[Attributor] Propagate alignment through ptrmask #150158
There are no files selected for viewing
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -5203,6 +5203,32 @@ static unsigned getKnownAlignForUse(Attributor &A, AAAlign &QueryingAA, | |
| TrackUse = true; | ||
| return 0; | ||
| } | ||
| if (auto *II = dyn_cast<IntrinsicInst>(I)) { | ||
| if (II->getIntrinsicID() == Intrinsic::ptrmask) { | ||
| auto *ConstVals = A.getAAFor<AAPotentialConstantValues>( | ||
| QueryingAA, IRPosition::value(*(II->getOperand(1))), | ||
| DepClassTy::NONE); | ||
| const AAAlign *AlignAA = A.getAAFor<AAAlign>( | ||
| QueryingAA, IRPosition::value(*(II)), DepClassTy::NONE); | ||
| if (ConstVals && ConstVals->isValidState()) { | ||
| if (ConstVals->isAtFixpoint()) { | ||
| uint64_t TrailingZeros = 64; | ||
| for (const auto &It : ConstVals->getAssumedSet()) | ||
| if (It.countTrailingZeros() < TrailingZeros) | ||
| TrailingZeros = It.countTrailingZeros(); | ||
| if (TrailingZeros < 64) { | ||
|
||
| uint64_t Mask = 1 << TrailingZeros; | ||
| if (Mask >= AlignAA->getKnownAlign().value()) { | ||
| return 0; | ||
| } | ||
| } | ||
| return AlignAA->getKnownAlign().value(); | ||
| } | ||
| } else if (AlignAA) { | ||
| return AlignAA->getKnownAlign().value(); | ||
| } | ||
| } | ||
| } | ||
|
|
||
| MaybeAlign MA; | ||
| if (const auto *CB = dyn_cast<CallBase>(I)) { | ||
|
|
@@ -5502,6 +5528,42 @@ struct AAAlignCallSiteReturned final | |
| AAAlignCallSiteReturned(const IRPosition &IRP, Attributor &A) | ||
| : Base(IRP, A) {} | ||
|
|
||
| ChangeStatus updateImpl(Attributor &A) override { | ||
| Instruction *I = getIRPosition().getCtxI(); | ||
shiltian marked this conversation as resolved.
Show resolved
Hide resolved
|
||
| if (auto *II = dyn_cast<IntrinsicInst>(I)) { | ||
| if (II->getIntrinsicID() == Intrinsic::ptrmask) { | ||
| const AAPotentialConstantValues *ConstVals = | ||
| A.getAAFor<AAPotentialConstantValues>( | ||
| *this, IRPosition::value(*(II->getOperand(1))), | ||
| DepClassTy::REQUIRED); | ||
| const AAAlign *AlignAA = | ||
| A.getAAFor<AAAlign>(*this, IRPosition::value(*(II->getOperand(0))), | ||
| DepClassTy::REQUIRED); | ||
| uint64_t Alignment = 0; | ||
| if (ConstVals && ConstVals->isValidState()) { | ||
| unsigned TrailingZeros = 64; | ||
|
||
| for (const auto &It : ConstVals->getAssumedSet()) | ||
| if (It.countTrailingZeros() < TrailingZeros) | ||
| TrailingZeros = It.countTrailingZeros(); | ||
| if (TrailingZeros < 64) | ||
| Alignment = 1 << TrailingZeros; | ||
| } | ||
| if (AlignAA && AlignAA->isValidState() && | ||
| Alignment < AlignAA->getAssumedAlign().value()) | ||
| Alignment = AlignAA->getAssumedAlign().value(); | ||
|
|
||
| if (Alignment != 0) { | ||
| uint64_t OldAssumed = getAssumed(); | ||
| takeAssumedMinimum(Alignment); | ||
| return OldAssumed == getAssumed() ? ChangeStatus::UNCHANGED | ||
| : ChangeStatus::CHANGED; | ||
| } else { | ||
|
||
| return ChangeStatus::UNCHANGED; | ||
| } | ||
| } | ||
| } | ||
| return Base::updateImpl(A); | ||
|
There was a problem hiding this comment. I'm not sure if we want to simply give up here if it is an intrinsic call and There was a problem hiding this comment. Hi @shiltian , same as known alignment I cannot tell from the return value that if it is correctly handled. But what is different here is if it is intrinsic, the call will be equivalently handled as call to deceleration, which will create a pessimistic alignment attribute for the caller. so the following handle can be ignored and directly return here. There was a problem hiding this comment. You can compare with There was a problem hiding this comment. Hi @shiltian , There was a problem hiding this comment. I'm lost here. There was a problem hiding this comment. Hi @shiltian , There was a problem hiding this comment. I'm so confused. Inside There was a problem hiding this comment. Hi @shiltian , if the return is However if I'm not wrong, let it continue will simply return the alignment of the intrinsic. Since intrinsics are function declaration, so it will initialize but not update and directly indicate pessimistic and make the "assumed = known". May be we can implement the logic inside There was a problem hiding this comment. When it returns I understand for now it will not make any difference but I'd just like to be a little bit future proof here. |
||
| }; | ||
| /// See AbstractAttribute::trackStatistics() | ||
| void trackStatistics() const override { STATS_DECLTRACK_CS_ATTR(align); } | ||
| }; | ||
|
|
||
| Original file line number | Diff line number | Diff line change | ||||
|---|---|---|---|---|---|---|
| @@ -0,0 +1,101 @@ | ||||||
| ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5 | ||||||
| ; RUN: opt -passes=attributor -S < %s | FileCheck %s | ||||||
|
|
||||||
| define float @align_ptrmask_back_no_prop(ptr align 2 %x, i1 %cmp1, i1 %cmp2) { | ||||||
| ; CHECK-LABEL: define float @align_ptrmask_back_no_prop( | ||||||
| ; CHECK-SAME: ptr nofree readonly align 2 [[X:%.*]], i1 [[CMP1:%.*]], i1 [[CMP2:%.*]]) #[[ATTR0:[0-9]+]] { | ||||||
| ; CHECK-NEXT: [[SEL:%.*]] = select i1 [[CMP1]], i64 -32, i64 -8 | ||||||
| ; CHECK-NEXT: [[SEL1:%.*]] = select i1 [[CMP2]], i64 [[SEL]], i64 -16 | ||||||
| ; CHECK-NEXT: [[P:%.*]] = tail call align 8 ptr @llvm.ptrmask.p0.i64(ptr [[X]], i64 noundef [[SEL1]]) #[[ATTR2:[0-9]+]] | ||||||
| ; CHECK-NEXT: [[RES:%.*]] = load float, ptr [[P]], align 8, !invariant.load [[META0:![0-9]+]] | ||||||
| ; CHECK-NEXT: ret float [[RES]] | ||||||
| ; | ||||||
| %sel = select i1 %cmp1, i64 -32, i64 -8 | ||||||
| %sel1 = select i1 %cmp2, i64 %sel, i64 -16 | ||||||
| %p = tail call ptr @llvm.ptrmask(ptr %x, i64 %sel1) | ||||||
| %res = load float, ptr %p, align 8 | ||||||
| ret float %res | ||||||
| } | ||||||
|
|
||||||
| define float @align_ptrmask_back_prop(ptr align 2 %x, i1 %cmp1, i1 %cmp2) { | ||||||
| ; CHECK-LABEL: define float @align_ptrmask_back_prop( | ||||||
| ; CHECK-SAME: ptr nofree readonly align 16 [[X:%.*]], i1 [[CMP1:%.*]], i1 [[CMP2:%.*]]) #[[ATTR0]] { | ||||||
| ; CHECK-NEXT: [[SEL:%.*]] = select i1 [[CMP1]], i64 -32, i64 -8 | ||||||
| ; CHECK-NEXT: [[SEL1:%.*]] = select i1 [[CMP2]], i64 [[SEL]], i64 -16 | ||||||
| ; CHECK-NEXT: [[P:%.*]] = tail call align 16 ptr @llvm.ptrmask.p0.i64(ptr [[X]], i64 noundef [[SEL1]]) #[[ATTR2]] | ||||||
| ; CHECK-NEXT: [[RES:%.*]] = load float, ptr [[P]], align 16, !invariant.load [[META0]] | ||||||
| ; CHECK-NEXT: ret float [[RES]] | ||||||
| ; | ||||||
| %sel = select i1 %cmp1, i64 -32, i64 -8 | ||||||
| %sel1 = select i1 %cmp2, i64 %sel, i64 -16 | ||||||
| %p = tail call ptr @llvm.ptrmask(ptr %x, i64 %sel1) | ||||||
| %res = load float, ptr %p, align 16 | ||||||
| ret float %res | ||||||
| } | ||||||
|
|
||||||
| define float @align_ptrmask_forward_mask(ptr align 2 %x, i1 %cmp1, i1 %cmp2) { | ||||||
| ; CHECK-LABEL: define float @align_ptrmask_forward_mask( | ||||||
| ; CHECK-SAME: ptr nofree readonly align 2 [[X:%.*]], i1 [[CMP1:%.*]], i1 [[CMP2:%.*]]) #[[ATTR0]] { | ||||||
| ; CHECK-NEXT: [[SEL:%.*]] = select i1 [[CMP1]], i64 -32, i64 -8 | ||||||
| ; CHECK-NEXT: [[SEL1:%.*]] = select i1 [[CMP2]], i64 [[SEL]], i64 -16 | ||||||
| ; CHECK-NEXT: [[P:%.*]] = tail call align 8 ptr @llvm.ptrmask.p0.i64(ptr [[X]], i64 noundef [[SEL1]]) #[[ATTR2]] | ||||||
| ; CHECK-NEXT: [[RES:%.*]] = load float, ptr [[P]], align 8, !invariant.load [[META0]] | ||||||
| ; CHECK-NEXT: ret float [[RES]] | ||||||
| ; | ||||||
| %sel = select i1 %cmp1, i64 -32, i64 -8 | ||||||
| %sel1 = select i1 %cmp2, i64 %sel, i64 -16 | ||||||
| %p = tail call ptr @llvm.ptrmask(ptr %x, i64 %sel1) | ||||||
| %res = load float, ptr %p, align 4 | ||||||
| ret float %res | ||||||
| } | ||||||
|
|
||||||
| define float @align_ptrmask_forward_ptr(ptr align 16 %x, i1 %cmp1, i1 %cmp2) { | ||||||
| ; CHECK-LABEL: define float @align_ptrmask_forward_ptr( | ||||||
| ; CHECK-SAME: ptr nofree readonly align 16 [[X:%.*]], i1 [[CMP1:%.*]], i1 [[CMP2:%.*]]) #[[ATTR0]] { | ||||||
| ; CHECK-NEXT: [[SEL:%.*]] = select i1 [[CMP1]], i64 -32, i64 -8 | ||||||
| ; CHECK-NEXT: [[SEL1:%.*]] = select i1 [[CMP2]], i64 [[SEL]], i64 -16 | ||||||
| ; CHECK-NEXT: [[P:%.*]] = tail call align 16 ptr @llvm.ptrmask.p0.i64(ptr [[X]], i64 noundef [[SEL1]]) #[[ATTR2]] | ||||||
| ; CHECK-NEXT: [[RES:%.*]] = load float, ptr [[P]], align 16, !invariant.load [[META0]] | ||||||
| ; CHECK-NEXT: ret float [[RES]] | ||||||
| ; | ||||||
| %sel = select i1 %cmp1, i64 -32, i64 -8 | ||||||
| %sel1 = select i1 %cmp2, i64 %sel, i64 -16 | ||||||
| %p = tail call ptr @llvm.ptrmask(ptr %x, i64 %sel1) | ||||||
| %res = load float, ptr %p, align 4 | ||||||
| ret float %res | ||||||
| } | ||||||
|
|
||||||
| define float @align_ptrmask_forward_nonconst_mask(ptr align 8 %x, i64 %y, i1 %cmp1, i1 %cmp2) { | ||||||
| ; CHECK-LABEL: define float @align_ptrmask_forward_nonconst_mask( | ||||||
| ; CHECK-SAME: ptr nofree readonly align 8 [[X:%.*]], i64 [[Y:%.*]], i1 [[CMP1:%.*]], i1 [[CMP2:%.*]]) #[[ATTR0]] { | ||||||
| ; CHECK-NEXT: [[SEL:%.*]] = select i1 [[CMP1]], i64 -32, i64 [[Y]] | ||||||
| ; CHECK-NEXT: [[SEL1:%.*]] = select i1 [[CMP2]], i64 [[SEL]], i64 -16 | ||||||
| ; CHECK-NEXT: [[P:%.*]] = tail call align 8 ptr @llvm.ptrmask.p0.i64(ptr [[X]], i64 [[SEL1]]) #[[ATTR2]] | ||||||
| ; CHECK-NEXT: [[RES:%.*]] = load float, ptr [[P]], align 8, !invariant.load [[META0]] | ||||||
| ; CHECK-NEXT: ret float [[RES]] | ||||||
| ; | ||||||
| %sel = select i1 %cmp1, i64 -32, i64 %y | ||||||
| %sel1 = select i1 %cmp2, i64 %sel, i64 -16 | ||||||
| %p = tail call ptr @llvm.ptrmask(ptr %x, i64 %sel1) | ||||||
| %res = load float, ptr %p, align 4 | ||||||
| ret float %res | ||||||
|
||||||
| } | ||||||
|
|
||||||
| define float @align_ptrmask_back_nonconst_mask(ptr align 4 %x, i64 %y, i1 %cmp1, i1 %cmp2) { | ||||||
| ; CHECK-LABEL: define float @align_ptrmask_back_nonconst_mask( | ||||||
| ; CHECK-SAME: ptr nofree readonly align 8 [[X:%.*]], i64 [[Y:%.*]], i1 [[CMP1:%.*]], i1 [[CMP2:%.*]]) #[[ATTR0]] { | ||||||
| ; CHECK-NEXT: [[SEL:%.*]] = select i1 [[CMP1]], i64 -32, i64 [[Y]] | ||||||
| ; CHECK-NEXT: [[SEL1:%.*]] = select i1 [[CMP2]], i64 [[SEL]], i64 -16 | ||||||
| ; CHECK-NEXT: [[P:%.*]] = tail call align 8 ptr @llvm.ptrmask.p0.i64(ptr [[X]], i64 [[SEL1]]) #[[ATTR2]] | ||||||
| ; CHECK-NEXT: [[RES:%.*]] = load float, ptr [[P]], align 8, !invariant.load [[META0]] | ||||||
| ; CHECK-NEXT: ret float [[RES]] | ||||||
| ; | ||||||
| %sel = select i1 %cmp1, i64 -32, i64 %y | ||||||
| %sel1 = select i1 %cmp2, i64 %sel, i64 -16 | ||||||
| %p = tail call ptr @llvm.ptrmask(ptr %x, i64 %sel1) | ||||||
|
||||||
| %p = tail call ptr @llvm.ptrmask(ptr %x, i64 %sel1) | |
| %p = tail call ptr @llvm.ptrmask.p0.i64(ptr %x, i64 %sel1) |
Throughout
There was a problem hiding this comment.
Can you add tests with these degenerate cases:
- Input mask value is a small positive constant, such that it cannot increase alignment.
- Mask that exactly matches maximum alignment, i.e. -4294967296
- Mask with 1 more bit than maximum alignment, i.e. -8589934592
- Poison mask
- ptrmask with align attributes on callsite
Outdated
There was a problem hiding this comment.
Test cases with constant values, which is the common case. Also test some unknowable cases. Also vectors
There was a problem hiding this comment.
Hi @arsenm , I'm not sure if I can find a vector case. The current frame work seems stop it from generating AAAlign attribute for vector case:
Also extractelement cannot propagate alignment, so I also cannot use store/load to test
There was a problem hiding this comment.
extractelement can propagate alignment, whether or not it does currently is another question
There was a problem hiding this comment.
Hi @arsenm , yep, sorry for how I said it. The current framework seems not supporting propagate alignment through extractelement. Would you like me to add simple support in this PR in order to add vector tests??
There was a problem hiding this comment.
No, add the vector test that shows it doesn't propagate for later improvement
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Can you split the intrinsic case into a separate function, and switch over the intrinsic ID