pimhwang
/
ILSpy
mirror of https://github.com/icsharpcode/ILSpy.git
1
0
Fork 0
Code Issues Projects Releases Wiki Activity
Browse Source

Properly fix the integer conversion mess. 

It's not actually all that hard once you realize:
a) the ECMA-335 specification is incorrect, whether to sign- or zero-extend does not always depends on the source type
b) the C# compilator generates completely weird code when casting between (U)IntPtr and integers.
   Map all IntPtr casts to a few semi-sane cases, and we can obtain the correct semantics without introducing any helper methods.
pull/728/head
Daniel Grunwald 9 years ago
parent
95c8e4bd00
commit
d07298b747
6 changed files with 161 additions and 104 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 32
      ICSharpCode.Decompiler/CSharp/CSharpDecompiler.cs
  2. 54
      ICSharpCode.Decompiler/CSharp/ExpressionBuilder.cs
  3. 70
      ICSharpCode.Decompiler/CSharp/TranslatedExpression.cs
  4. 2
      ICSharpCode.Decompiler/Tests/RoundtripAssembly.cs
  5. 1
      ILSpy.sln
  6. 94
      doc/IntPtr.txt

32
ICSharpCode.Decompiler/CSharp/CSharpDecompiler.cs
Unescape View File

@ -485,11 +485,6 @@ namespace ICSharpCode.Decompiler.CSharp @@ -485,11 +485,6 @@ namespace ICSharpCode.Decompiler.CSharp
}
}
/// <summary>
/// Whether we need to generate helper methods for u4->i or u->i8 conversions.
/// </summary>
internal bool needs_conv_i_ovf_un, needs_conv_i8_ovf_un;
EntityDeclaration DoDecompile(ITypeDefinition typeDef, ITypeResolveContext decompilationContext)
{
Debug.Assert(decompilationContext.CurrentTypeDefinition == typeDef);
@ -547,30 +542,6 @@ namespace ICSharpCode.Decompiler.CSharp @@ -547,30 +542,6 @@ namespace ICSharpCode.Decompiler.CSharp
section.Remove();
}
}
if (needs_conv_i_ovf_un) {
typeDecl.Members.Add(GenerateConvHelper(
"conv_i_ovf_un", KnownTypeCode.UInt32, KnownTypeCode.IntPtr, typeSystemAstBuilder,
// on 32-bit, 'conv.ovf u4->i' is like 'conv.ovf u4->i4'
new CheckedExpression(new CastExpression(
new NRefactory.CSharp.PrimitiveType("int"),
new IdentifierExpression("input")
)),
// on 64-bit, 'conv.ovf u4->i' is like 'conv.ovf u4->i8'
new IdentifierExpression("input")
));
needs_conv_i_ovf_un = false;
}
if (needs_conv_i8_ovf_un) {
typeDecl.Members.Add(GenerateConvHelper(
"conv_i8_ovf_un", KnownTypeCode.UIntPtr, KnownTypeCode.Int64, typeSystemAstBuilder,
// on 32-bit, 'conv.ovf u->i8' is like 'conv.ovf u4->i8'
new IdentifierExpression("input"),
// on 64-bit, 'conv.ovf u->i8' is like 'conv.ovf u8->i8'
new IdentifierExpression("input")
));
needs_conv_i8_ovf_un = false;
}
return typeDecl;
}
@ -675,9 +646,6 @@ namespace ICSharpCode.Decompiler.CSharp @@ -675,9 +646,6 @@ namespace ICSharpCode.Decompiler.CSharp
var statementBuilder = new StatementBuilder(decompilationContext, method);
var body = statementBuilder.ConvertAsBlock(function.Body);
needs_conv_i_ovf_un |= statementBuilder.exprBuilder.needs_conv_i_ovf_un;
needs_conv_i8_ovf_un |= statementBuilder.exprBuilder.needs_conv_i8_ovf_un;
entityDecl.AddChild(body, Roles.Body);
}

54
ICSharpCode.Decompiler/CSharp/ExpressionBuilder.cs
Unescape View File

@ -609,11 +609,6 @@ namespace ICSharpCode.Decompiler.CSharp @@ -609,11 +609,6 @@ namespace ICSharpCode.Decompiler.CSharp
return result;
}
/// <summary>
/// Whether we need to generate helper methods for u4->i or u->i8 conversions.
/// </summary>
internal bool needs_conv_i_ovf_un, needs_conv_i8_ovf_un;
protected internal override TranslatedExpression VisitConv(Conv inst)
{
var arg = Translate(inst.Argument);
@ -627,7 +622,9 @@ namespace ICSharpCode.Decompiler.CSharp @@ -627,7 +622,9 @@ namespace ICSharpCode.Decompiler.CSharp
// Also, we need to be very careful with regards to the conversions we emit:
// In C#, zero vs. sign-extension depends on the input type,
// but in the ILAst Conv instruction it depends on the output type.
// but in the ILAst conv instruction it depends on the output type.
// However, in the conv.ovf instructions, the .NET runtime behavior seems to depend on the input type,
// in violation of the ECMA-335 spec!
if (inst.CheckForOverflow || inst.Kind == ConversionKind.IntToFloat) {
// We need to first convert the argument to the expected sign.
@ -636,52 +633,11 @@ namespace ICSharpCode.Decompiler.CSharp @@ -636,52 +633,11 @@ namespace ICSharpCode.Decompiler.CSharp
if (arg.Type.GetSize() > inputStackType.GetSize() || arg.Type.GetSign() != inst.InputSign) {
arg = arg.ConvertTo(compilation.FindType(inputStackType.ToKnownTypeCode(inst.InputSign)), this);
}
if (inst.Kind == ConversionKind.ZeroExtend) {
// Zero extension with overflow check -> throws if the input value is negative.
// C# sign/zero extension depends on the source type, so we can't directly
// cast from the signed input type to the target type.
// Instead, perform a checked cast to the unsigned version of the input type
// (to throw an exception for negative values).
// Then the actual zero extension can be left to our parent instruction
// (due to the ExpressionBuilder post-condition being flexible with regards to the integer type width).
return arg.ConvertTo(compilation.FindType(inputStackType.ToKnownTypeCode(Sign.Unsigned)), this, true)
.WithILInstruction(inst);
} else if (inst.Kind == ConversionKind.SignExtend) {
// Sign extension with overflow check.
// Sign-extending conversions can fail when the "larger type" isn't actually larger, that is, in exactly two cases:
// * U4 -> I on 32-bit
// * U -> I8 on 64-bit
if (inst.InputSign == Sign.Unsigned && inputStackType == StackType.I4 && inst.TargetType == IL.PrimitiveType.I) {
// conv u4->i
// on 32-bit, this is a sign-changing conversion with overflow-check
// on 64-bit, this is a sign extension
needs_conv_i_ovf_un = true;
arg = arg.ConvertTo(compilation.FindType(KnownTypeCode.UInt32), this);
return new InvocationExpression(new IdentifierExpression("conv_i_ovf_un"), arg.Expression)
.WithRR(new ResolveResult(compilation.FindType(KnownTypeCode.IntPtr)))
.WithILInstruction(inst);
} else if (inst.InputSign == Sign.Unsigned && (inputStackType == StackType.I && inst.TargetType == IL.PrimitiveType.I8)) {
// conv u->i8
// on 32-bit, this is a sign extension
// on 64-bit, this is a sign-changing conversion with overflow-check
needs_conv_i8_ovf_un = true;
arg = arg.ConvertTo(compilation.FindType(KnownTypeCode.UIntPtr), this);
return new InvocationExpression(new IdentifierExpression("conv_i8_ovf_un"), arg.Expression)
.WithRR(new ResolveResult(compilation.FindType(KnownTypeCode.IntPtr)))
.WithILInstruction(inst);
} else {
// The overflow check cannot actually fail, so we can take the simple solution of performing
// an unchecked cast to signed int and let the parent instruction handle the actual sign extension.
return arg.ConvertTo(compilation.FindType(inputStackType.ToKnownTypeCode(Sign.Signed)), this)
.WithILInstruction(inst);
}
} else {
// Size-preserving sign-changing conversion, or int-to-float conversion:
// We can directly cast to the target type.
// Because casts with overflow check match C# semantics (zero/sign-extension depends on source type),
// we can just directly cast to the target type.
return arg.ConvertTo(compilation.FindType(inst.TargetType.ToKnownTypeCode()), this, true)
.WithILInstruction(inst);
}
}
switch (inst.Kind) {
case ConversionKind.StopGCTracking:

70
ICSharpCode.Decompiler/CSharp/TranslatedExpression.cs
Unescape View File

@ -164,6 +164,9 @@ namespace ICSharpCode.Decompiler.CSharp @@ -164,6 +164,9 @@ namespace ICSharpCode.Decompiler.CSharp
/// sign of the source type.
/// This fits with the ExpressionBuilder's post-condition, so e.g. an assignment can simply
/// call <c>Translate(stloc.Value).ConvertTo(stloc.Variable.Type)</c> and have the overall C# semantics match the IL semantics.
///
/// From the caller's perspective, IntPtr/UIntPtr behave like normal C# integers except that they have native int size.
/// All the special cases necessary to make IntPtr/UIntPtr behave sanely are handled internally in ConvertTo().
/// </remarks>
public TranslatedExpression ConvertTo(IType targetType, ExpressionBuilder expressionBuilder, bool checkForOverflow = false)
{
@ -179,31 +182,66 @@ namespace ICSharpCode.Decompiler.CSharp @@ -179,31 +182,66 @@ namespace ICSharpCode.Decompiler.CSharp
LdcI4(compilation, 0).ConvertTo(targetType, expressionBuilder, checkForOverflow)
).WithoutILInstruction().WithRR(new ResolveResult(targetType));
}
// Special-case IntPtr and UIntPtr: they behave slightly weird, e.g. converting to them always checks for overflow,
// but converting from them never checks for overflow.
if (checkForOverflow && type.IsKnownType(KnownTypeCode.IntPtr) && !targetType.IsKnownType(KnownTypeCode.Int64)) {
// Convert through `long` instead.
if (targetType.IsKnownType(KnownTypeCode.Boolean)) {
// convert to boolean through byte, to simulate the truncation to 8 bits
return this.ConvertTo(compilation.FindType(KnownTypeCode.Byte), expressionBuilder, checkForOverflow)
.ConvertToBoolean(expressionBuilder);
}
// Special-case IntPtr and UIntPtr: they behave extremely weird, see IntPtr.txt for details.
if (type.IsKnownType(KnownTypeCode.IntPtr)) { // Conversion from IntPtr
// Direct cast only works correctly for IntPtr -> long.
// IntPtr -> int works correctly only in checked context.
// Everything else can be worked around by casting via long.
if (!(targetType.IsKnownType(KnownTypeCode.Int64) || checkForOverflow && targetType.IsKnownType(KnownTypeCode.Int32))) {
return this.ConvertTo(compilation.FindType(KnownTypeCode.Int64), expressionBuilder, checkForOverflow)
.ConvertTo(targetType, expressionBuilder, checkForOverflow);
} else if (checkForOverflow && type.IsKnownType(KnownTypeCode.UIntPtr) && !targetType.IsKnownType(KnownTypeCode.UInt64)) {
// Convert through `ulong` instead.
}
} else if (type.IsKnownType(KnownTypeCode.UIntPtr)) { // Conversion from UIntPtr
// Direct cast only works correctly for UIntPtr -> ulong.
// UIntPtr -> uint works correctly only in checked context.
// Everything else can be worked around by casting via ulong.
if (!(targetType.IsKnownType(KnownTypeCode.UInt64) || checkForOverflow && targetType.IsKnownType(KnownTypeCode.UInt32))) {
return this.ConvertTo(compilation.FindType(KnownTypeCode.UInt64), expressionBuilder, checkForOverflow)
.ConvertTo(targetType, expressionBuilder, checkForOverflow);
}
if (targetType.IsKnownType(KnownTypeCode.Boolean)) {
// convert to boolean through byte, to simulate the truncation to 8 bits
return this.ConvertTo(compilation.FindType(KnownTypeCode.Byte), expressionBuilder, checkForOverflow)
.ConvertToBoolean(expressionBuilder);
}
if ((targetType.IsKnownType(KnownTypeCode.IntPtr) || targetType.IsKnownType(KnownTypeCode.UIntPtr))
&& type.Kind != TypeKind.Pointer && !checkForOverflow)
{
// (u)long -> (U)IntPtr casts in C# can throw overflow exceptions in 32-bit mode, even in unchecked context.
// To avoid those, convert via `void*`.
if (targetType.IsKnownType(KnownTypeCode.IntPtr)) { // Conversion to IntPtr
if (type.IsKnownType(KnownTypeCode.Int32)) {
// normal casts work for int (both in checked and unchecked context)
} else if (checkForOverflow) {
// if overflow-checking is enabled, we can simply cast via long:
// (and long itself works directly in checked context)
if (!type.IsKnownType(KnownTypeCode.Int64)) {
return this.ConvertTo(compilation.FindType(KnownTypeCode.Int64), expressionBuilder, checkForOverflow)
.ConvertTo(targetType, expressionBuilder, checkForOverflow);
}
} else {
// If overflow-checking is disabled, the only way to truncate to native size
// without throwing an exception in 32-bit mode is to use a pointer type.
if (type.Kind != TypeKind.Pointer) {
return this.ConvertTo(new PointerType(compilation.FindType(KnownTypeCode.Void)), expressionBuilder, checkForOverflow)
.ConvertTo(targetType, expressionBuilder, checkForOverflow);
}
}
} else if (targetType.IsKnownType(KnownTypeCode.UIntPtr)) { // Conversion to UIntPtr
if (type.IsKnownType(KnownTypeCode.UInt32) || type.Kind == TypeKind.Pointer) {
// normal casts work for uint and pointers (both in checked and unchecked context)
} else if (checkForOverflow) {
// if overflow-checking is enabled, we can simply cast via ulong:
// (and ulong itself works directly in checked context)
if (!type.IsKnownType(KnownTypeCode.UInt64)) {
return this.ConvertTo(compilation.FindType(KnownTypeCode.UInt64), expressionBuilder, checkForOverflow)
.ConvertTo(targetType, expressionBuilder, checkForOverflow);
}
} else {
// If overflow-checking is disabled, the only way to truncate to native size
// without throwing an exception in 32-bit mode is to use a pointer type.
return this.ConvertTo(new PointerType(compilation.FindType(KnownTypeCode.Void)), expressionBuilder, checkForOverflow)
.ConvertTo(targetType, expressionBuilder, checkForOverflow);
}
}
if (targetType.Kind == TypeKind.Pointer && type.Kind == TypeKind.Enum) {
// enum to pointer: C# doesn't allow such casts
// -> convert via underlying type

2
ICSharpCode.Decompiler/Tests/RoundtripAssembly.cs
Unescape View File

@ -98,7 +98,7 @@ namespace ICSharpCode.Decompiler.Tests @@ -98,7 +98,7 @@ namespace ICSharpCode.Decompiler.Tests
{
try {
RunWithOutput("Random Tests\\TestCases", "TestCase-1.exe");
} catch (AssertionException ex) {
} catch (CompilationFailedException ex) {
Assert.Ignore(ex.Message);
}
}

1
ILSpy.sln
Unescape View File

@ -8,6 +8,7 @@ Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "doc", "doc", "{F45DB999-7E7 @@ -8,6 +8,7 @@ Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "doc", "doc", "{F45DB999-7E7
ProjectSection(SolutionItems) = preProject
doc\Command Line.txt = doc\Command Line.txt
doc\ILAst.txt = doc\ILAst.txt
doc\IntPtr.txt = doc\IntPtr.txt
EndProjectSection
EndProject
Project("{FAE04EC0-301F-11D3-BF4B-00C04F79EFBC}") = "ILSpy", "ILSpy\ILSpy.csproj", "{1E85EFF9-E370-4683-83E4-8A3D063FF791}"

94
doc/IntPtr.txt
Unescape View File

@ -0,0 +1,94 @@ @@ -0,0 +1,94 @@
C# casts to/from IntPtr and UIntPtr don't behave like one would expect from the normal C# primitive types.
For example, they don't fully respect the checked/unchecked context.
First, let's consider what methods we have available for converting between (U)IntPtr and normal C# types.
Primitives for constructing IntPtr/UIntPtr:
* new IntPtr(int) equivalent to: conv i4->i <sign extend>
* new IntPtr(long) equivalent to: conv.ovf i8->i
* new IntPtr(void*) equivalent to: nop
* new UIntPtr(uint) equivalent to: conv u4->u <zero extend>
* new UIntPtr(ulong) equivalent to: conv.ovf u8->u
* new UIntPtr(void*) equivalent to: nop
Primitives for getting the value back out:
* IntPtr.ToInt32() equivalent to: conv.ovf i->i4
* IntPtr.ToInt64() equivalent to: conv i->i8 <sign extend>
* IntPtr.ToPointer() equivalent to: nop
* UIntPtr.ToUInt32() equivalent to: conv.ovf u->u4
* UIntPtr.ToUInt64() equivalent to: conv u->u8 <zero extend>
* UIntPtr.ToPointer() equivalent to: nop
The (U)IntPtr.op_Explicit implementations are equivalent to the corresponding primitives.
(void*) is a useful type because all (U)IntPtr<->void* conversions are no-ops.
C# pointer types act like a normal C# unsigned integer type (of 'native int' size), so we can
use `void*` whenever the target type is unsigned or the sign does not matter (overflow checking disabled).
Next, we'll consider what the C# compiler does when casting between integer types and IntPtr.
I tried all these conversions in both checked and unchecked mode, and the C# compiler was
always generating the same code in both modes!
OK = cast behavior is as if IntPtr was a built-in type and the context does not matter:
* sign/zero extension depending on source type
* never throws OverflowException and never is supposed to
CC = cast behavior is as if IntPtr was a built-in type and we are in a checked context:
* sign/zero extension depending on source type
* performs correct overflowing checking as in a direct cast to native (u)int
from -> to = C# cast
generated opcode sequence = what csc.exe produces for that cast
from -> to : generated opcode sequence overall effect equivalent to
OK short -> IntPtr: call op_Explicit(int32) conv i2->i <sign extend>
OK ushort -> IntPtr: call op_Explicit(int32) conv u2->u <zero extend>
Sign extension in op_Explicit does not matter because sign bit is always 0 at that point.
OK int -> IntPtr: call op_Explicit(int32) conv i4->i <sign extend>
CC uint -> IntPtr: conv.u8 + call op_Explicit(int64) conv.ovf u4->i <zero extend>
CC long -> IntPtr: call op_Explicit(int64) conv.ovf i8->i
ulong -> IntPtr: call op_Explicit(int64) conv.ovf i8->i
short -> UIntPtr: conv.i8 + call op_Explicit(uint64) 32-bit: conv.ovf i2->u <sign extend>; 64-bit: conv i2->i <sign extend>
OverflowException for negative input values only on 32-bit!
OK ushort -> UIntPtr: call op_Explicit(uint32) conv u2->u <zero extend>
int -> UIntPtr: conv.i8 + call op_Explicit(uint64) 32-bit: conv.ovf i4->u <sign extend>; 64-bit: conv i4->i <sign extend>
OverflowException for negative input values only on 32-bit!
OK uint -> UIntPtr: call op_Explicit(uint32) conv u4->u <zero extend>
long -> UIntPtr: call op_Explicit(uint64) conv.ovf u8->u
CC ulong -> UIntPtr: call op_Explicit(uint64) conv.ovf u8->u
If an unchecked conversion is desired and the desired entry is not marked 'OK',
we work around the problem by casting sourceType->void*->(U)IntPtr.
If a checked conversion is desired and the desired entry is not marked 'OK' or 'CC', we have to find a replacement.
signed type -> UIntPtr: (UIntPtr)(void*)value
ulong -> IntPtr: (IntPtr)(long)value
Continuing the conversion table for the other direction, (UIntPtr) to primitive types:
from -> to : generated opcode sequence overall effect equivalent to
IntPtr -> short: call int32 op_Explicit + conv.i2 conv.ovf i->i4; conv i4->i2
IntPtr -> ushort: call int32 op_Explicit + conv.u2 conv.ovf i->i4; conv i4->u2
CC IntPtr -> int: call int32 op_Explicit conv.ovf i->i4
IntPtr -> uint: call int32 op_Explicit conv.ovf i->i4
OK IntPtr -> long: call int64 op_Explicit conv i->i8 <sign extend>
IntPtr -> ulong: call int64 op_Explicit conv i->i8 <sign extend>
UIntPtr -> short: call uint32 op_Explicit + conv.i2 conv.ovf u->u4; conv u4->i2
UIntPtr -> ushort: call uint32 op_Explicit + conv.u2 conv.ovf u->u4; conv u4->u2
UIntPtr -> int: call uint32 op_Explicit conv.ovf u->u4
CC UIntPtr -> uint: call uint32 op_Explicit conv.ovf u->u4
UIntPtr -> long: call uint64 op_Explicit conv u->u8 <zero extend>
OK UIntPtr -> ulong: call uint64 op_Explicit conv u->u8 <zero extend>
If an unchecked conversion is desired and the desired entry is not marked 'OK',
we work around the problem by casting (U)IntPtr->(u)long->targetType.
(`void*` would also work instead of `ulong`/`long`, but let's avoid unsafe code where possible)
If a checked conversion is desired and the desired entry is not marked 'OK' or 'CC',
we also have to work around the problem, and this again works by casting via (u)long: (U)IntPtr->(u)long->targetType
(note that `void*` is not always a valid alternative in this case)
Finally, conversions between IntPtr and (U)IntPtr, or IntPtr and `void*` need special consideration:
* C# does not allow directly casting IntPtr <-> UIntPtr
* Casting via `void*` works but is always unchecked.
* These should work for checked conversions:
IntPtr -> UIntPtr: cast IntPtr->long->ulong->UIntPtr
IntPtr -> void*: cast IntPtr->long->void*
UIntPtr -> IntPtr: cast UIntPtr->ulong->long->IntPtr
void* -> IntPtr: cast void*->long->IntPtr
Write Preview
Loading…
Cancel
Save