[GVN][NewGVN][Local] Handle attributes for function calls after CSE #114011

dtcxzyw · 2024-10-29T07:32:15Z

This patch intersects attributes of two calls to avoid introducing UB. It also skips incompatible call pairs in GVN/NewGVN. However, I cannot provide negative tests for these changes.

Fixes #113997.

llvmbot · 2024-10-29T07:32:52Z

@llvm/pr-subscribers-llvm-transforms

Author: Yingwei Zheng (dtcxzyw)

Changes

This patch intersects attributes of two calls to avoid introducing UB. It also skips incompatible call pairs in GVN/NewGVN. However, I cannot provide negative tests for these changes.

Fixes #113997.

Full diff: https://github.com/llvm/llvm-project/pull/114011.diff

5 Files Affected:

(modified) llvm/lib/Transforms/Scalar/GVN.cpp (+12-2)
(modified) llvm/lib/Transforms/Scalar/NewGVN.cpp (+25-5)
(modified) llvm/lib/Transforms/Utils/Local.cpp (+11)
(added) llvm/test/Transforms/GVN/pr113997.ll (+33)
(added) llvm/test/Transforms/NewGVN/pr113997.ll (+33)

diff --git a/llvm/lib/Transforms/Scalar/GVN.cpp b/llvm/lib/Transforms/Scalar/GVN.cpp
index 2ba600497e00d3..ad9b1217089d7d 100644
--- a/llvm/lib/Transforms/Scalar/GVN.cpp
+++ b/llvm/lib/Transforms/Scalar/GVN.cpp
@@ -2189,6 +2189,16 @@ bool GVNPass::processAssumeIntrinsic(AssumeInst *IntrinsicI) {
   return Changed;
 }
 
+// Return true iff V1 can be replaced with V2.
+static bool canBeReplacedBy(Value *V1, Value *V2) {
+  if (auto *CB1 = dyn_cast<CallBase>(V1))
+    if (auto *CB2 = dyn_cast<CallBase>(V2))
+      return CB1->getAttributes()
+          .intersectWith(CB2->getContext(), CB2->getAttributes())
+          .has_value();
+  return true;
+}
+
 static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) {
   patchReplacementInstruction(I, Repl);
   I->replaceAllUsesWith(Repl);
@@ -2734,7 +2744,7 @@ bool GVNPass::processInstruction(Instruction *I) {
   // Perform fast-path value-number based elimination of values inherited from
   // dominators.
   Value *Repl = findLeader(I->getParent(), Num);
-  if (!Repl) {
+  if (!Repl || !canBeReplacedBy(I, Repl)) {
     // Failure, just remember this instance for future use.
     LeaderTable.insert(Num, I, I->getParent());
     return false;
@@ -3000,7 +3010,7 @@ bool GVNPass::performScalarPRE(Instruction *CurInst) {
 
     uint32_t TValNo = VN.phiTranslate(P, CurrentBlock, ValNo, *this);
     Value *predV = findLeader(P, TValNo);
-    if (!predV) {
+    if (!predV || !canBeReplacedBy(CurInst, predV)) {
       predMap.push_back(std::make_pair(static_cast<Value *>(nullptr), P));
       PREPred = P;
       ++NumWithout;
diff --git a/llvm/lib/Transforms/Scalar/NewGVN.cpp b/llvm/lib/Transforms/Scalar/NewGVN.cpp
index 13d9e8f186b47c..6800ad51cc0a8f 100644
--- a/llvm/lib/Transforms/Scalar/NewGVN.cpp
+++ b/llvm/lib/Transforms/Scalar/NewGVN.cpp
@@ -3854,6 +3854,16 @@ Value *NewGVN::findPHIOfOpsLeader(const Expression *E,
   return nullptr;
 }
 
+// Return true iff V1 can be replaced with V2.
+static bool canBeReplacedBy(Value *V1, Value *V2) {
+  if (auto *CB1 = dyn_cast<CallBase>(V1))
+    if (auto *CB2 = dyn_cast<CallBase>(V2))
+      return CB1->getAttributes()
+          .intersectWith(CB2->getContext(), CB2->getAttributes())
+          .has_value();
+  return true;
+}
+
 bool NewGVN::eliminateInstructions(Function &F) {
   // This is a non-standard eliminator. The normal way to eliminate is
   // to walk the dominator tree in order, keeping track of available
@@ -3963,6 +3973,9 @@ bool NewGVN::eliminateInstructions(Function &F) {
           MembersLeft.insert(Member);
           continue;
         }
+        if (!canBeReplacedBy(Member, Leader))
+          continue;
+
         LLVM_DEBUG(dbgs() << "Found replacement " << *(Leader) << " for "
                           << *Member << "\n");
         auto *I = cast<Instruction>(Member);
@@ -4069,8 +4082,11 @@ bool NewGVN::eliminateInstructions(Function &F) {
               if (DominatingLeader != Def) {
                 // Even if the instruction is removed, we still need to update
                 // flags/metadata due to downstreams users of the leader.
-                if (!match(DefI, m_Intrinsic<Intrinsic::ssa_copy>()))
+                if (!match(DefI, m_Intrinsic<Intrinsic::ssa_copy>())) {
+                  if (!canBeReplacedBy(DefI, DominatingLeader))
+                    continue;
                   patchReplacementInstruction(DefI, DominatingLeader);
+                }
 
                 markInstructionForDeletion(DefI);
               }
@@ -4112,17 +4128,21 @@ bool NewGVN::eliminateInstructions(Function &F) {
           // Don't replace our existing users with ourselves.
           if (U->get() == DominatingLeader)
             continue;
-          LLVM_DEBUG(dbgs()
-                     << "Found replacement " << *DominatingLeader << " for "
-                     << *U->get() << " in " << *(U->getUser()) << "\n");
 
           // If we replaced something in an instruction, handle the patching of
           // metadata.  Skip this if we are replacing predicateinfo with its
           // original operand, as we already know we can just drop it.
           auto *ReplacedInst = cast<Instruction>(U->get());
           auto *PI = PredInfo->getPredicateInfoFor(ReplacedInst);
-          if (!PI || DominatingLeader != PI->OriginalOp)
+          if (!PI || DominatingLeader != PI->OriginalOp) {
+            if (!canBeReplacedBy(ReplacedInst, DominatingLeader))
+              continue;
             patchReplacementInstruction(ReplacedInst, DominatingLeader);
+          }
+
+          LLVM_DEBUG(dbgs()
+                     << "Found replacement " << *DominatingLeader << " for "
+                     << *U->get() << " in " << *(U->getUser()) << "\n");
           U->set(DominatingLeader);
           // This is now a use of the dominating leader, which means if the
           // dominating leader was dead, it's now live!
diff --git a/llvm/lib/Transforms/Utils/Local.cpp b/llvm/lib/Transforms/Utils/Local.cpp
index 65c1669f92b4d3..47a70492559610 100644
--- a/llvm/lib/Transforms/Utils/Local.cpp
+++ b/llvm/lib/Transforms/Utils/Local.cpp
@@ -3508,6 +3508,17 @@ void llvm::patchReplacementInstruction(Instruction *I, Value *Repl) {
   else if (!isa<LoadInst>(I))
     ReplInst->andIRFlags(I);
 
+  // Handle attributes.
+  if (auto *CB1 = dyn_cast<CallBase>(ReplInst)) {
+    if (auto *CB2 = dyn_cast<CallBase>(I)) {
+      bool Success = CB1->tryIntersectAttributes(CB2);
+      assert(Success && "We should not be trying to sink callbases "
+                        "with non-intersectable attributes");
+      // For NDEBUG Compile.
+      (void)Success;
+    }
+  }
+
   // FIXME: If both the original and replacement value are part of the
   // same control-flow region (meaning that the execution of one
   // guarantees the execution of the other), then we can combine the
diff --git a/llvm/test/Transforms/GVN/pr113997.ll b/llvm/test/Transforms/GVN/pr113997.ll
new file mode 100644
index 00000000000000..35e73b1a4439b3
--- /dev/null
+++ b/llvm/test/Transforms/GVN/pr113997.ll
@@ -0,0 +1,33 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -S -passes=gvn < %s | FileCheck %s
+
+; Make sure attributes in function calls are intersected correctly.
+
+define i1 @bucket(i32 noundef %x) {
+; CHECK-LABEL: define i1 @bucket(
+; CHECK-SAME: i32 noundef [[X:%.*]]) {
+; CHECK-NEXT:    [[CMP1:%.*]] = icmp sgt i32 [[X]], 0
+; CHECK-NEXT:    [[CTPOP1:%.*]] = tail call range(i32 0, 33) i32 @llvm.ctpop.i32(i32 [[X]])
+; CHECK-NEXT:    [[CMP2:%.*]] = icmp samesign ult i32 [[CTPOP1]], 2
+; CHECK-NEXT:    [[COND:%.*]] = select i1 [[CMP1]], i1 [[CMP2]], i1 false
+; CHECK-NEXT:    br i1 [[COND]], label %[[IF_THEN:.*]], label %[[IF_ELSE:.*]]
+; CHECK:       [[IF_ELSE]]:
+; CHECK-NEXT:    [[RES:%.*]] = icmp eq i32 [[CTPOP1]], 1
+; CHECK-NEXT:    ret i1 [[RES]]
+; CHECK:       [[IF_THEN]]:
+; CHECK-NEXT:    ret i1 false
+;
+  %cmp1 = icmp sgt i32 %x, 0
+  %ctpop1 = tail call range(i32 1, 32) i32 @llvm.ctpop.i32(i32 %x)
+  %cmp2 = icmp samesign ult i32 %ctpop1, 2
+  %cond = select i1 %cmp1, i1 %cmp2, i1 false
+  br i1 %cond, label %if.then, label %if.else
+
+if.else:
+  %ctpop2 = tail call range(i32 0, 33) i32 @llvm.ctpop.i32(i32 %x)
+  %res = icmp eq i32 %ctpop2, 1
+  ret i1 %res
+
+if.then:
+  ret i1 false
+}
diff --git a/llvm/test/Transforms/NewGVN/pr113997.ll b/llvm/test/Transforms/NewGVN/pr113997.ll
new file mode 100644
index 00000000000000..a919c8c304b1b9
--- /dev/null
+++ b/llvm/test/Transforms/NewGVN/pr113997.ll
@@ -0,0 +1,33 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -S -passes=newgvn < %s | FileCheck %s
+
+; Make sure attributes in function calls are intersected correctly.
+
+define i1 @bucket(i32 noundef %x) {
+; CHECK-LABEL: define i1 @bucket(
+; CHECK-SAME: i32 noundef [[X:%.*]]) {
+; CHECK-NEXT:    [[CMP1:%.*]] = icmp sgt i32 [[X]], 0
+; CHECK-NEXT:    [[CTPOP1:%.*]] = tail call range(i32 0, 33) i32 @llvm.ctpop.i32(i32 [[X]])
+; CHECK-NEXT:    [[CMP2:%.*]] = icmp samesign ult i32 [[CTPOP1]], 2
+; CHECK-NEXT:    [[COND:%.*]] = select i1 [[CMP1]], i1 [[CMP2]], i1 false
+; CHECK-NEXT:    br i1 [[COND]], label %[[IF_THEN:.*]], label %[[IF_ELSE:.*]]
+; CHECK:       [[IF_ELSE]]:
+; CHECK-NEXT:    [[RES:%.*]] = icmp eq i32 [[CTPOP1]], 1
+; CHECK-NEXT:    ret i1 [[RES]]
+; CHECK:       [[IF_THEN]]:
+; CHECK-NEXT:    ret i1 false
+;
+  %cmp1 = icmp sgt i32 %x, 0
+  %ctpop1 = tail call range(i32 1, 32) i32 @llvm.ctpop.i32(i32 %x)
+  %cmp2 = icmp samesign ult i32 %ctpop1, 2
+  %cond = select i1 %cmp1, i1 %cmp2, i1 false
+  br i1 %cond, label %if.then, label %if.else
+
+if.else:
+  %ctpop2 = tail call range(i32 0, 33) i32 @llvm.ctpop.i32(i32 %x)
+  %res = icmp eq i32 %ctpop2, 1
+  ret i1 %res
+
+if.then:
+  ret i1 false
+}

llvm/test/Transforms/GVN/pr113997.ll

goldsteinn · 2024-10-29T14:55:06Z

This looks fine to me, although I think an easier way to implement this would be make patchReplacementInstruction return a bool to indicate I and Repl where successfully patched.

dtcxzyw · 2024-10-29T15:04:18Z

an easier way to implement this would be make patchReplacementInstruction return a bool to indicate I and Repl where successfully patched.

I tried that before. But it may be too late in GVNPass::performScalarPRE :(

pranavk · 2024-10-29T15:16:34Z

LGTM. I verified that this patch fixes our test case that was broken earlier.

pranavk

LGTM but someone more knowledgable than me should also take a look.

asmok-g · 2024-10-30T09:29:13Z

We need this to unblock us in google. When can we expect a merge ?

dtcxzyw · 2024-10-31T00:28:33Z

We need this to unblock us in google. When can we expect a merge ?

I will land this patch tomorrow if no more comments.

nikic

The general implementation approach here looks incorrect to me. You can't really restrict replacements after the fact in GVN, you need to do so during value numbering.

Consider this minor variation of the test case:

define i1 @bucket2(i32 noundef %x) {
; CHECK-LABEL: define i1 @bucket2(
; CHECK-SAME: i32 noundef [[X:%.*]]) {
; CHECK-NEXT:    [[CMP1:%.*]] = icmp sgt i32 [[X]], 0
; CHECK-NEXT:    [[CTPOP1:%.*]] = tail call range(i32 1, 32) i32 @llvm.ctpop.i32(i32 zeroext [[X]])
; CHECK-NEXT:    [[CTPOP1INC:%.*]] = add i32 [[CTPOP1]], 1
; CHECK-NEXT:    [[CMP2:%.*]] = icmp samesign ult i32 [[CTPOP1INC]], 3
; CHECK-NEXT:    [[COND:%.*]] = select i1 [[CMP1]], i1 [[CMP2]], i1 false
; CHECK-NEXT:    br i1 [[COND]], label %[[IF_THEN:.*]], label %[[IF_ELSE:.*]]
; CHECK:       [[IF_ELSE]]:
; CHECK-NEXT:    [[CTPOP2:%.*]] = tail call range(i32 0, 33) i32 @llvm.ctpop.i32(i32 [[X]])
; CHECK-NEXT:    [[RES:%.*]] = icmp eq i32 [[CTPOP1INC]], 2
; CHECK-NEXT:    ret i1 [[RES]]
; CHECK:       [[IF_THEN]]:
; CHECK-NEXT:    ret i1 false
;
  %cmp1 = icmp sgt i32 %x, 0
  %ctpop1 = tail call range(i32 1, 32) i32 @llvm.ctpop.i32(i32 zeroext %x)
  %ctpop1inc = add i32 %ctpop1, 1
  %cmp2 = icmp samesign ult i32 %ctpop1inc, 3
  %cond = select i1 %cmp1, i1 %cmp2, i1 false
  br i1 %cond, label %if.then, label %if.else

if.else:
  %ctpop2 = tail call range(i32 0, 33) i32 @llvm.ctpop.i32(i32 %x)
  %ctpop2inc = add i32 %ctpop2, 1
  %res = icmp eq i32 %ctpop2inc, 2
  ret i1 %res

if.then:
  ret i1 false
}

This just adds a zeroext dummy attribute to make sure the two ctpop calls are not combined directly, and an extra add 1 after them. Note that the adds will get CSEd, because the canBeReplacedBy check is specifically on the calls, not on any later instructions that are numbered based on the incorrect assumption that the calls are equivalent.

dtcxzyw · 2024-11-01T09:35:17Z

You can't really restrict replacements after the fact in GVN, you need to do so during value numbering.

I've considered this approach before. But it is too heavy to add Attributes to GVNPass::Expression.

Note that the adds will get CSEd, because the canBeReplacedBy check is specifically on the calls, not on any later instructions that are numbered based on the incorrect assumption that the calls are equivalent.

You are right. I missed this case.

nikic · 2024-11-01T12:03:13Z

You can't really restrict replacements after the fact in GVN, you need to do so during value numbering.

I've considered this approach before. But it is too heavy to add Attributes to GVNPass::Expression.

I don't have any particularly good ideas on how to fix this properly. Given that this is a CSE transform, I think we could get away with only dropping the poison-generating attributes, and otherwise keeping the attributes of the first call, to avoid the problematic !canBeReplacedBy case.

…lvm#114011) This patch intersects attributes of two calls to avoid introducing UB. It also skips incompatible call pairs in GVN/NewGVN. However, I cannot provide negative tests for these changes. Fixes llvm#113997.

Drop `canBeReplacedBy` and take call attributes into account in expressions. Address comment #114011 (review).

…114545) Drop `canBeReplacedBy` and take call attributes into account in expressions. Address comment llvm#114011 (review).

dtcxzyw added 2 commits October 29, 2024 15:10

[GVN][NewGVN] Add pre-commit tests. NFC.

f303e01

[GVN][NewGVN][Local] Handle attributes for function calls after CSE

fe37aad

dtcxzyw requested review from nikic, fhahn, andjo403, pranavk and goldsteinn October 29, 2024 07:32

llvmbot added the llvm:transforms label Oct 29, 2024

andjo403 reviewed Oct 29, 2024

View reviewed changes

llvm/test/Transforms/GVN/pr113997.ll Show resolved Hide resolved

This was referenced Oct 29, 2024

Task submission dtcxzyw/llvm-opt-benchmark#1312

Open

pre-commit: PR114011 dtcxzyw/llvm-opt-benchmark#1566

Closed

pranavk approved these changes Oct 29, 2024

View reviewed changes

dtcxzyw requested a review from jayfoad October 30, 2024 11:19

dtcxzyw merged commit f16bff1 into llvm:main Nov 1, 2024
10 checks passed

dtcxzyw deleted the perf/fix-pr113997 branch November 1, 2024 04:44

nikic reviewed Nov 1, 2024

View reviewed changes

dtcxzyw mentioned this pull request Nov 1, 2024

[GVN][NewGVN] Take call attributes into account in expressions #114545

Merged

dtcxzyw added a commit that referenced this pull request Nov 8, 2024

[GVN][NewGVN] Take call attributes into account in expressions (#114545)

984bca9

Drop `canBeReplacedBy` and take call attributes into account in expressions. Address comment #114011 (review).

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[GVN][NewGVN][Local] Handle attributes for function calls after CSE #114011

[GVN][NewGVN][Local] Handle attributes for function calls after CSE #114011

Uh oh!

dtcxzyw commented Oct 29, 2024

Uh oh!

llvmbot commented Oct 29, 2024

Uh oh!

Uh oh!

goldsteinn commented Oct 29, 2024

Uh oh!

dtcxzyw commented Oct 29, 2024

Uh oh!

pranavk commented Oct 29, 2024

Uh oh!

pranavk left a comment

Uh oh!

asmok-g commented Oct 30, 2024

Uh oh!

dtcxzyw commented Oct 31, 2024

Uh oh!

Uh oh!

nikic left a comment

Uh oh!

dtcxzyw commented Nov 1, 2024

Uh oh!

nikic commented Nov 1, 2024

Uh oh!

Uh oh!

[GVN][NewGVN][Local] Handle attributes for function calls after CSE #114011

[GVN][NewGVN][Local] Handle attributes for function calls after CSE #114011

Uh oh!

Conversation

dtcxzyw commented Oct 29, 2024

Uh oh!

llvmbot commented Oct 29, 2024

Uh oh!

Uh oh!

goldsteinn commented Oct 29, 2024

Uh oh!

dtcxzyw commented Oct 29, 2024

Uh oh!

pranavk commented Oct 29, 2024

Uh oh!

pranavk left a comment

Choose a reason for hiding this comment

Uh oh!

asmok-g commented Oct 30, 2024

Uh oh!

dtcxzyw commented Oct 31, 2024

Uh oh!

Uh oh!

nikic left a comment

Choose a reason for hiding this comment

Uh oh!

dtcxzyw commented Nov 1, 2024

Uh oh!

nikic commented Nov 1, 2024

Uh oh!

Uh oh!