libraries/base/GHC/Int.hs¶
Note [Order of tests]¶
(See #3065, #5161.) Suppose we had a definition like:
- quot x y
- Note in particular that the
- x == minBound
- test comes before the
- y == (-1)
test.
this expands to something like:
case y of 0 -> divZeroError _ -> case x of
- -9223372036854775808 ->
- case y of -1 -> overflowError _ -> x primQuot y
_ -> x primQuot y
Now if we have the call (x quot 2), and quot gets inlined, then we get:
case 2 of 0 -> divZeroError _ -> case x of
- -9223372036854775808 ->
- case 2 of -1 -> overflowError _ -> x primQuot 2
_ -> x primQuot 2
which simplifies to:
case x of -9223372036854775808 -> x primQuot 2 _ -> x primQuot 2
Now we have a case with two identical branches, which would be eliminated (assuming it doesn’t affect strictness, which it doesn’t in this case), leaving the desired:
x `primQuot` 2
except in the minBound branch we know what x is, and GHC cleverly does the division at compile time, giving:
case x of -9223372036854775808 -> -4611686018427387904 _ -> x primQuot 2
So instead we use a definition like:
- quot x y
which gives us:
case y of 0 -> divZeroError -1 ->
case x of -9223372036854775808 -> overflowError _ -> x primQuot y_ -> x primQuot y
for which our call (x quot 2) expands to:
case 2 of 0 -> divZeroError -1 ->
case x of -9223372036854775808 -> overflowError _ -> x primQuot 2_ -> x primQuot 2
which simplifies to:
x `primQuot` 2
as required.
But we now have the same problem with a constant numerator: the call (2 quot y) expands to
case y of 0 -> divZeroError -1 ->
case 2 of -9223372036854775808 -> overflowError _ -> 2 primQuot y_ -> 2 primQuot y
which simplifies to:
case y of 0 -> divZeroError -1 -> 2 primQuot y _ -> 2 primQuot y
which simplifies to:
case y of 0 -> divZeroError -1 -> -2 _ -> 2 primQuot y
However, constant denominators are more common than constant numerators, so the
y == (-1) && x == minBound
order gives us better code in the common case.
