[InstCombine] Add ctpop(A | B) + ctpop(A & B) -> ctpop(A) + ctpop(B)

[AtariDreams]

No description provided.

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: AtariDreams (AtariDreams)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/79089.diff

2 Files Affected:

• (modified) llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp (+19)
• (modified) llvm/test/Transforms/InstCombine/ctpop.ll (+15)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp b/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
index 8a00b75a1f74042..d8ef770d6fcef04 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
@@ -1718,6 +1718,25 @@ Instruction *InstCombinerImpl::visitAdd(BinaryOperator &I) {
         Builder.CreateIntrinsic(Intrinsic::umax, {I.getType()}, {A, B}));
   }
 
+  // ctpop(A | B) + ctpop(A & B) -> ctpop(A) + ctpop(B)
+  if ((match(LHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(
+                      m_And(m_Value(A), m_Value(B))))) &&
+       (match(RHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(
+                       m_Or(m_Specific(A), m_Specific(B))))) ||
+        match(RHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(
+                       m_Or(m_Specific(B), m_Specific(A))))))) ||
+      (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(
+                      m_Or(m_Value(A), m_Value(B))))) &&
+       (match(RHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(
+                       m_And(m_Specific(A), m_Specific(B))))) ||
+        match(RHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(
+                       m_And(m_Specific(B), m_Specific(A)))))))) {
+    return replaceInstUsesWith(
+        I, Builder.CreateAdd(
+               Builder.CreateIntrinsic(Intrinsic::ctpop, {A->getType()}, {A}),
+               Builder.CreateIntrinsic(Intrinsic::ctpop, {B->getType()}, {B})));
+  }
+
   // ctpop(A) + ctpop(B) => ctpop(A | B) if A and B have no bits set in common.
   if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(A)))) &&
       match(RHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(B)))) &&
diff --git a/llvm/test/Transforms/InstCombine/ctpop.ll b/llvm/test/Transforms/InstCombine/ctpop.ll
index dcea5fa87479eb4..6dd8faf65a6ff8c 100644
--- a/llvm/test/Transforms/InstCombine/ctpop.ll
+++ b/llvm/test/Transforms/InstCombine/ctpop.ll
@@ -264,6 +264,21 @@ define <2 x i32> @ctpop_add_no_common_bits_vec_use2(<2 x i32> %a, <2 x i32> %b,
   ret <2 x i32> %res
 }
 
+define i32 @ctpop_and_or_combine(i32 %a, i32 %b) {
+; CHECK-LABEL: @ctpop_and_or_combine(
+; CHECK-NEXT:    [[TMP1:%.*]] = call i32 @llvm.ctpop.i32(i32 [[A:%.*]]), !range [[RNG1]]
+; CHECK-NEXT:    [[TMP2:%.*]] = call i32 @llvm.ctpop.i32(i32 [[B:%.*]]), !range [[RNG1]]
+; CHECK-NEXT:    [[RES:%.*]] = add nuw nsw i32 [[TMP1]], [[TMP2]]
+; CHECK-NEXT:    ret i32 [[RES]]
+;
+  %and = and i32 %a, %b
+  %ctpop1 = tail call i32 @llvm.ctpop.i32(i32 %and)
+  %or = or i32 %a, %b
+  %ctpop2 = tail call i32 @llvm.ctpop.i32(i32 %or)
+  %res = add nuw nsw i32 %ctpop1, %ctpop2
+  ret i32 %res
+}
+
 define i8 @ctpop_rotate_left(i8 %a, i8 %amt)  {
 ; CHECK-LABEL: @ctpop_rotate_left(
 ; CHECK-NEXT:    [[RES:%.*]] = tail call i8 @llvm.ctpop.i8(i8 [[A:%.*]]), !range [[RNG0]]

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[AtariDreams]

@dtcxzyw Done!

[DavidSpickett]

(since you asked about this on Discord)

If you click through to the results and click a whole bunch more (it's not super easy to find), you'll get to the end of the log and see:

MT: command "C:\Program Files (x86)\Windows Kits\10\bin\10.0.19041.0\x64\mt.exe /nologo /manifest unittests\TargetParser\TargetParserTests.exe.manifest /outputresource:unittests\TargetParser\TargetParserTests.exe;#1" failed (exit code 0x1f) with the following output:
mt.exe : general error c101008d: Failed to write the updated manifest to the resource of file "unittests\TargetParser\TargetParserTests.exe". Operation did not complete successfully because the file contains a virus or potentially unwanted software.

Which is not related to your patch. Ideally you'd get a clean run at some point, maybe by rebasing once the patch is approved, but if that's not possible maybe your reviewers will allow a merge without a Windows CI run.

[dtcxzyw]

@AtariDreams Does this pattern exist in some real-world applications?

[AtariDreams]

@AtariDreams Does this pattern exist in some real-world applications?

Given the popcount(A|B) does exist, it makes sense that popcount(A&B) + popcount(A|B) exists.

[dtcxzyw]

The implementation LGTM. But I am sorry I cannot give the approval unless you can prove that it benefits some real-world applications.

There are an infinite number of atomic simplifications. But obviously you cannot handle all of them in LLVM.

I am not willing to waste my time reviewing these valueless patches. I recommend you look at issues created by @XChy, which are extracted from real-world applications.
https://github.com/llvm/llvm-project/issues/created_by/XChy

[nikic]

I agree with @dtcxzyw. The pattern here is simple enough that we could accept it without knowing about a specific use-case, but at the same time I have a really hard time imaging where this could possibly come up. I'm going to decline it for now.