Use correct rounding mode on x86 and x86-64 recompiled versions of PF2ID. Fixes rendering errors on Unreal 3DNow! software renderer

src/codegen_backend_x86-64.c

@@ -310,6 +310,7 @@ void codegen_backend_init()
                 "stmxcsr %0\n"
                 : "=m" (cpu_state.old_fp_control)
         );
+        cpu_state.trunc_fp_control = cpu_state.old_fp_control | 0x6000;
 }
 
 void codegen_set_rounding_mode(int mode)

src/codegen_backend_x86-64_ops.c

@@ -541,6 +541,11 @@ void host_x86_LDMXCSR(codeblock_t *block, void *p)
         {
                 codegen_addbyte4(block, 0x0f, 0xae, 0x50 | REG_EBP, offset); /*LDMXCSR offset[EBP]*/
         }
+        else if (offset < (1ull << 32))
+        {
+                codegen_addbyte3(block, 0x0f, 0xae, 0x90 | REG_EBP); /*LDMXCSR offset[EBP]*/
+                codegen_addlong(block, offset);
+        }
         else
         {
                 fatal("host_x86_LDMXCSR - out of range %p\n", p);

src/codegen_backend_x86-64_uops.c

@@ -1576,7 +1576,9 @@ static int codegen_PF2ID(codeblock_t *block, uop_t *uop)
 
         if (REG_IS_Q(dest_size) && REG_IS_Q(src_size_a))
         {
+                host_x86_LDMXCSR(block, &cpu_state.trunc_fp_control);
                 host_x86_CVTPS2DQ_XREG_XREG(block, dest_reg, src_reg_a);
+                host_x86_LDMXCSR(block, &cpu_state.old_fp_control);
         }
         else
                 fatal("PF2ID %02x %02x\n", uop->dest_reg_a_real);

src/codegen_backend_x86.c

@@ -333,6 +333,7 @@ pclog("  offsetof(codeblock_t, next_2)=%i\n", offsetof(codeblock_t, next_2));
                 : "=m" (cpu_state.old_fp_control2),
                   "=m" (cpu_state.old_fp_control)
         );
+        cpu_state.trunc_fp_control = cpu_state.old_fp_control | 0x6000;
 }
 
 void codegen_set_rounding_mode(int mode)

src/codegen_backend_x86_uops.c

@@ -1579,7 +1579,9 @@ static int codegen_PF2ID(codeblock_t *block, uop_t *uop)
 
         if (REG_IS_Q(dest_size) && REG_IS_Q(src_size_a))
         {
+                host_x86_LDMXCSR(block, &cpu_state.trunc_fp_control);
                 host_x86_CVTPS2DQ_XREG_XREG(block, dest_reg, src_reg_a);
+                host_x86_LDMXCSR(block, &cpu_state.old_fp_control);
         }
         else
                 fatal("PF2ID %02x %02x\n", uop->dest_reg_a_real);

src/ibm.h

@@ -181,6 +181,9 @@ struct
         uint32_t old_fp_control, new_fp_control;
 #if defined i386 || defined __i386 || defined __i386__ || defined _X86_
         uint16_t old_fp_control2, new_fp_control2;
+#endif
+#if defined i386 || defined __i386 || defined __i386__ || defined _X86_ || defined __amd64__
+        uint32_t trunc_fp_control;
 #endif
         x86seg seg_cs,seg_ds,seg_es,seg_ss,seg_fs,seg_gs;