SF_SET
 
 #define READFLAGS_FUSION(A, s1, s2, s3, s4, s5)                                                                 \
-    if (BOX64ENV(dynarec_nativeflags) && ninst > 0 && !dyn->insts[ninst - 1].nat_flags_nofusion) {                  \
+    if (BOX64ENV(dynarec_nativeflags) && ninst > 0 && !dyn->insts[ninst - 1].nat_flags_nofusion) {              \
         if ((A) == (X_ZF))                                                                                      \
             dyn->insts[ninst].nat_flags_fusion = 1;                                                             \
         else if (dyn->insts[ninst - 1].nat_flags_carry && ((A) == (X_CF) || (A) == (X_CF | X_ZF)))              \
diff --git a/src/dynarec/rv64/rv64_emitter.h b/src/dynarec/rv64/rv64_emitter.h
index e274d9af..c1dc6fc7 100644
--- a/src/dynarec/rv64/rv64_emitter.h
+++ b/src/dynarec/rv64/rv64_emitter.h
@@ -200,6 +200,57 @@
 #define BGTU(rs1, rs2, imm13) BLTU(rs2, rs1, imm13)
 #define BLEU(rs1, rs2, imm13) BGEU(rs2, rs1, imm13)
 
+#define SEQ(rd, rs1, rs2)        \
+    do {                         \
+        if (rs1 == xZR) {        \
+            SEQZ(rd, rs2);       \
+        } else if (rs2 == xZR) { \
+            SEQZ(rd, rs1);       \
+        } else {                 \
+            XOR(rd, rs1, rs2);   \
+            SEQZ(rd, rd);        \
+        }                        \
+    } while (0)
+
+#define SNE(rd, rs1, rs2)        \
+    do {                         \
+        if (rs1 == xZR) {        \
+            SNEZ(rd, rs2);       \
+        } else if (rs2 == xZR) { \
+            SNEZ(rd, rs1);       \
+        } else {                 \
+            XOR(rd, rs1, rs2);   \
+            SNEZ(rd, rd);        \
+        }                        \
+    } while (0)
+
+#define SGE(rd, rs1, rs2)      \
+    do {                       \
+        if (rs1 == xZR) {      \
+            SLTI(rd, rs2, 1);  \
+        } else {               \
+            SLT(rd, rs1, rs2); \
+            XORI(rd, rd, 1);   \
+        }                      \
+    } while (0)
+
+#define SGEU(rd, rs1, rs2)       \
+    do {                         \
+        if (rs1 == xZR) {        \
+            SEQZ(rd, rs2);       \
+        } else if (rs2 == xZR) { \
+            ADDI(rd, xZR, 1);    \
+        } else {                 \
+            SLTU(rd, rs1, rs2);  \
+            XORI(rd, rd, 1);     \
+        }                        \
+    } while (0)
+
+#define SGT(rd, rs1, rs2)  SLT(rd, rs2, rs1);
+#define SLE(rd, rs1, rs2)  SGE(rd, rs2, rs1);
+#define SGTU(rd, rs1, rs2) SLTU(rd, rs2, rs1);
+#define SLEU(rd, rs1, rs2) SGEU(rd, rs2, rs1);
+
 #define BEQ_safe(rs1, rs2, imm)              \
     if ((imm) > -0x1000 && (imm) < 0x1000) { \
         BEQ(rs1, rs2, imm);                  \