base: Make Foreign.Marshal.Alloc.allocBytes[Aligned] NOINLINE

As noted in #14346, touch# may be optimized away when the simplifier can see
that the continuation passed to allocaBytes will not return. Marking CPS-style
functions with NOINLINE ensures that the simplier can't draw any unsound
conclusions.

Ultimately the right solution here will be to do away with touch# and instead
introduce a scoped primitive as is suggested in #14375.

Note: This was present in 8.2 but was never merged to 8.4 in hopes that
we would have #14375 implemented in time. This meant that the issue
regressed again in 8.4. Thankfully we caught it in time to fix it for
8.6.

(cherry picked from commit 404bf05e)
(cherry picked from commit 56590db0)

libraries/base/Foreign/Marshal/Alloc.hs

@@ -136,6 +136,19 @@ alloca  = doAlloca undefined
 -- away with touch# in favor of a primitive that will capture the scoping left
 -- implicit in the case of touch#.
 
+-- Note [NOINLINE for touch#]
+-- ~~~~~~~~~~~~~~~~~~~~~~~~~~
+-- Both allocaBytes and allocaBytesAligned use the touch#, which is notoriously
+-- fragile in the presence of simplification (see #14346). In particular, the
+-- simplifier may drop the continuation containing the touch# if it can prove
+-- that the action passed to allocaBytes will not return. The hack introduced to
+-- fix this for 8.2.2 is to mark allocaBytes as NOINLINE, ensuring that the
+-- simplifier can't see the divergence.
+--
+-- These can be removed once #14375 is fixed, which suggests that we instead do
+-- away with touch# in favor of a primitive that will capture the scoping left
+-- implicit in the case of touch#.
+
 -- |@'allocaBytes' n f@ executes the computation @f@, passing as argument
 -- a pointer to a temporarily allocated block of memory of @n@ bytes.
 -- The block of memory is sufficiently aligned for any of the basic

testsuite/tests/perf/should_run/all.T

@@ -454,11 +454,12 @@ test('T9203',
                       # 2016-04-06     84345136 (i386/Debian) not sure
                       # 2017-03-24     77969268 (x86/Linux, 64-bit machine) probably join points
 
-                      , (wordsize(64), 84620888, 5) ]),
+                      , (wordsize(64), 98360576, 5) ]),
                       # was            95747304
                       # 2019-09-10     94547280 post-AMP cleanup
                       # 2015-10-28     95451192 emit Typeable at definition site
                       # 2016-12-19     84620888 Join points
+                      # 2018-07-30     98360576 it's unclear
       only_ways(['normal'])],
      compile_and_run,
      ['-O2'])