Make LLVM output robust to -dead_strip on mach-o platforms

This reverses commit 1686f30951292e94bf3076ce8b3eafb0bcbba91d (Mangle
.subsections_via_symbols away., D3287), and implements proper support
for `-dead_strip` via the injection of `.alt_entry` symbols for the
function definition pointing to the beginning of the prefix data.

This is the result of a lengthy discussion with rwbarton, and the
following llvm-dev mailing list thread:
http://lists.llvm.org/pipermail/llvm-dev/2017-March/110733.html

The essential problem is that there is no reference from a function to
its info table.  This combined with `.subsections_via_symbols`, which
llvm emits unconditionally, leads the linker to believe that the prefix
data is unnecessary and stripping it away if presented with the
`-dead_strip` flag.

The NCG has for this purpose special $dsp (dead strip preventer) symbols
and adds a relocation to the end of each function body pointing to that
function's $dsp symbol. We cannot easily do the same thing via LLVM.
Instead we use the `.alt_entry` directive on the function symbol, which
causes the linker to treat it as a continuation of the previous symbol,
namely the $dsp symbol. As a result the function body will not be
separated from its info table.

Reviewers: erikd, austin, rwbarton, bgamari

Reviewed By: bgamari

Subscribers: michalt, thomie

Differential Revision: https://phabricator.haskell.org/D3290

1 files changed, 70 insertions, 2 deletions

diff --git a/compiler/llvmGen/LlvmCodeGen/Ppr.hs b/compiler/llvmGen/LlvmCodeGen/Ppr.hs
index 37d1391519..8b6340d6f8 100644
--- a/compiler/llvmGen/LlvmCodeGen/Ppr.hs
+++ b/compiler/llvmGen/LlvmCodeGen/Ppr.hs
@@ -21,6 +21,7 @@ import FastString
 import Outputable
 import Unique
 
+import DynFlags (targetPlatform)
 
 -- ----------------------------------------------------------------------------
 -- * Top level
@@ -150,8 +151,75 @@ pprLlvmCmmDecl (CmmProc mb_info entry_lbl live (ListGraph blks))
            alias = LMGlobal funVar
                             (Just $ LMBitc (LMStaticPointer defVar)
                                            (LMPointer $ LMInt 8))
-
-       return (ppLlvmGlobal alias $+$ ppLlvmFunction fun', [])
+       -- our beloved dead_strip preventer.
+       -- the idea here is to inject
+       --
+       -- module asm "_symbol$dsp = _symbol-24" -- assuming prefix
+       --                                          of <{i64, i64, i64}>
+       -- module asm ".no_dead_strip _symbol$dsp"
+       --
+       -- and thereby generating a second symbol
+       -- at the start of the info table, which is dead strip prevented.
+       --
+       -- ideally, llvm should generate these for us, but as
+       -- things stand, this is the least hacky solution to
+       -- prevent dead_stripping of the prefix data, while
+       -- retaining dead stripping in general.
+       --
+       -- The general layout of the above code results in the following:
+       --
+       --   .------------. <- @<name>$def$dsp
+       --   | Info Table |
+       --   |------------| <- @<name>, @<name>$def
+       --   | Fn Body    |
+       --   '------------'
+       --
+       -- Why this @<name> and @<name>$def?  As the textual llvm ir
+       -- generator is only handed typeless labes, it often does not
+       -- know the type of the label (e.g. function to call), until
+       -- the actual call happens.  However, llvm requires symbol
+       -- lookups to be typed.  Therfore we create the actual function
+       -- as @<name>$def, and alias a bitcast to i8* as @<name>.
+       --   Any subsequent lookup can lookup @<name> as i8* and
+       -- bitcast it to the required type once we want to call it.
+       --
+       -- Why .no_dead_strip? Doesn't this prevent the linker from
+       -- -dead_strip'ing anything? Yes, it does. And we'll have to
+       -- live with this wart until a better solution is found that
+       -- ensures that all symbols that are used directly or
+       -- indirectly are marked used.
+       --
+       -- This is all rather annoying. ghc 8.2 uses the infamous
+       -- Mangler to drop the .subsections_via_symbols directive
+       -- from the assembly.  LLVM ingeniously emits said directive
+       -- unconditionally for mach-o files.  To lift the need for
+       -- extra mangler step, we explicitly mark every symbol
+       -- .no_dead_strip.
+       --
+       -- We are making a few assumptions here:
+       -- - the symbols end up being name _<symbol> in the final
+       --   assembly file.
+       --
+       dsp <- case mb_info of
+         Nothing -> pure empty
+         Just (Statics _ statics)
+           | platformHasSubsectionsViaSymbols (targetPlatform dflags) -> do
+               infoStatics <- mapM genData statics
+               -- remember, the prefix_size is in bits!
+               let prefix_size = sum (map (llvmWidthInBits dflags . getStatType)
+                                          infoStatics)
+                   dspName = defName `appendFS` fsLit "$dsp"
+                   defSymbol = text "_" <> ftext defName
+                   dspSymbol = text "_" <> ftext dspName
+                   moduleAsm s = text "module asm" <+> doubleQuotes s
+               return $ text "; insert dead_strip preventer"
+                      $+$ moduleAsm (dspSymbol <+> text "=" <+> defSymbol
+                                     <> text "-" <> int (prefix_size `div` 8))
+                      $+$ moduleAsm (text ".no_dead_strip" <+> dspSymbol)
+                      $+$ text "; end dead_strip preventer"
+           | otherwise -> pure empty
+
+       return (ppLlvmGlobal alias $+$ ppLlvmFunction fun' $+$ dsp, [])
 
 
 -- | The section we are putting info tables and their entry code into, should