From 986476b9db31de92b6a3fe473b2b62526db6fda5 Mon Sep 17 00:00:00 2001
From: TomW <TomW@localhost>
Date: Mon, 21 Sep 2015 21:54:43 +0100
Subject: [PATCH] FPU rounding modes now have effect in FADD double precision.
 Fixes rendering in Interstate '76.

---
 src/codegen_ops_x86-64.h | 33 +++++++++++++++++++++++++++++++++
 src/codegen_ops_x86.h    | 35 +++++++++++++++++++++++++++++++++++
 src/x87_ops.h            |  3 +++
 src/x87_ops_arith.h      |  4 ++++
 4 files changed, 75 insertions(+)

diff --git a/src/codegen_ops_x86-64.h b/src/codegen_ops_x86-64.h
index 1efb279..7b52213 100644
--- a/src/codegen_ops_x86-64.h
+++ b/src/codegen_ops_x86-64.h
@@ -4080,7 +4080,40 @@ static void FP_OP_D(int op)
         addbyte(0x0f);
         addbyte(0x6e);
         addbyte(0xc8);
+        if (((npxc >> 10) & 3) && op == FPU_ADD)
+        {
+                addbyte(0x0f); /*STMXCSR [ESP+8]*/
+                addbyte(0xae);
+                addbyte(0x5c);
+                addbyte(0x24);
+                addbyte(0x08);
+                addbyte(0x8b); /*MOV EAX, [ESP+8]*/
+                addbyte(0x44);
+                addbyte(0x24);
+                addbyte(0x08);
+                addbyte(0x25); /*AND EAX, ~(3 << 13)*/
+                addlong(~(3 << 10));
+                addbyte(0x0d); /*OR EAX, (npxc & (3 << 10)) << 3*/
+                addlong((npxc & (3 << 10)) << 3);
+                addbyte(0x89); /*MOV [RSP+12], EAX*/
+                addbyte(0x44);
+                addbyte(0x24);
+                addbyte(0x0c);
+                addbyte(0x0f); /*LDMXCSR [RSP+12]*/
+                addbyte(0xae);
+                addbyte(0x54);
+                addbyte(0x24);
+                addbyte(0x0c);
+        }
         FP_OP_MEM(op);
+        if (((npxc >> 10) & 3) && op == FPU_ADD)
+        {
+                addbyte(0x0f); /*LDMXCSR [RSP+8]*/
+                addbyte(0xae);
+                addbyte(0x54);
+                addbyte(0x24);
+                addbyte(0x08);
+        }
 }
 static void FP_OP_IW(int op)
 {
diff --git a/src/codegen_ops_x86.h b/src/codegen_ops_x86.h
index 1097eed..e8695ff 100644
--- a/src/codegen_ops_x86.h
+++ b/src/codegen_ops_x86.h
@@ -2496,6 +2496,34 @@ static void FP_OP_D(int op)
         addbyte(0x89); /*MOV [ESP], EAX*/
         addbyte(0x04);
         addbyte(0x24);
+        if (((npxc >> 10) & 3) && op == FPU_ADD)
+        {
+                addbyte(0x9b); /*FSTCW [ESP+8]*/
+                addbyte(0xd9);
+                addbyte(0x7c);
+                addbyte(0x24);
+                addbyte(0x08);
+                addbyte(0x66); /*MOV AX, [ESP+8]*/
+                addbyte(0x8b);
+                addbyte(0x44);
+                addbyte(0x24);
+                addbyte(0x08);
+                addbyte(0x66); /*AND AX, ~(3 << 10)*/
+                addbyte(0x25);
+                addword(~(3 << 10));
+                addbyte(0x66); /*OR AX, npxc & (3 << 10)*/
+                addbyte(0x0d);
+                addword(npxc & (3 << 10));
+                addbyte(0x66); /*MOV [ESP+12], AX*/
+                addbyte(0x89);
+                addbyte(0x44);
+                addbyte(0x24);
+                addbyte(0x0c);
+                addbyte(0xd9); /*FLDCW [ESP+12]*/
+                addbyte(0x6c);
+                addbyte(0x24);
+                addbyte(0x0c);
+        }
         addbyte(0x89); /*MOV [ESP+4], EDX*/
         addbyte(0x54);
         addbyte(0x24);
@@ -2515,6 +2543,13 @@ static void FP_OP_D(int op)
         addbyte(0x1c);
         addbyte(0xdd);
         addlong((uintptr_t)ST);
+        if (((npxc >> 10) & 3) && op == FPU_ADD)
+        {
+                addbyte(0xd9); /*FLDCW [ESP+8]*/
+                addbyte(0x6c);
+                addbyte(0x24);
+                addbyte(0x08);
+        }        
 }
 static void FP_OP_IW(int op)
 {
diff --git a/src/x87_ops.h b/src/x87_ops.h
index e17998f..188b9cc 100644
--- a/src/x87_ops.h
+++ b/src/x87_ops.h
@@ -1,7 +1,10 @@
 #include <math.h>
+#include <fenv.h>
 
 #define fplog 0
 
+static int rounding_modes[4] = {FE_TONEAREST, FE_DOWNWARD, FE_UPWARD, FE_TOWARDZERO};
+
 #define ST(x) ST[((TOP+(x))&7)]
 
 #define C0 (1<<8)
diff --git a/src/x87_ops_arith.h b/src/x87_ops_arith.h
index 7536040..c198289 100644
--- a/src/x87_ops_arith.h
+++ b/src/x87_ops_arith.h
@@ -5,7 +5,11 @@ static int opFADD ## name ## _a ## a_size(uint32_t fetchdat)    \
         FP_ENTER();                                             \
         fetch_ea_ ## a_size(fetchdat);                          \
         load_var = get(); if (abrt) return 1;                   \
+        if ((npxc >> 10) & 3)                                   \
+                fesetround(rounding_modes[(npxc >> 10) & 3]);   \
         ST(0) += use_var;                                       \
+        if ((npxc >> 10) & 3)                                   \
+                fesetround(FE_TONEAREST);                       \
         tag[TOP] &= ~TAG_UINT64;                                \
         CLOCK_CYCLES(8);                                        \
         return 0;                                               \