[DAG] add (~a | x) & (a | y) -> (a & (x ^ y)) ^y for foldMaskedMerge #144342
Conversation
This file contains hidden or 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
|
@llvm/pr-subscribers-llvm-selectiondag @llvm/pr-subscribers-backend-x86 Author: woruyu (woruyu) ChangesSummaryThis PR resolves #143864 Add (~a | x) & (a | y) -> (a & (x ^ y)) ^y for foldMaskedMerge func using SDPatternMatch aftering adding this pattern, run Full diff: https://github.com/llvm/llvm-project/pull/144342.diff 2 Files Affected:
diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 5d62ded171f4f..f131822e11468 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -7204,6 +7204,38 @@ static SDValue foldLogicTreeOfShifts(SDNode *N, SDValue LeftHand,
return DAG.getNode(LogicOpcode, DL, VT, CombinedShifts, W);
}
+/// Fold "masked merge" expressions like `(m & x) | (~m & y)` and its DeMorgan
+/// variant `(~m | x) & (m | y)` into the equivalent `((x ^ y) & m) ^ y)`
+/// pattern. This is typically a better representation for targets without a
+/// fused "and-not" operation.
+static SDValue foldMaskedMerge(SDNode *Node, SelectionDAG &DAG,
+ const TargetLowering &TLI, const SDLoc &DL) {
+ // Note that masked-merge variants using XOR or ADD expressions are
+ // normalized to OR by InstCombine so we only check for OR or AND.
+ assert(Node->getOpcode() == ISD::OR ||
+ Node->getOpcode() == ISD::AND &&
+ "Must be called with ISD::OR or ISD::AND node");
+
+ // If the target supports and-not, don't fold this.
+ if (TLI.hasAndNot(SDValue(Node, 0)))
+ return SDValue();
+
+ SDValue M, X, Y;
+
+ if (sd_match(Node,
+ m_Or(m_OneUse(m_And(m_OneUse(m_Not(m_Value(M))), m_Value(Y))),
+ m_OneUse(m_And(m_Deferred(M), m_Value(X))))) ||
+ sd_match(Node, m_And(m_OneUse(m_Or(m_Not(m_Value(M)), m_Value(X))),
+ m_OneUse(m_Or(m_Deferred(M), m_Value(Y)))))) {
+ EVT VT = M.getValueType();
+ SDValue Xor = DAG.getNode(ISD::XOR, DL, VT, X, Y);
+ SDValue And = DAG.getNode(ISD::AND, DL, VT, Xor, M);
+ return DAG.getNode(ISD::XOR, DL, VT, And, Y);
+ }
+
+ return SDValue();
+}
+
SDValue DAGCombiner::visitAND(SDNode *N) {
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
@@ -7644,6 +7676,10 @@ SDValue DAGCombiner::visitAND(SDNode *N) {
if (SDValue R = foldLogicTreeOfShifts(N, N0, N1, DAG))
return R;
+ if (VT.isScalarInteger() && VT != MVT::i1)
+ if (SDValue R = foldMaskedMerge(N, DAG, TLI, DL))
+ return R;
+
return SDValue();
}
@@ -8128,32 +8164,6 @@ static SDValue visitORCommutative(SelectionDAG &DAG, SDValue N0, SDValue N1,
return SDValue();
}
-/// Fold "masked merge" expressions like `(m & x) | (~m & y)` into the
-/// equivalent `((x ^ y) & m) ^ y)` pattern.
-/// This is typically a better representation for targets without a fused
-/// "and-not" operation.
-static SDValue foldMaskedMerge(SDNode *Node, SelectionDAG &DAG,
- const TargetLowering &TLI, const SDLoc &DL) {
- // Note that masked-merge variants using XOR or ADD expressions are
- // normalized to OR by InstCombine so we only check for OR.
- assert(Node->getOpcode() == ISD::OR && "Must be called with ISD::OR node");
-
- // If the target supports and-not, don't fold this.
- if (TLI.hasAndNot(SDValue(Node, 0)))
- return SDValue();
-
- SDValue M, X, Y;
- if (sd_match(Node,
- m_Or(m_OneUse(m_And(m_OneUse(m_Not(m_Value(M))), m_Value(Y))),
- m_OneUse(m_And(m_Deferred(M), m_Value(X)))))) {
- EVT VT = M.getValueType();
- SDValue Xor = DAG.getNode(ISD::XOR, DL, VT, X, Y);
- SDValue And = DAG.getNode(ISD::AND, DL, VT, Xor, M);
- return DAG.getNode(ISD::XOR, DL, VT, And, Y);
- }
- return SDValue();
-}
-
SDValue DAGCombiner::visitOR(SDNode *N) {
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
diff --git a/llvm/test/CodeGen/X86/fold-masked-merge-demorgan.ll b/llvm/test/CodeGen/X86/fold-masked-merge-demorgan.ll
new file mode 100644
index 0000000000000..6357680f7586e
--- /dev/null
+++ b/llvm/test/CodeGen/X86/fold-masked-merge-demorgan.ll
@@ -0,0 +1,267 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -o - %s -mtriple=x86_64-- | FileCheck %s --check-prefixes=CHECK,NOBMI
+; RUN: llc -o - %s -mtriple=x86_64-- -mattr=+bmi | FileCheck %s --check-prefixes=CHECK,BMI
+;
+; test that masked-merge code is generated as "xor;and;xor" sequence or
+; "andn ; and; or" if and-not is available.
+
+define i32 @masked_merge0_demorgan(i32 %a0, i32 %a1, i32 %a2) {
+; NOBMI-LABEL: masked_merge0_demorgan:
+; NOBMI: # %bb.0:
+; NOBMI-NEXT: movl %esi, %eax
+; NOBMI-NEXT: xorl %edx, %eax
+; NOBMI-NEXT: andl %edi, %eax
+; NOBMI-NEXT: xorl %edx, %eax
+; NOBMI-NEXT: retq
+;
+; BMI-LABEL: masked_merge0_demorgan:
+; BMI: # %bb.0:
+; BMI-NEXT: orl %edi, %edx
+; BMI-NEXT: andnl %edi, %esi, %eax
+; BMI-NEXT: andnl %edx, %eax, %eax
+; BMI-NEXT: retq
+ %not = xor i32 %a0, -1
+ %or0 = or i32 %not, %a1
+ %or1 = or i32 %a0, %a2
+ %and = and i32 %or0, %or1
+ ret i32 %and
+}
+
+define i16 @masked_merge1_demorgan(i16 %a0, i16 %a1, i16 %a2) {
+; NOBMI-LABEL: masked_merge1_demorgan:
+; NOBMI: # %bb.0:
+; NOBMI-NEXT: movl %esi, %eax
+; NOBMI-NEXT: xorl %edx, %eax
+; NOBMI-NEXT: andl %edi, %eax
+; NOBMI-NEXT: xorl %edx, %eax
+; NOBMI-NEXT: # kill: def $ax killed $ax killed $eax
+; NOBMI-NEXT: retq
+;
+; BMI-LABEL: masked_merge1_demorgan:
+; BMI: # %bb.0:
+; BMI-NEXT: andnl %edx, %edi, %eax
+; BMI-NEXT: andl %edi, %esi
+; BMI-NEXT: orl %esi, %eax
+; BMI-NEXT: # kill: def $ax killed $ax killed $eax
+; BMI-NEXT: retq
+ %not = xor i16 %a0, -1
+ %or0 = or i16 %not, %a1
+ %or1 = or i16 %a0, %a2
+ %and = and i16 %or0, %or1
+ ret i16 %and
+}
+
+define i8 @masked_merge2_demorgan(i8 %a0, i8 %a1, i8 %a2) {
+; CHECK-LABEL: masked_merge2_demorgan:
+; CHECK: # %bb.0:
+; CHECK-NEXT: movl %esi, %eax
+; CHECK-NEXT: # kill: def $al killed $al killed $eax
+; CHECK-NEXT: retq
+ %not = xor i8 %a0, -1
+ %or0 = or i8 %not, %a1
+ %or1 = or i8 %a0, %a1
+ %and = and i8 %or0, %or1
+ ret i8 %and
+}
+
+define i64 @masked_merge3_demorgan(i64 %a0, i64 %a1, i64 %a2) {
+; NOBMI-LABEL: masked_merge3_demorgan:
+; NOBMI: # %bb.0:
+; NOBMI-NEXT: movq %rsi, %rax
+; NOBMI-NEXT: notq %rdx
+; NOBMI-NEXT: xorq %rdx, %rax
+; NOBMI-NEXT: notq %rax
+; NOBMI-NEXT: andq %rdi, %rax
+; NOBMI-NEXT: xorq %rdx, %rax
+; NOBMI-NEXT: retq
+;
+; BMI-LABEL: masked_merge3_demorgan:
+; BMI: # %bb.0:
+; BMI-NEXT: andnq %rdx, %rdi, %rax
+; BMI-NEXT: andq %rdi, %rsi
+; BMI-NEXT: notq %rsi
+; BMI-NEXT: andnq %rsi, %rax, %rax
+; BMI-NEXT: retq
+ %not_a0 = xor i64 %a0, -1
+ %not_a1 = xor i64 %a1, -1
+ %not_a2 = xor i64 %a2, -1
+ %or0 = or i64 %not_a0, %not_a1
+ %or1 = or i64 %a0, %not_a2
+ %and = and i64 %or0, %or1
+ ret i64 %and
+}
+
+define i32 @not_a_masked_merge0_demorgan(i32 %a0, i32 %a1, i32 %a2) {
+; CHECK-LABEL: not_a_masked_merge0_demorgan:
+; CHECK: # %bb.0:
+; CHECK-NEXT: orl %edi, %edx
+; CHECK-NEXT: movl %edi, %eax
+; CHECK-NEXT: negl %eax
+; CHECK-NEXT: orl %esi, %eax
+; CHECK-NEXT: andl %edx, %eax
+; CHECK-NEXT: retq
+ %not_a_not = sub i32 0, %a0
+ %or0 = or i32 %not_a_not, %a1
+ %or1 = or i32 %a0, %a2
+ %and = and i32 %or0, %or1
+ ret i32 %and
+}
+
+; not a masked merge: `not` operand does not match another `and`-operand.
+define i32 @not_a_masked_merge1_demorgan(i32 %a0, i32 %a1, i32 %a2, i32 %a3) {
+; NOBMI-LABEL: not_a_masked_merge1_demorgan:
+; NOBMI: # %bb.0:
+; NOBMI-NEXT: movl %ecx, %eax
+; NOBMI-NEXT: orl %edx, %edi
+; NOBMI-NEXT: notl %eax
+; NOBMI-NEXT: orl %esi, %eax
+; NOBMI-NEXT: andl %edi, %eax
+; NOBMI-NEXT: retq
+;
+; BMI-LABEL: not_a_masked_merge1_demorgan:
+; BMI: # %bb.0:
+; BMI-NEXT: orl %edx, %edi
+; BMI-NEXT: andnl %ecx, %esi, %eax
+; BMI-NEXT: andnl %edi, %eax, %eax
+; BMI-NEXT: retq
+ %or1 = or i32 %a0, %a2
+ %not = xor i32 %a3, -1
+ %or0 = or i32 %not, %a1
+ %and = and i32 %or0, %or1
+ ret i32 %and
+}
+
+; not a masked merge: one of the operands of `and` is not an `or`.
+define i32 @not_a_masked_merge2_demorgan(i32 %a0, i32 %a1, i32 %a2) {
+; NOBMI-LABEL: not_a_masked_merge2_demorgan:
+; NOBMI: # %bb.0:
+; NOBMI-NEXT: movl %edi, %eax
+; NOBMI-NEXT: andl %edi, %edx
+; NOBMI-NEXT: notl %eax
+; NOBMI-NEXT: orl %esi, %eax
+; NOBMI-NEXT: andl %edx, %eax
+; NOBMI-NEXT: retq
+;
+; BMI-LABEL: not_a_masked_merge2_demorgan:
+; BMI: # %bb.0:
+; BMI-NEXT: andl %edi, %edx
+; BMI-NEXT: andnl %edi, %esi, %eax
+; BMI-NEXT: andnl %edx, %eax, %eax
+; BMI-NEXT: retq
+ %not_an_or1 = and i32 %a0, %a2
+ %not = xor i32 %a0, -1
+ %or0 = or i32 %not, %a1
+ %and = and i32 %or0, %not_an_or1
+ ret i32 %and
+}
+
+define i32 @not_a_masked_merge3_demorgan(i32 %a0, i32 %a1, i32 %a2) {
+; NOBMI-LABEL: not_a_masked_merge3_demorgan:
+; NOBMI: # %bb.0:
+; NOBMI-NEXT: movl %esi, %eax
+; NOBMI-NEXT: orl %edi, %edx
+; NOBMI-NEXT: xorl %edi, %eax
+; NOBMI-NEXT: notl %eax
+; NOBMI-NEXT: andl %edx, %eax
+; NOBMI-NEXT: retq
+;
+; BMI-LABEL: not_a_masked_merge3_demorgan:
+; BMI: # %bb.0:
+; BMI-NEXT: orl %edi, %edx
+; BMI-NEXT: xorl %edi, %esi
+; BMI-NEXT: andnl %edx, %esi, %eax
+; BMI-NEXT: retq
+ %or1 = or i32 %a0, %a2
+ %not = xor i32 %a0, -1
+ %not_an_or0 = xor i32 %not, %a1
+ %and = and i32 %not_an_or0, %or1
+ ret i32 %and
+}
+
+; not a masked merge: `not` operand must not be on same `or`.
+define i32 @not_a_masked_merge4_demorgan(i32 %a0, i32 %a1, i32 %a2) {
+; CHECK-LABEL: not_a_masked_merge4_demorgan:
+; CHECK: # %bb.0:
+; CHECK-NEXT: movl %edi, %eax
+; CHECK-NEXT: orl %edx, %eax
+; CHECK-NEXT: retq
+ %or1 = or i32 %a0, %a2
+ %not = xor i32 %a1, -1
+ %or0 = or i32 %not, %a1
+ %and = and i32 %or0, %or1
+ ret i32 %and
+}
+
+; should not transform when operands have multiple users.
+define i32 @masked_merge_no_transform0_demorgan(i32 %a0, i32 %a1, i32 %a2, ptr %p1) {
+; NOBMI-LABEL: masked_merge_no_transform0_demorgan:
+; NOBMI: # %bb.0:
+; NOBMI-NEXT: orl %edi, %edx
+; NOBMI-NEXT: movl %edi, %eax
+; NOBMI-NEXT: notl %eax
+; NOBMI-NEXT: orl %esi, %eax
+; NOBMI-NEXT: andl %edx, %eax
+; NOBMI-NEXT: movl %edx, (%rcx)
+; NOBMI-NEXT: retq
+;
+; BMI-LABEL: masked_merge_no_transform0_demorgan:
+; BMI: # %bb.0:
+; BMI-NEXT: orl %edi, %edx
+; BMI-NEXT: andnl %edi, %esi, %eax
+; BMI-NEXT: andnl %edx, %eax, %eax
+; BMI-NEXT: movl %edx, (%rcx)
+; BMI-NEXT: retq
+ %not = xor i32 %a0, -1
+ %or0 = or i32 %not, %a1
+ %or1 = or i32 %a0, %a2
+ %and = and i32 %or0, %or1
+ store i32 %or1, ptr %p1
+ ret i32 %and
+}
+
+; should not transform when operands have multiple users.
+define i32 @masked_merge_no_transform1_demorgan(i32 %a0, i32 %a1, i32 %a2, ptr %p1) {
+; NOBMI-LABEL: masked_merge_no_transform1_demorgan:
+; NOBMI: # %bb.0:
+; NOBMI-NEXT: movl %esi, %eax
+; NOBMI-NEXT: xorl %edx, %eax
+; NOBMI-NEXT: andl %edi, %eax
+; NOBMI-NEXT: notl %edi
+; NOBMI-NEXT: xorl %edx, %eax
+; NOBMI-NEXT: movl %edi, (%rcx)
+; NOBMI-NEXT: retq
+;
+; BMI-LABEL: masked_merge_no_transform1_demorgan:
+; BMI: # %bb.0:
+; BMI-NEXT: orl %edi, %edx
+; BMI-NEXT: andnl %edi, %esi, %eax
+; BMI-NEXT: notl %edi
+; BMI-NEXT: andnl %edx, %eax, %eax
+; BMI-NEXT: movl %edi, (%rcx)
+; BMI-NEXT: retq
+ %not = xor i32 %a0, -1
+ %or0 = or i32 %not, %a1
+ %or1 = or i32 %a0, %a2
+ %and = and i32 %or0, %or1
+ store i32 %not, ptr %p1
+ ret i32 %and
+}
+
+; should not transform when operands have multiple users.
+define i32 @masked_merge_no_transform2_demorgan(i32 %a0, i32 %a1, i32 %a2, ptr %p1) {
+; CHECK-LABEL: masked_merge_no_transform2_demorgan:
+; CHECK: # %bb.0:
+; CHECK-NEXT: movl %edx, %eax
+; CHECK-NEXT: orl %edi, %eax
+; CHECK-NEXT: notl %edi
+; CHECK-NEXT: orl %esi, %edi
+; CHECK-NEXT: andl %edi, %eax
+; CHECK-NEXT: movl %edi, (%rcx)
+; CHECK-NEXT: retq
+ %not = xor i32 %a0, -1
+ %or0 = or i32 %not, %a1
+ %or1 = or i32 %a0, %a2
+ %and = and i32 %or0, %or1
+ store i32 %or0, ptr %p1
+ ret i32 %and
+}
\ No newline at end of file
|
|
Based on last experience(#143855), I know these test cases are related to (m & x) | (~m & y) in order to wirte (~a | x) & (a | y) testcase, i pick up fold-masked-merge.ll as reference to write fold-masked-merge-demorgan.ll, any suggestions on testcase coverage here? |
|
@RKSimon, hello, any suggestions for modifications? |
RKSimon
reviewed
Jun 16, 2025
woruyu
force-pushed
the
feat/add-foldMaskedMerge-demorgan
branch
from
8610ae3 to
dd1ad26
Compare
|
@RKSimon , hello, pre-commit patch is in the first commit, codegen changes is on the second. |
|
Thanks @woruyu - I've moved foldMaskedMerge into place so the diff is clearer - please can you merge against trunk? |
ajaden-codes
pushed a commit
to Jaddyen/llvm-project
that referenced
this pull request
Jun 17, 2025
Reduces diff in llvm#144342
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.
Summary
This PR resolves #143864
Add (~a | x) & (a | y) -> (a & (x ^ y)) ^y for foldMaskedMerge func using SDPatternMatch
aftering adding this pattern, run
ninja check-llvm-codegen, all other cases remain unchanged, so I add a testcase(fold-masked-merge-demorgan.ll) for it