Factor out the multiply analysis code in ComputeMaskedBits and apply it to the

overflow checking multiply intrinsic as well.

Add a test for this, updating the test from grep to FileCheck.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@153028 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Nick Lewycky 2012-03-18 23:28:48 +00:00
parent 97327dc6ef
commit f201a06662
2 changed files with 150 additions and 68 deletions

View File

@ -130,6 +130,71 @@ static void ComputeMaskedBitsAddSub(bool Add, Value *Op0, Value *Op1, bool NSW,
}
}
static void ComputeMaskedBitsMul(Value *Op0, Value *Op1, bool NSW,
const APInt &Mask,
APInt &KnownZero, APInt &KnownOne,
APInt &KnownZero2, APInt &KnownOne2,
const TargetData *TD, unsigned Depth) {
unsigned BitWidth = Mask.getBitWidth();
APInt Mask2 = APInt::getAllOnesValue(BitWidth);
ComputeMaskedBits(Op1, Mask2, KnownZero, KnownOne, TD, Depth+1);
ComputeMaskedBits(Op0, Mask2, KnownZero2, KnownOne2, TD, Depth+1);
assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
bool isKnownNegative = false;
bool isKnownNonNegative = false;
// If the multiplication is known not to overflow, compute the sign bit.
if (Mask.isNegative() && NSW) {
if (Op0 == Op1) {
// The product of a number with itself is non-negative.
isKnownNonNegative = true;
} else {
bool isKnownNonNegativeOp1 = KnownZero.isNegative();
bool isKnownNonNegativeOp0 = KnownZero2.isNegative();
bool isKnownNegativeOp1 = KnownOne.isNegative();
bool isKnownNegativeOp0 = KnownOne2.isNegative();
// The product of two numbers with the same sign is non-negative.
isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) ||
(isKnownNonNegativeOp1 && isKnownNonNegativeOp0);
// The product of a negative number and a non-negative number is either
// negative or zero.
if (!isKnownNonNegative)
isKnownNegative = (isKnownNegativeOp1 && isKnownNonNegativeOp0 &&
isKnownNonZero(Op0, TD, Depth)) ||
(isKnownNegativeOp0 && isKnownNonNegativeOp1 &&
isKnownNonZero(Op1, TD, Depth));
}
}
// If low bits are zero in either operand, output low known-0 bits.
// Also compute a conserative estimate for high known-0 bits.
// More trickiness is possible, but this is sufficient for the
// interesting case of alignment computation.
KnownOne.clearAllBits();
unsigned TrailZ = KnownZero.countTrailingOnes() +
KnownZero2.countTrailingOnes();
unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
KnownZero2.countLeadingOnes(),
BitWidth) - BitWidth;
TrailZ = std::min(TrailZ, BitWidth);
LeadZ = std::min(LeadZ, BitWidth);
KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
APInt::getHighBitsSet(BitWidth, LeadZ);
KnownZero &= Mask;
// Only make use of no-wrap flags if we failed to compute the sign bit
// directly. This matters if the multiplication always overflows, in
// which case we prefer to follow the result of the direct computation,
// though as the program is invoking undefined behaviour we can choose
// whatever we like here.
if (isKnownNonNegative && !KnownOne.isNegative())
KnownZero.setBit(BitWidth - 1);
else if (isKnownNegative && !KnownZero.isNegative())
KnownOne.setBit(BitWidth - 1);
}
/// ComputeMaskedBits - Determine which of the bits specified in Mask are
/// known to be either zero or one and return them in the KnownZero/KnownOne
/// bit sets. This code only analyzes bits in Mask, in order to short-circuit
@ -294,68 +359,11 @@ void llvm::ComputeMaskedBits(Value *V, const APInt &Mask,
return;
}
case Instruction::Mul: {
APInt Mask2 = APInt::getAllOnesValue(BitWidth);
ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero, KnownOne, TD,Depth+1);
ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
Depth+1);
assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
bool isKnownNegative = false;
bool isKnownNonNegative = false;
// If the multiplication is known not to overflow, compute the sign bit.
if (Mask.isNegative() &&
cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap()) {
Value *Op1 = I->getOperand(1), *Op2 = I->getOperand(0);
if (Op1 == Op2) {
// The product of a number with itself is non-negative.
isKnownNonNegative = true;
} else {
bool isKnownNonNegative1 = KnownZero.isNegative();
bool isKnownNonNegative2 = KnownZero2.isNegative();
bool isKnownNegative1 = KnownOne.isNegative();
bool isKnownNegative2 = KnownOne2.isNegative();
// The product of two numbers with the same sign is non-negative.
isKnownNonNegative = (isKnownNegative1 && isKnownNegative2) ||
(isKnownNonNegative1 && isKnownNonNegative2);
// The product of a negative number and a non-negative number is either
// negative or zero.
if (!isKnownNonNegative)
isKnownNegative = (isKnownNegative1 && isKnownNonNegative2 &&
isKnownNonZero(Op2, TD, Depth)) ||
(isKnownNegative2 && isKnownNonNegative1 &&
isKnownNonZero(Op1, TD, Depth));
}
}
// If low bits are zero in either operand, output low known-0 bits.
// Also compute a conserative estimate for high known-0 bits.
// More trickiness is possible, but this is sufficient for the
// interesting case of alignment computation.
KnownOne.clearAllBits();
unsigned TrailZ = KnownZero.countTrailingOnes() +
KnownZero2.countTrailingOnes();
unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
KnownZero2.countLeadingOnes(),
BitWidth) - BitWidth;
TrailZ = std::min(TrailZ, BitWidth);
LeadZ = std::min(LeadZ, BitWidth);
KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
APInt::getHighBitsSet(BitWidth, LeadZ);
KnownZero &= Mask;
// Only make use of no-wrap flags if we failed to compute the sign bit
// directly. This matters if the multiplication always overflows, in
// which case we prefer to follow the result of the direct computation,
// though as the program is invoking undefined behaviour we can choose
// whatever we like here.
if (isKnownNonNegative && !KnownOne.isNegative())
KnownZero.setBit(BitWidth - 1);
else if (isKnownNegative && !KnownZero.isNegative())
KnownOne.setBit(BitWidth - 1);
return;
bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
ComputeMaskedBitsMul(I->getOperand(0), I->getOperand(1), NSW,
Mask, KnownZero, KnownOne, KnownZero2, KnownOne2,
TD, Depth);
break;
}
case Instruction::UDiv: {
// For the purposes of computing leading zeros we can conservatively
@ -777,6 +785,12 @@ void llvm::ComputeMaskedBits(Value *V, const APInt &Mask,
KnownZero, KnownOne, KnownZero2, KnownOne2,
TD, Depth);
break;
case Intrinsic::umul_with_overflow:
case Intrinsic::smul_with_overflow:
ComputeMaskedBitsMul(II->getArgOperand(0), II->getArgOperand(1),
false, Mask, KnownZero, KnownOne,
KnownZero2, KnownOne2, TD, Depth);
break;
}
}
}

View File

@ -1,116 +1,184 @@
; This test makes sure that mul instructions are properly eliminated.
; RUN: opt < %s -instcombine -S | not grep mul
; RUN: opt < %s -instcombine -S | FileCheck %s
define i32 @test1(i32 %A) {
; CHECK: @test1
%B = mul i32 %A, 1 ; <i32> [#uses=1]
ret i32 %B
; CHECK: ret i32 %A
}
define i32 @test2(i32 %A) {
; CHECK: @test2
; Should convert to an add instruction
%B = mul i32 %A, 2 ; <i32> [#uses=1]
ret i32 %B
; CHECK: shl i32 %A, 1
}
define i32 @test3(i32 %A) {
; CHECK: @test3
; This should disappear entirely
%B = mul i32 %A, 0 ; <i32> [#uses=1]
ret i32 %B
; CHECK: ret i32 0
}
define double @test4(double %A) {
; CHECK: @test4
; This is safe for FP
%B = fmul double 1.000000e+00, %A ; <double> [#uses=1]
ret double %B
; CHECK: ret double %A
}
define i32 @test5(i32 %A) {
; CHECK: @test5
%B = mul i32 %A, 8 ; <i32> [#uses=1]
ret i32 %B
; CHECK: shl i32 %A, 3
}
define i8 @test6(i8 %A) {
; CHECK: @test6
%B = mul i8 %A, 8 ; <i8> [#uses=1]
%C = mul i8 %B, 8 ; <i8> [#uses=1]
ret i8 %C
; CHECK: shl i8 %A, 6
}
define i32 @test7(i32 %i) {
; CHECK: @test7
%tmp = mul i32 %i, -1 ; <i32> [#uses=1]
ret i32 %tmp
; CHECK: sub i32 0, %i
}
define i64 @test8(i64 %i) {
; tmp = sub 0, %i
; CHECK: @test8
%j = mul i64 %i, -1 ; <i64> [#uses=1]
ret i64 %j
; CHECK: sub i64 0, %i
}
define i32 @test9(i32 %i) {
; %j = sub 0, %i
; CHECK: @test9
%j = mul i32 %i, -1 ; <i32> [#uses=1]
ret i32 %j
; CHECJ: sub i32 0, %i
}
define i32 @test10(i32 %a, i32 %b) {
; CHECK: @test10
%c = icmp slt i32 %a, 0 ; <i1> [#uses=1]
%d = zext i1 %c to i32 ; <i32> [#uses=1]
; e = b & (a >> 31)
%e = mul i32 %d, %b ; <i32> [#uses=1]
ret i32 %e
; CHECK: [[TEST10:%.*]] = ashr i32 %a, 31
; CHECK-NEXT: %e = and i32 [[TEST10]], %b
; CHECK-NEXT: ret i32 %e
}
define i32 @test11(i32 %a, i32 %b) {
; CHECK: @test11
%c = icmp sle i32 %a, -1 ; <i1> [#uses=1]
%d = zext i1 %c to i32 ; <i32> [#uses=1]
; e = b & (a >> 31)
%e = mul i32 %d, %b ; <i32> [#uses=1]
ret i32 %e
; CHECK: [[TEST11:%.*]] = ashr i32 %a, 31
; CHECK-NEXT: %e = and i32 [[TEST11]], %b
; CHECK-NEXT: ret i32 %e
}
define i32 @test12(i8 %a, i32 %b) {
%c = icmp ugt i8 %a, 127 ; <i1> [#uses=1]
define i32 @test12(i32 %a, i32 %b) {
; CHECK: @test12
%c = icmp ugt i32 %a, 2147483647 ; <i1> [#uses=1]
%d = zext i1 %c to i32 ; <i32> [#uses=1]
; e = b & (a >> 31)
%e = mul i32 %d, %b ; <i32> [#uses=1]
ret i32 %e
; CHECK: [[TEST12:%.*]] = ashr i32 %a, 31
; CHECK-NEXT: %e = and i32 [[TEST12]], %b
; CHECK-NEXT: ret i32 %e
}
; PR2642
define internal void @test13(<4 x float>*) {
; CHECK: @test13
load <4 x float>* %0, align 1
fmul <4 x float> %2, < float 1.000000e+00, float 1.000000e+00, float 1.000000e+00, float 1.000000e+00 >
store <4 x float> %3, <4 x float>* %0, align 1
ret void
; CHECK-NEXT: ret void
}
define <16 x i8> @test14(<16 x i8> %a) {
; CHECK: @test14
%b = mul <16 x i8> %a, zeroinitializer
ret <16 x i8> %b
; CHECK-NEXT: ret <16 x i8> zeroinitializer
}
; rdar://7293527
define i32 @test15(i32 %A, i32 %B) {
; CHECK: @test15
entry:
%shl = shl i32 1, %B
%m = mul i32 %shl, %A
ret i32 %m
; CHECK: shl i32 %A, %B
}
; X * Y (when Y is 0 or 1) --> x & (0-Y)
define i32 @test16(i32 %b, i1 %c) {
; CHECK: @test16
%d = zext i1 %c to i32 ; <i32> [#uses=1]
; e = b & (a >> 31)
%e = mul i32 %d, %b ; <i32> [#uses=1]
ret i32 %e
; CHECK: [[TEST16:%.*]] = sext i1 %c to i32
; CHECK-NEXT: %e = and i32 [[TEST16]], %b
; CHECK-NEXT: ret i32 %e
}
; X * Y (when Y is 0 or 1) --> x & (0-Y)
define i32 @test17(i32 %a, i32 %b) {
; CHECK: @test17
%a.lobit = lshr i32 %a, 31
%e = mul i32 %a.lobit, %b
ret i32 %e
; CHECK: [[TEST17:%.*]] = ashr i32 %a, 31
; CHECK-NEXT: %e = and i32 [[TEST17]], %b
; CHECK-NEXT: ret i32 %e
}
define i32 @test18(i32 %A, i32 %B) {
; CHECK: @test18
%C = and i32 %A, 1
%D = and i32 %B, 1
%E = mul i32 %C, %D
%F = and i32 %E, 16
ret i32 %F
; CHECK-NEXT: ret i32 0
}
declare {i32, i1} @llvm.smul.with.overflow.i32(i32, i32)
declare void @use(i1)
define i32 @test19(i32 %A, i32 %B) {
; CHECK: @test19
%C = and i32 %A, 1
%D = and i32 %B, 1
; It would be nice if we also started proving that this doesn't overflow.
%E = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %C, i32 %D)
%F = extractvalue {i32, i1} %E, 0
%G = extractvalue {i32, i1} %E, 1
call void @use(i1 %G)
%H = and i32 %F, 16
ret i32 %H
; CHECK: ret i32 0
}