-
-
Notifications
You must be signed in to change notification settings - Fork 5.5k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
fix incorrect results in expm1(::Union{Float16, Float32})
#50989
Conversation
`unsafe_trunc(UInt, -1.0)` is UB but worked fine on apple and AMD so we didn't notice??? This has been very broken since 1.7
Note that there are no tests for this because we were already pretty extensively testing these functions and this bug only shows up on intel and none of our CI machines are intel aparently... |
I don't think that's a justification for not adding tests, nothing prevents us from having CI agents running on Intel CPUs in the future, and having tests that can potentially catch bugs on users machine is useful regardless of what current CI setup is. |
I was mis-remembering. I thought we had a ton of tests for |
expm1(::Union{Float16, Float32})
expm1(::Union{Float16, Float32})
@@ -460,7 +460,7 @@ function expm1(x::Float32) | |||
end | |||
x = Float64(x) | |||
N_float = round(x*Ln2INV(Float64)) | |||
N = unsafe_trunc(UInt64, N_float) | |||
N = unsafe_trunc(Int64, N_float) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I imagine MAX_EXP is much smaller than the maximum integer here? If so LGTM
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
yeah. N
will end up as the exponent for the float so it will be in [-128,128]
AFAICT the |
`unsafe_trunc(UInt, -1.0)` is unspecified behavior but worked fine on apple and AMD so we didn't notice??? This has been very broken since 1.7. (cherry picked from commit 61ebaf6)
For reference, a bit of previous discussion about |
`unsafe_trunc(UInt, -1.0)` is unspecified behavior but worked fine on apple and AMD so we didn't notice??? This has been very broken since 1.7. (cherry picked from commit 61ebaf6)
Backported PRs: - [x] #50932 <!-- types: fix hash values of Vararg --> - [x] #50975 <!-- Use rr-safe `nopl; rdtsc` sequence --> - [x] #50989 <!-- fix incorrect results in `expm1(::Union{Float16, Float32})` --> - [x] #51284 <!-- Avoid infinite loop when doing SIGTRAP in arm64-apple --> - [x] #51332 <!-- Add s4 field to Xoshiro --> - [x] #51397 <!-- call Pkg precompile hook in latest world --> - [x] #51405 <!-- Remove fallback that assigns a module to inlined frames. --> - [x] #51491 <!-- Throw clearer ArgumentError for strip with two string args --> - [x] #51531 <!-- fix `_tryonce_download_from_cache` (busybox.exe download error) --> - [x] #51541 <!-- Fix string index error in tab completion code --> - [x] #51530 <!-- Don't mark nonlocal symbols as hidden --> - [x] #51557 <!-- Fix last startup & shutdown precompiles --> - [x] #51512 <!-- avoid limiting Type{Any} to Type --> - [x] #51595 <!-- reset `maxprobe` on `empty!` --> - [x] #51582 <!-- Aggressive constprop in LinearAlgebra.wrap --> - [x] #51592 <!-- correctly track element pointer in heap snapshot --> - [x] #51326 <!-- complete false & true more generally as vals --> - [x] #51376 <!-- make `hash(::Xoshiro)` compatible with `==` --> - [x] #51557 <!-- Fix last startup & shutdown precompiles --> - [x] #51845 - [x] #51840 - [x] #50663 <!-- Fix Expr(:loopinfo) codegen --> - [x] #51863 <!-- LLVM 15.0.7-9 --> Contains multiple commits, manual intervention needed: - [ ] #51035 <!-- refactor GC scanning code to reflect jl_binding_t are now first class --> - [ ] #51092 <!-- inference: fix bad effects for recursion --> Non-merged PRs with backport label: - [ ] #51479 <!-- prevent code loading from lookin in the versioned environment when building Julia --> - [ ] #51414 <!-- improvements on GC scheduler shutdown --> - [ ] #51366 <!-- Handle infix operators in REPL completion --> - [ ] #50919 <!-- Code loading: do the "skipping mtime check for stdlib" check regardless of the value of `ispath(f)` --> - [ ] #50824 <!-- Add some aliasing warnings to docstrings for mutating functions in Base --> - [ ] #49805 <!-- Limit TimeType subtraction to AbstractDateTime -->
`unsafe_trunc(UInt, -1.0)` is unspecified behavior but worked fine on apple and AMD so we didn't notice??? This has been very broken since 1.7. (cherry picked from commit 61ebaf6)
Backported PRs: - [x] #49357 <!-- Fix unclosed code fence in src/manual/methods.md --> - [x] #50842 <!-- Avoid race conditions with recursive rm --> - [x] #50858 <!-- Add a `threadpool` parameter to `Channel` constructor --> - [x] #50730 <!-- Fix integer overflow in `isapprox` --> - [x] #50823 <!-- Make ranges more robust with unsigned indexes. --> - [x] #50915 <!-- Add note the `Task` about sticky bit --> - [x] #50989 <!-- fix incorrect results in `expm1(::Union{Float16, Float32})` --> - [x] #50912 <!-- Separate foreign threads into a :foreign threadpool --> - [x] #51019 <!-- fix a case of potentially use of undefined variable when handling error in distributed message processing --> - [x] #51222 <!-- Check again if the tty is open inside the IO lock --> - [x] #51254 <!-- Ryu: make sure adding zeros does not overwrite trailing dot --> - [x] #51284 <!-- Avoid infinite loop when doing SIGTRAP in arm64-apple --> - [x] #51491 <!-- Throw clearer ArgumentError for strip with two string args --> - [x] #51531 <!-- fix `_tryonce_download_from_cache` (busybox.exe download error) -->
unsafe_trunc(UInt, -1.0)
is unspecified behavior but worked fine on apple and AMD so we didn't notice??? This has been very broken since 1.7