Skip to content

Commit

Permalink
Match the logic in mini_emit_sext_index_reg and OP_BOUNDS_CHECK decom…
Browse files Browse the repository at this point in the history
…position (#108469)

Match the logic in mini_emit_sext_index_reg and OP_BOUNDS_CHECK decomposition. This was already done for LLVM but not for any other target. Instead there was an assert which was hit early on startup when the optimization was enabled (tested on ARM64).

The optimization is not enabled by default but that's separate issue. There are potentially some performance wins to have. For example, the `Burgers` benchmarks in dotnet/performance shows this:

Default:
```
| Method    | Mean       | Error   | StdDev  | Median     | Min        | Max        | Gen0       | Gen1       | Gen2       | Allocated    |
|---------- |-----------:|--------:|--------:|-----------:|-----------:|-----------:|-----------:|-----------:|-----------:|-------------:|
| Burgers_3 |   487.9 ms | 2.33 ms | 2.06 ms |   488.2 ms |   483.5 ms |   490.7 ms |          - |          - |          - |    156.76 KB |
| Burgers_0 | 1,266.8 ms | 3.19 ms | 2.83 ms | 1,265.5 ms | 1,264.5 ms | 1,273.6 ms | 47000.0000 | 47000.0000 | 47000.0000 | 781750.05 KB |
| Burgers_1 | 1,180.1 ms | 2.98 ms | 2.49 ms | 1,180.4 ms | 1,176.5 ms | 1,185.3 ms |          - |          - |          - |    156.74 KB |
| Burgers_2 |   286.9 ms | 0.17 ms | 0.15 ms |   286.9 ms |   286.6 ms |   287.0 ms |          - |          - |          - |    156.74 KB |
```

With -O=abcrem:
```
| Method    | Mean       | Error   | StdDev  | Median     | Min        | Max        | Gen0       | Gen1       | Gen2       | Allocated    |
|---------- |-----------:|--------:|--------:|-----------:|-----------:|-----------:|-----------:|-----------:|-----------:|-------------:|
| Burgers_3 |   395.5 ms | 1.25 ms | 1.11 ms |   395.6 ms |   394.0 ms |   397.8 ms |          - |          - |          - |    156.76 KB |
| Burgers_0 | 1,167.8 ms | 5.45 ms | 4.83 ms | 1,167.5 ms | 1,161.5 ms | 1,176.8 ms | 47000.0000 | 47000.0000 | 47000.0000 | 781750.05 KB |
| Burgers_1 | 1,104.2 ms | 2.82 ms | 2.64 ms | 1,102.9 ms | 1,100.4 ms | 1,108.5 ms |          - |          - |          - |    156.74 KB |
| Burgers_2 |   192.2 ms | 0.36 ms | 0.30 ms |   192.2 ms |   191.8 ms |   192.8 ms |          - |          - |          - |    156.74 KB |
```
  • Loading branch information
filipnavara authored Nov 2, 2024
1 parent 1e41404 commit 408caa4
Showing 1 changed file with 8 additions and 9 deletions.
17 changes: 8 additions & 9 deletions src/mono/mono/mini/decompose.c
Original file line number Diff line number Diff line change
Expand Up @@ -1565,18 +1565,17 @@ mono_decompose_array_access_opts (MonoCompile *cfg)
case OP_BOUNDS_CHECK: {
gboolean need_sext = ins->backend.need_sext;
MONO_EMIT_NULL_CHECK (cfg, ins->sreg1, FALSE);
int index2_reg;
if (need_sext) {
index2_reg = alloc_preg (cfg);
MONO_EMIT_NEW_UNALU (cfg, OP_SEXT_I4, index2_reg, ins->sreg2);
} else {
index2_reg = ins->sreg2;
}
if (COMPILE_LLVM (cfg)) {
int index2_reg;
if (need_sext) {
index2_reg = alloc_preg (cfg);
MONO_EMIT_NEW_UNALU (cfg, OP_SEXT_I4, index2_reg, ins->sreg2);
} else {
index2_reg = ins->sreg2;
}
MONO_EMIT_DEFAULT_BOUNDS_CHECK (cfg, ins->sreg1, GINT32_TO_UINT32(ins->inst_imm), index2_reg, ins->flags & MONO_INST_FAULT, ins->inst_p0);
} else {
g_assert (!need_sext);
MONO_ARCH_EMIT_BOUNDS_CHECK (cfg, ins->sreg1, ins->inst_imm, ins->sreg2, ins->inst_p0);
MONO_ARCH_EMIT_BOUNDS_CHECK (cfg, ins->sreg1, ins->inst_imm, index2_reg, ins->inst_p0);
}
break;
}
Expand Down

0 comments on commit 408caa4

Please sign in to comment.