[InstCombine] Improve TryToSinkInstruction with multiple uses

This patch allows sinking an instruction which can have multiple uses in a
single user. We were previously over-restrictive by looking for exactly one use,
rather than one user.

Also added an API for retrieving a unique undroppable user.

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D109700

diff --git a/llvm/include/llvm/IR/Value.h b/llvm/include/llvm/IR/Value.h
index d19c5af..253c97c 100644 (file)
--- a/llvm/include/llvm/IR/Value.h
+++ b/llvm/include/llvm/IR/Value.h
@@ -459,6 +459,13 @@ public:
     return const_cast<Value *>(this)->getSingleUndroppableUse();
   }
 
+  /// Return true if there is exactly one unique user of this value that cannot be
+  /// dropped (that user can have multiple uses of this value).
+  User *getUniqueUndroppableUser();
+  const User *getUniqueUndroppableUser() const {
+    return const_cast<Value *>(this)->getUniqueUndroppableUser();
+  }
+
   /// Return true if there this value.
   ///
   /// This is specialized because it is a common request and does not require

diff --git a/llvm/lib/IR/Value.cpp b/llvm/lib/IR/Value.cpp
index a13380b..0f11bf6 100644 (file)
--- a/llvm/lib/IR/Value.cpp
+++ b/llvm/lib/IR/Value.cpp
@@ -176,6 +176,18 @@ Use *Value::getSingleUndroppableUse() {
   return Result;
 }
 
+User *Value::getUniqueUndroppableUser() {
+  User *Result = nullptr;
+  for (auto *U : users()) {
+    if (!U->isDroppable()) {
+      if (Result && Result != U)
+        return nullptr;
+      Result = U;
+    }
+  }
+  return Result;
+}
+
 bool Value::hasNUndroppableUses(unsigned int N) const {
   return hasNItems(user_begin(), user_end(), N, isUnDroppableUser);
 }

diff --git a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
index 6ba4e1c..249915b 100644 (file)
--- a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
@@ -3613,7 +3613,7 @@ Instruction *InstCombinerImpl::visitFreeze(FreezeInst &I) {
 /// instruction past all of the instructions between it and the end of its
 /// block.
 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
-  assert(I->getSingleUndroppableUse() && "Invariants didn't hold!");
+  assert(I->getUniqueUndroppableUser() && "Invariants didn't hold!");
   BasicBlock *SrcBlock = I->getParent();
 
   // Cannot move control-flow-involving, volatile loads, vaarg, etc.
@@ -3772,18 +3772,27 @@ bool InstCombinerImpl::run() {
         [this](Instruction *I) -> Optional<BasicBlock *> {
       if (!EnableCodeSinking)
         return None;
-      Use *SingleUse = I->getSingleUndroppableUse();
-      if (!SingleUse)
+      auto *UserInst = cast_or_null<Instruction>(I->getUniqueUndroppableUser());
+      if (!UserInst)
         return None;
 
       BasicBlock *BB = I->getParent();
-      Instruction *UserInst = cast<Instruction>(SingleUse->getUser());
-      BasicBlock *UserParent;
-
-      // Get the block the use occurs in.
-      if (PHINode *PN = dyn_cast<PHINode>(UserInst))
-        UserParent = PN->getIncomingBlock(*SingleUse);
-      else
+      BasicBlock *UserParent = nullptr;
+
+      // Special handling for Phi nodes - get the block the use occurs in.
+      if (PHINode *PN = dyn_cast<PHINode>(UserInst)) {
+        for (unsigned i = 0; i < PN->getNumIncomingValues(); i++) {
+          if (PN->getIncomingValue(i) == I) {
+            // Bail out if we have uses in different blocks. We don't do any
+            // sophisticated analysis (i.e finding NearestCommonDominator of these
+            // use blocks).
+            if (UserParent && UserParent != PN->getIncomingBlock(i))
+              return None;
+            UserParent = PN->getIncomingBlock(i);
+          }
+        }
+        assert(UserParent && "expected to find user block!");
+      } else
         UserParent = UserInst->getParent();
 
       // Try sinking to another block. If that block is unreachable, then do

diff --git a/llvm/test/Transforms/InstCombine/icmp-mul-zext.ll b/llvm/test/Transforms/InstCombine/icmp-mul-zext.ll
index 9b4d994..e70d070 100644 (file)
--- a/llvm/test/Transforms/InstCombine/icmp-mul-zext.ll
+++ b/llvm/test/Transforms/InstCombine/icmp-mul-zext.ll
@@ -56,9 +56,9 @@ lor.end:
 
 define void @PR33765(i8 %beth) {
 ; CHECK-LABEL: @PR33765(
-; CHECK-NEXT:    [[CONV:%.*]] = zext i8 [[BETH:%.*]] to i32
 ; CHECK-NEXT:    br i1 false, label [[IF_THEN9:%.*]], label [[IF_THEN9]]
 ; CHECK:       if.then9:
+; CHECK-NEXT:    [[CONV:%.*]] = zext i8 [[BETH:%.*]] to i32
 ; CHECK-NEXT:    [[MUL:%.*]] = mul nuw nsw i32 [[CONV]], [[CONV]]
 ; CHECK-NEXT:    [[TINKY:%.*]] = load i16, i16* @glob, align 2
 ; CHECK-NEXT:    [[TMP1:%.*]] = trunc i32 [[MUL]] to i16

diff --git a/llvm/test/Transforms/InstCombine/sink_instruction.ll b/llvm/test/Transforms/InstCombine/sink_instruction.ll
index 4ac0057..a9aba31 100644 (file)
--- a/llvm/test/Transforms/InstCombine/sink_instruction.ll
+++ b/llvm/test/Transforms/InstCombine/sink_instruction.ll
@@ -114,13 +114,13 @@ sw.epilog:                                        ; preds = %entry, %sw.bb
 }
 
 declare i32 @foo(i32, i32)
-; TODO: Two uses in a single user. We can still sink the instruction (tmp.9).
+; Two uses in a single user. We can still sink the instruction (tmp.9).
 define i32 @test4(i32 %A, i32 %B, i1 %C) {
 ; CHECK-LABEL: @test4(
 ; CHECK-NEXT:  entry:
-; CHECK-NEXT:    [[TMP_9:%.*]] = add i32 [[B:%.*]], [[A:%.*]]
 ; CHECK-NEXT:    br i1 [[C:%.*]], label [[THEN:%.*]], label [[ENDIF:%.*]]
 ; CHECK:       then:
+; CHECK-NEXT:    [[TMP_9:%.*]] = add i32 [[B:%.*]], [[A:%.*]]
 ; CHECK-NEXT:    [[RES:%.*]] = call i32 @foo(i32 [[TMP_9]], i32 [[TMP_9]])
 ; CHECK-NEXT:    ret i32 [[RES]]
 ; CHECK:       endif:
@@ -175,16 +175,16 @@ sw.epilog:                                        ; preds = %entry, %sw.bb
   ret i32 %sum.0
 }
 
-; TODO: Multiple uses but from same BB. We can sink.
+; Multiple uses but from same BB. We can sink.
 define i32 @test6(i32* nocapture readonly %P, i32 %i, i1 %cond) {
 ; CHECK-LABEL: @test6(
 ; CHECK-NEXT:  entry:
-; CHECK-NEXT:    [[IDXPROM:%.*]] = sext i32 [[I:%.*]] to i64
-; CHECK-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i32, i32* [[P:%.*]], i64 [[IDXPROM]]
-; CHECK-NEXT:    [[TMP0:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
 ; CHECK-NEXT:    [[ADD:%.*]] = shl nsw i32 [[I]], 1
 ; CHECK-NEXT:    br label [[DISPATCHBB:%.*]]
 ; CHECK:       dispatchBB:
+; CHECK-NEXT:    [[IDXPROM:%.*]] = sext i32 [[I:%.*]] to i64
+; CHECK-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i32, i32* [[P:%.*]], i64 [[IDXPROM]]
+; CHECK-NEXT:    [[TMP0:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
 ; CHECK-NEXT:    switch i32 [[I]], label [[SW_BB:%.*]] [
 ; CHECK-NEXT:    i32 5, label [[SW_EPILOG:%.*]]
 ; CHECK-NEXT:    i32 2, label [[SW_EPILOG]]

diff --git a/llvm/test/Transforms/SLPVectorizer/X86/blending-shuffle-inseltpoison.ll b/llvm/test/Transforms/SLPVectorizer/X86/blending-shuffle-inseltpoison.ll
index 1c806c6..f23eb38 100644 (file)
--- a/llvm/test/Transforms/SLPVectorizer/X86/blending-shuffle-inseltpoison.ll
+++ b/llvm/test/Transforms/SLPVectorizer/X86/blending-shuffle-inseltpoison.ll
@@ -139,9 +139,9 @@ define i8 @k(<4 x i8> %x) {
 
 define i8 @k_bb(<4 x i8> %x) {
 ; CHECK-LABEL: @k_bb(
-; CHECK-NEXT:    [[X0:%.*]] = extractelement <4 x i8> [[X:%.*]], i32 0
 ; CHECK-NEXT:    br label [[BB1:%.*]]
 ; CHECK:       bb1:
+; CHECK-NEXT:    [[X0:%.*]] = extractelement <4 x i8> [[X:%.*]], i32 0
 ; CHECK-NEXT:    [[X3:%.*]] = extractelement <4 x i8> [[X]], i32 3
 ; CHECK-NEXT:    [[X1:%.*]] = extractelement <4 x i8> [[X]], i32 1
 ; CHECK-NEXT:    [[X2:%.*]] = extractelement <4 x i8> [[X]], i32 2

diff --git a/llvm/test/Transforms/SLPVectorizer/X86/blending-shuffle.ll b/llvm/test/Transforms/SLPVectorizer/X86/blending-shuffle.ll
index 0fcc096..6bbfb53 100644 (file)
--- a/llvm/test/Transforms/SLPVectorizer/X86/blending-shuffle.ll
+++ b/llvm/test/Transforms/SLPVectorizer/X86/blending-shuffle.ll
@@ -139,9 +139,9 @@ define i8 @k(<4 x i8> %x) {
 
 define i8 @k_bb(<4 x i8> %x) {
 ; CHECK-LABEL: @k_bb(
-; CHECK-NEXT:    [[X0:%.*]] = extractelement <4 x i8> [[X:%.*]], i32 0
 ; CHECK-NEXT:    br label [[BB1:%.*]]
 ; CHECK:       bb1:
+; CHECK-NEXT:    [[X0:%.*]] = extractelement <4 x i8> [[X:%.*]], i32 0
 ; CHECK-NEXT:    [[X3:%.*]] = extractelement <4 x i8> [[X]], i32 3
 ; CHECK-NEXT:    [[X1:%.*]] = extractelement <4 x i8> [[X]], i32 1
 ; CHECK-NEXT:    [[X2:%.*]] = extractelement <4 x i8> [[X]], i32 2