[ValueTracking] Add support for llvm.vector.reduce.{xor,or,and} ops.
#88320
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
goldsteinn
changed the title
llvm.vector.reduce.* ops.
goldsteinn
force-pushed
the
goldsteinn/vec-reduce-support
branch
from
ec8857a
to
4e949bd
Compare
|
The missing intrins are |
dtcxzyw
reviewed
Apr 11, 2024
llvm/lib/Analysis/ValueTracking.cpp
Outdated
| dyn_cast<FixedVectorType>(I->getOperand(0)->getType())) { | ||
| computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | ||
| KnownBits SingleKnown = Known; | ||
| for (unsigned i = 1, e = VecTy->getNumElements(); i < e; ++i) { |
There was a problem hiding this comment.
You can reduce the computational complexity from O(n) to O(lgn) with exponentiation by squaring.
There was a problem hiding this comment.
yeah, I figured most likely N isn't too big here and this isn't a particularly hot path so it wasn't worth to code complexity, but no real strong opinions.
nikic
reviewed
Apr 11, 2024
llvm/lib/Analysis/ValueTracking.cpp
Outdated
| case Intrinsic::vector_reduce_fminimum: | ||
| computeKnownFPClass(II->getArgOperand(0), Known, InterestedClasses, | ||
| Depth + 1, Q); | ||
| break; |
There was a problem hiding this comment.
Can you please split this into a separate PR for review by FP folks? I'm not sure things are quite as simple as that (e.g., do we need to account for denormal flush like the scalar min/max handling above)?
There was a problem hiding this comment.
Done, see: #88408
llvm/lib/Analysis/ValueTracking.cpp
Outdated
| @@ -2904,6 +2942,41 @@ static bool isKnownNonZeroFromOperator(const Operator *I, | |||
| case Intrinsic::vector_reduce_smax: | |||
| case Intrinsic::vector_reduce_smin: | |||
| return isKnownNonZero(II->getArgOperand(0), Depth, Q); | |||
There was a problem hiding this comment.
Can't we easily handle "or" reduction for isKnownNonZero()?
llvm/lib/Analysis/ValueTracking.cpp
Outdated
| break; | ||
| return isKnownNonZero(II->getArgOperand(0), Depth, Q); | ||
| } | ||
| break; |
There was a problem hiding this comment.
Yeah, I don't think this is worthwhile.
goldsteinn
force-pushed
the
goldsteinn/vec-reduce-support
branch
from
4e949bd
to
95dbc18
Compare
|
@llvm/pr-subscribers-llvm-transforms @llvm/pr-subscribers-llvm-analysis Author: None (goldsteinn) Changes
Full diff: https://github.com/llvm/llvm-project/pull/88320.diff 3 Files Affected:
diff --git a/llvm/lib/Analysis/ValueTracking.cpp b/llvm/lib/Analysis/ValueTracking.cpp
index 3a10de72a27562..4e9d446ddb1807 100644
--- a/llvm/lib/Analysis/ValueTracking.cpp
+++ b/llvm/lib/Analysis/ValueTracking.cpp
@@ -1621,14 +1621,32 @@ static void computeKnownBitsFromOperator(const Operator *I,
computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
Known = KnownBits::ssub_sat(Known, Known2);
break;
- // for min/max reduce, any bit common to each element in the input vec
- // is set in the output.
+ // for min/max/and/or reduce, any bit common to each element in the
+ // input vec is set in the output.
+ case Intrinsic::vector_reduce_and:
+ case Intrinsic::vector_reduce_or:
case Intrinsic::vector_reduce_umax:
case Intrinsic::vector_reduce_umin:
case Intrinsic::vector_reduce_smax:
case Intrinsic::vector_reduce_smin:
computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
break;
+ case Intrinsic::vector_reduce_xor:
+ computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
+ // The zeros common to all vecs are zero in the output.
+ // If the number of elements is odd, then the common ones remain. If the
+ // number of elements is even, then the common ones becomes zeros.
+ if (auto *VecTy =
+ dyn_cast<FixedVectorType>(I->getOperand(0)->getType())) {
+ // Even, so the ones become zeros.
+ if ((VecTy->getNumElements() % 2) == 0) {
+ Known.Zero |= Known.One;
+ Known.One.clearAllBits();
+ }
+
+ } else
+ Known.One.clearAllBits();
+ break;
case Intrinsic::umin:
computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
@@ -2898,7 +2916,8 @@ static bool isKnownNonZeroFromOperator(const Operator *I,
return isNonZeroAdd(DemandedElts, Depth, Q, BitWidth,
II->getArgOperand(0), II->getArgOperand(1),
/*NSW=*/true, /* NUW=*/false);
- // umin/smin/smax/smin of all non-zero elements is always non-zero.
+ // umin/smin/smax/smin/or of all non-zero elements is always non-zero.
+ case Intrinsic::vector_reduce_or:
case Intrinsic::vector_reduce_umax:
case Intrinsic::vector_reduce_umin:
case Intrinsic::vector_reduce_smax:
diff --git a/llvm/test/Transforms/InstCombine/known-bits.ll b/llvm/test/Transforms/InstCombine/known-bits.ll
index d210b19bb7faf2..6c3701a32f227f 100644
--- a/llvm/test/Transforms/InstCombine/known-bits.ll
+++ b/llvm/test/Transforms/InstCombine/known-bits.ll
@@ -999,5 +999,130 @@ define i1 @extract_value_smul_fail(i8 %xx, i8 %yy) {
ret i1 %r
}
+define i8 @known_reduce_or(<2 x i8> %xx) {
+; CHECK-LABEL: @known_reduce_or(
+; CHECK-NEXT: ret i8 1
+;
+ %x = or <2 x i8> %xx, <i8 5, i8 3>
+ %v = call i8 @llvm.vector.reduce.or(<2 x i8> %x)
+ %r = and i8 %v, 1
+ ret i8 %r
+}
+
+define i8 @known_reduce_or_fail(<2 x i8> %xx) {
+; CHECK-LABEL: @known_reduce_or_fail(
+; CHECK-NEXT: [[X:%.*]] = or <2 x i8> [[XX:%.*]], <i8 5, i8 3>
+; CHECK-NEXT: [[V:%.*]] = call i8 @llvm.vector.reduce.or.v2i8(<2 x i8> [[X]])
+; CHECK-NEXT: [[R:%.*]] = and i8 [[V]], 4
+; CHECK-NEXT: ret i8 [[R]]
+;
+ %x = or <2 x i8> %xx, <i8 5, i8 3>
+ %v = call i8 @llvm.vector.reduce.or(<2 x i8> %x)
+ %r = and i8 %v, 4
+ ret i8 %r
+}
+
+define i8 @known_reduce_and(<2 x i8> %xx) {
+; CHECK-LABEL: @known_reduce_and(
+; CHECK-NEXT: ret i8 1
+;
+ %x = or <2 x i8> %xx, <i8 5, i8 3>
+ %v = call i8 @llvm.vector.reduce.or(<2 x i8> %x)
+ %r = and i8 %v, 1
+ ret i8 %r
+}
+
+define i8 @known_reduce_and_fail(<2 x i8> %xx) {
+; CHECK-LABEL: @known_reduce_and_fail(
+; CHECK-NEXT: [[X:%.*]] = or <2 x i8> [[XX:%.*]], <i8 5, i8 3>
+; CHECK-NEXT: [[V:%.*]] = call i8 @llvm.vector.reduce.or.v2i8(<2 x i8> [[X]])
+; CHECK-NEXT: [[R:%.*]] = and i8 [[V]], 2
+; CHECK-NEXT: ret i8 [[R]]
+;
+ %x = or <2 x i8> %xx, <i8 5, i8 3>
+ %v = call i8 @llvm.vector.reduce.or(<2 x i8> %x)
+ %r = and i8 %v, 2
+ ret i8 %r
+}
+
+define i8 @known_reduce_xor_even(<2 x i8> %xx) {
+; CHECK-LABEL: @known_reduce_xor_even(
+; CHECK-NEXT: ret i8 0
+;
+ %x = or <2 x i8> %xx, <i8 5, i8 3>
+ %v = call i8 @llvm.vector.reduce.xor(<2 x i8> %x)
+ %r = and i8 %v, 1
+ ret i8 %r
+}
+
+define i8 @known_reduce_xor_even2(<2 x i8> %xx) {
+; CHECK-LABEL: @known_reduce_xor_even2(
+; CHECK-NEXT: ret i8 0
+;
+ %x = and <2 x i8> %xx, <i8 15, i8 15>
+ %v = call i8 @llvm.vector.reduce.xor(<2 x i8> %x)
+ %r = and i8 %v, 16
+ ret i8 %r
+}
+
+define i8 @known_reduce_xor_even_fail(<2 x i8> %xx) {
+; CHECK-LABEL: @known_reduce_xor_even_fail(
+; CHECK-NEXT: [[X:%.*]] = or <2 x i8> [[XX:%.*]], <i8 5, i8 3>
+; CHECK-NEXT: [[V:%.*]] = call i8 @llvm.vector.reduce.xor.v2i8(<2 x i8> [[X]])
+; CHECK-NEXT: [[R:%.*]] = and i8 [[V]], 2
+; CHECK-NEXT: ret i8 [[R]]
+;
+ %x = or <2 x i8> %xx, <i8 5, i8 3>
+ %v = call i8 @llvm.vector.reduce.xor(<2 x i8> %x)
+ %r = and i8 %v, 2
+ ret i8 %r
+}
+
+define i8 @known_reduce_xor_odd(<3 x i8> %xx) {
+; CHECK-LABEL: @known_reduce_xor_odd(
+; CHECK-NEXT: ret i8 1
+;
+ %x = or <3 x i8> %xx, <i8 5, i8 3, i8 9>
+ %v = call i8 @llvm.vector.reduce.xor.v3i8(<3 x i8> %x)
+ %r = and i8 %v, 1
+ ret i8 %r
+}
+
+define i8 @known_reduce_xor_odd2(<3 x i8> %xx) {
+; CHECK-LABEL: @known_reduce_xor_odd2(
+; CHECK-NEXT: ret i8 0
+;
+ %x = and <3 x i8> %xx, <i8 15, i8 15, i8 31>
+ %v = call i8 @llvm.vector.reduce.xor.v3i8(<3 x i8> %x)
+ %r = and i8 %v, 32
+ ret i8 %r
+}
+
+define i8 @known_reduce_xor_odd2_fail(<3 x i8> %xx) {
+; CHECK-LABEL: @known_reduce_xor_odd2_fail(
+; CHECK-NEXT: [[X:%.*]] = and <3 x i8> [[XX:%.*]], <i8 15, i8 15, i8 31>
+; CHECK-NEXT: [[V:%.*]] = call i8 @llvm.vector.reduce.xor.v3i8(<3 x i8> [[X]])
+; CHECK-NEXT: [[R:%.*]] = and i8 [[V]], 16
+; CHECK-NEXT: ret i8 [[R]]
+;
+ %x = and <3 x i8> %xx, <i8 15, i8 15, i8 31>
+ %v = call i8 @llvm.vector.reduce.xor.v3i8(<3 x i8> %x)
+ %r = and i8 %v, 16
+ ret i8 %r
+}
+
+define i8 @known_reduce_xor_odd_fail(<3 x i8> %xx) {
+; CHECK-LABEL: @known_reduce_xor_odd_fail(
+; CHECK-NEXT: [[X:%.*]] = or <3 x i8> [[XX:%.*]], <i8 5, i8 3, i8 9>
+; CHECK-NEXT: [[V:%.*]] = call i8 @llvm.vector.reduce.xor.v3i8(<3 x i8> [[X]])
+; CHECK-NEXT: [[R:%.*]] = and i8 [[V]], 2
+; CHECK-NEXT: ret i8 [[R]]
+;
+ %x = or <3 x i8> %xx, <i8 5, i8 3, i8 9>
+ %v = call i8 @llvm.vector.reduce.xor.v3i8(<3 x i8> %x)
+ %r = and i8 %v, 2
+ ret i8 %r
+}
+
declare void @use(i1)
declare void @sink(i8)
diff --git a/llvm/test/Transforms/InstSimplify/known-non-zero.ll b/llvm/test/Transforms/InstSimplify/known-non-zero.ll
index d9b8f5eed32390..fd2862eb04a24d 100644
--- a/llvm/test/Transforms/InstSimplify/known-non-zero.ll
+++ b/llvm/test/Transforms/InstSimplify/known-non-zero.ll
@@ -377,3 +377,26 @@ define <2 x i1> @insert_nonzero_any_idx_fail(<2 x i8> %xx, i8 %yy, i32 %idx) {
%r = icmp eq <2 x i8> %ins, zeroinitializer
ret <2 x i1> %r
}
+
+define i1 @nonzero_reduce_or(<2 x i8> %xx) {
+; CHECK-LABEL: @nonzero_reduce_or(
+; CHECK-NEXT: ret i1 false
+;
+ %x = add nuw <2 x i8> %xx, <i8 1, i8 1>
+ %v = call i8 @llvm.vector.reduce.or(<2 x i8> %x)
+ %r = icmp eq i8 %v, 0
+ ret i1 %r
+}
+
+define i1 @nonzero_reduce_or_fail(<2 x i8> %xx) {
+; CHECK-LABEL: @nonzero_reduce_or_fail(
+; CHECK-NEXT: [[X:%.*]] = add nsw <2 x i8> [[XX:%.*]], <i8 1, i8 1>
+; CHECK-NEXT: [[V:%.*]] = call i8 @llvm.vector.reduce.or.v2i8(<2 x i8> [[X]])
+; CHECK-NEXT: [[R:%.*]] = icmp eq i8 [[V]], 0
+; CHECK-NEXT: ret i1 [[R]]
+;
+ %x = add nsw <2 x i8> %xx, <i8 1, i8 1>
+ %v = call i8 @llvm.vector.reduce.or(<2 x i8> %x)
+ %r = icmp eq i8 %v, 0
+ ret i1 %r
+}
|
goldsteinn
changed the title
llvm.vector.reduce.* ops.llvm.vector.reduce.{xor,or,and} ops.
goldsteinn
changed the title
llvm.vector.reduce.{xor,or,and} ops.llvm.vector.reduce.{xor,or,and} ops.
arsenm
reviewed
Apr 12, 2024
llvm/lib/Analysis/ValueTracking.cpp
Outdated
Comment on lines
1639
to
1640
| if (auto *VecTy = | ||
| dyn_cast<FixedVectorType>(I->getOperand(0)->getType())) { |
There was a problem hiding this comment.
Can also handle scalable if you need to know an even count?
There was a problem hiding this comment.
How? IIUC the only thing we know about scalable is that sizeof(vec) == N * sizeof(scalar) where N >= MinCount. How can we get odd/even?
There was a problem hiding this comment.
It's <scale * N * element>, if you know N is even it doesn't matter what scale is
There was a problem hiding this comment.
Ah, thanks. Done.
goldsteinn
force-pushed
the
goldsteinn/vec-reduce-support
branch
from
95dbc18
to
ed99b2a
Compare
nikic
reviewed
Apr 13, 2024
llvm/lib/Analysis/ValueTracking.cpp
Outdated
| // number of elements is even, then the common ones becomes zeros. | ||
| auto *VecTy = cast<VectorType>(I->getOperand(0)->getType()); | ||
| // Even, so the ones become zeros. | ||
| if (VecTy->getElementCount().isKnownEven()) { |
There was a problem hiding this comment.
Shouldn't this have an else that sets the ones to unknown? Otherwise what about something like vscale x 3 which is not known even but might be.
(Actually, I'd just condition the whole xor code on isKnownEven, we don't care about odd-size vectors anyway.)
There was a problem hiding this comment.
Yeah, you're right. Pushed fix/test.
Its very minimal code complexity to handle odd, so left it in for now (although think I agree its moreso for completeness than practical applicability). LMK if you feel strongly about dropping.
…ce.{or,and}`; NFC
goldsteinn
force-pushed
the
goldsteinn/vec-reduce-support
branch
from
ed99b2a
to
c6ed849
Compare
nikic
approved these changes
Apr 13, 2024
| // If the number of elements is odd, then the common ones remain. If the | ||
| // number of elements is even, then the common ones becomes zeros. | ||
| auto *VecTy = cast<VectorType>(I->getOperand(0)->getType()); | ||
| // Even, so the ones become zeros. |
There was a problem hiding this comment.
Redundant with the preceding comment?
There was a problem hiding this comment.
Guess the idea is the top comment explains what the block will do. The lower comments explain what the next few lines will do. With this minimal code though yeah it does become a bit redundant, but comments are like sex so...
bazuzi
pushed a commit
to bazuzi/llvm-project
that referenced
this pull request
Apr 15, 2024
aniplcc
pushed a commit
to aniplcc/llvm-project
that referenced
this pull request
Apr 15, 2024
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
computeKnownBitsofllvm.vector.reduce.{or,and}; NFCcomputeKnownBitsforllvm.vector.reduce.{or,and}computeKnownBitsofllvm.vector.reduce.xor; NFCcomputeKnownBitsforllvm.vector.reduce.xorisKnownNonZeroofllvm.vector.reduce.or; NFCisKnownNonZeroforllvm.vector.reduce.or