ARM: Fix up nans to match ARM's expected behavior.

2 files changed, 106 insertions, 26 deletions

diff --git a/src/arch/arm/insts/vfp.hh b/src/arch/arm/insts/vfp.hh
index b0fc8b6dc..b69f93598 100644
--- a/src/arch/arm/insts/vfp.hh
+++ b/src/arch/arm/insts/vfp.hh
@@ -45,6 +45,9 @@
 #include <fenv.h>
 #include <cmath>
 
+namespace ArmISA
+{
+
 enum VfpMicroMode {
     VfpNotAMicroop,
     VfpMicroop,
@@ -122,6 +125,81 @@ vfpFlushToZero(uint32_t &fpscr, fpType &op1, fpType &op2)
     vfpFlushToZero(fpscr, op2);
 }
 
+static inline uint32_t
+fpToBits(float fp)
+{
+    union
+    {
+        float fp;
+        uint32_t bits;
+    } val;
+    val.fp = fp;
+    return val.bits;
+}
+
+static inline uint64_t
+fpToBits(double fp)
+{
+    union
+    {
+        double fp;
+        uint64_t bits;
+    } val;
+    val.fp = fp;
+    return val.bits;
+}
+
+static inline float
+bitsToFp(uint64_t bits, float junk)
+{
+    union
+    {
+        float fp;
+        uint32_t bits;
+    } val;
+    val.bits = bits;
+    return val.fp;
+}
+
+static inline double
+bitsToFp(uint64_t bits, double junk)
+{
+    union
+    {
+        double fp;
+        uint64_t bits;
+    } val;
+    val.bits = bits;
+    return val.fp;
+}
+
+template <class fpType>
+static inline fpType
+fixNan(FPSCR fpscr, fpType val, fpType op1, fpType op2)
+{
+    if (std::isnan(val)) {
+        const bool single = (sizeof(val) == sizeof(float));
+        const uint64_t qnan = single ? 0x7fc00000 : ULL(0x7ff8000000000000);
+        const bool nan1 = std::isnan(op1);
+        const bool nan2 = std::isnan(op2);
+        const bool signal1 = nan1 && ((fpToBits(op1) & qnan) != qnan);
+        const bool signal2 = nan2 && ((fpToBits(op2) & qnan) != qnan);
+        fpType junk = 0.0;
+        if ((!nan1 && !nan2) || (fpscr.dn == 1)) {
+            val = bitsToFp(qnan, junk);
+        } else if (signal1) {
+            val = bitsToFp(fpToBits(op1) | qnan, junk);
+        } else if (signal2) {
+            val = bitsToFp(fpToBits(op2) | qnan, junk);
+        } else if (nan1) {
+            val = op1;
+        } else if (nan2) {
+            val = op2;
+        }
+    }
+    return val;
+}
+
 static inline uint64_t
 vfpFpSToFixed(float val, bool isSigned, bool half, uint8_t imm)
 {
@@ -481,4 +559,6 @@ class VfpRegRegRegOp : public RegRegRegOp
     }
 };
 
+}
+
 #endif //__ARCH_ARM_INSTS_VFP_HH__
diff --git a/src/arch/arm/isa/insts/fp.isa b/src/arch/arm/isa/insts/fp.isa
index cd1ddc498..6cd3c1345 100644
--- a/src/arch/arm/isa/insts/fp.isa
+++ b/src/arch/arm/isa/insts/fp.isa
@@ -386,7 +386,7 @@ let {{
         vfpFlushToZero(Fpscr, FpOp1, FpOp2);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (FpOp1) : "m" (FpOp1));
-        FpDest = FpOp1 * FpOp2;
+        FpDest = fixNan(Fpscr, FpOp1 * FpOp2, FpOp1, FpOp2);
         __asm__ __volatile__("" :: "m" (FpDest));
         Fpscr = setVfpFpscr(Fpscr, state);
         if ((isinf(FpOp1) && FpOp2 == 0) || (isinf(FpOp2) && FpOp1 == 0)) {
@@ -407,7 +407,7 @@ let {{
         vfpFlushToZero(Fpscr, cOp1.fp, cOp2.fp);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (cOp1.fp) : "m" (cOp1.fp));
-        cDest.fp = cOp1.fp * cOp2.fp;
+        cDest.fp = fixNan(Fpscr, cOp1.fp * cOp2.fp, cOp1.fp, cOp2.fp);
         __asm__ __volatile__("" :: "m" (cDest.fp));
         Fpscr = setVfpFpscr(Fpscr, state);
         if ((isinf(cOp1.fp) && cOp2.fp == 0) ||
@@ -476,7 +476,7 @@ let {{
         vfpFlushToZero(Fpscr, FpOp1, FpOp2);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (FpOp1) : "m" (FpOp1));
-        FpDest = FpOp1 + FpOp2;
+        FpDest = fixNan(Fpscr, FpOp1 + FpOp2, FpOp1, FpOp2);
         __asm__ __volatile__("" :: "m" (FpDest));
         Fpscr = setVfpFpscr(Fpscr, state);
     '''
@@ -494,7 +494,7 @@ let {{
         vfpFlushToZero(Fpscr, cOp1.fp, cOp2.fp);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (cOp1.fp) : "m" (cOp1.fp));
-        cDest.fp = cOp1.fp + cOp2.fp;
+        cDest.fp = fixNan(Fpscr, cOp1.fp + cOp2.fp, cOp1.fp, cOp2.fp);
         __asm__ __volatile__("" :: "m" (cDest.fp));
         Fpscr = setVfpFpscr(Fpscr, state);
         FpDestP0.uw = cDest.bits;
@@ -511,7 +511,7 @@ let {{
         vfpFlushToZero(Fpscr, FpOp1, FpOp2);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (FpOp1) : "m" (FpOp1));
-        FpDest = FpOp1 - FpOp2;
+        FpDest = fixNan(Fpscr, FpOp1 - FpOp2, FpOp1, FpOp2);
         __asm__ __volatile__("" :: "m" (FpDest));
         Fpscr = setVfpFpscr(Fpscr, state)
     '''
@@ -529,7 +529,7 @@ let {{
         vfpFlushToZero(Fpscr, cOp1.fp, cOp2.fp);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (cOp1.fp) : "m" (cOp1.fp));
-        cDest.fp = cOp1.fp - cOp2.fp;
+        cDest.fp = fixNan(Fpscr, cOp1.fp - cOp2.fp, cOp1.fp, cOp2.fp);
         __asm__ __volatile__("" :: "m" (cDest.fp));
         Fpscr = setVfpFpscr(Fpscr, state);
         FpDestP0.uw = cDest.bits;
@@ -546,7 +546,7 @@ let {{
         vfpFlushToZero(Fpscr, FpOp1, FpOp2);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (FpOp1) : "m" (FpOp1));
-        FpDest = FpOp1 / FpOp2;
+        FpDest = fixNan(Fpscr, FpOp1 / FpOp2, FpOp1, FpOp2);
         __asm__ __volatile__("" :: "m" (FpDest));
         Fpscr = setVfpFpscr(Fpscr, state);
     '''
@@ -564,7 +564,7 @@ let {{
         vfpFlushToZero(Fpscr, cOp1.fp, cOp2.fp);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (cOp1.fp) : "m" (cDest.fp));
-        cDest.fp = cOp1.fp / cOp2.fp;
+        cDest.fp = fixNan(Fpscr, cOp1.fp / cOp2.fp, cOp1.fp, cOp2.fp);
         __asm__ __volatile__("" :: "m" (cDest.fp));
         Fpscr = setVfpFpscr(Fpscr, state);
         FpDestP0.uw = cDest.bits;
@@ -628,12 +628,12 @@ let {{
         vfpFlushToZero(Fpscr, FpOp1, FpOp2);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (FpOp1) : "m" (FpOp1));
-        float mid = FpOp1 * FpOp2;
+        float mid = fixNan(Fpscr, FpOp1 * FpOp2, FpOp1, FpOp2);
         if ((isinf(FpOp1) && FpOp2 == 0) || (isinf(FpOp2) && FpOp1 == 0)) {
             mid = NAN;
         }
         vfpFlushToZero(Fpscr, FpDest, mid);
-        FpDest = FpDest + mid;
+        FpDest = fixNan(Fpscr, FpDest + mid, FpDest, mid);
         __asm__ __volatile__("" :: "m" (FpDest));
         Fpscr = setVfpFpscr(Fpscr, state);
     '''
@@ -652,13 +652,13 @@ let {{
         vfpFlushToZero(Fpscr, cOp1.fp, cOp2.fp);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (cOp1.fp) : "m" (cOp1.fp));
-        double mid = cOp1.fp * cOp2.fp;
+        double mid = fixNan(Fpscr, cOp1.fp * cOp2.fp, cOp1.fp, cOp2.fp);
         if ((isinf(cOp1.fp) && cOp2.fp == 0) ||
                 (isinf(cOp2.fp) && cOp1.fp == 0)) {
             mid = NAN;
         }
         vfpFlushToZero(Fpscr, cDest.fp, mid);
-        cDest.fp = cDest.fp + mid;
+        cDest.fp = fixNan(Fpscr, cDest.fp + mid, cDest.fp, mid);
         __asm__ __volatile__("" :: "m" (cDest.fp));
         Fpscr = setVfpFpscr(Fpscr, state);
         FpDestP0.uw = cDest.bits;
@@ -675,12 +675,12 @@ let {{
         vfpFlushToZero(Fpscr, FpOp1, FpOp2);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (FpOp1) : "m" (FpOp1));
-        float mid = FpOp1 * FpOp2;
+        float mid = fixNan(Fpscr, FpOp1 * FpOp2, FpOp1, FpOp2);
         if ((isinf(FpOp1) && FpOp2 == 0) || (isinf(FpOp2) && FpOp1 == 0)) {
             mid = NAN;
         }
         vfpFlushToZero(Fpscr, FpDest, mid);
-        FpDest = FpDest - mid;
+        FpDest = fixNan(Fpscr, FpDest - mid, FpDest, mid);
         __asm__ __volatile__("" :: "m" (FpDest));
         Fpscr = setVfpFpscr(Fpscr, state);
     '''
@@ -699,12 +699,12 @@ let {{
         vfpFlushToZero(Fpscr, cOp1.fp, cOp2.fp);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (cOp1.fp) : "m" (cOp1.fp));
-        double mid = cOp1.fp * cOp2.fp;
+        double mid = fixNan(Fpscr, cOp1.fp * cOp2.fp, cOp1.fp, cOp2.fp);
         if ((isinf(cOp1.fp) && cOp2.fp == 0) ||
                 (isinf(cOp2.fp) && cOp1.fp == 0)) {
             mid = NAN;
         }
-        cDest.fp = cDest.fp - mid;
+        cDest.fp = fixNan(Fpscr, cDest.fp - mid, cDest.fp, mid);
         vfpFlushToZero(Fpscr, cDest.fp, mid);
         __asm__ __volatile__("" :: "m" (cDest.fp));
         Fpscr = setVfpFpscr(Fpscr, state);
@@ -722,12 +722,12 @@ let {{
         vfpFlushToZero(Fpscr, FpOp1, FpOp2);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (FpOp1) : "m" (FpOp1));
-        float mid = FpOp1 * FpOp2;
+        float mid = fixNan(Fpscr, FpOp1 * FpOp2, FpOp1, FpOp2);
         if ((isinf(FpOp1) && FpOp2 == 0) || (isinf(FpOp2) && FpOp1 == 0)) {
             mid = NAN;
         }
         vfpFlushToZero(Fpscr, FpDest, mid);
-        FpDest = -FpDest - mid;
+        FpDest = fixNan(Fpscr, -FpDest - mid, FpDest, mid);
         __asm__ __volatile__("" :: "m" (FpDest));
         Fpscr = setVfpFpscr(Fpscr, state);
     '''
@@ -746,13 +746,13 @@ let {{
         vfpFlushToZero(Fpscr, cOp1.fp, cOp2.fp);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (cOp1.fp) : "m" (cOp1.fp));
-        double mid = cOp1.fp * cOp2.fp;
+        double mid = fixNan(Fpscr, cOp1.fp * cOp2.fp, cOp1.fp, cOp2.fp);
         if ((isinf(cOp1.fp) && cOp2.fp == 0) ||
                 (isinf(cOp2.fp) && cOp1.fp == 0)) {
             mid = NAN;
         }
         vfpFlushToZero(Fpscr, cDest.fp, mid);
-        cDest.fp = -cDest.fp - mid;
+        cDest.fp = fixNan(Fpscr, -cDest.fp - mid, cDest.fp, mid);
         __asm__ __volatile__("" :: "m" (cDest.fp));
         Fpscr = setVfpFpscr(Fpscr, state);
         FpDestP0.uw = cDest.bits;
@@ -769,12 +769,12 @@ let {{
         vfpFlushToZero(Fpscr, FpOp1, FpOp2);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (FpOp1) : "m" (FpOp1));
-        float mid = FpOp1 * FpOp2;
+        float mid = fixNan(Fpscr, FpOp1 * FpOp2, FpOp1, FpOp2);
         if ((isinf(FpOp1) && FpOp2 == 0) || (isinf(FpOp2) && FpOp1 == 0)) {
             mid = NAN;
         }
         vfpFlushToZero(Fpscr, FpDest, mid);
-        FpDest = -FpDest + mid;
+        FpDest = fixNan(Fpscr, -FpDest + mid, FpDest, mid);
         __asm__ __volatile__("" :: "m" (FpDest));
         Fpscr = setVfpFpscr(Fpscr, state);
     '''
@@ -793,13 +793,13 @@ let {{
         vfpFlushToZero(Fpscr, cOp1.fp, cOp2.fp);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (cOp1.fp) : "m" (cOp1.fp));
-        double mid = cOp1.fp * cOp2.fp;
+        double mid = fixNan(Fpscr, cOp1.fp * cOp2.fp, cOp1.fp, cOp2.fp);
         if ((isinf(cOp1.fp) && cOp2.fp == 0) ||
                 (isinf(cOp2.fp) && cOp1.fp == 0)) {
             mid = NAN;
         }
         vfpFlushToZero(Fpscr, cDest.fp, mid);
-        cDest.fp = -cDest.fp + mid;
+        cDest.fp = fixNan(Fpscr, -cDest.fp + mid, cDest.fp, mid);
         __asm__ __volatile__("" :: "m" (cDest.fp));
         Fpscr = setVfpFpscr(Fpscr, state);
         FpDestP0.uw = cDest.bits;
@@ -816,7 +816,7 @@ let {{
         vfpFlushToZero(Fpscr, FpOp1, FpOp2);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (FpOp1) : "m" (FpOp1));
-        float mid = FpOp1 * FpOp2;
+        float mid = fixNan(Fpscr, FpOp1 * FpOp2, FpOp1, FpOp2);
         if ((isinf(FpOp1) && FpOp2 == 0) || (isinf(FpOp2) && FpOp1 == 0)) {
             mid = NAN;
         }
@@ -839,7 +839,7 @@ let {{
         vfpFlushToZero(Fpscr, cOp1.fp, cOp2.fp);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (cOp1.fp) : "m" (cOp1.fp));
-        double mid = cOp1.fp * cOp2.fp;
+        double mid = fixNan(Fpscr, cOp1.fp * cOp2.fp, cOp1.fp, cOp2.fp);
         if ((isinf(cOp1.fp) && cOp2.fp == 0) ||
                 (isinf(cOp2.fp) && cOp1.fp == 0)) {
             mid = NAN;