[all-commits] [llvm/llvm-project] 8e8c95: [GISel] Erase the root instruction after emitting ...

Sergei Barannikov via All-commits all-commits at lists.llvm.org
Sat Jan 13 00:17:53 PST 2024


  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 8e8c954a173e0e814de1207513f357e99dff2e85
      https://github.com/llvm/llvm-project/commit/8e8c954a173e0e814de1207513f357e99dff2e85
  Author: Sergei Barannikov <barannikov88 at gmail.com>
  Date:   2024-01-13 (Sat, 13 Jan 2024)

  Changed paths:
    M llvm/test/TableGen/DefaultOpsGlobalISel.td
    M llvm/test/TableGen/GlobalISelCombinerEmitter/builtins/match-table-replacerreg.td
    M llvm/test/TableGen/GlobalISelCombinerEmitter/match-table-operand-types.td
    M llvm/test/TableGen/GlobalISelCombinerEmitter/match-table-permutations.td
    M llvm/test/TableGen/GlobalISelCombinerEmitter/match-table-typeof.td
    M llvm/test/TableGen/GlobalISelCombinerEmitter/match-table.td
    M llvm/test/TableGen/GlobalISelEmitter-input-discard.td
    M llvm/test/TableGen/GlobalISelEmitter-multiple-output-discard.td
    M llvm/test/TableGen/GlobalISelEmitter-multiple-output.td
    M llvm/test/TableGen/GlobalISelEmitter-nested-subregs.td
    M llvm/test/TableGen/GlobalISelEmitter-output-discard.td
    M llvm/test/TableGen/GlobalISelEmitter-zero-reg.td
    M llvm/test/TableGen/GlobalISelEmitter.td
    M llvm/test/TableGen/GlobalISelEmitterMatchTableOptimizer.td
    M llvm/test/TableGen/GlobalISelEmitterMatchTableOptimizerSameOperand-invalid.td
    M llvm/test/TableGen/GlobalISelEmitterRegSequence.td
    M llvm/test/TableGen/GlobalISelEmitterSubreg.td
    M llvm/test/TableGen/HasNoUse.td
    M llvm/test/TableGen/gisel-physreg-input.td
    M llvm/utils/TableGen/GlobalISelCombinerEmitter.cpp
    M llvm/utils/TableGen/GlobalISelEmitter.cpp
    M llvm/utils/TableGen/GlobalISelMatchTable.cpp

  Log Message:
  -----------
  [GISel] Erase the root instruction after emitting all its potential uses (#77494)

This tries to fix a bug by resolving a few FIXMEs. The bug is that
`EraseInstAction` is emitted after emitting the _first_ `BuildMIAction`,
which is too early because the erased instruction may still be used by
subsequent `BuildMIAction`s (in particular, by `CopyRenderer`).

An example of the bug (from `match-table-operand-types.td`):
```
def InstTest0 : GICombineRule<
  (defs root:$a),
  (match  (G_MUL i32:$x, i32:$b, i32:$c),
          (G_MUL $a, i32:$b, i32:$x)),
  (apply  (G_ADD i64:$tmp, $b, i32:$c),
          (G_ADD i8:$a, $b, i64:$tmp))>;

GIR_EraseFromParent, /*InsnID*/0,
GIR_BuildMI, /*InsnID*/1, /*Opcode*/GIMT_Encode2(TargetOpcode::G_ADD),
GIR_Copy, /*NewInsnID*/1, /*OldInsnID*/0, /*OpIdx*/0, // a
GIR_Copy, /*NewInsnID*/1, /*OldInsnID*/0, /*OpIdx*/1, // b
GIR_AddSimpleTempRegister, /*InsnID*/1, /*TempRegID*/0,
```

Here, the root instruction is destroyed before copying its operands ('a'
and 'b') to the new instruction.

The solution is to emit `EraseInstAction` for the root instruction as
the last action in the emission pipeline.




More information about the All-commits mailing list