A.1 Floating-Point Round to Single-Precision Model
The following describes algorithmically the operation of the Floating Round to Single-Precision instruction.
def FRSP(FRB):
if ((FRB)[1:11] <u 897) & ((FRB)[1:63] >u 0) then
if FPSCR[UE] = 0 then
return FRSP_Disabled_Exponent_Underflow(FRB)
if FPSCR[UE] = 1 then
return FRSP_Enabled_Exponent_Underflow(FRB)
if ((FRB)[1:11] >u 1150) & ((FRB)[1:11] <u 2047) then
if FPSCR[OE] = 0 then
return FRSP_Disabled_Exponent_Overflow(FRB)
if FPSCR[OE] = 1 then
return FRSP_Enabled_Exponent_Overflow(FRB)
if ((FRB)[1:11] >u 896) & ((FRB)[1:11] <u 1151) then
return FRSP_Normal_Operand(FRB)
if (FRB)[1:63] = 0 then
return FRSP_Zero_Operand(FRB)
if (FRB)[1:11] = 2047 then
if (FRB)[12:63] = 0 then
return FRSP_Infinity_Operand(FRB)
if (FRB)[12] = 1 then
return FRSP_QNaN_Operand(FRB)
if ((FRB)[12] = 0) & ((FRB)[13:63] >u 0) then
return FRSP_SNaN_Operand(FRB)
def FRSP_Disabled_Exponent_Underflow(FRB):
sign <- (FRB)[0]
frac[0:52] <- 0
exp <- 0
if (FRB)[1:11] = 0 then
exp <- -1022
frac[0:52] <- 0b0 || (FRB)[12:63]
if (FRB)[1:11] >u 0 then
exp <- (FRB)[1:11] - 1023
frac[0:52] <- 0b1 || (FRB)[12:63]
# Denormalize operand:
G <- 0b0
R <- 0b0
X <- 0b0
do while exp < -126
exp <- exp + 1
X <- X | R
R <- G
G <- frac[52]
frac[0:52] <- 0b0 || frac[0:51]
FPSCR[UX] <- (frac[24:52] || G || R || X) >u 0
exp, frac <- Round_Single(sign, exp, frac[0:52], G, R, X)
FPSCR[XX] <- FPSCR[XX] | FPSCR[FI]
FRT <- [0b0] * 64
if frac[0:52] = 0 then
FRT[0] <- sign
FRT[1:63] <- 0
if sign = 0 then FPSCR[FPRF] <- '+ zero'
if sign = 1 then FPSCR[FPRF] <- '- zero'
if frac[0:52] >u 0 then
if frac[0] = 1 then
if sign = 0 then FPSCR[FPRF] <- '+ normal number'
if sign = 1 then FPSCR[FPRF] <- '- normal number'
if frac[0] = 0 then
if sign = 0 then FPSCR[FPRF] <- '+ denormalized number'
if sign = 1 then FPSCR[FPRF] <- '- denormalized number'
# Normalize operand:
do while frac[0] = 0
exp <- exp-1
frac[0:52] <- frac[1:52] || 0b0
FRT[0] <- sign
FRT[1:11] <- exp + 1023
FRT[12:63] <- frac[1:52]
return FRT
def FRSP_Enabled_Exponent_Underflow(FRB):
FPSCR[UX] <- 1
sign <- (FRB)[0]
frac <- [0b0] * 53
exp <- 0
if (FRB)[1:11] = 0 then
exp <- -1022
frac[0:52] <- 0b0 || (FRB)[12:63]
if (FRB)[1:11] >u 0 then
exp <- (FRB)[1:11] - 1023
frac[0:52] <- 0b1 || (FRB)[12:63]
# Normalize operand:
do while frac[0] = 0
exp <- exp - 1
frac[0:52] <- frac[1:52] || 0b0
exp, frac <- Round_Single(sign, exp, frac[0:52], 0b0, 0b0, 0b0)
FPSCR[XX] <- FPSCR[XX] | FPSCR[FI]
exp <- exp + 192
FRT <- [0b0] * 64
FRT[0] <- sign
FRT[1:11] <- exp + 1023
FRT[12:63] <- frac[1:52]
if sign = 0 then FPSCR[FPRF] <- '+ normal number'
if sign = 1 then FPSCR[FPRF] <- '- normal number'
return FRT
def FRSP_Disabled_Exponent_Overflow(FRB):
FPSCR[OX] <- 1
FRT <- [0b0] * 64
if FPSCR[RN] = 0b00 then # Round to Nearest
if (FRB)[0] = 0 then FRT <- 0x7FF0_0000_0000_0000
if (FRB)[0] = 1 then FRT <- 0xFFF0_0000_0000_0000
if (FRB)[0] = 0 then FPSCR[FPRF] <- '+ infinity'
if (FRB)[0] = 1 then FPSCR[FPRF] <- '- infinity'
if FPSCR[RN] = 0b01 then # Round toward Zero
if (FRB)[0] = 0 then FRT <- 0x47EF_FFFF_E000_0000
if (FRB)[0] = 1 then FRT <- 0xC7EF_FFFF_E000_0000
if (FRB)[0] = 0 then FPSCR[FPRF] <- '+ normal number'
if (FRB)[0] = 1 then FPSCR[FPRF] <- '- normal number'
if FPSCR[RN] = 0b10 then # Round toward +Infinity
if (FRB)[0] = 0 then FRT <- 0x7FF0_0000_0000_0000
if (FRB)[0] = 1 then FRT <- 0xC7EF_FFFF_E000_0000
if (FRB)[0] = 0 then FPSCR[FPRF] <- '+ infinity'
if (FRB)[0] = 1 then FPSCR[FPRF] <- '- normal number'
if FPSCR[RN] = 0b11 then # Round toward -Infinity
if (FRB)[0] = 0 then FRT <- 0x47EF_FFFF_E000_0000
if (FRB)[0] = 1 then FRT <- 0xFFF0_0000_0000_0000
if (FRB)[0] = 0 then FPSCR[FPRF] <- '+ normal number'
if (FRB)[0] = 1 then FPSCR[FPRF] <- '- infinity'
FPSCR[FR] <- undefined(0) # FIXME: figure out what values POWER9 uses
FPSCR[FI] <- 1
FPSCR[XX] <- 1
return FRT
def FRSP_Enabled_Exponent_Overflow(FRB):
sign <- (FRB)[0]
exp <- (FRB)[1:11] - 1023
frac <- [0b0] * 53
frac[0:52] <- 0b1 || (FRB)[12:63]
exp, frac <- Round_Single(sign, exp, frac[0:52], 0b0, 0b0, 0b0)
FPSCR[XX] <- FPSCR[XX] | FPSCR[FI]
# Enabled Overflow:
FPSCR[OX] <- 1
exp <- exp - 192
FRT <- [0b0] * 64
FRT[0] <- sign
FRT[1:11] <- exp + 1023
FRT[12:63] <- frac[1:52]
if sign = 0 then FPSCR[FPRF] <- '+ normal number'
if sign = 1 then FPSCR[FPRF] <- '- normal number'
return FRT
def FRSP_Zero_Operand(FRB):
FRT <- (FRB)
if (FRB)[0] = 0 then FPSCR[FPRF] <- '+ zero'
if (FRB)[0] = 1 then FPSCR[FPRF] <- '- zero'
FPSCR[FRFI] <- 0b00
return FRT
def FRSP_Infinity_Operand(FRB):
FRT <- (FRB)
if (FRB)[0] = 0 then FPSCR[FPRF] <- '+ infinity'
if (FRB)[0] = 1 then FPSCR[FPRF] <- '- infinity'
FPSCR[FRFI] <- 0b00
return FRT
def FRSP_QNaN_Operand(FRB):
FRT <- (FRB)[0:34] || [0b0] * 29
FPSCR[FPRF] <- 'QNaN'
FPSCR[FR] <- 0b0
FPSCR[FI] <- 0b0
return FRT
def FRSP_SNaN_Operand(FRB):
FPSCR[VXSNAN] <- 1
FRT <- [0b0] * 64
if FPSCR[VE] = 0 then
FRT[0:11] <- (FRB)[0:11]
FRT[12] <- 1
FRT[13:63] <- (FRB)[13:34] || [0b0] * 29
FPSCR[FPRF] <- 'QNaN'
FPSCR[FR] <- 0b0
FPSCR[FI] <- 0b0
return FRT
def FRSP_Normal_Operand(FRB):
sign <- (FRB)[0]
exp <- (FRB)[1:11] - 1023
frac <- [0b0] * 53
frac[0:52] <- 0b1 || (FRB)[12:63]
exp, frac <- Round_Single(sign, exp, frac[0:52], 0b0, 0b0, 0b0)
FPSCR[XX] <- FPSCR[XX] | FPSCR[FI]
if (exp > 127) & (FPSCR[OE] = 0) then
return FRSP_Disabled_Exponent_Overflow(FRB)
if (exp > 127) & (FPSCR[OE] = 1) then
return FRSP_Enabled_Overflow(FRB)
FRT <- [0b0] * 64
FRT[0] <- sign
FRT[1:11] <- exp + 1023
FRT[12:63] <- frac[1:52]
if sign = 0 then FPSCR[FPRF] <- '+ normal number'
if sign = 1 then FPSCR[FPRF] <- '- normal number'
return FRT
def Round_Single(sign, exp, frac, G, R, X):
inc <- 0
lsb <- frac[23]
gbit <- frac[24]
rbit <- frac[25]
xbit <- (frac[26:52]||G||R||X) != 0
if FPSCR[RN] = 0b00 then # Round to Nearest
# comparisons ignore u bits
if (lsb || gbit) = 0b11 then inc <- 1
if (lsb || gbit || rbit) = 0b011 then inc <- 1
if (lsb || gbit || xbit) = 0b011 then inc <- 1
if FPSCR[RN] = 0b10 then # Round toward + Infinity
# comparisons ignore u bits
if (sign || gbit) = 0b01 then inc <- 1
if (sign || rbit) = 0b01 then inc <- 1
if (sign || xbit) = 0b01 then inc <- 1
if FPSCR[RN] = 0b11 then # Round toward - Infinity
# comparisons ignore u bits
if (sign || gbit) = 0b11 then inc <- 1
if (sign || rbit) = 0b11 then inc <- 1
if (sign || xbit) = 0b11 then inc <- 1
frac[0:23] <- frac[0:23] + inc
if (inc = 1) & (frac[0:23] = 0) then
frac[0:23] <- 0b1 || frac[0:22]
exp <- exp + 1
frac[24:52] <- [0b0] * 29
FPSCR[FR] <- inc
FPSCR[FI] <- gbit | rbit | xbit
return exp, frac
def DOUBLE(WORD):
exp <- [0] * 11
frac <- [0] * 53
sign <- 0b0
FRT <- [0] * 64
# Normalized Operand
if (WORD[1:8] >u 0) & (WORD[1:8] <u 255) then
FRT[0:1] <- WORD[0:1]
FRT[2] <- ¬WORD[1]
FRT[3] <- ¬WORD[1]
FRT[4] <- ¬WORD[1]
FRT[5:63] <- WORD[2:31] || [0]*29
# Denormalized Operand
if (WORD[1:8] = 0) & (WORD[9:31] != 0) then
sign <- WORD[0]
exp <- -126
frac[0:52] <- 0b0 || WORD[9:31] || [0]*29
#normalize the operand
do while frac[0] = 0
frac[0:52] <- frac[1:52] || 0b0
exp <- exp - 1
FRT[0] <- sign
FRT[1:11] <- exp + 1023
FRT[12:63] <- frac[1:52]
# Zero / Infinity / NaN
if (WORD[1:8] = 255) | (WORD[1:31] = 0) then
FRT[0:1] <- WORD[0:1]
FRT[2] <- WORD[1]
FRT[3] <- WORD[1]
FRT[4] <- WORD[1]
FRT[5:63] <- WORD[2:31] || [0]*29
return FRT