66545da2008cd9e806e41b74522ded259cd64f86 Remove support for double serialization (Pieter Wuille)
fff1cae43af959a601cf2558cb3c77f3c2b1aa80 Convert uses of double-serialization to {En,De}codeDouble (Pieter Wuille)
afd964d70b6f7583ecf89c380f80db07f5b66a60 Convert existing float encoding tests (Pieter Wuille)
bda33f98e2f32f2411fb0a8f5fb4f0a32abdf7d4 Add unit tests for serfloat module (Pieter Wuille)
2be4cd94f4c7d92a4287971233a20d68db81c9c9 Add platform-independent float encoder/decoder (Pieter Wuille)
e40224d0c77674348bf0a518365208bc118f39a4 Remove unused float serialization (MarcoFalke)
Pull request description:
Based on #21981.
This adds a software-based platform-independent float/double encoder/decoder (platform independent in the sense that it only uses arithmetic and library calls, but never inspects the binary representation). This should strengthen our guarantee that encoded float/double values are portable across platforms. It then removes the functionality to serialize doubles from serialize.h, and replaces its only (non-test) use for fee estimation data serialization with the software encoder.
At least on x86/ARM, the only difference should be how certain NaN values are encoded/decoded (but not *whether* they are NaN or not).
It comes with tests that verify on is_iec559 platforms (which are the only ones we support, at least for now) that the serialized bytes exactly match the binary representation of floats in memory (for non-NaN).
ACKs for top commit:
laanwj:
Code review re-ACK 66545da2008cd9e806e41b74522ded259cd64f86
practicalswift:
cr re-ACK 66545da2008cd9e806e41b74522ded259cd64f86
Tree-SHA512: 62ad9adc26e28707b2eb12a919feefd4fd10cf9032652dbb1ca1cc97638ac21de89e240858e80d293d5112685c623e58affa3d316a9783ff0e6d291977a141f5
707ba8692b