Fix problem with scaling general loop blocks; add dumpers #99742
Conversation
1. Fix a problem where scaling block weights of a range of basic blocks that compose a general loop could encounter a block that is unreachable. This can happen for various kinds of blocks that the JIT doesn't like to remove even if unreachable. Simply skip scaling those blocks. 2. Add dumpers for `FlowGraphDfsTree` and `BlockToNaturalLoopMap`. These can be called from the debugger (perhaps they should be called to output these things to the JitDump). 3. Update `fgDfsBlocksAndRemove()` to dump any blocks that it did not remove, even if they were unreachable. This was found as part of fixing JitOptRepeat: dotnet#94250
dotnet-issue-labeler
bot
added
the
area-CodeGen-coreclr
|
@jakobbotsch PTAL |
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch |
This includes various fixes that are being separately PR'ed: 1. dotnet#99744: Introduce HandleKindDataIsInvariant helper 2. dotnet#99743: Add basic support for TYP_MASK constants 3. dotnet#99742: Fix problem with scaling general loop blocks; add dumpers 4. dotnet#99740: Add edge likelihood dumping; fix one edge likelihood update case Also: 1. Add support for running JitOptRepeat under JitStress. This is still over-written by JitOptRepeat being forced on at 4 iterations.
| printf("%02u %02u -> " FMT_BB "[%u, %u]\n", i, rpoNum, block->bbNum, block->bbPreorderNum, | ||
| block->bbPostorderNum); |
There was a problem hiding this comment.
i and block->bbPostorderNum is the same thing here.
There was a problem hiding this comment.
Well, this is a dumper, so this can help verify that in the output.
|
|
||
| // Should we call this using the phase method: | ||
| // DoPhase(this, PHASE_SET_BLOCK_WEIGHTS, &Compiler::optSetBlockWeights); | ||
| // ? It could be called multiple times. | ||
| optSetBlockWeights(); | ||
|
|
There was a problem hiding this comment.
Hopefully we'll just end up removing optSetBlockWeights entirely in favor of some profile maintenance that means that this call will just go away at some point.
| #ifdef DEBUG | ||
| // Did we actually remove all the blocks we said we were going to? | ||
| if (verbose) | ||
| { | ||
| if (m_dfsTree->GetPostOrderCount() != fgBBcount) | ||
| { | ||
| printf("%u unreachable blocks were not removed:\n", fgBBcount - m_dfsTree->GetPostOrderCount()); | ||
| for (BasicBlock* block : Blocks()) | ||
| { | ||
| if (!m_dfsTree->Contains(block)) | ||
| { | ||
| printf(" " FMT_BB "\n", block->bbNum); | ||
| } | ||
| } | ||
| } | ||
| } | ||
| #endif // DEBUG |
There was a problem hiding this comment.
Luckily the blocks are still removed in the "graph" sense -- they don't have any preds/succs left anymore, so they don't interact with reachable blocks in any way. I.e. asserts like these are ok:
runtime/src/coreclr/jit/morph.cpp
Lines 13833 to 13835 in b60a541
But yeah, we still have this annoying case to check for when iterating the block list.
Co-authored-by: Jakob Botsch Nielsen <[email protected]>
Fix a problem where scaling block weights of a range of basic blocks that compose a general loop could encounter a block that is unreachable. This can happen for various kinds of blocks that the JIT doesn't like to remove even if unreachable. Simply skip scaling those blocks.
Add dumpers for
FlowGraphDfsTreeandBlockToNaturalLoopMap. These can be called from the debugger (perhaps they should be called to output these things to the JitDump).Update
fgDfsBlocksAndRemove()to dump any blocks that it did not remove, even if they were unreachable.This was found as part of fixing JitOptRepeat: #94250