This is an archive of the discontinued LLVM Phabricator instance.

Differential D152336

[libc] Fix undefined behavior of left shifting signed integer in exp2f.cpp.
ClosedPublic

Authored by lntue on Jun 6 2023, 9:34 PM.

Details

Reviewers
michaelrj
sivachandra
brooksmoses
Commits
rGc0a751ae3da3: [libc] Fix undefined behavior of left shifting signed integer in exp2f.cpp.
Summary

Fix undefined behavior of left shifting signed integer in exp2f.cpp.

Diff Detail

Event Timeline

lntue created this revision.Jun 6 2023, 9:34 PM
Herald added projects: Restricted Project, Restricted Project. · View Herald TranscriptJun 6 2023, 9:34 PM
lntue requested review of this revision.Jun 6 2023, 9:34 PM
Harbormaster completed remote builds in B237164: Diff 529150.Jun 6 2023, 9:40 PM
sivachandra accepted this revision.Jun 6 2023, 9:53 PM
This revision is now accepted and ready to land.Jun 6 2023, 9:53 PM
Closed by commit rGc0a751ae3da3: [libc] Fix undefined behavior of left shifting signed integer in exp2f.cpp. (authored by lntue). · Explain WhyJun 6 2023, 10:15 PM
This revision was automatically updated to reflect the committed changes.
lntue added a commit: rGc0a751ae3da3: [libc] Fix undefined behavior of left shifting signed integer in exp2f.cpp..

Revision Contents

PathSize
libc/
src/
math/
generic/
5 lines

Diff 529158

libc/src/math/generic/exp2f.cpp

Show First 20 Lines • Show All 101 Lines • ▼ Show 20 LinesLLVM_LIBC_FUNCTION(float, exp2f, (float x)) {
// kf = (hi + mid) * 2^5 = round(x * 2^5) // kf = (hi + mid) * 2^5 = round(x * 2^5)
float kf = fputil::nearest_integer(x * 32.0f); float kf = fputil::nearest_integer(x * 32.0f);
// dx = lo = x - (hi + mid) = x - kf * 2^(-5) // dx = lo = x - (hi + mid) = x - kf * 2^(-5)
double dx = fputil::multiply_add(-0x1.0p-5f, kf, x); double dx = fputil::multiply_add(-0x1.0p-5f, kf, x);
int k = static_cast<int>(kf); int k = static_cast<int>(kf);
// hi = floor(kf * 2^(-4)) // hi = floor(kf * 2^(-4))
// exp_hi = shift hi to the exponent field of double precision. // exp_hi = shift hi to the exponent field of double precision.
int64_t exp_hi = static_cast<int64_t>(k >> ExpBase::MID_BITS) int64_t exp_hi =
<< fputil::FloatProperties<double>::MANTISSA_WIDTH; static_cast<int64_t>(static_cast<uint64_t>(k >> ExpBase::MID_BITS)
<< fputil::FloatProperties<double>::MANTISSA_WIDTH);
// mh = 2^hi * 2^mid // mh = 2^hi * 2^mid
// mh_bits = bit field of mh // mh_bits = bit field of mh
int64_t mh_bits = ExpBase::EXP_2_MID[k & ExpBase::MID_MASK] + exp_hi; int64_t mh_bits = ExpBase::EXP_2_MID[k & ExpBase::MID_MASK] + exp_hi;
double mh = fputil::FPBits<double>(uint64_t(mh_bits)).get_val(); double mh = fputil::FPBits<double>(uint64_t(mh_bits)).get_val();
// Degree-5 polynomial approximating (2^x - 1)/x generating by Sollya with: // Degree-5 polynomial approximating (2^x - 1)/x generating by Sollya with:
// > P = fpminimax((2^x - 1)/x, 5, [|D...|], [-1/32. 1/32]); // > P = fpminimax((2^x - 1)/x, 5, [|D...|], [-1/32. 1/32]);
constexpr double COEFFS[5] = {0x1.62e42fefa39efp-1, 0x1.ebfbdff8131c4p-3, constexpr double COEFFS[5] = {0x1.62e42fefa39efp-1, 0x1.ebfbdff8131c4p-3,
Show All 15 Lines