mirrors/Bochs
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Quite different implementation of DAZ feature

Browse Source 
Fixed bugs in execution of float64 packed instructions (almost all instructions affected)
This commit is contained in:
Stanislav Shwartsman 2003-05-26 18:06:09 +00:00
parent c82060a215
commit 0cb847f6d3
4 changed files with 399 additions and 202 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
bochs/cpu/softfloat-specialize.h
View File

@ -65,16 +65,6 @@ BX_CPP_INLINE int get_float_nan_handling_mode(float_status_t &status)
return status.float_nan_handling_mode;
}
/*----------------------------------------------------------------------------
| Returns 1 if the <denormals-are-zeroes> feature is supported;
| otherwise returns 0.
*----------------------------------------------------------------------------*/
BX_CPP_INLINE int get_DAZ(float_status_t &status)
{
return status.denormals_are_zeroes;
}
/*----------------------------------------------------------------------------
| Returns 1 if the <flush-underflow-to-zero> feature is supported;
| otherwise returns 0.
@ -107,7 +97,7 @@ typedef struct {
| otherwise returns 0.
*----------------------------------------------------------------------------*/
flag float32_is_nan(float32 a)
int float32_is_nan(float32 a)
{
return (0xFF000000 < (Bit32u) (a<<1));
}
@ -117,7 +107,7 @@ flag float32_is_nan(float32 a)
| NaN; otherwise returns 0.
*----------------------------------------------------------------------------*/
flag float32_is_signaling_nan(float32 a)
int float32_is_signaling_nan(float32 a)
{
return (((a>>22) & 0x1FF) == 0x1FE) && (a & 0x003FFFFF);
}
@ -199,7 +189,7 @@ static float32 propagateFloat32NaN(float32 a, float32 b, float_status_t &status)
| otherwise returns 0.
*----------------------------------------------------------------------------*/
flag float64_is_nan(float64 a)
int float64_is_nan(float64 a)
{
return (BX_CONST64(0xFFE0000000000000) < (Bit64u) (a<<1));
}
@ -209,7 +199,7 @@ flag float64_is_nan(float64 a)
| NaN; otherwise returns 0.
*----------------------------------------------------------------------------*/
flag float64_is_signaling_nan(float64 a)
int float64_is_signaling_nan(float64 a)
{
return (((a>>51) & 0xFFF) == 0xFFE) && (a & BX_CONST64(0x0007FFFFFFFFFFFF));
}

262
bochs/cpu/softfloat.cc
View File

@ -766,7 +766,7 @@ float64 float32_to_float64(float32 a, float_status_t &status)
return packFloat64(aSign, 0x7FF, 0);
}
if (aExp == 0) {
if (aSig == 0 || get_DAZ(status)) return packFloat64(aSign, 0, 0);
if (aSig == 0) return packFloat64(aSign, 0, 0);
float_raise(status, float_flag_denormal);
normalizeFloat32Subnormal(aSig, &aExp, &aSig);
--aExp;
@ -852,16 +852,10 @@ static float32 addFloat32Sigs(float32 a, float32 b, flag zSign, float_status_t &
bExp = extractFloat32Exp(b);
if ((aExp == 0) && aSig) {
if (get_DAZ(status)) {
aSig = 0;
}
else float_raise(status, float_flag_denormal);
float_raise(status, float_flag_denormal);
}
if ((bExp == 0) && bSig) {
if (get_DAZ(status)) {
bSig = 0;
}
else float_raise(status, float_flag_denormal);
float_raise(status, float_flag_denormal);
}
expDiff = aExp - bExp;
@ -937,16 +931,10 @@ static float32 subFloat32Sigs(float32 a, float32 b, flag zSign, float_status_t &
bExp = extractFloat32Exp(b);
if ((aExp == 0) && aSig) {
if (get_DAZ(status)) {
aSig = 0;
}
else float_raise(status, float_flag_denormal);
float_raise(status, float_flag_denormal);
}
if ((bExp == 0) && bSig) {
if (get_DAZ(status)) {
bSig = 0;
}
else float_raise(status, float_flag_denormal);
float_raise(status, float_flag_denormal);
}
expDiff = aExp - bExp;
@ -1085,12 +1073,12 @@ float32 float32_mul(float32 a, float32 b, float_status_t &status)
return packFloat32(zSign, 0xFF, 0);
}
if (aExp == 0) {
if (aSig == 0 || get_DAZ(status)) return packFloat32(zSign, 0, 0);
if (aSig == 0) return packFloat32(zSign, 0, 0);
float_raise(status, float_flag_denormal);
normalizeFloat32Subnormal(aSig, &aExp, &aSig);
}
if (bExp == 0) {
if (bSig == 0 || get_DAZ(status)) return packFloat32(zSign, 0, 0);
if (bSig == 0) return packFloat32(zSign, 0, 0);
float_raise(status, float_flag_denormal);
normalizeFloat32Subnormal(bSig, &bExp, &bSig);
}
@ -1139,12 +1127,10 @@ float32 float32_div(float32 a, float32 b, float_status_t &status)
return packFloat32(zSign, 0, 0);
}
if (bExp == 0) {
if (bSig == 0 || get_DAZ(status)) {
if (aExp == 0) {
if (aSig == 0 || get_DAZ(status)) {
float_raise(status, float_flag_invalid);
return float32_default_nan;
}
if (bSig == 0) {
if ((aExp | aSig) == 0) {
float_raise(status, float_flag_invalid);
return float32_default_nan;
}
float_raise(status, float_flag_divbyzero);
return packFloat32(zSign, 0xFF, 0);
@ -1153,7 +1139,7 @@ float32 float32_div(float32 a, float32 b, float_status_t &status)
normalizeFloat32Subnormal(bSig, &bExp, &bSig);
}
if (aExp == 0) {
if (aSig == 0 || get_DAZ(status)) return packFloat32(zSign, 0, 0);
if (aSig == 0) return packFloat32(zSign, 0, 0);
float_raise(status, float_flag_denormal);
normalizeFloat32Subnormal(aSig, &aExp, &aSig);
}
@ -1205,7 +1191,7 @@ float32 float32_rem(float32 a, float32 b, float_status_t &status)
return a;
}
if (bExp == 0) {
if (bSig == 0 || get_DAZ(status)) {
if (bSig == 0) {
float_raise(status, float_flag_invalid);
return float32_default_nan;
}
@ -1213,7 +1199,7 @@ float32 float32_rem(float32 a, float32 b, float_status_t &status)
normalizeFloat32Subnormal(bSig, &bExp, &bSig);
}
if (aExp == 0) {
if (aSig == 0 || get_DAZ(status)) return packFloat32(aSign, 0, 0);
if (aSig == 0) return packFloat32(aSign, 0, 0);
float_raise(status, float_flag_denormal);
normalizeFloat32Subnormal(aSig, &aExp, &aSig);
}
@ -1295,14 +1281,12 @@ float32 float32_sqrt(float32 a, float_status_t &status)
return float32_default_nan;
}
if (aSign) {
if (aExp == 0) {
if (aSig == 0 || get_DAZ(status)) return packFloat32(aSign, 0, 0);
}
if ((aExp | aSig) == 0) return a;
float_raise(status, float_flag_invalid);
return float32_default_nan;
}
if (aExp == 0) {
if (aSig == 0 || get_DAZ(status)) return 0;
if (aSig == 0) return 0;
float_raise(status, float_flag_denormal);
normalizeFloat32Subnormal(aSig, &aExp, &aSig);
}
@ -1348,13 +1332,9 @@ int float32_eq(float32 a, float32 b, float_status_t &status)
return 0;
}
if (aClass == float_denormal) {
if (get_DAZ(status)) a = 0;
else float_raise(status, float_flag_denormal);
}
if (bClass == float_denormal) {
if (get_DAZ(status)) b = 0;
else float_raise(status, float_flag_denormal);
if (aClass == float_denormal || bClass == float_denormal)
{
float_raise(status, float_flag_denormal);
}
return (a == b) || ((Bit32u) ((a | b)<<1) == 0);
@ -1377,13 +1357,9 @@ int float32_le(float32 a, float32 b, float_status_t &status)
return 0;
}
if (aClass == float_denormal) {
if (get_DAZ(status)) a = 0;
else float_raise(status, float_flag_denormal);
}
if (bClass == float_denormal) {
if (get_DAZ(status)) b = 0;
else float_raise(status, float_flag_denormal);
if (aClass == float_denormal || bClass == float_denormal)
{
float_raise(status, float_flag_denormal);
}
flag aSign = extractFloat32Sign(a);
@ -1408,13 +1384,9 @@ int float32_lt(float32 a, float32 b, float_status_t &status)
return 0;
}
if (aClass == float_denormal) {
if (get_DAZ(status)) a = 0;
else float_raise(status, float_flag_denormal);
}
if (bClass == float_denormal) {
if (get_DAZ(status)) b = 0;
else float_raise(status, float_flag_denormal);
if (aClass == float_denormal || bClass == float_denormal)
{
float_raise(status, float_flag_denormal);
}
flag aSign = extractFloat32Sign(a);
@ -1440,13 +1412,9 @@ int float32_eq_signaling(float32 a, float32 b, float_status_t &status)
return 0;
}
if (aClass == float_denormal) {
if (get_DAZ(status)) a = 0;
else float_raise(status, float_flag_denormal);
}
if (bClass == float_denormal) {
if (get_DAZ(status)) b = 0;
else float_raise(status, float_flag_denormal);
if (aClass == float_denormal || bClass == float_denormal)
{
float_raise(status, float_flag_denormal);
}
return (a == b) || ((Bit32u) ((a | b)<<1) == 0);
@ -1473,13 +1441,9 @@ int float32_le_quiet(float32 a, float32 b, float_status_t &status)
return 0;
}
if (aClass == float_denormal) {
if (get_DAZ(status)) a = 0;
else float_raise(status, float_flag_denormal);
}
if (bClass == float_denormal) {
if (get_DAZ(status)) b = 0;
else float_raise(status, float_flag_denormal);
if (aClass == float_denormal || bClass == float_denormal)
{
float_raise(status, float_flag_denormal);
}
flag aSign = extractFloat32Sign(a);
@ -1509,13 +1473,9 @@ int float32_lt_quiet(float32 a, float32 b, float_status_t &status)
return 0;
}
if (aClass == float_denormal) {
if (get_DAZ(status)) a = 0;
else float_raise(status, float_flag_denormal);
}
if (bClass == float_denormal) {
if (get_DAZ(status)) b = 0;
else float_raise(status, float_flag_denormal);
if (aClass == float_denormal || bClass == float_denormal)
{
float_raise(status, float_flag_denormal);
}
flag aSign = extractFloat32Sign(a);
@ -1545,8 +1505,7 @@ int float32_unordered(float32 a, float32 b, float_status_t &status)
if (aClass == float_denormal || bClass == float_denormal)
{
if (! get_DAZ(status))
float_raise(status, float_flag_denormal);
float_raise(status, float_flag_denormal);
}
return 0;
@ -1570,13 +1529,9 @@ int float32_compare(float32 a, float32 b, float_status_t &status)
return float_relation_unordered;
}
if (aClass == float_denormal) {
if (get_DAZ(status)) a = 0;
else float_raise(status, float_flag_denormal);
}
if (bClass == float_denormal) {
if (get_DAZ(status)) b = 0;
else float_raise(status, float_flag_denormal);
if (aClass == float_denormal || bClass == float_denormal)
{
float_raise(status, float_flag_denormal);
}
if ((a == b) || ((Bit32u) ((a | b)<<1) == 0)) return float_relation_equal;
@ -1613,13 +1568,9 @@ int float32_compare_quiet(float32 a, float32 b, float_status_t &status)
return float_relation_unordered;
}
if (aClass == float_denormal) {
if (get_DAZ(status)) a = 0;
else float_raise(status, float_flag_denormal);
}
if (bClass == float_denormal) {
if (get_DAZ(status)) b = 0;
else float_raise(status, float_flag_denormal);
if (aClass == float_denormal || bClass == float_denormal)
{
float_raise(status, float_flag_denormal);
}
if ((a == b) || ((Bit32u) ((a | b)<<1) == 0)) return float_relation_equal;
@ -1816,7 +1767,7 @@ float32 float64_to_float32(float64 a, float_status_t &status)
return packFloat32(aSign, 0xFF, 0);
}
if (aExp == 0) {
if (aSig == 0 || get_DAZ(status)) return packFloat32(aSign, 0, 0);
if (aSig == 0) return packFloat32(aSign, 0, 0);
float_raise(status, float_flag_denormal);
}
shift64RightJamming(aSig, 22, &aSig);
@ -1907,16 +1858,10 @@ static float64 addFloat64Sigs(float64 a, float64 b, flag zSign, float_status_t &
bExp = extractFloat64Exp(b);
if ((aExp == 0) && aSig) {
if (get_DAZ(status)) {
aSig = 0;
}
else float_raise(status, float_flag_denormal);
float_raise(status, float_flag_denormal);
}
if ((bExp == 0) && bSig) {
if (get_DAZ(status)) {
bSig = 0;
}
else float_raise(status, float_flag_denormal);
float_raise(status, float_flag_denormal);
}
expDiff = aExp - bExp;
@ -1991,16 +1936,10 @@ static float64 subFloat64Sigs(float64 a, float64 b, flag zSign, float_status_t &
bExp = extractFloat64Exp(b);
if ((aExp == 0) && aSig) {
if (get_DAZ(status)) {
aSig = 0;
}
else float_raise(status, float_flag_denormal);
float_raise(status, float_flag_denormal);
}
if ((bExp == 0) && bSig) {
if (get_DAZ(status)) {
bSig = 0;
}
else float_raise(status, float_flag_denormal);
float_raise(status, float_flag_denormal);
}
expDiff = aExp - bExp;
@ -2137,12 +2076,12 @@ float64 float64_mul(float64 a, float64 b, float_status_t &status)
return packFloat64(zSign, 0x7FF, 0);
}
if (aExp == 0) {
if (aSig == 0 || get_DAZ(status)) return packFloat64(zSign, 0, 0);
if (aSig == 0) return packFloat64(zSign, 0, 0);
float_raise(status, float_flag_denormal);
normalizeFloat64Subnormal(aSig, &aExp, &aSig);
}
if (bExp == 0) {
if (bSig == 0 || get_DAZ(status)) return packFloat64(zSign, 0, 0);
if (bSig == 0) return packFloat64(zSign, 0, 0);
float_raise(status, float_flag_denormal);
normalizeFloat64Subnormal(bSig, &bExp, &bSig);
}
@ -2193,12 +2132,10 @@ float64 float64_div(float64 a, float64 b, float_status_t &status)
return packFloat64(zSign, 0, 0);
}
if (bExp == 0) {
if (bSig == 0 || get_DAZ(status)) {
if (aExp == 0) {
if (aSig == 0 || get_DAZ(status)) {
float_raise(status, float_flag_invalid);
return float64_default_nan;
}
if (bSig == 0) {
if ((aExp | aSig) == 0) {
float_raise(status, float_flag_invalid);
return float64_default_nan;
}
float_raise(status, float_flag_divbyzero);
return packFloat64(zSign, 0x7FF, 0);
@ -2207,7 +2144,7 @@ float64 float64_div(float64 a, float64 b, float_status_t &status)
normalizeFloat64Subnormal(bSig, &bExp, &bSig);
}
if (aExp == 0) {
if (aSig == 0 || get_DAZ(status)) return packFloat64(zSign, 0, 0);
if (aSig == 0) return packFloat64(zSign, 0, 0);
float_raise(status, float_flag_denormal);
normalizeFloat64Subnormal(aSig, &aExp, &aSig);
}
@ -2263,7 +2200,7 @@ float64 float64_rem(float64 a, float64 b, float_status_t &status)
return a;
}
if (bExp == 0) {
if (bSig == 0 || get_DAZ(status)) {
if (bSig == 0) {
float_raise(status, float_flag_invalid);
return float64_default_nan;
}
@ -2271,7 +2208,7 @@ float64 float64_rem(float64 a, float64 b, float_status_t &status)
normalizeFloat64Subnormal(bSig, &bExp, &bSig);
}
if (aExp == 0) {
if (aSig == 0 || get_DAZ(status)) return packFloat64(aSign, 0, 0);
if (aSig == 0) return packFloat64(aSign, 0, 0);
float_raise(status, float_flag_denormal);
normalizeFloat64Subnormal(aSig, &aExp, &aSig);
}
@ -2341,14 +2278,12 @@ float64 float64_sqrt(float64 a, float_status_t &status)
return float64_default_nan;
}
if (aSign) {
if (aExp == 0) {
if (aSig == 0 || get_DAZ(status)) return packFloat64(aSign, 0, 0);
}
if ((aExp | aSig) == 0) return a;
float_raise(status, float_flag_invalid);
return float64_default_nan;
}
if (aExp == 0) {
if (aSig == 0 || get_DAZ(status)) return 0;
if (aSig == 0) return 0;
float_raise(status, float_flag_denormal);
normalizeFloat64Subnormal(aSig, &aExp, &aSig);
}
@ -2391,13 +2326,9 @@ int float64_eq(float64 a, float64 b, float_status_t &status)
return 0;
}
if (aClass == float_denormal) {
if (get_DAZ(status)) a = 0;
else float_raise(status, float_flag_denormal);
}
if (bClass == float_denormal) {
if (get_DAZ(status)) b = 0;
else float_raise(status, float_flag_denormal);
if (aClass == float_denormal || bClass == float_denormal)
{
float_raise(status, float_flag_denormal);
}
return (a == b) || ((Bit64u) ((a | b)<<1) == 0);
@ -2420,13 +2351,9 @@ int float64_le(float64 a, float64 b, float_status_t &status)
return 0;
}
if (aClass == float_denormal) {
if (get_DAZ(status)) a = 0;
else float_raise(status, float_flag_denormal);
}
if (bClass == float_denormal) {
if (get_DAZ(status)) b = 0;
else float_raise(status, float_flag_denormal);
if (aClass == float_denormal || bClass == float_denormal)
{
float_raise(status, float_flag_denormal);
}
flag aSign = extractFloat64Sign(a);
@ -2451,13 +2378,9 @@ int float64_lt(float64 a, float64 b, float_status_t &status)
return 0;
}
if (aClass == float_denormal) {
if (get_DAZ(status)) a = 0;
else float_raise(status, float_flag_denormal);
}
if (bClass == float_denormal) {
if (get_DAZ(status)) b = 0;
else float_raise(status, float_flag_denormal);
if (aClass == float_denormal || bClass == float_denormal)
{
float_raise(status, float_flag_denormal);
}
flag aSign = extractFloat64Sign(a);
@ -2483,13 +2406,9 @@ int float64_eq_signaling(float64 a, float64 b, float_status_t &status)
return 0;
}
if (aClass == float_denormal) {
if (get_DAZ(status)) a = 0;
else float_raise(status, float_flag_denormal);
}
if (bClass == float_denormal) {
if (get_DAZ(status)) b = 0;
else float_raise(status, float_flag_denormal);
if (aClass == float_denormal || bClass == float_denormal)
{
float_raise(status, float_flag_denormal);
}
return (a == b) || ((Bit64u) ((a | b)<<1) == 0);
@ -2516,13 +2435,9 @@ int float64_le_quiet(float64 a, float64 b, float_status_t &status)
return 0;
}
if (aClass == float_denormal) {
if (get_DAZ(status)) a = 0;
else float_raise(status, float_flag_denormal);
}
if (bClass == float_denormal) {
if (get_DAZ(status)) b = 0;
else float_raise(status, float_flag_denormal);
if (aClass == float_denormal || bClass == float_denormal)
{
float_raise(status, float_flag_denormal);
}
flag aSign = extractFloat64Sign(a);
@ -2552,13 +2467,9 @@ int float64_lt_quiet(float64 a, float64 b, float_status_t &status)
return 0;
}
if (aClass == float_denormal) {
if (get_DAZ(status)) a = 0;
else float_raise(status, float_flag_denormal);
}
if (bClass == float_denormal) {
if (get_DAZ(status)) b = 0;
else float_raise(status, float_flag_denormal);
if (aClass == float_denormal || bClass == float_denormal)
{
float_raise(status, float_flag_denormal);
}
flag aSign = extractFloat64Sign(a);
@ -2588,8 +2499,7 @@ int float64_unordered(float64 a, float64 b, float_status_t &status)
if (aClass == float_denormal || bClass == float_denormal)
{
if (! get_DAZ(status))
float_raise(status, float_flag_denormal);
float_raise(status, float_flag_denormal);
}
return 0;
@ -2613,13 +2523,9 @@ int float64_compare(float64 a, float64 b, float_status_t &status)
return float_relation_unordered;
}
if (aClass == float_denormal) {
if (get_DAZ(status)) a = 0;
else float_raise(status, float_flag_denormal);
}
if (bClass == float_denormal) {
if (get_DAZ(status)) b = 0;
else float_raise(status, float_flag_denormal);
if (aClass == float_denormal || bClass == float_denormal)
{
float_raise(status, float_flag_denormal);
}
if ((a == b) || ((Bit64u) ((a | b)<<1) == 0)) return float_relation_equal;
@ -2656,13 +2562,9 @@ int float64_compare_quiet(float64 a, float64 b, float_status_t &status)
return float_relation_unordered;
}
if (aClass == float_denormal) {
if (get_DAZ(status)) a = 0;
else float_raise(status, float_flag_denormal);
}
if (bClass == float_denormal) {
if (get_DAZ(status)) b = 0;
else float_raise(status, float_flag_denormal);
if (aClass == float_denormal || bClass == float_denormal)
{
float_raise(status, float_flag_denormal);
}
if ((a == b) || ((Bit64u) ((a | b)<<1) == 0)) return float_relation_equal;

2
bochs/cpu/softfloat.h
View File

@ -36,7 +36,6 @@ these four paragraphs for those parts of this code that are retained.
* Still to be fixed:
* -----------------
* * Denormals support for add/sub/compare/convert operations (implemented)
* * DAZ (denormals-are-zeroes) feature support
* * Flush underflow to zero feature support (implemented)
* * Floating point reciprocal and sqrt reciprocal operations
*/
@ -119,7 +118,6 @@ struct float_status_t
int float_rounding_mode;
int float_exception_flags;
int float_nan_handling_mode;
int denormals_are_zeroes; /* flag register */
int flush_underflow_to_zero; /* flag register */
};
typedef struct float_status_t softfloat_status_word_t;

319
bochs/cpu/sse_pfp.cc
View File

@ -47,12 +47,23 @@ static void MXCSR_to_softfloat_status_word(softfloat_status_word_t &status, bx_m
status.float_exception_flags = 0; // clear exceptions before execution
status.float_nan_handling_mode = float_first_operand_nan;
status.float_rounding_mode = mxcsr.get_rounding_mode();
status.denormals_are_zeroes = mxcsr.get_DAZ();
// if underflow is masked and FUZ is 1, set it to 1, else to 0
status.flush_underflow_to_zero =
(mxcsr.get_flush_masked_underflow() && mxcsr.get_UM()) ? 1 : 0;
}
float32 handleDAZ(float32 op)
{
if (float32_class(op) == float_denormal) op &= ((Bit32u)(1) << 31);
return op;
}
float64 handleDAZ(float64 op)
{
if (float64_class(op) == float_denormal) op &= ((Bit64u)(1) << 63);
return op;
}
/* Comparison predicate for CMPSS/CMPPS instructions */
static float32_compare_method compare32[4] = {
float32_eq,
@ -629,6 +640,11 @@ void BX_CPU_C::CVTPS2PD_VpsWps(bxInstruction_c *i)
Float32 r0 = (Float32)(op & 0xFFFFFFFF);
Float32 r1 = (Float32)(op >> 32);
if (MXCSR.get_DAZ()) {
r0 = handleDAZ(r0);
r1 = handleDAZ(r1);
}
result.xmm32u(0) =
float32_to_float64(r0, status_word);
result.xmm32u(1) =
@ -669,6 +685,12 @@ void BX_CPU_C::CVTPD2PS_VpdWpd(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ())
{
op.xmm64u(0) = handleDAZ(op.xmm64u(0));
op.xmm64u(0) = handleDAZ(op.xmm64u(1));
}
result.xmm32u(0) =
float64_to_float32(op.xmm64u(0), status_word);
result.xmm32u(1) =
@ -710,6 +732,7 @@ void BX_CPU_C::CVTSD2SS_VsdWsd(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ()) op = handleDAZ(op);
result = float64_to_float32(op, status_word);
BX_CPU_THIS_PTR check_exceptionsSSE(status_word.float_exception_flags);
BX_WRITE_XMM_REG_LO_DWORD(i->nnn(), result);
@ -744,6 +767,7 @@ void BX_CPU_C::CVTSS2SD_VssWss(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ()) op = handleDAZ(op);
result = float32_to_float64(op, status_word);
BX_CPU_THIS_PTR check_exceptionsSSE(status_word.float_exception_flags);
BX_WRITE_XMM_REG_LO_QWORD(i->nnn(), result);
@ -1020,6 +1044,12 @@ void BX_CPU_C::UCOMISS_VssWss(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ())
{
op1 = handleDAZ(op1);
op2 = handleDAZ(op2);
}
int rc = float32_compare_quiet(op1, op2, status_word);
BX_CPU_THIS_PTR check_exceptionsSSE(status_word.float_exception_flags);
@ -1071,6 +1101,12 @@ void BX_CPU_C::UCOMISD_VsdWsd(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ())
{
op1 = handleDAZ(op1);
op2 = handleDAZ(op2);
}
int rc = float64_compare_quiet(op1, op2, status_word);
BX_CPU_THIS_PTR check_exceptionsSSE(status_word.float_exception_flags);
@ -1122,6 +1158,12 @@ void BX_CPU_C::COMISS_VpsWps(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ())
{
op1 = handleDAZ(op1);
op2 = handleDAZ(op2);
}
int rc = float32_compare(op1, op2, status_word);
BX_CPU_THIS_PTR check_exceptionsSSE(status_word.float_exception_flags);
@ -1173,6 +1215,12 @@ void BX_CPU_C::COMISD_VpdWpd(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ())
{
op1 = handleDAZ(op1);
op2 = handleDAZ(op2);
}
int rc = float64_compare(op1, op2, status_word);
BX_CPU_THIS_PTR check_exceptionsSSE(status_word.float_exception_flags);
@ -1224,6 +1272,14 @@ void BX_CPU_C::SQRTPS_VpsWps(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ())
{
op.xmm32u(0) = handleDAZ(op.xmm32u(0));
op.xmm32u(1) = handleDAZ(op.xmm32u(1));
op.xmm32u(2) = handleDAZ(op.xmm32u(2));
op.xmm32u(3) = handleDAZ(op.xmm32u(3));
}
result.xmm32u(0) =
float32_sqrt(op.xmm32u(0), status_word);
result.xmm32u(1) =
@ -1266,10 +1322,16 @@ void BX_CPU_C::SQRTPD_VpdWpd(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ())
{
op.xmm64u(0) = handleDAZ(op.xmm64u(0));
op.xmm64u(1) = handleDAZ(op.xmm64u(1));
}
result.xmm64u(0) =
float64_sqrt(op.xmm64u(0), status_word);
result.xmm64u(1) =
float32_sqrt(op.xmm64u(1), status_word);
float64_sqrt(op.xmm64u(1), status_word);
BX_CPU_THIS_PTR check_exceptionsSSE(status_word.float_exception_flags);
BX_WRITE_XMM_REG(i->nnn(), result);
@ -1303,6 +1365,7 @@ void BX_CPU_C::SQRTSD_VsdWsd(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ()) op = handleDAZ(op);
result = float64_sqrt(op, status_word);
BX_CPU_THIS_PTR check_exceptionsSSE(status_word.float_exception_flags);
BX_WRITE_XMM_REG_LO_QWORD(i->nnn(), result);
@ -1337,6 +1400,7 @@ void BX_CPU_C::SQRTSS_VssWss(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
result = float32_sqrt(op, status_word);
if (MXCSR.get_DAZ()) op = handleDAZ(op);
BX_CPU_THIS_PTR check_exceptionsSSE(status_word.float_exception_flags);
BX_WRITE_XMM_REG_LO_DWORD(i->nnn(), result);
@ -1418,6 +1482,18 @@ void BX_CPU_C::ADDPS_VpsWps(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ()) {
op1.xmm32u(0) = handleDAZ(op1.xmm32u(0));
op1.xmm32u(1) = handleDAZ(op1.xmm32u(1));
op1.xmm32u(2) = handleDAZ(op1.xmm32u(2));
op1.xmm32u(3) = handleDAZ(op1.xmm32u(3));
op2.xmm32u(0) = handleDAZ(op2.xmm32u(0));
op2.xmm32u(1) = handleDAZ(op2.xmm32u(1));
op2.xmm32u(2) = handleDAZ(op2.xmm32u(2));
op2.xmm32u(3) = handleDAZ(op2.xmm32u(3));
}
result.xmm32u(0) =
float32_add(op1.xmm32u(0), op2.xmm32u(0), status_word);
result.xmm32u(1) =
@ -1460,10 +1536,19 @@ void BX_CPU_C::ADDPD_VpdWpd(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ())
{
op1.xmm64u(0) = handleDAZ(op1.xmm64u(0));
op1.xmm64u(1) = handleDAZ(op1.xmm64u(1));
op2.xmm64u(0) = handleDAZ(op2.xmm64u(0));
op2.xmm64u(1) = handleDAZ(op2.xmm64u(1));
}
result.xmm64u(0) =
float64_add(op1.xmm64u(0), op2.xmm64u(0), status_word);
result.xmm64u(1) =
float32_add(op1.xmm64u(1), op2.xmm64u(1), status_word);
float64_add(op1.xmm64u(1), op2.xmm64u(1), status_word);
BX_CPU_THIS_PTR check_exceptionsSSE(status_word.float_exception_flags);
BX_WRITE_XMM_REG(i->nnn(), result);
@ -1497,6 +1582,13 @@ void BX_CPU_C::ADDSD_VsdWsd(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ())
{
op1 = handleDAZ(op1);
op2 = handleDAZ(op2);
}
result = float64_add(op1, op2, status_word);
BX_CPU_THIS_PTR check_exceptionsSSE(status_word.float_exception_flags);
BX_WRITE_XMM_REG_LO_QWORD(i->nnn(), result);
@ -1530,6 +1622,13 @@ void BX_CPU_C::ADDSS_VssWss(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ())
{
op1 = handleDAZ(op1);
op2 = handleDAZ(op2);
}
result = float32_add(op1, op2, status_word);
BX_CPU_THIS_PTR check_exceptionsSSE(status_word.float_exception_flags);
BX_WRITE_XMM_REG_LO_DWORD(i->nnn(), result);
@ -1564,6 +1663,18 @@ void BX_CPU_C::MULPS_VpsWps(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ()) {
op1.xmm32u(0) = handleDAZ(op1.xmm32u(0));
op1.xmm32u(1) = handleDAZ(op1.xmm32u(1));
op1.xmm32u(2) = handleDAZ(op1.xmm32u(2));
op1.xmm32u(3) = handleDAZ(op1.xmm32u(3));
op2.xmm32u(0) = handleDAZ(op2.xmm32u(0));
op2.xmm32u(1) = handleDAZ(op2.xmm32u(1));
op2.xmm32u(2) = handleDAZ(op2.xmm32u(2));
op2.xmm32u(3) = handleDAZ(op2.xmm32u(3));
}
result.xmm32u(0) =
float32_mul(op1.xmm32u(0), op2.xmm32u(0), status_word);
result.xmm32u(1) =
@ -1606,10 +1717,19 @@ void BX_CPU_C::MULPD_VpdWpd(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ())
{
op1.xmm64u(0) = handleDAZ(op1.xmm64u(0));
op1.xmm64u(1) = handleDAZ(op1.xmm64u(1));
op2.xmm64u(0) = handleDAZ(op2.xmm64u(0));
op2.xmm64u(1) = handleDAZ(op2.xmm64u(1));
}
result.xmm64u(0) =
float64_mul(op1.xmm64u(0), op2.xmm64u(0), status_word);
result.xmm64u(1) =
float32_mul(op1.xmm64u(1), op2.xmm64u(1), status_word);
float64_mul(op1.xmm64u(1), op2.xmm64u(1), status_word);
BX_CPU_THIS_PTR check_exceptionsSSE(status_word.float_exception_flags);
BX_WRITE_XMM_REG(i->nnn(), result);
@ -1643,6 +1763,13 @@ void BX_CPU_C::MULSD_VsdWsd(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ())
{
op1 = handleDAZ(op1);
op2 = handleDAZ(op2);
}
result = float64_mul(op1, op2, status_word);
BX_CPU_THIS_PTR check_exceptionsSSE(status_word.float_exception_flags);
BX_WRITE_XMM_REG_LO_QWORD(i->nnn(), result);
@ -1676,6 +1803,13 @@ void BX_CPU_C::MULSS_VssWss(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ())
{
op1 = handleDAZ(op1);
op2 = handleDAZ(op2);
}
result = float32_mul(op1, op2, status_word);
BX_CPU_THIS_PTR check_exceptionsSSE(status_word.float_exception_flags);
BX_WRITE_XMM_REG_LO_DWORD(i->nnn(), result);
@ -1710,6 +1844,18 @@ void BX_CPU_C::SUBPS_VpsWps(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ()) {
op1.xmm32u(0) = handleDAZ(op1.xmm32u(0));
op1.xmm32u(1) = handleDAZ(op1.xmm32u(1));
op1.xmm32u(2) = handleDAZ(op1.xmm32u(2));
op1.xmm32u(3) = handleDAZ(op1.xmm32u(3));
op2.xmm32u(0) = handleDAZ(op2.xmm32u(0));
op2.xmm32u(1) = handleDAZ(op2.xmm32u(1));
op2.xmm32u(2) = handleDAZ(op2.xmm32u(2));
op2.xmm32u(3) = handleDAZ(op2.xmm32u(3));
}
result.xmm32u(0) =
float32_sub(op1.xmm32u(0), op2.xmm32u(0), status_word);
result.xmm32u(1) =
@ -1752,10 +1898,19 @@ void BX_CPU_C::SUBPD_VpdWpd(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ())
{
op1.xmm64u(0) = handleDAZ(op1.xmm64u(0));
op1.xmm64u(1) = handleDAZ(op1.xmm64u(1));
op2.xmm64u(0) = handleDAZ(op2.xmm64u(0));
op2.xmm64u(1) = handleDAZ(op2.xmm64u(1));
}
result.xmm64u(0) =
float64_sub(op1.xmm64u(0), op2.xmm64u(0), status_word);
result.xmm64u(1) =
float32_sub(op1.xmm64u(1), op2.xmm64u(1), status_word);
float64_sub(op1.xmm64u(1), op2.xmm64u(1), status_word);
BX_CPU_THIS_PTR check_exceptionsSSE(status_word.float_exception_flags);
BX_WRITE_XMM_REG(i->nnn(), result);
@ -1789,6 +1944,13 @@ void BX_CPU_C::SUBSD_VsdWsd(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ())
{
op1 = handleDAZ(op1);
op2 = handleDAZ(op2);
}
result = float64_sub(op1, op2, status_word);
BX_CPU_THIS_PTR check_exceptionsSSE(status_word.float_exception_flags);
BX_WRITE_XMM_REG_LO_QWORD(i->nnn(), result);
@ -1822,6 +1984,13 @@ void BX_CPU_C::SUBSS_VssWss(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ())
{
op1 = handleDAZ(op1);
op2 = handleDAZ(op2);
}
result = float32_sub(op1, op2, status_word);
BX_CPU_THIS_PTR check_exceptionsSSE(status_word.float_exception_flags);
BX_WRITE_XMM_REG_LO_DWORD(i->nnn(), result);
@ -1857,6 +2026,18 @@ void BX_CPU_C::MINPS_VpsWps(bxInstruction_c *i)
MXCSR_to_softfloat_status_word(status_word, MXCSR);
int rc;
if (MXCSR.get_DAZ()) {
op1.xmm32u(0) = handleDAZ(op1.xmm32u(0));
op1.xmm32u(1) = handleDAZ(op1.xmm32u(1));
op1.xmm32u(2) = handleDAZ(op1.xmm32u(2));
op1.xmm32u(3) = handleDAZ(op1.xmm32u(3));
op2.xmm32u(0) = handleDAZ(op2.xmm32u(0));
op2.xmm32u(1) = handleDAZ(op2.xmm32u(1));
op2.xmm32u(2) = handleDAZ(op2.xmm32u(2));
op2.xmm32u(3) = handleDAZ(op2.xmm32u(3));
}
rc = float32_compare(op1.xmm32u(0), op2.xmm32u(0), status_word);
result.xmm32u(0) =
(rc == float_relation_less) ? op1.xmm32u(0) : op2.xmm32u(0);
@ -1904,6 +2085,15 @@ void BX_CPU_C::MINPD_VpdWpd(bxInstruction_c *i)
MXCSR_to_softfloat_status_word(status_word, MXCSR);
int rc;
if (MXCSR.get_DAZ())
{
op1.xmm64u(0) = handleDAZ(op1.xmm64u(0));
op1.xmm64u(1) = handleDAZ(op1.xmm64u(1));
op2.xmm64u(0) = handleDAZ(op2.xmm64u(0));
op2.xmm64u(1) = handleDAZ(op2.xmm64u(1));
}
rc = float64_compare(op1.xmm64u(0), op2.xmm64u(0), status_word);
result.xmm64u(0) =
(rc == float_relation_less) ? op1.xmm64u(0) : op2.xmm64u(0);
@ -1943,6 +2133,13 @@ void BX_CPU_C::MINSD_VsdWsd(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ())
{
op1 = handleDAZ(op1);
op2 = handleDAZ(op2);
}
int rc = float64_compare(op1, op2, status_word);
BX_CPU_THIS_PTR check_exceptionsSSE(status_word.float_exception_flags);
BX_WRITE_XMM_REG_LO_QWORD(i->nnn(),
@ -1977,6 +2174,13 @@ void BX_CPU_C::MINSS_VssWss(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ())
{
op1 = handleDAZ(op1);
op2 = handleDAZ(op2);
}
int rc = float32_compare(op1, op2, status_word);
BX_CPU_THIS_PTR check_exceptionsSSE(status_word.float_exception_flags);
BX_WRITE_XMM_REG_LO_DWORD(i->nnn(),
@ -2012,6 +2216,18 @@ void BX_CPU_C::DIVPS_VpsWps(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ()) {
op1.xmm32u(0) = handleDAZ(op1.xmm32u(0));
op1.xmm32u(1) = handleDAZ(op1.xmm32u(1));
op1.xmm32u(2) = handleDAZ(op1.xmm32u(2));
op1.xmm32u(3) = handleDAZ(op1.xmm32u(3));
op2.xmm32u(0) = handleDAZ(op2.xmm32u(0));
op2.xmm32u(1) = handleDAZ(op2.xmm32u(1));
op2.xmm32u(2) = handleDAZ(op2.xmm32u(2));
op2.xmm32u(3) = handleDAZ(op2.xmm32u(3));
}
result.xmm32u(0) =
float32_div(op1.xmm32u(0), op2.xmm32u(0), status_word);
result.xmm32u(1) =
@ -2054,10 +2270,19 @@ void BX_CPU_C::DIVPD_VpdWpd(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ())
{
op1.xmm64u(0) = handleDAZ(op1.xmm64u(0));
op1.xmm64u(1) = handleDAZ(op1.xmm64u(1));
op2.xmm64u(0) = handleDAZ(op2.xmm64u(0));
op2.xmm64u(1) = handleDAZ(op2.xmm64u(1));
}
result.xmm64u(0) =
float64_div(op1.xmm64u(0), op2.xmm64u(0), status_word);
result.xmm64u(1) =
float32_div(op1.xmm64u(1), op2.xmm64u(1), status_word);
float64_div(op1.xmm64u(1), op2.xmm64u(1), status_word);
BX_CPU_THIS_PTR check_exceptionsSSE(status_word.float_exception_flags);
BX_WRITE_XMM_REG(i->nnn(), result);
@ -2091,6 +2316,13 @@ void BX_CPU_C::DIVSD_VsdWsd(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ())
{
op1 = handleDAZ(op1);
op2 = handleDAZ(op2);
}
result = float64_div(op1, op2, status_word);
BX_CPU_THIS_PTR check_exceptionsSSE(status_word.float_exception_flags);
BX_WRITE_XMM_REG_LO_QWORD(i->nnn(), result);
@ -2124,6 +2356,13 @@ void BX_CPU_C::DIVSS_VssWss(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ())
{
op1 = handleDAZ(op1);
op2 = handleDAZ(op2);
}
result = float32_div(op1, op2, status_word);
BX_CPU_THIS_PTR check_exceptionsSSE(status_word.float_exception_flags);
BX_WRITE_XMM_REG_LO_DWORD(i->nnn(), result);
@ -2159,6 +2398,18 @@ void BX_CPU_C::MAXPS_VpsWps(bxInstruction_c *i)
MXCSR_to_softfloat_status_word(status_word, MXCSR);
int rc;
if (MXCSR.get_DAZ()) {
op1.xmm32u(0) = handleDAZ(op1.xmm32u(0));
op1.xmm32u(1) = handleDAZ(op1.xmm32u(1));
op1.xmm32u(2) = handleDAZ(op1.xmm32u(2));
op1.xmm32u(3) = handleDAZ(op1.xmm32u(3));
op2.xmm32u(0) = handleDAZ(op2.xmm32u(0));
op2.xmm32u(1) = handleDAZ(op2.xmm32u(1));
op2.xmm32u(2) = handleDAZ(op2.xmm32u(2));
op2.xmm32u(3) = handleDAZ(op2.xmm32u(3));
}
rc = float32_compare(op1.xmm32u(0), op2.xmm32u(0), status_word);
result.xmm32u(0) =
(rc == float_relation_greater) ? op1.xmm32u(0) : op2.xmm32u(0);
@ -2206,6 +2457,15 @@ void BX_CPU_C::MAXPD_VpdWpd(bxInstruction_c *i)
MXCSR_to_softfloat_status_word(status_word, MXCSR);
int rc;
if (MXCSR.get_DAZ())
{
op1.xmm64u(0) = handleDAZ(op1.xmm64u(0));
op1.xmm64u(1) = handleDAZ(op1.xmm64u(1));
op2.xmm64u(0) = handleDAZ(op2.xmm64u(0));
op2.xmm64u(1) = handleDAZ(op2.xmm64u(1));
}
rc = float64_compare(op1.xmm64u(0), op2.xmm64u(0), status_word);
result.xmm64u(0) =
(rc == float_relation_greater) ? op1.xmm64u(0) : op2.xmm64u(0);
@ -2245,6 +2505,13 @@ void BX_CPU_C::MAXSD_VsdWsd(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ())
{
op1 = handleDAZ(op1);
op2 = handleDAZ(op2);
}
int rc = float64_compare(op1, op2, status_word);
BX_CPU_THIS_PTR check_exceptionsSSE(status_word.float_exception_flags);
BX_WRITE_XMM_REG_LO_QWORD(i->nnn(),
@ -2279,6 +2546,13 @@ void BX_CPU_C::MAXSS_VssWss(bxInstruction_c *i)
softfloat_status_word_t status_word;
MXCSR_to_softfloat_status_word(status_word, MXCSR);
if (MXCSR.get_DAZ())
{
op1 = handleDAZ(op1);
op2 = handleDAZ(op2);
}
int rc = float32_compare(op1, op2, status_word);
BX_CPU_THIS_PTR check_exceptionsSSE(status_word.float_exception_flags);
BX_WRITE_XMM_REG_LO_DWORD(i->nnn(),
@ -2315,6 +2589,18 @@ void BX_CPU_C::CMPPS_VpsWpsIb(bxInstruction_c *i)
MXCSR_to_softfloat_status_word(status, MXCSR);
int ib = i->Ib();
if (MXCSR.get_DAZ()) {
op1.xmm32u(0) = handleDAZ(op1.xmm32u(0));
op1.xmm32u(1) = handleDAZ(op1.xmm32u(1));
op1.xmm32u(2) = handleDAZ(op1.xmm32u(2));
op1.xmm32u(3) = handleDAZ(op1.xmm32u(3));
op2.xmm32u(0) = handleDAZ(op2.xmm32u(0));
op2.xmm32u(1) = handleDAZ(op2.xmm32u(1));
op2.xmm32u(2) = handleDAZ(op2.xmm32u(2));
op2.xmm32u(3) = handleDAZ(op2.xmm32u(3));
}
if(ib < 4)
{
result.xmm32u(0) =
@ -2377,6 +2663,15 @@ void BX_CPU_C::CMPPD_VpdWpdIb(bxInstruction_c *i)
MXCSR_to_softfloat_status_word(status, MXCSR);
int ib = i->Ib();
if (MXCSR.get_DAZ())
{
op1.xmm64u(0) = handleDAZ(op1.xmm64u(0));
op1.xmm64u(1) = handleDAZ(op1.xmm64u(1));
op2.xmm64u(0) = handleDAZ(op2.xmm64u(0));
op2.xmm64u(1) = handleDAZ(op2.xmm64u(1));
}
if(ib < 4)
{
result.xmm64u(0) = compare64[ib](op1.xmm64u(0), op2.xmm64u(0), status) ?
@ -2431,6 +2726,12 @@ void BX_CPU_C::CMPSD_VsdWsdIb(bxInstruction_c *i)
MXCSR_to_softfloat_status_word(status_word, MXCSR);
int ib = i->Ib();
if (MXCSR.get_DAZ())
{
op1 = handleDAZ(op1);
op2 = handleDAZ(op2);
}
if(ib < 4) {
if(compare64[ib](op1, op2, status_word)) {
result = BX_CONST64(0xFFFFFFFFFFFFFFFF);
@ -2481,6 +2782,12 @@ void BX_CPU_C::CMPSS_VssWssIb(bxInstruction_c *i)
MXCSR_to_softfloat_status_word(status_word, MXCSR);
int ib = i->Ib();
if (MXCSR.get_DAZ())
{
op1 = handleDAZ(op1);
op2 = handleDAZ(op2);
}
if(ib < 4) {
if(compare32[ib](op1, op2, status_word)) {
result = 0xFFFFFFFF;