Commit 42b8aa9e authored by's avatar

Documentation for INLINABLE

parent 2b8358cf
......@@ -1685,11 +1685,9 @@ pseudoop "inline"
{\tt inline} function expands to the identity function in Phase zero; so its
use imposes no overhead.
If the function is defined in another module, GHC only exposes its inlining
in the interface file if the function is sufficiently small that it might be
inlined by the automatic mechanism. There is currently no way to tell GHC to
expose arbitrarily-large functions in the interface file. (This shortcoming
is something that could be fixed, with some kind of pragma.) }
It is good practice to mark the function with an INLINABLE pragma at
its definition, (a) so that GHC guarantees to expose its unfolding regardless
of size, and (b) so that you have control over exactly what is inlined. }
pseudoop "lazy"
a -> a
......@@ -7594,21 +7594,68 @@ key_function :: Int -> String -> (Bool, Double)
function "<literal>f</literal>" has a number of other effects:
No functions are inlined into <literal>f</literal>. Otherwise
GHC might inline a big function into <literal>f</literal>'s right hand side,
making <literal>f</literal> big; and then inline <literal>f</literal> blindly.
While GHC is keen to inline the function, it does not do so
blindly. For example, if you write
map key_function xs
there really isn't any point in inlining <literal>key_function</literal> to get
map (\x -> <replaceable>body</replaceable>) xs
In general, GHC only inlines the function if there is some reason (no matter
how slight) to supose that it is useful to do so.
The float-in, float-out, and common-sub-expression transformations are not
applied to the body of <literal>f</literal>.
Moreover, GHC will only inline the function if it is <emphasis>fully applied</emphasis>,
where "fully applied"
means applied to as many arguments as appear (syntactically)
on the LHS of the function
definition. For example:
comp1 :: (b -> c) -> (a -> b) -> a -> c
{-# INLINE comp1 #-}
comp1 f g = \x -> f (g x)
comp2 :: (b -> c) -> (a -> b) -> a -> c
{-# INLINE comp2 #-}
comp2 f g x = f (g x)
The two functions <literal>comp1</literal> and <literal>comp2</literal> have the
same semantics, but <literal>comp1</literal> will be inlined when applied
to <emphasis>two</emphasis> arguments, while <literal>comp2</literal> requires
<emphasis>three</emphasis>. This might make a big difference if you say
map (not `comp1` not) xs
which will optimise better than the corresponding use of `comp2`.
It is useful for GHC to optimise the definition of an
INLINE function <literal>f</literal> just like any other non-INLINE function,
in case the non-inlined version of <literal>f</literal> is
ultimately called. But we don't want to inline
the <emphasis>optimised</emphasis> version
of <literal>f</literal>;
a major reason for INLINE pragmas is to expose functions
in <literal>f</literal>'s RHS that have
rewrite rules, and it's no good if those functions have been optimised
So <emphasis>GHC guarantees to inline precisely the code that you wrote</emphasis>, no more
and no less. It does this by capturing a copy of the definition of the function to use
for inlining (we call this the "inline-RHS"), which it leaves untouched,
while optimising the ordinarly RHS as usual. For externally-visible functions
the inline-RHS (not the optimised RHS) is recorded in the interface file.
An INLINE function is not worker/wrappered by strictness analysis.
It's going to be inlined wholesale instead.
All of these effects are aimed at ensuring that what gets inlined is
exactly what you asked for, no more and no less.
<para>GHC ensures that inlining cannot go on forever: every mutually-recursive
group is cut by one or more <emphasis>loop breakers</emphasis> that is never inlined
......@@ -7636,8 +7683,9 @@ itself, so an INLINE pragma is always ignored.</para>
{-# INLINE returnUs #-}
<para>See also the <literal>NOINLINE</literal> pragma (<xref
<para>See also the <literal>NOINLINE</literal> (<xref linkend="inlinable-pragma"/>)
and <literal>INLINABLE</literal> (<xref linkend="noinline-pragma"/>)
<para>Note: the HBC compiler doesn't like <literal>INLINE</literal> pragmas,
so if you want your code to be HBC-compatible you'll have to surround
......@@ -7646,6 +7694,36 @@ itself, so an INLINE pragma is always ignored.</para>
<sect3 id="inlinable-pragma">
<title>INLINABLE pragma</title>
<para>An INLINABLE pragma works very like an INLINE pragma, except that:
INLINE says "please inline me", but INLINABLE says "feel free to inline me; use your
discretion". In other words the choice is left to GHC, which uses the same
rules as for pragma-free functions. Unlike INLINE, That decision is made at
the <emphasis>call site</emphasis>, and
will therefore be affected by the inlining threshold, optimisation level etc.
Like INLINE, the INLINABLE pragma retains a copy of the original RHS for
inlining purposes, and persists it in the interface file, regardless of
the size of the RHS.
If you use the special function <literal>inline</literal> (<xref linkend="special-ids"/>)
to force inlining at a
call site, you will get a copy of the the original RHS.
Indeed, if you intend to use <literal>inline f</literal> it
is a good idea to mark the definition of <literal>f</literal> INLINABLE,
so that GHC guarantees to expose an unfolding regardless of how big it is.
<sect3 id="noinline-pragma">
<title>NOINLINE pragma</title>
......@@ -8760,7 +8838,23 @@ r) ->
<para>GHC has a few built-in functions with special behaviour. These
are now described in the module <ulink
in the library documentation.</para>
in the library documentation.
In particular:
<ulink url="&libraryGhcPrimLocation;/GHC-Prim.html#v%3Ainline"><literal>inline</literal></ulink>
allows control over inlining on a per-call-site basis.
<ulink url="&libraryGhcPrimLocation;/GHC-Prim.html#v%3Alazy"><literal>lazy</literal></ulink>
restrains the strictness analyser.
<ulink url="&libraryGhcPrimLocation;/GHC-Prim.html#v%3AunsafeCoerce%23"><literal>lazy</literal></ulink>
allows you to fool the type checker.
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment