tint: implement updated spec rules for shift left of concrete values

Rules for shift left of concrete values are now split between signed and
unsigned. Shifting unsigned values no longer fails with "sign change"
errors. Furthermore, shifting unsigned values must only shift out 0s.

See https://github.com/gpuweb/gpuweb/pull/3539.

Bug: tint:1701
Bug: tint:1717
Change-Id: Iba2799f4b02cdc77cc58a6c7c104aaa408f0f0f9
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/106381
Reviewed-by: Dan Sinclair <dsinclair@chromium.org>
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: Antonio Maiorano <amaiorano@google.com>
This commit is contained in:
Antonio Maiorano 2022-10-19 17:46:12 +00:00 committed by Dawn LUCI CQ
parent 3fa6887679
commit 1c94938726
10 changed files with 135 additions and 25 deletions

View File

@ -1491,13 +1491,25 @@ ConstEval::Result ConstEval::OpShiftLeft(const sem::Type* ty,
return nullptr;
}
// The e2 + 1 most significant bits of e1 must have the same bit value, otherwise
// sign change (overflow) would occur.
size_t must_match_msb = e2u + 1;
UT mask = ~UT{0} << (bit_width - must_match_msb);
if ((e1u & mask) != 0 && (e1u & mask) != mask) {
AddError("shift left operation results in sign change", source);
return nullptr;
if constexpr (std::is_signed_v<T>) {
// If T is a signed integer type, and the e2+1 most significant bits of e1 do
// not have the same bit value, then error.
size_t must_match_msb = e2u + 1;
UT mask = ~UT{0} << (bit_width - must_match_msb);
if ((e1u & mask) != 0 && (e1u & mask) != mask) {
AddError("shift left operation results in sign change", source);
return nullptr;
}
} else {
// If T is an unsigned integer type, and any of the e2 most significant bits of
// e1 are 1, then error.
if (e2u > 0) {
size_t must_be_zero_msb = e2u;
UT mask = ~UT{0} << (bit_width - must_be_zero_msb);
if ((e1u & mask) != 0) {
AddError(OverflowErrorMessage(e1, "<<", e2), source);
}
}
}
}

View File

@ -611,19 +611,16 @@ std::vector<Case> ShiftLeftCases() {
using ST = std::conditional_t<IsAbstract<T>, T, u32>;
using B = BitValues<T>;
auto r = std::vector<Case>{
C(T{0b1010}, ST{0}, T{0b0000'0000'1010}), //
C(T{0b1010}, ST{1}, T{0b0000'0001'0100}), //
C(T{0b1010}, ST{2}, T{0b0000'0010'1000}), //
C(T{0b1010}, ST{3}, T{0b0000'0101'0000}), //
C(T{0b1010}, ST{4}, T{0b0000'1010'0000}), //
C(T{0b1010}, ST{5}, T{0b0001'0100'0000}), //
C(T{0b1010}, ST{6}, T{0b0010'1000'0000}), //
C(T{0b1010}, ST{7}, T{0b0101'0000'0000}), //
C(T{0b1010}, ST{8}, T{0b1010'0000'0000}), //
C(B::LeftMost, ST{0}, B::LeftMost), //
C(B::TwoLeftMost, ST{1}, B::LeftMost), // No overflow
C(B::All, ST{1}, B::AllButRightMost), // No overflow
C(B::All, ST{B::NumBits - 1}, B::LeftMost), // No overflow
C(T{0b1010}, ST{0}, T{0b0000'0000'1010}), //
C(T{0b1010}, ST{1}, T{0b0000'0001'0100}), //
C(T{0b1010}, ST{2}, T{0b0000'0010'1000}), //
C(T{0b1010}, ST{3}, T{0b0000'0101'0000}), //
C(T{0b1010}, ST{4}, T{0b0000'1010'0000}), //
C(T{0b1010}, ST{5}, T{0b0001'0100'0000}), //
C(T{0b1010}, ST{6}, T{0b0010'1000'0000}), //
C(T{0b1010}, ST{7}, T{0b0101'0000'0000}), //
C(T{0b1010}, ST{8}, T{0b1010'0000'0000}), //
C(B::LeftMost, ST{0}, B::LeftMost), //
C(Vec(T{0b1010}, T{0b1010}), //
Vec(ST{0}, ST{1}), //
@ -641,7 +638,7 @@ std::vector<Case> ShiftLeftCases() {
// Only abstract 0 can be shifted left as much as we like. For concrete 0 (and any number), it
// cannot be shifted equal or more than the number of bits of the lhs (see
// ResolverConstEvalShiftLeftConcreteGeqBitWidthError)
// ResolverConstEvalShiftLeftConcreteGeqBitWidthError for negative tests)
ConcatIntoIf<IsAbstract<T>>( //
r, std::vector<Case>{
C(T{0}, ST{64}, T{0}),
@ -656,6 +653,31 @@ std::vector<Case> ShiftLeftCases() {
C(Negate(T{0}), T::Highest(), Negate(T{0})),
});
// Cases that are fine for signed values (no sign change), but would overflow unsigned values.
// See ResolverConstEvalBinaryOpTest_Overflow for negative tests.
ConcatIntoIf<IsSignedIntegral<T>>( //
r, std::vector<Case>{
C(B::TwoLeftMost, ST{1}, B::LeftMost), //
C(B::All, ST{1}, B::AllButRightMost), //
C(B::All, ST{B::NumBits - 1}, B::LeftMost) //
});
// Cases that are fine for unsigned values, but would overflow (sign change) signed
// values. See ShiftLeftSignChangeErrorCases() for negative tests.
ConcatIntoIf<IsUnsignedIntegral<T>>( //
r, std::vector<Case>{
C(T{0b0001}, ST{B::NumBits - 1}, B::Lsh(0b0001, B::NumBits - 1)),
C(T{0b0010}, ST{B::NumBits - 2}, B::Lsh(0b0010, B::NumBits - 2)),
C(T{0b0100}, ST{B::NumBits - 3}, B::Lsh(0b0100, B::NumBits - 3)),
C(T{0b1000}, ST{B::NumBits - 4}, B::Lsh(0b1000, B::NumBits - 4)),
C(T{0b0011}, ST{B::NumBits - 2}, B::Lsh(0b0011, B::NumBits - 2)),
C(T{0b0110}, ST{B::NumBits - 3}, B::Lsh(0b0110, B::NumBits - 3)),
C(T{0b1100}, ST{B::NumBits - 4}, B::Lsh(0b1100, B::NumBits - 4)),
C(B::AllButLeftMost, ST{1}, B::AllButRightMost),
});
return r;
}
INSTANTIATE_TEST_SUITE_P(ShiftLeft,
@ -795,7 +817,21 @@ INSTANTIATE_TEST_SUITE_P(
Val(AInt{BitValues<AInt>::NumBits + 1})}, //
OverflowCase{ast::BinaryOp::kShiftLeft, //
Val(AInt{BitValues<AInt>::AllButLeftMost}), //
Val(AInt{BitValues<AInt>::NumBits + 1000})}
Val(AInt{BitValues<AInt>::NumBits + 1000})},
// ShiftLeft of u32s that overflow (non-zero bits get shifted out)
OverflowCase{ast::BinaryOp::kShiftLeft, Val(0b00010_u), Val(31_u)},
OverflowCase{ast::BinaryOp::kShiftLeft, Val(0b00100_u), Val(30_u)},
OverflowCase{ast::BinaryOp::kShiftLeft, Val(0b01000_u), Val(29_u)},
OverflowCase{ast::BinaryOp::kShiftLeft, Val(0b10000_u), Val(28_u)},
// ...
OverflowCase{ast::BinaryOp::kShiftLeft, Val(u32(1u << 28)), Val(4_u)},
OverflowCase{ast::BinaryOp::kShiftLeft, Val(u32(1u << 29)), Val(3_u)},
OverflowCase{ast::BinaryOp::kShiftLeft, Val(u32(1u << 30)), Val(2_u)},
OverflowCase{ast::BinaryOp::kShiftLeft, Val(u32(1u << 31)), Val(1_u)},
// And some more
OverflowCase{ast::BinaryOp::kShiftLeft, Val(BitValues<u32>::All), Val(1_u)},
OverflowCase{ast::BinaryOp::kShiftLeft, Val(BitValues<u32>::AllButLeftMost), Val(2_u)}
));
@ -945,8 +981,7 @@ INSTANTIATE_TEST_SUITE_P(Test,
ResolverConstEvalShiftLeftSignChangeError,
testing::ValuesIn(Concat( //
ShiftLeftSignChangeErrorCases<AInt>(),
ShiftLeftSignChangeErrorCases<i32>(),
ShiftLeftSignChangeErrorCases<u32>())));
ShiftLeftSignChangeErrorCases<i32>())));
} // namespace
} // namespace tint::resolver

View File

@ -281,7 +281,7 @@ struct BitValues {
/// @returns the shifted value
template <typename U, typename V>
static constexpr NumberT Lsh(U val, V shiftBy) {
return NumberT{T{val} << T{shiftBy}};
return NumberT{static_cast<T>(val) << static_cast<T>(shiftBy)};
}
};

View File

@ -0,0 +1,3 @@
fn f() -> u32 {
return 0x1u << 31u;
}

View File

@ -0,0 +1,8 @@
[numthreads(1, 1, 1)]
void unused_entry_point() {
return;
}
uint f() {
return 2147483648u;
}

View File

@ -0,0 +1,8 @@
[numthreads(1, 1, 1)]
void unused_entry_point() {
return;
}
uint f() {
return 2147483648u;
}

View File

@ -0,0 +1,10 @@
#version 310 es
layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
void unused_entry_point() {
return;
}
uint f() {
return 2147483648u;
}

View File

@ -0,0 +1,7 @@
#include <metal_stdlib>
using namespace metal;
uint f() {
return 2147483648u;
}

View File

@ -0,0 +1,24 @@
; SPIR-V
; Version: 1.3
; Generator: Google Tint Compiler; 0
; Bound: 10
; Schema: 0
OpCapability Shader
OpMemoryModel Logical GLSL450
OpEntryPoint GLCompute %unused_entry_point "unused_entry_point"
OpExecutionMode %unused_entry_point LocalSize 1 1 1
OpName %unused_entry_point "unused_entry_point"
OpName %f "f"
%void = OpTypeVoid
%1 = OpTypeFunction %void
%uint = OpTypeInt 32 0
%5 = OpTypeFunction %uint
%uint_2147483648 = OpConstant %uint 2147483648
%unused_entry_point = OpFunction %void None %1
%4 = OpLabel
OpReturn
OpFunctionEnd
%f = OpFunction %uint None %5
%8 = OpLabel
OpReturnValue %uint_2147483648
OpFunctionEnd

View File

@ -0,0 +1,3 @@
fn f() -> u32 {
return (1u << 31u);
}