Reapply [CVP] Simplify non-overflowing saturating add/sub

If we can determine that a saturating add/sub will not overflow based
on range analysis, convert it into a simple binary operation. This is
a sibling transform to the existing with.overflow handling.

Reapplying this with an additional check that the saturating intrinsic
has integer type, as LVI currently does not support vector types.

Differential Revision: https://reviews.llvm.org/D62703

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@362263 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Transforms/Scalar/CorrelatedValuePropagation.cpp b/lib/Transforms/Scalar/CorrelatedValuePropagation.cpp
index 9ec2afb99d683af226cdc725201e93b94b1a395b..5bfdf17c9a03f78a1375b3ad6494830530ec4a91 100644 (file)
--- a/lib/Transforms/Scalar/CorrelatedValuePropagation.cpp
+++ b/lib/Transforms/Scalar/CorrelatedValuePropagation.cpp
@@ -63,6 +63,8 @@ STATISTIC(NumUDivs,     "Number of udivs whose width was decreased");
 STATISTIC(NumAShrs,     "Number of ashr converted to lshr");
 STATISTIC(NumSRems,     "Number of srem converted to urem");
 STATISTIC(NumOverflows, "Number of overflow checks removed");
+STATISTIC(NumSaturating,
+    "Number of saturating arithmetics converted to normal arithmetics");
 
 static cl::opt<bool> DontAddNoWrapFlags("cvp-dont-add-nowrap-flags", cl::init(true));
 
@@ -413,7 +415,7 @@ static void processOverflowIntrinsic(WithOverflowInst *WO) {
   IRBuilder<> B(WO);
   Value *NewOp = B.CreateBinOp(
       WO->getBinaryOp(), WO->getLHS(), WO->getRHS(), WO->getName());
-  // Constant-holing could have happened.
+  // Constant-folding could have happened.
   if (auto *Inst = dyn_cast<Instruction>(NewOp)) {
     if (WO->isSigned())
       Inst->setHasNoSignedWrap();
@@ -428,6 +430,20 @@ static void processOverflowIntrinsic(WithOverflowInst *WO) {
   ++NumOverflows;
 }
 
+static void processSaturatingInst(SaturatingInst *SI) {
+  BinaryOperator *BinOp = BinaryOperator::Create(
+      SI->getBinaryOp(), SI->getLHS(), SI->getRHS(), SI->getName(), SI);
+  BinOp->setDebugLoc(SI->getDebugLoc());
+  if (SI->isSigned())
+    BinOp->setHasNoSignedWrap();
+  else
+    BinOp->setHasNoUnsignedWrap();
+
+  SI->replaceAllUsesWith(BinOp);
+  SI->eraseFromParent();
+  ++NumSaturating;
+}
+
 /// Infer nonnull attributes for the arguments at the specified callsite.
 static bool processCallSite(CallSite CS, LazyValueInfo *LVI) {
   SmallVector<unsigned, 4> ArgNos;
@@ -440,6 +456,13 @@ static bool processCallSite(CallSite CS, LazyValueInfo *LVI) {
     }
   }
 
+  if (auto *SI = dyn_cast<SaturatingInst>(CS.getInstruction())) {
+    if (SI->getType()->isIntegerTy() && willNotOverflow(SI, LVI)) {
+      processSaturatingInst(SI);
+      return true;
+    }
+  }
+
   // Deopt bundle operands are intended to capture state with minimal
   // perturbance of the code otherwise.  If we can find a constant value for
   // any such operand and remove a use of the original value, that's

diff --git a/test/Transforms/CorrelatedValuePropagation/overflows.ll b/test/Transforms/CorrelatedValuePropagation/overflows.ll
index 66cd85abe9751d52e46e2f9d2a35b57737542b8d..04b1471ebb44eaf96cbb321bc749b449b3234960 100644 (file)
--- a/test/Transforms/CorrelatedValuePropagation/overflows.ll
+++ b/test/Transforms/CorrelatedValuePropagation/overflows.ll
@@ -752,8 +752,8 @@ define i8 @uadd_sat_no_overflow(i8 %x) {
 ; CHECK-NEXT:    call void @llvm.trap()
 ; CHECK-NEXT:    unreachable
 ; CHECK:       cont:
-; CHECK-NEXT:    [[RES:%.*]] = call i8 @llvm.uadd.sat.i8(i8 [[X]], i8 100)
-; CHECK-NEXT:    ret i8 [[RES]]
+; CHECK-NEXT:    [[RES1:%.*]] = add nuw i8 [[X]], 100
+; CHECK-NEXT:    ret i8 [[RES1]]
 ;
   %cmp = icmp ugt i8 %x, 100
   br i1 %cmp, label %trap, label %cont
@@ -775,8 +775,8 @@ define i8 @sadd_sat_no_overflow(i8 %x) {
 ; CHECK-NEXT:    call void @llvm.trap()
 ; CHECK-NEXT:    unreachable
 ; CHECK:       cont:
-; CHECK-NEXT:    [[RES:%.*]] = call i8 @llvm.sadd.sat.i8(i8 [[X]], i8 20)
-; CHECK-NEXT:    ret i8 [[RES]]
+; CHECK-NEXT:    [[RES1:%.*]] = add nsw i8 [[X]], 20
+; CHECK-NEXT:    ret i8 [[RES1]]
 ;
   %cmp = icmp sgt i8 %x, 100
   br i1 %cmp, label %trap, label %cont
@@ -798,8 +798,8 @@ define i8 @usub_sat_no_overflow(i8 %x) {
 ; CHECK-NEXT:    call void @llvm.trap()
 ; CHECK-NEXT:    unreachable
 ; CHECK:       cont:
-; CHECK-NEXT:    [[RES:%.*]] = call i8 @llvm.usub.sat.i8(i8 [[X]], i8 100)
-; CHECK-NEXT:    ret i8 [[RES]]
+; CHECK-NEXT:    [[RES1:%.*]] = sub nuw i8 [[X]], 100
+; CHECK-NEXT:    ret i8 [[RES1]]
 ;
   %cmp = icmp ult i8 %x, 100
   br i1 %cmp, label %trap, label %cont
@@ -821,8 +821,8 @@ define i8 @ssub_sat_no_overflow(i8 %x) {
 ; CHECK-NEXT:    call void @llvm.trap()
 ; CHECK-NEXT:    unreachable
 ; CHECK:       cont:
-; CHECK-NEXT:    [[RES:%.*]] = call i8 @llvm.ssub.sat.i8(i8 [[X]], i8 20)
-; CHECK-NEXT:    ret i8 [[RES]]
+; CHECK-NEXT:    [[RES1:%.*]] = sub nsw i8 [[X]], 20
+; CHECK-NEXT:    ret i8 [[RES1]]
 ;
   %cmp = icmp slt i8 %x, -100
   br i1 %cmp, label %trap, label %cont