Cheng Shao pushed to branch wip/ubsan at Glasgow Haskell Compiler / GHC

Commits:

5 changed files:

Changes:

  • hadrian/doc/flavours.md
    ... ... @@ -238,6 +238,10 @@ The supported transformers are listed below:
    238 238 
             <td><code>thread_sanitizer</code></td>
    239 239 
             <td>Build the runtime system with ThreadSanitizer support</td>
    240 240 
         </tr>
    241 
    +    <tr>
    242 
    +        <td><code>ubsan</code></td>
    243 
    +        <td>Build all stage1+ C/C++ code with UndefinedBehaviorSanitizer support</td>
    244 
    +    </tr>
    241 245 
         <tr>
    242 246 
             <td><code>llvm</code></td>
    243 247 
             <td>Use GHC's LLVM backend (`-fllvm`) for all stage1+ compilation.</td>

  • hadrian/src/Flavour.hs
    ... ... @@ -7,6 +7,7 @@ module Flavour
    7 7 
       , addArgs
    8 8 
       , splitSections
    9 9 
       , enableThreadSanitizer
    10 
    +  , enableUBSan
    10 11 
       , enableLateCCS
    11 12 
       , enableHashUnitIds
    12 13 
       , enableDebugInfo, enableTickyGhc
    ... ... @@ -33,6 +34,9 @@ import Data.Either
    33 34 
     import Data.Map (Map)
    34 35 
     import qualified Data.Map as M
    35 36 
     import qualified Data.Set as Set
    37 
    +import GHC.Platform.ArchOS
    38 
    +import Oracles.Flag
    39 
    +import Oracles.Setting
    36 40 
     import Packages
    37 41 
     import Flavour.Type
    38 42 
     import Settings.Parser
    ... ... @@ -53,6 +57,7 @@ flavourTransformers = M.fromList
    53 57 
         , "no_split_sections" =: noSplitSections
    54 58 
         , "thread_sanitizer" =: enableThreadSanitizer False
    55 59 
         , "thread_sanitizer_cmm" =: enableThreadSanitizer True
    60 
    +    , "ubsan"            =: enableUBSan
    56 61 
         , "llvm"             =: viaLlvmBackend
    57 62 
         , "profiled_ghc"     =: enableProfiledGhc
    58 63 
         , "no_dynamic_ghc"   =: disableDynamicGhcPrograms
    ... ... @@ -258,6 +263,66 @@ enableThreadSanitizer instrumentCmm = addArgs $ notStage0 ? mconcat
    258 263 
             ]
    259 264 
         ]
    260 265
    266 
    +-- | Whether or not -shared-libsan should be passed to clang at
    267 
    +-- link-time.
    268 
    +--
    269 
    +-- See
    270 
    +-- https://github.com/llvm/llvm-project/blob/llvmorg-21.1.5/clang/lib/Driver/SanitizerArgs.cpp#L1008,
    271 
    +-- clang defaults to -shared-libsan on darwin/windows and
    272 
    +-- -static-libsan on linux. In general, -static-libsan is incredibly
    273 
    +-- problematic when multiple copies of the sanitizer runtimes coexist
    274 
    +-- in the same address space due to being linked into multiple Haskell
    275 
    +-- libraries. So we should explicitly specify `-shared-libsan` if
    276 
    +-- needed.
    277 
    +--
    278 
    +-- A small downside of -shared-libsan is the clang-specific sanitizer
    279 
    +-- runtime shared library path needs to be manually specified via
    280 
    +-- @export LD_LIBRARY_PATH=$(dirname $(clang -print-libgcc-file-name
    281 
    +-- -rtlib=compiler-rt))@ for ld.so to find it at runtime.
    282 
    +needSharedLibSAN :: Action Bool
    283 
    +needSharedLibSAN = do
    284 
    +  is_clang <- flag CcLlvmBackend
    285 
    +  is_default_shared_libsan <- anyTargetOs [OSDarwin, OSMinGW32]
    286 
    +  pure $ is_clang && not is_default_shared_libsan
    287 
    +
    288 
    +-- | Build all stage1+ C/C++ code with UndefinedBehaviorSanitizer
    289 
    +-- support:
    290 
    +-- https://clang.llvm.org/docs/UndefinedBehaviorSanitizer.html.
    291 
    +--
    292 
    +-- Note that we also pass -fno-sanitize=function to clang, since
    293 
    +-- "runtime call to function foo through pointer to incorrect function
    294 
    +-- type" is unfortunately pretty common (e.g. evac_fn in rts) and
    295 
    +-- impact the signal to noise ratio of UBSAN warnings. gcc doesn't
    296 
    +-- implement this instrumentation though.
    297 
    +enableUBSan :: Flavour -> Flavour
    298 
    +enableUBSan =
    299 
    +  addArgs $
    300 
    +    notStage0
    301 
    +      ? mconcat
    302 
    +        [ package rts
    303 
    +            ? builder (Cabal Flags)
    304 
    +            ? arg "+ubsan"
    305 
    +            <> (needSharedLibSAN ? arg "+shared-libsan"),
    306 
    +          builder (Ghc CompileHs)
    307 
    +            ? arg "-optc-fsanitize=undefined"
    308 
    +            <> (flag CcLlvmBackend ? arg "-optc-fno-sanitize=function"),
    309 
    +          builder (Ghc CompileCWithGhc)
    310 
    +            ? arg "-optc-fsanitize=undefined"
    311 
    +            <> (flag CcLlvmBackend ? arg "-optc-fno-sanitize=function"),
    312 
    +          builder (Ghc CompileCppWithGhc)
    313 
    +            ? arg "optcxx-fsanitize=undefined"
    314 
    +            <> (flag CcLlvmBackend ? arg "-optcxx-fno-sanitize=function"),
    315 
    +          builder (Ghc LinkHs)
    316 
    +            ? arg "-optc-fsanitize=undefined"
    317 
    +            <> arg "-optl-fsanitize=undefined"
    318 
    +            <> (needSharedLibSAN ? arg "-optl-shared-libsan")
    319 
    +            <> (flag CcLlvmBackend ? arg "-optc-fno-sanitize=function"),
    320 
    +          builder (Cc CompileC)
    321 
    +            ? arg "-fsanitize=undefined"
    322 
    +            <> (flag CcLlvmBackend ? arg "-fno-sanitize=function"),
    323 
    +          builder Testsuite ? arg "--config=have_ubsan=True"
    324 
    +        ]
    325 
    +
    261 326 
     -- | Use the LLVM backend in stages 1 and later.
    262 327 
     viaLlvmBackend :: Flavour -> Flavour
    263 328 
     viaLlvmBackend = addArgs $ notStage0 ? builder Ghc ? arg "-fllvm"

  • rts/rts.cabal
    ... ... @@ -91,6 +91,19 @@ flag thread-sanitizer
    91 91 
         in @rts/include/rts/TSANUtils.h@.
    92 92 
       default: False
    93 93 
       manual: True
    94 
    +flag ubsan
    95 
    +  description:
    96 
    +    Link with -fsanitize=undefined, to be enabled when building with
    97 
    +    UndefinedBehaviorSanitizer.
    98 
    +  default: False
    99 
    +  manual: True
    100 
    +flag shared-libsan
    101 
    +  description:
    102 
    +    Link with -shared-libsan, to guarantee only one copy of the
    103 
    +    sanitizer runtimes exist in the address space. See
    104 
    +    needSharedLibSAN in hadrian/src/Flavour.hs.
    105 
    +  default: False
    106 
    +  manual: True
    94 107
    95 108 
     library
    96 109 
         -- rts is a wired in package and
    ... ... @@ -200,6 +213,12 @@ library
    200 213 
             cc-options: -fsanitize=thread
    201 214 
             ld-options: -fsanitize=thread
    202 215
    216 
    +      if flag(ubsan)
    217 
    +        ld-options: -fsanitize=undefined
    218 
    +
    219 
    +      if flag(shared-libsan)
    220 
    +        ld-options: -shared-libsan
    221 
    +
    203 222 
           if os(linux)
    204 223 
              -- the RTS depends upon libc. while this dependency is generally
    205 224 
              -- implicitly added by `cc`, we must explicitly add it here to ensure

  • testsuite/driver/testglobals.py
    ... ... @@ -186,6 +186,9 @@ class TestConfig:
    186 186 
             # Are we running in a ThreadSanitizer-instrumented build?
    187 187 
             self.have_thread_sanitizer = False
    188 188
    189 
    +        # Are we running with UndefinedBehaviorSanitizer enabled?
    190 
    +        self.have_ubsan = False
    191 
    +
    189 192 
             # Do symbols use leading underscores?
    190 193 
             self.leading_underscore = False
    191 194

  • testsuite/driver/testlib.py
    ... ... @@ -1090,6 +1090,8 @@ def llvm_build ( ) -> bool:
    1090 1090 
     def have_thread_sanitizer( ) -> bool:
    1091 1091 
         return config.have_thread_sanitizer
    1092 1092
    1093 
    +def have_ubsan( ) -> bool:
    1094 
    +    return config.have_ubsan
    1093 1095
    1094 1096 
     def gcc_as_cmmp() -> bool:
    1095 1097 
         return config.cmm_cpp_is_gcc

    • —
    View it on GitLab.
    You're receiving this email because of your account on gitlab.haskell.org. Manage all notifications · Help