FCOM/FUCOM now use inline assembler on x86 targets. Unreal now works.

src/x87_ops.h

@@ -164,6 +164,74 @@ static inline void x87_stmmx(MMX_REG r)
         writememw(easeg, eaaddr + 8, 0xffff);
 }
 
+static inline uint16_t x87_compare(double a, double b)
+{
+#if defined i386 || defined __i386 || defined __i386__ || defined _X86_ || defined WIN32 || defined _WIN32 || defined _WIN32
+        uint32_t out;
+        
+        /* Memory barrier, to force GCC to write to the input parameters
+         * before the compare rather than after */
+        asm volatile ("" : : : "memory");
+        
+        asm(
+                "fldl %2\n"
+                "fldl %1\n"
+                "fclex\n"
+                "fcompp\n"                
+                "fnstsw %0\n"
+                : "=m" (out)
+                : "m" (a), "m" (b)
+        );
+
+        return out & (C0|C2|C3);
+#else
+        /* Generic C version is known to give incorrect results in some
+         * situations, eg comparison of infinity (Unreal) */
+        uint32_t out = 0;
+        
+        if (a == b)
+                out |= C3;
+        else if (a < b)
+                out |= C0;
+                
+        return out;
+#endif
+}
+
+static inline uint16_t x87_ucompare(double a, double b)
+{
+#if defined i386 || defined __i386 || defined __i386__ || defined _X86_ || defined WIN32 || defined _WIN32 || defined _WIN32
+        uint32_t out;
+        
+        /* Memory barrier, to force GCC to write to the input parameters
+         * before the compare rather than after */
+        asm volatile ("" : : : "memory");
+        
+        asm(
+                "fldl %2\n"
+                "fldl %1\n"
+                "fclex\n"
+                "fucompp\n"                
+                "fnstsw %0\n"
+                : "=m" (out)
+                : "m" (a), "m" (b)
+        );
+
+        return out & (C0|C2|C3);
+#else
+        /* Generic C version is known to give incorrect results in some
+         * situations, eg comparison of infinity (Unreal) */
+        uint32_t out = 0;
+        
+        if (a == b)
+                out |= C3;
+        else if (a < b)
+                out |= C0;
+                
+        return out;
+#endif
+}
+
 typedef union
 {
         float s;

src/x87_ops_arith.h

@@ -16,8 +16,7 @@ static int opFCOM ## name ## _a ## a_size(uint32_t fetchdat)    \
         fetch_ea_ ## a_size(fetchdat);                          \
         load_var = get(); if (abrt) return 1;                   \
         npxs &= ~(C0|C2|C3);                                    \
-        if (ST(0) == use_var)     npxs |= C3;                   \
+        npxs |= x87_compare(ST(0), (double)use_var);            \
-        else if (ST(0) < use_var) npxs |= C0;                   \
         CLOCK_CYCLES(4);                                        \
         return 0;                                               \
 }                                                               \
@@ -28,8 +27,7 @@ static int opFCOMP ## name ## _a ## a_size(uint32_t fetchdat)   \
         fetch_ea_ ## a_size(fetchdat);                          \
         load_var = get(); if (abrt) return 1;                   \
         npxs &= ~(C0|C2|C3);                                    \
-        if (ST(0) == use_var)     npxs |= C3;                   \
+        npxs |= x87_compare(ST(0), (double)use_var);            \
-        else if (ST(0) < use_var) npxs |= C0;                   \
         x87_pop();                                              \
         CLOCK_CYCLES(4);                                        \
         return 0;                                               \
@@ -146,8 +144,7 @@ static int opFCOMP(uint32_t fetchdat)
         pc++;
         if (fplog) pclog("FCOMP\n");
         npxs &= ~(C0|C2|C3);
-        if (ST(0) == ST(fetchdat & 7))     npxs |= C3;
+        npxs |= x87_compare(ST(0), ST(fetchdat & 7));
-        else if (ST(0) < ST(fetchdat & 7)) npxs |= C0;
         x87_pop();
         CLOCK_CYCLES(4);
         return 0;
@@ -161,10 +158,9 @@ static int opFCOMPP(uint32_t fetchdat)
         npxs &= ~(C0|C2|C3);
         if (*(uint64_t *)&ST(0) == ((uint64_t)1 << 63) && *(uint64_t *)&ST(1) == 0)
                 npxs |= C0; /*Nasty hack to fix 80387 detection*/
-        else if (ST(0) == ST(1))
+        else
-                npxs |= C3;
+                npxs |= x87_compare(ST(0), ST(1));
-        else if (ST(0) < ST(1))
+
-                npxs |= C0;
         x87_pop();
         x87_pop();
         CLOCK_CYCLES(4);
@@ -176,8 +172,7 @@ static int opFUCOMPP(uint32_t fetchdat)
         pc++;
         if (fplog) pclog("FUCOMPP\n", easeg, eaaddr);
         npxs &= ~(C0|C2|C3);
-        if (ST(0) == ST(1))     npxs |= C3;
+        npxs |= x87_ucompare(ST(0), ST(1));
-        else if (ST(0) < ST(1)) npxs |= C0;
         x87_pop();
         x87_pop();
         CLOCK_CYCLES(5);
@@ -335,8 +330,7 @@ static int opFUCOM(uint32_t fetchdat)
         pc++;
         if (fplog) pclog("FUCOM\n");
         npxs &= ~(C0|C2|C3);
-        if (ST(0) == ST(fetchdat&7))     npxs |= C3;
+        npxs |= x87_ucompare(ST(0), ST(fetchdat & 7));
-        else if (ST(0) < ST(fetchdat&7)) npxs |= C0;
         CLOCK_CYCLES(4);
         return 0;
 }
@@ -347,8 +341,7 @@ static int opFUCOMP(uint32_t fetchdat)
         pc++;
         if (fplog) pclog("FUCOMP\n");
         npxs &= ~(C0|C2|C3);
-        if (ST(0) == ST(fetchdat&7))     npxs |= C3;
+        npxs |= x87_ucompare(ST(0), ST(fetchdat & 7));
-        else if (ST(0) < ST(fetchdat&7)) npxs |= C0;
         x87_pop();
         CLOCK_CYCLES(4);
         return 0;