Add reduction to set operands to default null values
Details
Diff Detail
- Repository
- rG LLVM Github Monorepo
Event Timeline
I'm not sure if we should be using undef, poison, or the null value for reductions. I'll ask llvm-dev for opinions.
llvm/test/tools/llvm-reduce/remove-operands.ll | ||
---|---|---|
2 | we should add tests for things like making sure we don't crash on a br since basic blocks don't have a null value | |
llvm/tools/llvm-reduce/DeltaManager.cpp | ||
53 | reduceOperandsDeltaPass for consistency | |
llvm/tools/llvm-reduce/deltas/ReduceOperands.cpp | ||
32 | can we list types that don't have a null value instead? seems better than the current approach since the list is much smaller I'll run this patch over a larger file to see if we're missing some types in this check (as in we're trying to get the null value of some type that doesn't have a null value) | |
63 | these are always used together, can we create something like canReduceOperand()? |
llvm/tools/llvm-reduce/deltas/ReduceOperands.cpp | ||
---|---|---|
32 | I would rather keep it a list of types that work rather than types that don't work. Since if a new type is added that is not nullable in the future this pass would crash. But it would probably bet better for this logic to live in the same place as Constant::getNullValue |
Wouldn't it be better to leave the question of with what to replace to a separate patch, and stick to the current consistent practice here?
I don't think it matters too much if we've agreed in the thread that null is better, and we do a followup patch soon where we add a flag and consolidate what we replace things with into a helper method controlled by the flag.
But did we agree on null?
And i'm not really sure what "I don't think it matters too much" means - this either does introduce inconsistency or it does not.
Reading through the thread, there was a vague consensus that using undef was not ideal and null is probably better in most cases. https://groups.google.com/g/llvm-dev/c/ofvHCsXrLN8
And i'm not really sure what "I don't think it matters too much" means - this either does introduce inconsistency or it does not.
It does introduce inconsistency, but I don't think the inconsistency matters. We'll end up either having undef or null in the reduced repro, either one is probably for most cases.
May i propose that we do that either after this patch, or before this patch, but not overlap the two?
I am not seeing quite seeing from where the motivation for the latter comes from.
Sure I can add Constant::hasNullValue and remove ReduceOperands::typeIsNullable in a separate patch
I meant the part about reducing to undef vs null. Constant::hasNullValue() should be in this patch
Also no, that can be separated - place it before Constant::getNullValue() / base it on that function,
and change Constant::getNullValue() to first assert that said new function says there is such a constant.
But, i'm still not sold on null everywhere. E.g. div-by-zero is ub, gep inbounds of null w/ non-zero offset is null, etc.
Isn't it the same for undef? Since we could arbitrarily decide to replace all undefs with null. Using undef should always be at least the same amount of ub as using null.
I was thinking to change getNullValue to return nullptr when there is no null value instead of crashing.
Then make hasNullValue to something like the following so the code only needs to be updated in one place when future types are added.
bool Constant::hasNullValue(Type *Ty) { return getNullValue(Ty) != nullptr; }
Also for the null vs undef discussion, llvm-reduce should not reduce using this delta pass if it doesn't pass the interestingness test.
Is the worry that this pass will make the the reduction worse (longer or harder to read)?
I honestly don't understand what you're trying to say. I thought your point was that making everything null instead of undef would be bad because some operations may have UB when using null (e.g. your null GEP example). But that argument doesn't make sense since undef is more undefined than null. div by undef is still UB.
Or are you trying to say that we should be doing the operand reduction pass in the first place on some instructions?
@aeubanks I believe @lebedev.ri is asking you to use undef for the initial patch and then switch undef to 0 consistently for all reductions in llvm-reduce. Right now everything else uses undef, and this place is going to use 0.
When I said
I was expecting a separate patch to add a helper method to choose what value to replace things with given a type.
Since replacing most operands with undef seems not ideal when it comes to reducing, I'd rather that patch come first, but I wouldn't object to making this patch use undef first then having the patch to choose what value to reduce to coming after.
Separately, after rereading @lebedev.ri 's comments, I was getting the impression that maybe I misunderstood the comments about not wanting to replace every operand of every instruction with null (or undef). Perhaps you meant that we should skip replacing the pointer operand of a GEP or the second operand of a div with null/undef?
Precisely. I have no clue how else i need to reword my message to get my thought across,
Sorry for my bad communication, I was fine with doing what you wanted. That's what I meant with
almost there
llvm/test/tools/llvm-reduce/no-replace-intrinsic-callee-with-undef.ll | ||
---|---|---|
6 | this doesn't contain any operand bundles, so the pass shouldn't be necessary | |
llvm/test/tools/llvm-reduce/remove-invoked-functions.ll | ||
1 | I think the following is cleaner: diff --git a/llvm/test/tools/llvm-reduce/remove-invoked-functions.ll b/llvm/test/tools/llvm-reduce/remove-invoked-functions.ll index e1224b99e176..0d3f8f87fd4a 100644 --- a/llvm/test/tools/llvm-reduce/remove-invoked-functions.ll +++ b/llvm/test/tools/llvm-reduce/remove-invoked-functions.ll @@ -1,4 +1,4 @@ -; RUN: llvm-reduce --delta-passes=functions,function-bodies,arguments,attributes --test FileCheck --test-arg --check-prefixes=CHECK-ALL,CHECK-INTERESTINGNESS --test-arg %s --test-arg --input-file %s -o %t +; RUN: llvm-reduce --test FileCheck --test-arg --check-prefixes=CHECK-ALL,CHECK-INTERESTINGNESS --test-arg %s --test-arg --input-file %s -o %t ; RUN: cat %t | FileCheck --check-prefixes=CHECK-ALL,CHECK-FINAL %s ; CHECK-INTERESTINGNESS: define i32 @maybe_throwing_callee( @@ -22,8 +22,8 @@ declare void @thrown() define void @caller(i32 %arg) personality i8* bitcast (i32 (...)* @__gxx_personality_v0 to i8*) { ; CHECK-ALL: bb: bb: -; CHECK-INTERESTINGNESS: %i0 = invoke i32 -; CHECK-FINAL: %i0 = invoke i32 bitcast (i32 ()* @maybe_throwing_callee to i32 (i32)*)(i32 %arg) +; CHECK-INTERESTINGNESS: %i0 = invoke i32 {{.*}}@maybe_throwing_callee +; CHECK-FINAL: %i0 = invoke i32 bitcast (i32 ()* @maybe_throwing_callee to i32 (i32)*) ; CHECK-ALL: to label %bb3 unwind label %bb1 %i0 = invoke i32 @maybe_throwing_callee(i32 %arg) to label %bb3 unwind label %bb1 | |
llvm/tools/llvm-reduce/deltas/ReduceOperands.cpp | ||
17 | do we need this? | |
19 | do we need this? | |
23 | undef | |
37 | we actually can replace token values with an undef e.g. declare void @llvm.foo(token) define void @f() { call void @llvm.foo(token undef) ret void } verifies |
this doesn't contain any operand bundles, so the pass shouldn't be necessary