[PATCH] D60791: [CVP] Simplify umulo and smulo that cannot overflow

[Nikita Popov via Phabricator via llvm-commits]

nikic created this revision.
nikic added reviewers: lebedev.ri, spatel.
Herald added subscribers: llvm-commits, hiraditya.
Herald added a project: LLVM.

If a umul.with.overflow or smul.with.overflow operation cannot overflow, simplify it to a simple mul nuw / mul nsw. After the refactoring in D60668 <https://reviews.llvm.org/D60668> this is just a matter of removing an explicit check against multiplications.


Repository:
  rL LLVM

https://reviews.llvm.org/D60791

Files:
  llvm/lib/Transforms/Scalar/CorrelatedValuePropagation.cpp
  llvm/test/Transforms/CorrelatedValuePropagation/overflows.ll


Index: llvm/test/Transforms/CorrelatedValuePropagation/overflows.ll
===================================================================
--- llvm/test/Transforms/CorrelatedValuePropagation/overflows.ll
+++ llvm/test/Transforms/CorrelatedValuePropagation/overflows.ll
@@ -507,9 +507,11 @@
 ; CHECK-NEXT:    [[CMP:%.*]] = icmp ugt i32 [[X:%.*]], 10000
 ; CHECK-NEXT:    br i1 [[CMP]], label [[COND_END:%.*]], label [[COND_FALSE:%.*]]
 ; CHECK:       cond.false:
-; CHECK-NEXT:    [[MULO:%.*]] = tail call { i32, i1 } @llvm.umul.with.overflow.i32(i32 [[X]], i32 100)
-; CHECK-NEXT:    [[RES:%.*]] = extractvalue { i32, i1 } [[MULO]], 0
-; CHECK-NEXT:    [[OV:%.*]] = extractvalue { i32, i1 } [[MULO]], 1
+; CHECK-NEXT:    [[MULO1:%.*]] = mul nuw i32 [[X]], 100
+; CHECK-NEXT:    [[TMP0:%.*]] = insertvalue { i32, i1 } undef, i32 [[MULO1]], 0
+; CHECK-NEXT:    [[TMP1:%.*]] = insertvalue { i32, i1 } [[TMP0]], i1 false, 1
+; CHECK-NEXT:    [[RES:%.*]] = extractvalue { i32, i1 } [[TMP1]], 0
+; CHECK-NEXT:    [[OV:%.*]] = extractvalue { i32, i1 } [[TMP1]], 1
 ; CHECK-NEXT:    br i1 [[OV]], label [[TRAP:%.*]], label [[COND_END]]
 ; CHECK:       trap:
 ; CHECK-NEXT:    tail call void @llvm.trap()
@@ -545,9 +547,11 @@
 ; CHECK-NEXT:    [[CMP3:%.*]] = or i1 [[CMP1]], [[CMP2]]
 ; CHECK-NEXT:    br i1 [[CMP3]], label [[COND_END:%.*]], label [[COND_FALSE:%.*]]
 ; CHECK:       cond.false:
-; CHECK-NEXT:    [[MULO:%.*]] = tail call { i32, i1 } @llvm.smul.with.overflow.i32(i32 [[X]], i32 100)
-; CHECK-NEXT:    [[RES:%.*]] = extractvalue { i32, i1 } [[MULO]], 0
-; CHECK-NEXT:    [[OV:%.*]] = extractvalue { i32, i1 } [[MULO]], 1
+; CHECK-NEXT:    [[MULO1:%.*]] = mul nsw i32 [[X]], 100
+; CHECK-NEXT:    [[TMP0:%.*]] = insertvalue { i32, i1 } undef, i32 [[MULO1]], 0
+; CHECK-NEXT:    [[TMP1:%.*]] = insertvalue { i32, i1 } [[TMP0]], i1 false, 1
+; CHECK-NEXT:    [[RES:%.*]] = extractvalue { i32, i1 } [[TMP1]], 0
+; CHECK-NEXT:    [[OV:%.*]] = extractvalue { i32, i1 } [[TMP1]], 1
 ; CHECK-NEXT:    br i1 [[OV]], label [[TRAP:%.*]], label [[COND_END]]
 ; CHECK:       trap:
 ; CHECK-NEXT:    tail call void @llvm.trap()
Index: llvm/lib/Transforms/Scalar/CorrelatedValuePropagation.cpp
===================================================================
--- llvm/lib/Transforms/Scalar/CorrelatedValuePropagation.cpp
+++ llvm/lib/Transforms/Scalar/CorrelatedValuePropagation.cpp
@@ -400,15 +400,10 @@
 
 // See if we can prove that the given overflow intrinsic will not overflow.
 static bool willNotOverflow(WithOverflowInst *WO, LazyValueInfo *LVI) {
-  // TODO: Also support multiplication.
-  Instruction::BinaryOps BinOp = WO->getBinaryOp();
-  if (BinOp == Instruction::Mul)
-    return false;
-
   Value *RHS = WO->getRHS();
   ConstantRange RRange = LVI->getConstantRange(RHS, WO->getParent(), WO);
   ConstantRange NWRegion = ConstantRange::makeGuaranteedNoWrapRegion(
-      BinOp, RRange, WO->getNoWrapKind());
+      WO->getBinaryOp(), RRange, WO->getNoWrapKind());
   // As an optimization, do not compute LRange if we do not need it.
   if (NWRegion.isEmptySet())
     return false;


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D60791.195435.patch
Type: text/x-patch
Size: 3092 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20190416/3715c43e/attachment.bin>