Merge pull request #928 from wannacu/main

[ARM64_DYNAREC] Correct PF caculation

8 files changed, 17 insertions, 18 deletions

diff --git a/src/dynarec/arm64/dynarec_arm64_00.c b/src/dynarec/arm64/dynarec_arm64_00.c
index 8f2f6882..c760d9d4 100644
--- a/src/dynarec/arm64/dynarec_arm64_00.c
+++ b/src/dynarec/arm64/dynarec_arm64_00.c
@@ -1008,7 +1008,7 @@ uintptr_t dynarec64_00(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
             nextop=F8;
             GETGD;
             GETED(0);
-            emit_test32(dyn, ninst, rex, ed, gd, x3, x5);
+            emit_test32(dyn, ninst, rex, ed, gd, x3, x5, x6);
             break;
         case 0x86:
             INST_NAME("(LOCK)XCHG Eb, Gb");
@@ -1477,7 +1477,7 @@ uintptr_t dynarec64_00(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
             SETFLAGS(X_ALL, SF_SET_PENDING);
             i64 = F32S;
             MOV64xw(x2, i64);
-            emit_test32(dyn, ninst, rex, xRAX, x2, x3, x4);
+            emit_test32(dyn, ninst, rex, xRAX, x2, x3, x4, x5);
             break;
         case 0xAA:
             if(rep) {
@@ -2608,7 +2608,7 @@ uintptr_t dynarec64_00(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                     GETEDH(x1, 4);
                     i64 = F32S;
                     MOV64xw(x2, i64);
-                    emit_test32(dyn, ninst, rex, ed, x2, x3, x4);
+                    emit_test32(dyn, ninst, rex, ed, x2, x3, x4, x5);
                     break;
                 case 2:
                     INST_NAME("NOT Ed");
diff --git a/src/dynarec/arm64/dynarec_arm64_64.c b/src/dynarec/arm64/dynarec_arm64_64.c
index 540ab50e..d88f3b69 100644
--- a/src/dynarec/arm64/dynarec_arm64_64.c
+++ b/src/dynarec/arm64/dynarec_arm64_64.c
@@ -884,7 +884,7 @@ uintptr_t dynarec64_64(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                     GETEDO(x6, 4);
                     i64 = F32S;
                     MOV64xw(x2, i64);
-                    emit_test32(dyn, ninst, rex, ed, x2, x3, x4);
+                    emit_test32(dyn, ninst, rex, ed, x2, x3, x4, x5);
                     break;
                 case 2:
                     INST_NAME("NOT Ed");
diff --git a/src/dynarec/arm64/dynarec_arm64_67.c b/src/dynarec/arm64/dynarec_arm64_67.c
index 04065e7a..06e1e93a 100644
--- a/src/dynarec/arm64/dynarec_arm64_67.c
+++ b/src/dynarec/arm64/dynarec_arm64_67.c
@@ -990,7 +990,7 @@ uintptr_t dynarec64_67(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                     GETED32H(x1, 4);

                     i64 = F32S;

                     MOV64xw(x2, i64);

-                    emit_test32(dyn, ninst, rex, ed, x2, x3, x4);

+                    emit_test32(dyn, ninst, rex, ed, x2, x3, x4, x5);

                     break;

                 case 2:

                     INST_NAME("NOT Ed");

diff --git a/src/dynarec/arm64/dynarec_arm64_emit_tests.c b/src/dynarec/arm64/dynarec_arm64_emit_tests.c
index 447ce8c7..bf190586 100644
--- a/src/dynarec/arm64/dynarec_arm64_emit_tests.c
+++ b/src/dynarec/arm64/dynarec_arm64_emit_tests.c
@@ -265,8 +265,8 @@ void emit_cmp8_0(dynarec_arm_t* dyn, int ninst, int s1, int s3, int s4)
     }
 }
 
-// emit TEST32 instruction, from test s1, s2, using s3 and s4 as scratch
-void emit_test32(dynarec_arm_t* dyn, int ninst, rex_t rex, int s1, int s2, int s3, int s4)
+// emit TEST32 instruction, from test s1, s2, using s3, s4 and s5 as scratch
+void emit_test32(dynarec_arm_t* dyn, int ninst, rex_t rex, int s1, int s2, int s3, int s4, int s5)
 {
     MAYUSE(s1); MAYUSE(s2); MAYUSE(s3); MAYUSE(s4);
     IFX_PENDOR0 {
@@ -292,18 +292,11 @@ void emit_test32(dynarec_arm_t* dyn, int ninst, rex_t rex, int s1, int s2, int s
     }
     // PF: (((emu->x64emu_parity_tab[(res) / 32] >> ((res) % 32)) & 1) == 0)
     IFX(X_PF) {
-        ANDw_mask(s3, s3, 0b011011, 0b000010); // 0xE0
-        LSRw(s3, s3, 5);
-        MOV64x(s4, (uintptr_t)GetParityTab());
-        LDRw_REG_LSL2(s4, s4, s3);
-        ANDw_mask(s3, s1, 0, 0b000100);   // 0x1f
-        LSRw_REG(s4, s4, s3);
-        MVNx_REG(s4, s4);
-        BFIw(xFlags, s4, F_PF, 1);
+        emit_pf(dyn, ninst, s3, s4, s5);
     }
 }
 
-// emit TEST16 instruction, from test s1, s2, using s3 and s4 as scratch
+// emit TEST16 instruction, from test s1, s2, using s3, s4 and s5 as scratch
 void emit_test16(dynarec_arm_t* dyn, int ninst, int s1, int s2, int s3, int s4, int s5)
 {
     MAYUSE(s1); MAYUSE(s2);
@@ -334,7 +327,7 @@ void emit_test16(dynarec_arm_t* dyn, int ninst, int s1, int s2, int s3, int s4,
     }
 }
 
-// emit TEST8 instruction, from test s1, s2, using s3 and s4 as scratch
+// emit TEST8 instruction, from test s1, s2, using s3, s4 and s5 as scratch
 void emit_test8(dynarec_arm_t* dyn, int ninst, int s1, int s2, int s3, int s4, int s5)
 {
     MAYUSE(s1); MAYUSE(s2);
diff --git a/src/dynarec/arm64/dynarec_arm64_helper.h b/src/dynarec/arm64/dynarec_arm64_helper.h
index b6021585..5c7c21f3 100644
--- a/src/dynarec/arm64/dynarec_arm64_helper.h
+++ b/src/dynarec/arm64/dynarec_arm64_helper.h
@@ -1035,7 +1035,7 @@ void emit_cmp16_0(dynarec_arm_t* dyn, int ninst, int s1, int s3, int s4);
 void emit_cmp32_0(dynarec_arm_t* dyn, int ninst, rex_t rex, int s1, int s3, int s4);
 void emit_test8(dynarec_arm_t* dyn, int ninst, int s1, int s2, int s3, int s4, int s5);
 void emit_test16(dynarec_arm_t* dyn, int ninst, int s1, int s2, int s3, int s4, int s5);
-void emit_test32(dynarec_arm_t* dyn, int ninst, rex_t rex, int s1, int s2, int s3, int s4);
+void emit_test32(dynarec_arm_t* dyn, int ninst, rex_t rex, int s1, int s2, int s3, int s4, int s5);
 void emit_add32(dynarec_arm_t* dyn, int ninst, rex_t rex, int s1, int s2, int s3, int s4);
 void emit_add32c(dynarec_arm_t* dyn, int ninst, rex_t rex, int s1, int64_t c, int s3, int s4, int s5);
 void emit_add8(dynarec_arm_t* dyn, int ninst, int s1, int s2, int s3, int s4);
diff --git a/src/wrapped/generated/functions_list.txt b/src/wrapped/generated/functions_list.txt
index fc2c6348..db5e8e08 100644
--- a/src/wrapped/generated/functions_list.txt
+++ b/src/wrapped/generated/functions_list.txt
@@ -2175,6 +2175,7 @@
 #() vFppCuupp
 #() vFppuiiip
 #() vFppuupii
+#() vFppUuupp
 #() vFppddddu
 #() vFppddpiu
 #() vFpplpppi
diff --git a/src/wrapped/generated/wrapper.c b/src/wrapped/generated/wrapper.c
index cb85dc83..6ed5bbc8 100644
--- a/src/wrapped/generated/wrapper.c
+++ b/src/wrapped/generated/wrapper.c
@@ -2211,6 +2211,7 @@ typedef void (*vFppipppp_t)(void*, void*, int32_t, void*, void*, void*, void*);
 typedef void (*vFppCuupp_t)(void*, void*, uint8_t, uint32_t, uint32_t, void*, void*);
 typedef void (*vFppuiiip_t)(void*, void*, uint32_t, int32_t, int32_t, int32_t, void*);
 typedef void (*vFppuupii_t)(void*, void*, uint32_t, uint32_t, void*, int32_t, int32_t);
+typedef void (*vFppUuupp_t)(void*, void*, uint64_t, uint32_t, uint32_t, void*, void*);
 typedef void (*vFppddddu_t)(void*, void*, double, double, double, double, uint32_t);
 typedef void (*vFppddpiu_t)(void*, void*, double, double, void*, int32_t, uint32_t);
 typedef void (*vFpplpppi_t)(void*, void*, intptr_t, void*, void*, void*, int32_t);
@@ -3788,6 +3789,7 @@ void vFppip(x64emu_t *emu, uintptr_t fcn) { vFppip_t fn = (vFppip_t)fcn; fn((voi
 void vFppui(x64emu_t *emu, uintptr_t fcn) { vFppui_t fn = (vFppui_t)fcn; fn((void*)R_RDI, (void*)R_RSI, (uint32_t)R_RDX, (int32_t)R_RCX); }
 void vFppuu(x64emu_t *emu, uintptr_t fcn) { vFppuu_t fn = (vFppuu_t)fcn; fn((void*)R_RDI, (void*)R_RSI, (uint32_t)R_RDX, (uint32_t)R_RCX); }
 void vFppup(x64emu_t *emu, uintptr_t fcn) { vFppup_t fn = (vFppup_t)fcn; fn((void*)R_RDI, (void*)R_RSI, (uint32_t)R_RDX, (void*)R_RCX); }
+void vFppUu(x64emu_t *emu, uintptr_t fcn) { vFppUu_t fn = (vFppUu_t)fcn; fn((void*)R_RDI, (void*)R_RSI, (uint64_t)R_RDX, (uint32_t)R_RCX); }
 void vFppfi(x64emu_t *emu, uintptr_t fcn) { vFppfi_t fn = (vFppfi_t)fcn; fn((void*)R_RDI, (void*)R_RSI, emu->xmm[0].f[0], (int32_t)R_RDX); }
 void vFppff(x64emu_t *emu, uintptr_t fcn) { vFppff_t fn = (vFppff_t)fcn; fn((void*)R_RDI, (void*)R_RSI, emu->xmm[0].f[0], emu->xmm[1].f[0]); }
 void vFppdu(x64emu_t *emu, uintptr_t fcn) { vFppdu_t fn = (vFppdu_t)fcn; fn((void*)R_RDI, (void*)R_RSI, emu->xmm[0].d[0], (uint32_t)R_RDX); }
@@ -5108,6 +5110,7 @@ void vFppipppp(x64emu_t *emu, uintptr_t fcn) { vFppipppp_t fn = (vFppipppp_t)fcn
 void vFppCuupp(x64emu_t *emu, uintptr_t fcn) { vFppCuupp_t fn = (vFppCuupp_t)fcn; fn((void*)R_RDI, (void*)R_RSI, (uint8_t)R_RDX, (uint32_t)R_RCX, (uint32_t)R_R8, (void*)R_R9, *(void**)(R_RSP + 8)); }
 void vFppuiiip(x64emu_t *emu, uintptr_t fcn) { vFppuiiip_t fn = (vFppuiiip_t)fcn; fn((void*)R_RDI, (void*)R_RSI, (uint32_t)R_RDX, (int32_t)R_RCX, (int32_t)R_R8, (int32_t)R_R9, *(void**)(R_RSP + 8)); }
 void vFppuupii(x64emu_t *emu, uintptr_t fcn) { vFppuupii_t fn = (vFppuupii_t)fcn; fn((void*)R_RDI, (void*)R_RSI, (uint32_t)R_RDX, (uint32_t)R_RCX, (void*)R_R8, (int32_t)R_R9, *(int32_t*)(R_RSP + 8)); }
+void vFppUuupp(x64emu_t *emu, uintptr_t fcn) { vFppUuupp_t fn = (vFppUuupp_t)fcn; fn((void*)R_RDI, (void*)R_RSI, (uint64_t)R_RDX, (uint32_t)R_RCX, (uint32_t)R_R8, (void*)R_R9, *(void**)(R_RSP + 8)); }
 void vFppddddu(x64emu_t *emu, uintptr_t fcn) { vFppddddu_t fn = (vFppddddu_t)fcn; fn((void*)R_RDI, (void*)R_RSI, emu->xmm[0].d[0], emu->xmm[1].d[0], emu->xmm[2].d[0], emu->xmm[3].d[0], (uint32_t)R_RDX); }
 void vFppddpiu(x64emu_t *emu, uintptr_t fcn) { vFppddpiu_t fn = (vFppddpiu_t)fcn; fn((void*)R_RDI, (void*)R_RSI, emu->xmm[0].d[0], emu->xmm[1].d[0], (void*)R_RDX, (int32_t)R_RCX, (uint32_t)R_R8); }
 void vFpplpppi(x64emu_t *emu, uintptr_t fcn) { vFpplpppi_t fn = (vFpplpppi_t)fcn; fn((void*)R_RDI, (void*)R_RSI, (intptr_t)R_RDX, (void*)R_RCX, (void*)R_R8, (void*)R_R9, *(int32_t*)(R_RSP + 8)); }
@@ -6569,6 +6572,7 @@ int isSimpleWrapper(wrapper_t fun) {
 	if (fun == &vFppui) return 1;
 	if (fun == &vFppuu) return 1;
 	if (fun == &vFppup) return 1;
+	if (fun == &vFppUu) return 1;
 	if (fun == &vFppfi) return 2;
 	if (fun == &vFppff) return 3;
 	if (fun == &vFppdu) return 2;
diff --git a/src/wrapped/generated/wrapper.h b/src/wrapped/generated/wrapper.h
index d8a1fa40..67b07f74 100644
--- a/src/wrapped/generated/wrapper.h
+++ b/src/wrapped/generated/wrapper.h
@@ -2212,6 +2212,7 @@ void vFppipppp(x64emu_t *emu, uintptr_t fnc);
 void vFppCuupp(x64emu_t *emu, uintptr_t fnc);
 void vFppuiiip(x64emu_t *emu, uintptr_t fnc);
 void vFppuupii(x64emu_t *emu, uintptr_t fnc);
+void vFppUuupp(x64emu_t *emu, uintptr_t fnc);
 void vFppddddu(x64emu_t *emu, uintptr_t fnc);
 void vFppddpiu(x64emu_t *emu, uintptr_t fnc);
 void vFpplpppi(x64emu_t *emu, uintptr_t fnc);