The MoveForSingleUse or MoveAndTeeForMultiUse functions move wasm instructions,
however DBG_VALUE stay unchanged -- moving or cloning these.
Details
- Reviewers
dschuff - Commits
- rGb3857e4d35fe: [WebAssembly] Fix MRI.hasOneNonDBGUse assert in WebAssemblyRegStackify pass
rG7c18d6083ab5: [WebAssembly] Move/clone DBG_VALUE during WebAssemblyRegStackify pass
rL343154: [WebAssembly] Fix MRI.hasOneNonDBGUse assert in WebAssemblyRegStackify pass
rL343007: [WebAssembly] Move/clone DBG_VALUE during WebAssemblyRegStackify pass
Diff Detail
- Repository
- rL LLVM
Event Timeline
lib/Target/WebAssembly/WebAssemblyRegStackify.cpp | ||
---|---|---|
472 ↗ | (On Diff #154430) | I'm wondering if we should check that MI is nullptr here, or maybe assert? I know parent MI's can be initialized to nullptr and can be cleared, but I'm not sure if that would ever be the case at this point. |
534 ↗ | (On Diff #154430) | Minor nit: I'd suggest running clang-format here, usually formals split between two lines end up following the opening parenthesis, lining up with each other. |
585 ↗ | (On Diff #154430) | I would reword this to be imperative: "Move or clone corresponding DBG_VALUEs to the 'Insert' location. |
Thanks Yury, it looks like you addressed my immediate concerns. I didn't see anything glaringly wrong, when I looked at the original patch. You should add a few reviewers and tag this as "debug-info" by editing the revision and updating the "Tags" field. That should get your patch a bit more exposure. Do you happen to have a smaller/reduced reproducer? I'm wondering how much of the test can be trimmed down, for instance, you declared a variable 't' but it's unused.
- Reducing the test case further: remove attributes, lifetime markers and non-"a" variable metadata.
- Add test for MoveAndTeeForMultiUse changes
Do you happen to have a smaller/reduced reproducer?
It is really hard for me to create reduced test case. I modified the test cases to remove some metadata per https://reviews.llvm.org/D48994#1164085 recommendation.
Otherwise LGTM, although I'm not an expert in DebugInfo either.
lib/Target/WebAssembly/WebAssemblyRegStackify.cpp | ||
---|---|---|
470 ↗ | (On Diff #162014) | I think this can just be something like for (const auto &Op : MRI.reg_operands(Reg)) (and below) |
otherwise LGTM
test/DebugInfo/WebAssembly/dbg-value-move.ll | ||
---|---|---|
10 ↗ | (On Diff #166762) | Was this generated with any optimization flags in clang? (maybe just also add the clang command line as well as the sources in these files?) |
This seems to have broken some of the emscripten tests (binaryen2.test_poppler and binaryen2.test_the_bullet) that we run on the wasm waterfall:
https://logs.chromium.org/v/?s=chromium%2Fbb%2Fclient.wasm.llvm%2Flinux%2F35933%2F%2B%2Frecipes%2Fsteps%2FExecute_emscripten_testsuite__emwasm_%2F0%2Fstdout
Looks like MRI.hasOneUse checks also DBG_VALUE register usages (that causes assert to fail). I need to check if use_nodbg_begin() will fix the issue. (FWIW OneUseDominatesOtherUses uses xx_nodbg_yyy operations in some places)
We we revert this or get it fixed? Its currently breaking the wasm waterfall: https://wasm-stat.us/console
The fix just to pass tests is
diff --git a/lib/Target/WebAssembly/WebAssemblyRegStackify.cpp b/lib/Target/WebAssembly/WebAssemblyRegStackify.cpp index 5b6a91dbc9d..6583057e7fb 100644 --- a/lib/Target/WebAssembly/WebAssemblyRegStackify.cpp +++ b/lib/Target/WebAssembly/WebAssemblyRegStackify.cpp @@ -428,8 +428,10 @@ static bool OneUseDominatesOtherUses(unsigned Reg, const MachineOperand &OneUse, if (!TargetRegisterInfo::isVirtualRegister(DefReg) || !MFI.isVRegStackified(DefReg)) return false; - assert(MRI.hasOneUse(DefReg)); - const MachineOperand &NewUse = *MRI.use_begin(DefReg); + auto It = MRI.use_nodbg_begin(DefReg); + assert(It != MRI.use_nodbg_end()); + const MachineOperand &NewUse = *It; + assert(++It == MRI.use_nodbg_end()); const MachineInstr *NewUseInst = NewUse.getParent(); if (NewUseInst == OneUseInst) { if (&OneUse > &NewUse)
It will take care of WebAssemblyRegStackify, but it will be nice to see it at MRI::hasOneUse.
Let me know if I need to backup entire patch and attempt to fix hasOneUse first.
Not exactly sure what you mean; do you mean that hasOneUse should return true if there is only one non-debug use? That makes sense, maybe it should be hasOneNonDbgUse or something. But please either commit this fix and then follow up, or revert this patch and re-apply with hasOneUse fixed.
This still breaks two emscripten tests, test_the_bullet and test_poppler
In particular, it creates a wasm module that fails to typecheck, seemingly because of a reordering of instructions. Commenting out MoveDebugValues and CloneDebugValues is needed to avoid this.
The simplest reproduction I can find is:
wasm32-clang++ $EMSCRIPTEN/tests/bullet/src/BulletCollision/NarrowPhaseCollision/btPersistentManifold.cpp \ -D__EMSCRIPTEN__=1 \ -I$EMSCRIPTEN/tests/bullet/src/ \ -nostdlib \ -Xclang -nostdsysteminc \ -Xclang -isystem$EMSCRIPTEN/system/include/libc \ -Xlinker --no-entry \ -Xlinker --allow-undefined \ -O2 -g -o test.wasm
where EMSCRIPTEN is the path to emscripten's main directory (containing emcc.py et al). This creates a test.wasm that is invalid, quickly verifiable with wabt's wasm2wat
I think this has unfortunately caused a regression :( -- https://bugs.llvm.org/show_bug.cgi?id=39158
Forgot to add, need to add __attribute__((used, visibility("default"))) to btPersistentManifold::clearUserCache in that cpp file
It looks DBG_VALUE instruction handling in RegStackify has several problems.
- MoveDebugValues and UpdateDebugValuesReg do not seem to consider the case where there are multiple defs to a single register within a BB. Currently they treat all DBG_VALUE instruction to a register as a group and move / update them together whenever any one of the definitions of the register in a BB is stackified. A sample test case with multiple defs within a BB that fails is attached in PR40209.
2-3 are written inline.
I also filed PR40209 with the same contents for easy tracking.
llvm/trunk/lib/Target/WebAssembly/WebAssemblyRegStackify.cpp | ||
---|---|---|
486 | Here Op.setReg(NewReg) modifies the iterator MRI.reg_operands(Reg) while traversing it. So this ends up hitting only the first element and bailing out. This function has problems even without this bug because of the 1 above though. | |
542 | This function does not consider a case that an instruction to be rematerialized is in another BB. Even though RegStackify mostly work within a BB, when a def is trivially rematerializable, i.e., const_i32 instruction, that instruction can be another BB and can be copied to this BB. But CloneDebugValues limit the search within this BB by MI->getParent() == &MBB clause, so in that case the debug value for the rematerialized instruction will not be cloned. A test case can be made trivially, so I'm not gonna attach an mir file for that, but it should look like this: bb1: %3 = const_i32 3 ; This is trivially rematerializable ... bb2: ... use %3 |
Here Op.setReg(NewReg) modifies the iterator MRI.reg_operands(Reg) while traversing it. So this ends up hitting only the first element and bailing out. This function has problems even without this bug because of the 1 above though.