Marge Bot pushed to branch master at Glasgow Haskell Compiler / GHC

Commits:

5 changed files:

Changes:

  • changelog.d/T27722-cbe-entry-block.md
    1 1 
     section: cmm
    2 2 
     issues: #27722
    3 
    -mrs: !16592
    3 
    +mrs: !16592 !16861
    4 4 
     synopsis:
    5 5 
       Fix common block elimination dropping entry block info table in hand written cmm.

  • compiler/GHC/Cmm/CommonBlockElim.hs
    ... ... @@ -66,25 +66,23 @@ elimCommonBlocks g =
    66 66 
       where
    67 67 
          g' = replaceLabels env $ copyTicks env g
    68 68 
          env = iterate mapEmpty blocks_with_key
    69 
    -     -- The order of blocks doesn't matter here. While we could use
    70 
    -     -- revPostorder which drops unreachable blocks this is done in
    71 
    -     -- ContFlowOpt already which runs before this pass. So we use
    72 
    -     -- toBlockList since it is faster.
    73 
    -     -- One exception: The entry block most come first or we risk eliminating it
    74 
    -     -- in favour of another block. See Note [Retain entry block during common block elimination.]
    75 
    -     groups = groupByInt hash_block (toBlockListEntryFirst g) :: [[CmmBlock]]
    76 69 
          blocks_with_key = [ [ (successors b, [b]) | b <- bs] | bs <- groups]
    77 
    -
    78 
    --- Note [Retain entry block during common block elimination.]
    79 
    --- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    80 
    --- At the stage we run common block elimination (CBE) we only have one info
    81 
    --- table for the entry label. Which means we can get away without applying the
    82 
    --- block label substitution to the info table *as long as we keep the first block*.
    83 
    --- When combining blocks the first block in the list of blocks is kept, and the later
    84 
    --- one eliminated, so we can achieve this by simply using toBlockListEntryFirst.
    70 
    +     groups = groupByInt hash_block candidates :: [[CmmBlock]]
    71 
    +     -- See Note [Retain entry block during common block elimination]
    72 
    +     candidates = filter ((/= g_entry g) . entryLabel) (toBlockList g)
    73 
    +
    74 
    +-- Note [Retain entry block during common block elimination]
    75 
    +-- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    76 
    +-- When CBE runs, only the entry label has an info table, so we can get away
    77 
    +-- without applying the block label substitution to info tables, as long as we
    78 
    +-- preserve the entry block (#27722). We achieve this by simply keeping it out
    79 
    +-- of the list of candidates for elimination.
    85 80 
     --
    86 
    --- If we don't we end up with #27722 where the entry block was eliminated in favour
    87 
    --- of another block.
    81 
    +-- The downside is that any duplicates of the entry block stay unmerged. This
    82 
    +-- should be very rare though, and only affects hand-written Cmm: the entry
    83 
    +-- block of any proc generated from STG contains the stack check (see
    84 
    +-- entryHeapCheck in GHC.StgToCmm.Heap) and is therefore distinct from any
    85 
    +-- other block in the same proc.
    88 86
    89 87 
     -- Invariant: The blocks in the list are pairwise distinct
    90 88 
     -- (so avoid comparing them again)

  • testsuite/tests/cmm/should_run/T27722.hs
    1 
    +import Control.Monad
    2 
    +import Foreign.Ptr
    3 
    +import System.Exit
    4 
    +
    5 
    +foreign import ccall "&zork_info" zork_info :: Ptr ()
    6 
    +
    7 
    +main :: IO ()
    8 
    +main = when (zork_info == nullPtr) exitFailure

  • testsuite/tests/cmm/should_run/T27722_cmm.cmm
    1 
    +#include "Cmm.h"
    2 
    +
    3 
    +// Regression test for #27722: common block elimination (CBE) must not
    4 
    +// eliminate the entry block, because the proc's info table refers to it.
    5 
    +// With the bug, zork_info is never emitted and T27722.hs fails to link.
    6 
    +//
    7 
    +// The entry block only becomes a duplicate in the third round of CBE:
    8 
    +// round 1 merges the two then-blocks (both duplicates of the entry
    9 
    +// block), round 2 merges E1 and E2, and only then does the entry block
    10 
    +// match the surviving then-block.
    11 
    +//
    12 
    +// The then-blocks are braced if-bodies so that, like the entry block,
    13 
    +// they carry two source notes (braces and statement): CBE's block hash
    14 
    +// counts tick nodes (#27877), and blocks with different hashes are never
    15 
    +// compared.
    16 
    +INFO_TABLE_RET (zork, RET_SMALL)
    17 
    +    return ()
    18 
    +{
    19 
    +  bits64 x;
    20 
    +  x = 1 :: bits64;
    21 
    +  goto E1;
    22 
    +E1:
    23 
    +  if (x > 0 :: bits64) { x = 1 :: bits64; goto E2; } else { goto D; }
    24 
    +E2:
    25 
    +  if (x > 0 :: bits64) { x = 1 :: bits64; goto E2; } else { goto D; }
    26 
    +D:
    27 
    +  if (x > 1 :: bits64) goto E1;
    28 
    +  goto E2;
    29 
    +}

  • testsuite/tests/cmm/should_run/all.T
    ... ... @@ -60,3 +60,6 @@ test('T27447',
    60 60 
          ],
    61 61 
          multi_compile_and_run,
    62 62 
          ['T27447', [('T27447_cmm.cmm', '')], ''])
    63 
    +
    64 
    +test('T27722', [req_cmm], multi_compile_and_run,
    65 
    +     ['T27722', [('T27722_cmm.cmm', '')], '-O'])

    • —
    View it on GitLab.
    You're receiving this email because of your account on gitlab.haskell.org. Manage all notifications · Help Notification message regarding https://gitlab.haskell.org/ghc/ghc/-/commit/a2208b21fcb5b417e7dd535295eb823eb4c3527f at 1793707343