ARM: Compensate for ARM's underflow coming from -before- rounding, but x86's after.

2 files changed, 127 insertions, 25 deletions

diff --git a/src/arch/arm/insts/vfp.hh b/src/arch/arm/insts/vfp.hh
index e32aac721..11ae8ed96 100644
--- a/src/arch/arm/insts/vfp.hh
+++ b/src/arch/arm/insts/vfp.hh
@@ -175,6 +175,30 @@ bitsToFp(uint64_t bits, double junk)
 
 template <class fpType>
 static inline fpType
+fixDest(FPSCR fpscr, fpType val, fpType op1)
+{
+    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 nan = std::isnan(op1);
+        if (!nan || (fpscr.dn == 1)) {
+            val = bitsToFp(qnan, junk);
+        } else if (nan) {
+            val = bitsToFp(fpToBits(op1) | qnan, junk);
+        }
+    } 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;
+}
+
+template <class fpType>
+static inline fpType
 fixDest(FPSCR fpscr, fpType val, fpType op1, fpType op2)
 {
     int fpClass = std::fpclassify(val);
@@ -206,6 +230,84 @@ fixDest(FPSCR fpscr, fpType val, fpType op1, fpType op2)
     return val;
 }
 
+template <class fpType>
+static inline fpType
+fixMultDest(FPSCR fpscr, fpType val, fpType op1, fpType op2)
+{
+    fpType mid = fixDest(fpscr, val, op1, op2);
+    const bool single = (sizeof(fpType) == sizeof(float));
+    const fpType junk = 0.0;
+    if ((single && (val == bitsToFp(0x00800000, junk) ||
+                    val == bitsToFp(0x80800000, junk))) ||
+        (!single && (val == bitsToFp(ULL(0x0010000000000000), junk) ||
+                     val == bitsToFp(ULL(0x8010000000000000), junk)))
+        ) {
+        __asm__ __volatile__("" : "=m" (op1) : "m" (op1));
+        fesetround(FeRoundZero);
+        fpType temp = 0.0;
+        __asm__ __volatile__("" : "=m" (temp) : "m" (temp));
+        temp = op1 * op2;
+        if (!std::isnormal(temp)) {
+            feraiseexcept(FeUnderflow);
+        }
+        __asm__ __volatile__("" :: "m" (temp));
+    }
+    return mid;
+}
+
+template <class fpType>
+static inline fpType
+fixDivDest(FPSCR fpscr, fpType val, fpType op1, fpType op2)
+{
+    fpType mid = fixDest(fpscr, val, op1, op2);
+    const bool single = (sizeof(fpType) == sizeof(float));
+    const fpType junk = 0.0;
+    if ((single && (val == bitsToFp(0x00800000, junk) ||
+                    val == bitsToFp(0x80800000, junk))) ||
+        (!single && (val == bitsToFp(ULL(0x0010000000000000), junk) ||
+                     val == bitsToFp(ULL(0x8010000000000000), junk)))
+        ) {
+        __asm__ __volatile__("" : "=m" (op1) : "m" (op1));
+        fesetround(FeRoundZero);
+        fpType temp = 0.0;
+        __asm__ __volatile__("" : "=m" (temp) : "m" (temp));
+        temp = op1 / op2;
+        if (!std::isnormal(temp)) {
+            feraiseexcept(FeUnderflow);
+        }
+        __asm__ __volatile__("" :: "m" (temp));
+    }
+    return mid;
+}
+
+static inline float
+fixFpDFpSDest(FPSCR fpscr, double val)
+{
+    const float junk = 0.0;
+    float op1 = 0.0;
+    if (std::isnan(val)) {
+        uint64_t valBits = fpToBits(val);
+        uint32_t op1Bits = bits(valBits, 50, 29) |
+                           (mask(9) << 22) |
+                           (bits(valBits, 63) << 31);
+        op1 = bitsToFp(op1Bits, junk);
+    }
+    float mid = fixDest(fpscr, (float)val, op1);
+    if (mid == bitsToFp(0x00800000, junk) ||
+        mid == bitsToFp(0x80800000, junk)) {
+        __asm__ __volatile__("" : "=m" (val) : "m" (val));
+        fesetround(FeRoundZero);
+        float temp = 0.0;
+        __asm__ __volatile__("" : "=m" (temp) : "m" (temp));
+        temp = val;
+        if (!std::isnormal(temp)) {
+            feraiseexcept(FeUnderflow);
+        }
+        __asm__ __volatile__("" :: "m" (temp));
+    }
+    return mid;
+}
+
 static inline uint64_t
 vfpFpSToFixed(float val, bool isSigned, bool half, uint8_t imm)
 {
@@ -282,7 +384,7 @@ vfpFpSToFixed(float val, bool isSigned, bool half, uint8_t imm)
 }
 
 static inline float
-vfpUFixedToFpS(uint32_t val, bool half, uint8_t imm)
+vfpUFixedToFpS(FPSCR fpscr, uint32_t val, bool half, uint8_t imm)
 {
     fesetround(FeRoundNearest);
     if (half)
@@ -291,11 +393,11 @@ vfpUFixedToFpS(uint32_t val, bool half, uint8_t imm)
     __asm__ __volatile__("" : "=m" (scale) : "m" (scale));
     feclearexcept(FeAllExceptions);
     __asm__ __volatile__("" : "=m" (scale) : "m" (scale));
-    return val / scale;
+    return fixDivDest(fpscr, val / scale, (float)val, scale);
 }
 
 static inline float
-vfpSFixedToFpS(int32_t val, bool half, uint8_t imm)
+vfpSFixedToFpS(FPSCR fpscr, int32_t val, bool half, uint8_t imm)
 {
     fesetround(FeRoundNearest);
     if (half)
@@ -304,7 +406,7 @@ vfpSFixedToFpS(int32_t val, bool half, uint8_t imm)
     __asm__ __volatile__("" : "=m" (scale) : "m" (scale));
     feclearexcept(FeAllExceptions);
     __asm__ __volatile__("" : "=m" (scale) : "m" (scale));
-    return val / scale;
+    return fixDivDest(fpscr, val / scale, (float)val, scale);
 }
 
 static inline uint64_t
@@ -383,7 +485,7 @@ vfpFpDToFixed(double val, bool isSigned, bool half, uint8_t imm)
 }
 
 static inline double
-vfpUFixedToFpD(uint32_t val, bool half, uint8_t imm)
+vfpUFixedToFpD(FPSCR fpscr, uint32_t val, bool half, uint8_t imm)
 {
     fesetround(FeRoundNearest);
     if (half)
@@ -392,11 +494,11 @@ vfpUFixedToFpD(uint32_t val, bool half, uint8_t imm)
     __asm__ __volatile__("" : "=m" (scale) : "m" (scale));
     feclearexcept(FeAllExceptions);
     __asm__ __volatile__("" : "=m" (scale) : "m" (scale));
-    return val / scale;
+    return fixDivDest(fpscr, val / scale, (double)val, scale);
 }
 
 static inline double
-vfpSFixedToFpD(int32_t val, bool half, uint8_t imm)
+vfpSFixedToFpD(FPSCR fpscr, int32_t val, bool half, uint8_t imm)
 {
     fesetround(FeRoundNearest);
     if (half)
@@ -405,7 +507,7 @@ vfpSFixedToFpD(int32_t val, bool half, uint8_t imm)
     __asm__ __volatile__("" : "=m" (scale) : "m" (scale));
     feclearexcept(FeAllExceptions);
     __asm__ __volatile__("" : "=m" (scale) : "m" (scale));
-    return val / scale;
+    return fixDivDest(fpscr, val / scale, (double)val, scale);
 }
 
 typedef int VfpSavedState;
diff --git a/src/arch/arm/isa/insts/fp.isa b/src/arch/arm/isa/insts/fp.isa
index 4f693f46a..026e332d4 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 = fixDest(Fpscr, FpOp1 * FpOp2, FpOp1, FpOp2);
+        FpDest = fixMultDest(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 = fixDest(Fpscr, cOp1.fp * cOp2.fp, cOp1.fp, cOp2.fp);
+        cDest.fp = fixMultDest(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) ||
@@ -683,7 +683,7 @@ let {{
             mid = NAN;
         }
         vfpFlushToZero(Fpscr, FpDest, mid);
-        FpDest = fixDest(Fpscr, FpDest - mid, FpDest, mid);
+        FpDest = fixDest(Fpscr, FpDest - mid, FpDest, -mid);
         __asm__ __volatile__("" :: "m" (FpDest));
         Fpscr = setVfpFpscr(Fpscr, state);
     '''
@@ -707,7 +707,7 @@ let {{
                 (isinf(cOp2.fp) && cOp1.fp == 0)) {
             mid = NAN;
         }
-        cDest.fp = fixDest(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);
@@ -730,7 +730,7 @@ let {{
             mid = NAN;
         }
         vfpFlushToZero(Fpscr, FpDest, mid);
-        FpDest = fixDest(Fpscr, -FpDest - mid, FpDest, mid);
+        FpDest = fixDest(Fpscr, -FpDest - mid, -FpDest, -mid);
         __asm__ __volatile__("" :: "m" (FpDest));
         Fpscr = setVfpFpscr(Fpscr, state);
     '''
@@ -755,7 +755,7 @@ let {{
             mid = NAN;
         }
         vfpFlushToZero(Fpscr, cDest.fp, mid);
-        cDest.fp = fixDest(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;
@@ -777,7 +777,7 @@ let {{
             mid = NAN;
         }
         vfpFlushToZero(Fpscr, FpDest, mid);
-        FpDest = fixDest(Fpscr, -FpDest + mid, FpDest, mid);
+        FpDest = fixDest(Fpscr, -FpDest + mid, -FpDest, mid);
         __asm__ __volatile__("" :: "m" (FpDest));
         Fpscr = setVfpFpscr(Fpscr, state);
     '''
@@ -802,7 +802,7 @@ let {{
             mid = NAN;
         }
         vfpFlushToZero(Fpscr, cDest.fp, mid);
-        cDest.fp = fixDest(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;
@@ -1089,7 +1089,7 @@ let {{
         vfpFlushToZero(Fpscr, cOp1.fp);
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (cOp1.fp) : "m" (cOp1.fp));
-        FpDest = cOp1.fp;
+        FpDest = fixFpDFpSDest(Fpscr, cOp1.fp);
         __asm__ __volatile__("" :: "m" (FpDest));
         Fpscr = setVfpFpscr(Fpscr, state);
     '''
@@ -1267,7 +1267,7 @@ let {{
     vcvtSFixedFpSCode = '''
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (FpOp1.sw) : "m" (FpOp1.sw));
-        FpDest = vfpSFixedToFpS(FpOp1.sw, false, imm);
+        FpDest = vfpSFixedToFpS(Fpscr, FpOp1.sw, false, imm);
         __asm__ __volatile__("" :: "m" (FpDest));
         Fpscr = setVfpFpscr(Fpscr, state);
     '''
@@ -1283,7 +1283,7 @@ let {{
         uint64_t mid = ((uint64_t)FpOp1P0.uw | ((uint64_t)FpOp1P1.uw << 32));
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (mid) : "m" (mid));
-        cDest.fp = vfpSFixedToFpD(mid, false, imm);
+        cDest.fp = vfpSFixedToFpD(Fpscr, mid, false, imm);
         __asm__ __volatile__("" :: "m" (cDest.fp));
         Fpscr = setVfpFpscr(Fpscr, state);
         FpDestP0.uw = cDest.bits;
@@ -1299,7 +1299,7 @@ let {{
     vcvtUFixedFpSCode = '''
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (FpOp1.uw) : "m" (FpOp1.uw));
-        FpDest = vfpUFixedToFpS(FpOp1.uw, false, imm);
+        FpDest = vfpUFixedToFpS(Fpscr, FpOp1.uw, false, imm);
         __asm__ __volatile__("" :: "m" (FpDest));
         Fpscr = setVfpFpscr(Fpscr, state);
     '''
@@ -1315,7 +1315,7 @@ let {{
         uint64_t mid = ((uint64_t)FpOp1P0.uw | ((uint64_t)FpOp1P1.uw << 32));
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (mid) : "m" (mid));
-        cDest.fp = vfpUFixedToFpD(mid, false, imm);
+        cDest.fp = vfpUFixedToFpD(Fpscr, mid, false, imm);
         __asm__ __volatile__("" :: "m" (cDest.fp));
         Fpscr = setVfpFpscr(Fpscr, state);
         FpDestP0.uw = cDest.bits;
@@ -1403,7 +1403,7 @@ let {{
     vcvtSHFixedFpSCode = '''
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (FpOp1.sh) : "m" (FpOp1.sh));
-        FpDest = vfpSFixedToFpS(FpOp1.sh, true, imm);
+        FpDest = vfpSFixedToFpS(Fpscr, FpOp1.sh, true, imm);
         __asm__ __volatile__("" :: "m" (FpDest));
         Fpscr = setVfpFpscr(Fpscr, state);
     '''
@@ -1420,7 +1420,7 @@ let {{
         uint64_t mid = ((uint64_t)FpOp1P0.uw | ((uint64_t)FpOp1P1.uw << 32));
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (mid) : "m" (mid));
-        cDest.fp = vfpSFixedToFpD(mid, true, imm);
+        cDest.fp = vfpSFixedToFpD(Fpscr, mid, true, imm);
         __asm__ __volatile__("" :: "m" (cDest.fp));
         Fpscr = setVfpFpscr(Fpscr, state);
         FpDestP0.uw = cDest.bits;
@@ -1437,7 +1437,7 @@ let {{
     vcvtUHFixedFpSCode = '''
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (FpOp1.uh) : "m" (FpOp1.uh));
-        FpDest = vfpUFixedToFpS(FpOp1.uh, true, imm);
+        FpDest = vfpUFixedToFpS(Fpscr, FpOp1.uh, true, imm);
         __asm__ __volatile__("" :: "m" (FpDest));
         Fpscr = setVfpFpscr(Fpscr, state);
     '''
@@ -1454,7 +1454,7 @@ let {{
         uint64_t mid = ((uint64_t)FpOp1P0.uw | ((uint64_t)FpOp1P1.uw << 32));
         VfpSavedState state = prepVfpFpscr(Fpscr);
         __asm__ __volatile__("" : "=m" (mid) : "m" (mid));
-        cDest.fp = vfpUFixedToFpD(mid, true, imm);
+        cDest.fp = vfpUFixedToFpD(Fpscr, mid, true, imm);
         __asm__ __volatile__("" :: "m" (cDest.fp));
         Fpscr = setVfpFpscr(Fpscr, state);
         FpDestP0.uw = cDest.bits;