[llvm] r326908 - Re-land: Teach CorrelatedValuePropagation to reduce the width of udiv/urem instructions.
Justin Lebar via llvm-commits
llvm-commits at lists.llvm.org
Wed Mar 7 08:56:50 PST 2018
Author: jlebar
Date: Wed Mar 7 08:56:49 2018
New Revision: 326908
URL: http://llvm.org/viewvc/llvm-project?rev=326908&view=rev
Log:
Re-land: Teach CorrelatedValuePropagation to reduce the width of udiv/urem instructions.
Summary:
If the operands of a udiv/urem can be proved to fit within a smaller
power-of-two-sized type, reduce the width of the udiv/urem.
Backed out for failing an assert in clang bootstrap builds. Re-landing
with a fix for handling non-power-of-two inputs (e.g. udiv i24).
Original Differential Revision: https://reviews.llvm.org/D44102
Added:
llvm/trunk/test/Transforms/CorrelatedValuePropagation/udiv.ll
llvm/trunk/test/Transforms/CorrelatedValuePropagation/urem.ll
Modified:
llvm/trunk/lib/Transforms/Scalar/CorrelatedValuePropagation.cpp
Modified: llvm/trunk/lib/Transforms/Scalar/CorrelatedValuePropagation.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Scalar/CorrelatedValuePropagation.cpp?rev=326908&r1=326907&r2=326908&view=diff
==============================================================================
--- llvm/trunk/lib/Transforms/Scalar/CorrelatedValuePropagation.cpp (original)
+++ llvm/trunk/lib/Transforms/Scalar/CorrelatedValuePropagation.cpp Wed Mar 7 08:56:49 2018
@@ -58,6 +58,7 @@ STATISTIC(NumCmps, "Number of compa
STATISTIC(NumReturns, "Number of return values propagated");
STATISTIC(NumDeadCases, "Number of switch cases removed");
STATISTIC(NumSDivs, "Number of sdiv converted to udiv");
+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");
@@ -432,6 +433,48 @@ static bool hasPositiveOperands(BinaryOp
return true;
}
+/// Try to shrink a udiv/urem's width down to the smallest power of two that's
+/// sufficient to contain its operands.
+static bool processUDivOrURem(BinaryOperator *Instr, LazyValueInfo *LVI) {
+ assert(Instr->getOpcode() == Instruction::UDiv ||
+ Instr->getOpcode() == Instruction::URem);
+ if (Instr->getType()->isVectorTy())
+ return false;
+
+ // Find the smallest power of two bitwidth that's sufficient to hold Instr's
+ // operands.
+ auto OrigWidth = Instr->getType()->getIntegerBitWidth();
+ ConstantRange OperandRange(OrigWidth, /*isFullset=*/false);
+ for (Value *Operand : Instr->operands()) {
+ OperandRange = OperandRange.unionWith(
+ LVI->getConstantRange(Operand, Instr->getParent()));
+ }
+ // Don't shrink below 8 bits wide.
+ unsigned NewWidth = std::max<unsigned>(
+ PowerOf2Ceil(OperandRange.getUnsignedMax().getActiveBits()), 8);
+ // NewWidth might be greater than OrigWidth if OrigWidth is not a power of
+ // two.
+ if (NewWidth >= OrigWidth)
+ return false;
+
+ ++NumUDivs;
+ auto *TruncTy = Type::getIntNTy(Instr->getContext(), NewWidth);
+ auto *LHS = CastInst::Create(Instruction::Trunc, Instr->getOperand(0), TruncTy,
+ Instr->getName() + ".lhs.trunc", Instr);
+ auto *RHS = CastInst::Create(Instruction::Trunc, Instr->getOperand(1), TruncTy,
+ Instr->getName() + ".rhs.trunc", Instr);
+ auto *BO =
+ BinaryOperator::Create(Instr->getOpcode(), LHS, RHS, Instr->getName(), Instr);
+ auto *Zext = CastInst::Create(Instruction::ZExt, BO, Instr->getType(),
+ Instr->getName() + ".zext", Instr);
+ if (BO->getOpcode() == Instruction::UDiv)
+ BO->setIsExact(Instr->isExact());
+
+ Instr->replaceAllUsesWith(Zext);
+ Instr->eraseFromParent();
+ return true;
+}
+
static bool processSRem(BinaryOperator *SDI, LazyValueInfo *LVI) {
if (SDI->getType()->isVectorTy() || !hasPositiveOperands(SDI, LVI))
return false;
@@ -441,6 +484,10 @@ static bool processSRem(BinaryOperator *
SDI->getName(), SDI);
SDI->replaceAllUsesWith(BO);
SDI->eraseFromParent();
+
+ // Try to process our new urem.
+ processUDivOrURem(BO, LVI);
+
return true;
}
@@ -460,6 +507,9 @@ static bool processSDiv(BinaryOperator *
SDI->replaceAllUsesWith(BO);
SDI->eraseFromParent();
+ // Try to simplify our new udiv.
+ processUDivOrURem(BO, LVI);
+
return true;
}
@@ -595,6 +645,10 @@ static bool runImpl(Function &F, LazyVal
case Instruction::SDiv:
BBChanged |= processSDiv(cast<BinaryOperator>(II), LVI);
break;
+ case Instruction::UDiv:
+ case Instruction::URem:
+ BBChanged |= processUDivOrURem(cast<BinaryOperator>(II), LVI);
+ break;
case Instruction::AShr:
BBChanged |= processAShr(cast<BinaryOperator>(II), LVI);
break;
Added: llvm/trunk/test/Transforms/CorrelatedValuePropagation/udiv.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/CorrelatedValuePropagation/udiv.ll?rev=326908&view=auto
==============================================================================
--- llvm/trunk/test/Transforms/CorrelatedValuePropagation/udiv.ll (added)
+++ llvm/trunk/test/Transforms/CorrelatedValuePropagation/udiv.ll Wed Mar 7 08:56:49 2018
@@ -0,0 +1,95 @@
+; RUN: opt < %s -correlated-propagation -S | FileCheck %s
+
+; CHECK-LABEL: @test_nop
+define void @test_nop(i32 %n) {
+; CHECK udiv i32 %n, 100
+ %div = udiv i32 %n, 100
+ ret void
+}
+
+; CHECK-LABEL: @test1(
+define void @test1(i32 %n) {
+entry:
+ %cmp = icmp ule i32 %n, 65535
+ br i1 %cmp, label %bb, label %exit
+
+bb:
+; CHECK: udiv i16
+ %div = udiv i32 %n, 100
+ br label %exit
+
+exit:
+ ret void
+}
+
+; CHECK-LABEL: @test2(
+define void @test2(i32 %n) {
+entry:
+ %cmp = icmp ule i32 %n, 65536
+ br i1 %cmp, label %bb, label %exit
+
+bb:
+; CHECK: udiv i32 %n, 100
+ %div = udiv i32 %n, 100
+ br label %exit
+
+exit:
+ ret void
+}
+
+; CHECK-LABEL: @test3(
+define void @test3(i32 %m, i32 %n) {
+entry:
+ %cmp1 = icmp ult i32 %m, 65535
+ %cmp2 = icmp ult i32 %n, 65535
+ %cmp = and i1 %cmp1, %cmp2
+ br i1 %cmp, label %bb, label %exit
+
+bb:
+; CHECK: udiv i16
+ %div = udiv i32 %m, %n
+ br label %exit
+
+exit:
+ ret void
+}
+
+; CHECK-LABEL: @test4(
+define void @test4(i32 %m, i32 %n) {
+entry:
+ %cmp1 = icmp ult i32 %m, 65535
+ %cmp2 = icmp ule i32 %n, 65536
+ %cmp = and i1 %cmp1, %cmp2
+ br i1 %cmp, label %bb, label %exit
+
+bb:
+; CHECK: udiv i32 %m, %n
+ %div = udiv i32 %m, %n
+ br label %exit
+
+exit:
+ ret void
+}
+
+; CHECK-LABEL: @test5
+define void @test5(i32 %n) {
+ %trunc = and i32 %n, 65535
+ ; CHECK: udiv i16
+ %div = udiv i32 %trunc, 42
+ ret void
+}
+
+; CHECK-LABEL: @test6
+define void @test6(i32 %n) {
+entry:
+ %cmp = icmp ule i32 %n, 255
+ br i1 %cmp, label %bb, label %exit
+
+bb:
+; CHECK: udiv i8
+ %div = sdiv i32 %n, 100
+ br label %exit
+
+exit:
+ ret void
+}
Added: llvm/trunk/test/Transforms/CorrelatedValuePropagation/urem.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/CorrelatedValuePropagation/urem.ll?rev=326908&view=auto
==============================================================================
--- llvm/trunk/test/Transforms/CorrelatedValuePropagation/urem.ll (added)
+++ llvm/trunk/test/Transforms/CorrelatedValuePropagation/urem.ll Wed Mar 7 08:56:49 2018
@@ -0,0 +1,101 @@
+; RUN: opt < %s -correlated-propagation -S | FileCheck %s
+
+; CHECK-LABEL: @test_nop
+define void @test_nop(i32 %n) {
+; CHECK udiv i32 %n, 100
+ %div = udiv i32 %n, 100
+ ret void
+}
+
+; CHECK-LABEL: @test1(
+define void @test1(i32 %n) {
+entry:
+ %cmp = icmp ule i32 %n, 65535
+ br i1 %cmp, label %bb, label %exit
+
+bb:
+; CHECK: urem i16
+ %div = urem i32 %n, 100
+ br label %exit
+
+exit:
+ ret void
+}
+
+; CHECK-LABEL: @test2(
+define void @test2(i32 %n) {
+entry:
+ %cmp = icmp ule i32 %n, 65536
+ br i1 %cmp, label %bb, label %exit
+
+bb:
+; CHECK: urem i32 %n, 100
+ %div = urem i32 %n, 100
+ br label %exit
+
+exit:
+ ret void
+}
+
+; CHECK-LABEL: @test3(
+define void @test3(i32 %m, i32 %n) {
+entry:
+ %cmp1 = icmp ult i32 %m, 65535
+ %cmp2 = icmp ult i32 %n, 65535
+ %cmp = and i1 %cmp1, %cmp2
+ br i1 %cmp, label %bb, label %exit
+
+bb:
+; CHECK: urem i16
+ %div = urem i32 %m, %n
+ br label %exit
+
+exit:
+ ret void
+}
+
+; CHECK-LABEL: @test4(
+define void @test4(i32 %m, i32 %n) {
+entry:
+ %cmp1 = icmp ult i32 %m, 65535
+ %cmp2 = icmp ule i32 %n, 65536
+ %cmp = and i1 %cmp1, %cmp2
+ br i1 %cmp, label %bb, label %exit
+
+bb:
+; CHECK: urem i32 %m, %n
+ %div = urem i32 %m, %n
+ br label %exit
+
+exit:
+ ret void
+}
+
+; CHECK-LABEL: @test5
+define void @test5(i32 %n) {
+ %trunc = and i32 %n, 63
+ ; CHECK: urem i8
+ %div = urem i32 %trunc, 42
+ ret void
+}
+
+; CHECK-LABEL: @test6
+define void @test6(i32 %n) {
+entry:
+ %cmp = icmp ule i32 %n, 255
+ br i1 %cmp, label %bb, label %exit
+
+bb:
+; CHECK: urem i8
+ %div = srem i32 %n, 100
+ br label %exit
+
+exit:
+ ret void
+}
+
+; CHECK-LABEL: @non_power_of_2
+define void @non_power_of_2(i24 %n) {
+ %div = urem i24 %n, 42
+ ret void
+}
More information about the llvm-commits
mailing list