ARM: Implement flush to zero for destinations as well.

2 files changed, 38 insertions, 29 deletions

diff --git a/src/arch/arm/insts/vfp.hh b/src/arch/arm/insts/vfp.hh
index b69f93598..e32aac721 100644
--- a/src/arch/arm/insts/vfp.hh
+++ b/src/arch/arm/insts/vfp.hh
@@ -175,16 +175,17 @@ bitsToFp(uint64_t bits, double junk)
 
 template <class fpType>
 static inline fpType
-fixNan(FPSCR fpscr, fpType val, fpType op1, fpType op2)
+fixDest(FPSCR fpscr, fpType val, fpType op1, fpType op2)
 {
-    if (std::isnan(val)) {
+    int fpClass = std::fpclassify(val);
+    fpType junk = 0.0;
+    if (fpClass == FP_NAN) {
         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) {
@@ -196,6 +197,11 @@ fixNan(FPSCR fpscr, fpType val, fpType op1, fpType op2)
         } else if (nan2) {
             val = op2;
         }
+    } else if (fpClass == FP_SUBNORMAL && fpscr.fz == 1) {
+        // Turn val into a zero with the correct sign;
+        uint64_t bitMask = ULL(0x1) << (sizeof(fpType) * 8 - 1);
+        val = bitsToFp(fpToBits(val) & bitMask, junk);
+        feraiseexcept(FeUnderflow);
     }
     return val;
 }
diff --git a/src/arch/arm/isa/insts/fp.isa b/src/arch/arm/isa/insts/fp.isa
index 6cd3c1345..4f693f46a 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 = fixNan(Fpscr, FpOp1 * FpOp2, FpOp1, FpOp2);
+        FpDest = fixDest(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 = fixNan(Fpscr, cOp1.fp * cOp2.fp, cOp1.fp, cOp2.fp);
+        cDest.fp = fixDest(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 = fixNan(Fpscr, FpOp1 + FpOp2, FpOp1, FpOp2);
+        FpDest = fixDest(Fpscr, FpOp1 + FpOp2, FpOp1, FpOp2);
         __asm__ __volatile__("" :: "m" (FpDest));
         Fpscr = setVfpFpscr(Fpscr, state);
     '''
@@ -491,10 +491,13 @@ let {{
         IntDoubleUnion cOp1, cOp2, cDest;
         cOp1.bits = ((uint64_t)FpOp1P0.uw | ((uint64_t)FpOp1P1.uw << 32));
         cOp2.bits = ((uint64_t)FpOp2P0.uw | ((uint64_t)FpOp2P1.uw << 32));
+        DPRINTFN("cOp1.bits = %#x, cOp1.fp = %f.\\n", cOp1.bits, cOp1.fp);
+        DPRINTFN("cOp2.bits = %#x, cOp2.fp = %f.\\n", cOp2.bits, cOp2.fp);
         vfpFlushToZero(Fpscr, cOp1.fp, cOp2.fp);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (cOp1.fp) : "m" (cOp1.fp));
-        cDest.fp = fixNan(Fpscr, cOp1.fp + cOp2.fp, cOp1.fp, cOp2.fp);
+        cDest.fp = fixDest(Fpscr, cOp1.fp + cOp2.fp, cOp1.fp, cOp2.fp);
+        DPRINTFN("cDest.bits = %#x, cDest.fp = %f.\\n", cDest.bits, cDest.fp);
         __asm__ __volatile__("" :: "m" (cDest.fp));
         Fpscr = setVfpFpscr(Fpscr, state);
         FpDestP0.uw = cDest.bits;
@@ -511,7 +514,7 @@ let {{
         vfpFlushToZero(Fpscr, FpOp1, FpOp2);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (FpOp1) : "m" (FpOp1));
-        FpDest = fixNan(Fpscr, FpOp1 - FpOp2, FpOp1, FpOp2);
+        FpDest = fixDest(Fpscr, FpOp1 - FpOp2, FpOp1, FpOp2);
         __asm__ __volatile__("" :: "m" (FpDest));
         Fpscr = setVfpFpscr(Fpscr, state)
     '''
@@ -529,7 +532,7 @@ let {{
         vfpFlushToZero(Fpscr, cOp1.fp, cOp2.fp);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (cOp1.fp) : "m" (cOp1.fp));
-        cDest.fp = fixNan(Fpscr, cOp1.fp - cOp2.fp, cOp1.fp, cOp2.fp);
+        cDest.fp = fixDest(Fpscr, cOp1.fp - cOp2.fp, cOp1.fp, cOp2.fp);
         __asm__ __volatile__("" :: "m" (cDest.fp));
         Fpscr = setVfpFpscr(Fpscr, state);
         FpDestP0.uw = cDest.bits;
@@ -546,7 +549,7 @@ let {{
         vfpFlushToZero(Fpscr, FpOp1, FpOp2);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (FpOp1) : "m" (FpOp1));
-        FpDest = fixNan(Fpscr, FpOp1 / FpOp2, FpOp1, FpOp2);
+        FpDest = fixDest(Fpscr, FpOp1 / FpOp2, FpOp1, FpOp2);
         __asm__ __volatile__("" :: "m" (FpDest));
         Fpscr = setVfpFpscr(Fpscr, state);
     '''
@@ -564,7 +567,7 @@ let {{
         vfpFlushToZero(Fpscr, cOp1.fp, cOp2.fp);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (cOp1.fp) : "m" (cDest.fp));
-        cDest.fp = fixNan(Fpscr, cOp1.fp / cOp2.fp, cOp1.fp, cOp2.fp);
+        cDest.fp = fixDest(Fpscr, cOp1.fp / cOp2.fp, cOp1.fp, cOp2.fp);
         __asm__ __volatile__("" :: "m" (cDest.fp));
         Fpscr = setVfpFpscr(Fpscr, state);
         FpDestP0.uw = cDest.bits;
@@ -628,12 +631,12 @@ let {{
         vfpFlushToZero(Fpscr, FpOp1, FpOp2);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (FpOp1) : "m" (FpOp1));
-        float mid = fixNan(Fpscr, FpOp1 * FpOp2, FpOp1, FpOp2);
+        float mid = fixDest(Fpscr, FpOp1 * FpOp2, FpOp1, FpOp2);
         if ((isinf(FpOp1) && FpOp2 == 0) || (isinf(FpOp2) && FpOp1 == 0)) {
             mid = NAN;
         }
         vfpFlushToZero(Fpscr, FpDest, mid);
-        FpDest = fixNan(Fpscr, FpDest + mid, FpDest, mid);
+        FpDest = fixDest(Fpscr, FpDest + mid, FpDest, mid);
         __asm__ __volatile__("" :: "m" (FpDest));
         Fpscr = setVfpFpscr(Fpscr, state);
     '''
@@ -652,13 +655,13 @@ let {{
         vfpFlushToZero(Fpscr, cOp1.fp, cOp2.fp);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (cOp1.fp) : "m" (cOp1.fp));
-        double mid = fixNan(Fpscr, cOp1.fp * cOp2.fp, cOp1.fp, cOp2.fp);
+        double mid = fixDest(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 = fixNan(Fpscr, cDest.fp + mid, cDest.fp, mid);
+        cDest.fp = fixDest(Fpscr, cDest.fp + mid, cDest.fp, mid);
         __asm__ __volatile__("" :: "m" (cDest.fp));
         Fpscr = setVfpFpscr(Fpscr, state);
         FpDestP0.uw = cDest.bits;
@@ -675,12 +678,12 @@ let {{
         vfpFlushToZero(Fpscr, FpOp1, FpOp2);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (FpOp1) : "m" (FpOp1));
-        float mid = fixNan(Fpscr, FpOp1 * FpOp2, FpOp1, FpOp2);
+        float mid = fixDest(Fpscr, FpOp1 * FpOp2, FpOp1, FpOp2);
         if ((isinf(FpOp1) && FpOp2 == 0) || (isinf(FpOp2) && FpOp1 == 0)) {
             mid = NAN;
         }
         vfpFlushToZero(Fpscr, FpDest, mid);
-        FpDest = fixNan(Fpscr, FpDest - mid, FpDest, mid);
+        FpDest = fixDest(Fpscr, FpDest - mid, FpDest, mid);
         __asm__ __volatile__("" :: "m" (FpDest));
         Fpscr = setVfpFpscr(Fpscr, state);
     '''
@@ -699,12 +702,12 @@ let {{
         vfpFlushToZero(Fpscr, cOp1.fp, cOp2.fp);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (cOp1.fp) : "m" (cOp1.fp));
-        double mid = fixNan(Fpscr, cOp1.fp * cOp2.fp, cOp1.fp, cOp2.fp);
+        double mid = fixDest(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 = fixNan(Fpscr, cDest.fp - mid, cDest.fp, mid);
+        cDest.fp = fixDest(Fpscr, cDest.fp - mid, cDest.fp, mid);
         vfpFlushToZero(Fpscr, cDest.fp, mid);
         __asm__ __volatile__("" :: "m" (cDest.fp));
         Fpscr = setVfpFpscr(Fpscr, state);
@@ -722,12 +725,12 @@ let {{
         vfpFlushToZero(Fpscr, FpOp1, FpOp2);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (FpOp1) : "m" (FpOp1));
-        float mid = fixNan(Fpscr, FpOp1 * FpOp2, FpOp1, FpOp2);
+        float mid = fixDest(Fpscr, FpOp1 * FpOp2, FpOp1, FpOp2);
         if ((isinf(FpOp1) && FpOp2 == 0) || (isinf(FpOp2) && FpOp1 == 0)) {
             mid = NAN;
         }
         vfpFlushToZero(Fpscr, FpDest, mid);
-        FpDest = fixNan(Fpscr, -FpDest - mid, FpDest, mid);
+        FpDest = fixDest(Fpscr, -FpDest - mid, FpDest, mid);
         __asm__ __volatile__("" :: "m" (FpDest));
         Fpscr = setVfpFpscr(Fpscr, state);
     '''
@@ -746,13 +749,13 @@ let {{
         vfpFlushToZero(Fpscr, cOp1.fp, cOp2.fp);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (cOp1.fp) : "m" (cOp1.fp));
-        double mid = fixNan(Fpscr, cOp1.fp * cOp2.fp, cOp1.fp, cOp2.fp);
+        double mid = fixDest(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 = fixNan(Fpscr, -cDest.fp - mid, cDest.fp, mid);
+        cDest.fp = fixDest(Fpscr, -cDest.fp - mid, cDest.fp, mid);
         __asm__ __volatile__("" :: "m" (cDest.fp));
         Fpscr = setVfpFpscr(Fpscr, state);
         FpDestP0.uw = cDest.bits;
@@ -769,12 +772,12 @@ let {{
         vfpFlushToZero(Fpscr, FpOp1, FpOp2);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (FpOp1) : "m" (FpOp1));
-        float mid = fixNan(Fpscr, FpOp1 * FpOp2, FpOp1, FpOp2);
+        float mid = fixDest(Fpscr, FpOp1 * FpOp2, FpOp1, FpOp2);
         if ((isinf(FpOp1) && FpOp2 == 0) || (isinf(FpOp2) && FpOp1 == 0)) {
             mid = NAN;
         }
         vfpFlushToZero(Fpscr, FpDest, mid);
-        FpDest = fixNan(Fpscr, -FpDest + mid, FpDest, mid);
+        FpDest = fixDest(Fpscr, -FpDest + mid, FpDest, mid);
         __asm__ __volatile__("" :: "m" (FpDest));
         Fpscr = setVfpFpscr(Fpscr, state);
     '''
@@ -793,13 +796,13 @@ let {{
         vfpFlushToZero(Fpscr, cOp1.fp, cOp2.fp);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (cOp1.fp) : "m" (cOp1.fp));
-        double mid = fixNan(Fpscr, cOp1.fp * cOp2.fp, cOp1.fp, cOp2.fp);
+        double mid = fixDest(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 = fixNan(Fpscr, -cDest.fp + mid, cDest.fp, mid);
+        cDest.fp = fixDest(Fpscr, -cDest.fp + mid, cDest.fp, mid);
         __asm__ __volatile__("" :: "m" (cDest.fp));
         Fpscr = setVfpFpscr(Fpscr, state);
         FpDestP0.uw = cDest.bits;
@@ -816,7 +819,7 @@ let {{
         vfpFlushToZero(Fpscr, FpOp1, FpOp2);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (FpOp1) : "m" (FpOp1));
-        float mid = fixNan(Fpscr, FpOp1 * FpOp2, FpOp1, FpOp2);
+        float mid = fixDest(Fpscr, FpOp1 * FpOp2, FpOp1, FpOp2);
         if ((isinf(FpOp1) && FpOp2 == 0) || (isinf(FpOp2) && FpOp1 == 0)) {
             mid = NAN;
         }
@@ -839,7 +842,7 @@ let {{
         vfpFlushToZero(Fpscr, cOp1.fp, cOp2.fp);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (cOp1.fp) : "m" (cOp1.fp));
-        double mid = fixNan(Fpscr, cOp1.fp * cOp2.fp, cOp1.fp, cOp2.fp);
+        double mid = fixDest(Fpscr, cOp1.fp * cOp2.fp, cOp1.fp, cOp2.fp);
         if ((isinf(cOp1.fp) && cOp2.fp == 0) ||
                 (isinf(cOp2.fp) && cOp1.fp == 0)) {
             mid = NAN;