SH7750_08 Datasheet, PDF (283/1162 Pages) Renesas Technology Corp – 32-Bit RISC Microcomputer

Section 6 Floating-Point Unit (FPU)
6.7.3 Sign of Operation Result when Using FIPR or FTRV Instruction
When two or more data items used in an operation by the FIPR or FTRV instruction are infinity,
and all of the infinity items in the multiplication results have the same sign, the sign of the
operation result may be incorrect.
Workarounds
1. Do not use infinity. If conditions a. to c. below are satisfied, infinity is never used in
operations.
a. Use Round to Zero (FPSCR.RM = 01) as the rounding mode.
b. Do not divide by zero.
c. Do not transfer a value of positive or negative infinity to FR0 to FR15 or to XF0 to XF15.
2. Avoid using the FIPR and FTRV instructions, and use the FADD, FMUL, and FMAC
instructions instead.
6.7.4 Notes on Double-Precision FADD and FSUB Instructions
Description: If the input data for a double-precision FADD instruction or a double-precision
FSUB instruction satisfies all of the conditions listed below, the inexact bits (FPSCR.Flag.I and
FPSCR.Cause.I) may not be set even through the operation result is inexact.
Condition 1: The operation instruction is a double-precision FADD instruction or a double-
precision FSUB instruction.
Condition 2: The difference between the DRn and DRm exponents is between 43 and 50.
Condition 3: At least one of bits 31 to 24 of the mantissa portion of whichever of DRn and DRm
has the smaller absolute value is 1.
Condition 4: Bits 23 to 0 of the mantissa portion of whichever of DRn and DRm has the smaller
absolute value are all 0.
Condition 5: Bits 40 to 32 of the mantissa portion of whichever of DRn and DRm has the smaller
absolute value are all 0.
In addition, the result of an operation meeting the above conditions may have a rounding error.
Specifically, in a case where the closest expressible value less than the unrounded value should be
selected, the closest expressible value greater than the unrounded value is selected instead.
Conversely, in a case where the closest expressible value greater than the unrounded value should
be selected, the closest expressible value less than the unrounded value is selected instead.
Rev.7.00 Oct. 10, 2008 Page 199 of 1074
REJ09B0366-0700