From: Eric Botcazou <ebotca...@adacore.com>
This restores the rounding of the first scaled divide operation applied to
the extra digits, which yields more consistent results for smalls that are
reciprocal of integers.
gcc/ada/ChangeLog:
* libgnat/s-valuef.adb (Integer_to_Fixed): Restore rounding of the
first scaled divide operation.
Tested on x86_64-pc-linux-gnu, committed on master.
---
gcc/ada/libgnat/s-valuef.adb | 6 +++++-
1 file changed, 5 insertions(+), 1 deletion(-)
diff --git a/gcc/ada/libgnat/s-valuef.adb b/gcc/ada/libgnat/s-valuef.adb
index 6ea22117432..1743749f542 100644
--- a/gcc/ada/libgnat/s-valuef.adb
+++ b/gcc/ada/libgnat/s-valuef.adb
@@ -97,6 +97,8 @@ package body System.Value_F is
-- V = Q1 + (Q2 + R1) / Num + R2 / (Num * (Base ** 2))
+ -- but we get rid of the third term by using a rounding divide for (2a).
+
-- This works only if Den * (Base ** ScaleB) does not overflow for inputs
-- corresponding to 'Image. Let S = Num / Den, B = Base and N the scale in
-- base B of S, i.e. the smallest integer such that B**N * S >= 1. Then,
@@ -125,6 +127,8 @@ package body System.Value_F is
-- V = Q1 + (Q2 + R1) / (Num * (Base ** -ScaleB)) + R2 / (Num * (...))
+ -- but we get rid of the third term by using a rounding divide for (2b).
+
-- This works only if Num * (Base ** -ScaleB) does not overflow for inputs
-- corresponding to 'Image. With the determination of ScaleB above, we have
@@ -300,7 +304,7 @@ package body System.Value_F is
if E > 0 then
Scaled_Divide (To_Signed (V), Y, Z, Q1, R1, Round => False);
- Scaled_Divide (To_Signed (Uns (E)), Y, -B**2, Q2, R2, Round => False);
+ Scaled_Divide (To_Signed (Uns (E)), Y, -B**2, Q2, R2, Round => True);
-- Avoid an overflow during the subtraction. Note that Q2 is smaller
-- than Y and R1 smaller than Z in magnitude, so it is safe to take
--
2.43.0
