Skip to content

R
runtime

dotNET Platform / runtime
12 个月 前同步成功

通知 1
Star 1
Fork 0
R
runtime

体验新版 GitCode，发现更多精彩内容 >>

切换分支/标签
  1. 23 3月, 2021 10 次提交
    • J
      Don't try bypassing `uname` for CPU detection on Mac (#49992) · 727996e8
      由 Jo Shields 提交于 
      On Apple Silicon, the command `arch -x86_64 XYZ` allows executing XYZ
as if on an AMD64 CPU, rather than an ARMv8 CPU, for cases where both
are possible (e.g. a shell script like build.sh).

The deleted lines of code bypass the lies `arch` writes into uname,
and instead forcibly detects the true nature of the host system's CPU
regardless of whether or not the developer has reasons to lie.

Whilst this trick was important in the early days of bringing up
Apple Silicon support, it should hopefully now be obsolete.

This fixes linking of LLVM in Mono runtime builds, where
`arch -x86_64 ./build.sh -arch x64` would still attempt to link
against ARM64 LLVM.
      727996e8
    • R
      [wasm] windows build improvements (#49771) · ee2d515e
      由 Radek Doulik 提交于 
      Fix build on windows in folders containing space(s) in the name and also
in case where `EMSDK_PATH` contains space(s).

* Add few quotes

To fix windows build in directory containing spaces.

* Add quotes around link attributes paths

* More quotes and trimming

* More improvements related to EMSDK_PATH

* Use immediate set for EMSCRIPTEN_VERSION

* Do not set ExeExt twice

* Force ninja on Browser/windows

And check for cmake and ninja in CheckEnv

* Apply Ankit's quotation changes

* Simplify forcing wasm, fix typos

* Use -File instead of & "<ps1 script>"

* Fix typo

* Only add the escaped quotes on windows
Co-authored-by: NAnkit Jain <radical@gmail.com>
      ee2d515e
    • A
      use packages instead of package report for packageInstallation and closure verification (#49315) · 6f73cce0
      由 Anirudh Agnihotry 提交于 
      * use packages instead of package report for packageInstallation and closure verification

* add the package dependency

* correct the names

* excluding freebsd from testing

* Update Versions.props

* update the package versions, add support for exclude pakcages and packages to test

* update the package validation version

* Make package testing work indepent of version

* testpackages -> testpackagesPath

* fix restore or when no packages are produced

* correct the no packages produced condition

* use regex instead of major version
      6f73cce0
    • B
      Fix DEBUG/non-DEBUG asm diffs due to instruction group size estimation (#49947) · 7012cd7f
      由 Bruce Forstall 提交于 
      * Fix DEBUG/non-DEBUG asm diffs due to instruction group size estimation

This change fixes a long-existing difference between DEBUG and non-DEBUG codegen
that recently manifest as asm diffs in instruction encoding in a Checked/Release
SuperPMI run, mostly in the tests\JIT\Directed\arglist\vararg.cs test, specifically:

```
NativeVarargTest.VarArg:TestPassingManyFloats(System.Double[]):bool
NativeVarargTest.VarArg:TestPassingManyDoublesManaged(System.Double[]):bool
NativeVarargTest.VarArg:TestPassingManyDoubles(System.Double[]):bool
NativeVarargTest.VarArg:TestPassingManyFloatsManaged(System.Single[]):bool
double1:ProcessJagged3DArray(System.Double[][])
``
Example diff:
```
215c7758454: 76 6a                      jbe     106 // checked
215c77b8ab4: 0f 86 6a 00 00 00          jbe     106 // release
```

Note that the offset ends up exactly the same, but the encoding is different.

Some context on the emitter:

The emitter collects generated machine instructions as a set of variable-sized
`instrDesc` structures in instruction groups (IG, or `insGroup`). An IG is filled
with a variable number of instrDescs into a fixed-sized buffer, which, when full
or a new label (e.g., branch target) is encountered, is copied to dynamically
allocated precise-sized memory for the set of instructions collected. If a BasicBlock
required more instructions than fit in a single IG, multiple IG are generated, and
the ones besides the first are marked with the IGF_EXTEND flag, called "emitAdd"
or overflow IG. After codegen, all the IG are filled, and we have an estimated
size for each instruction, and each IG has an estimated code offset from the
beginning of the function. The estimated sizes could be wrong if we incorrectly
estimated the size of an instruction (which in most cases could be considered a
bug, but is particularly common with our Intel vector instructions currently) or
the offset required for a branch (especially a forward branch). The JIT only does
a single pass of branch tightening, so could have some cases of non-optimal estimates.
Note that estimates are required to be too large; it is fatal for an estimate to
be too small. During emission, we walk the IG list and instructions, encoding them
into the memory buffer provided by the VM. At this point, we might shrink an instruction
encoding, possibly because we can use a smaller branch encoding that expected.

Notably, the size of instrDescs and insGroups varies between DEBUG and non-DEBUG builds.
That means that the number of instrDescs that fit in an IG differs. This should have
no effect on generated code, as long as code properly handles overflow IG and doesn't
affect codegen based on IG size or number of contained instructions.

The problem:

During emission, we calculate an "offset adjustment" (in `emitOffsAdj`) that is used to
estimate the distance from a forward jump instruction to its target. This determines the
size of jump instruction encodings that we emit. For example, on x86/x64, a relative
conditional branch can have a 1 byte or a 4 byte target offset. A 1-byte offset can be
used if the target instruction is 127 bytes or fewer away. So, we estimate the distance
using the jump instruction's known offset and the estimated offset of the target, which
is the estimated offset of the instruction group (the label) calculated before emission
minus the "shrinkage" of instruction sizes that we've seen so far. This "shrinkage" is
due to over-estimation of instruction encoding sizes and branch estimated sizes.

The shrinkage in `emitOffsAdjs` is only computed at the beginning of an IG. As additional
instructions in an IG shrink, we don't take that into account when we reach a forward jump
instruction and try to determine if a "small" branch encoding can be used. Thus, it's
possible for a branch to be estimated to require a large 4-byte offset instead of a 1-byte
offset because the forward branch is barely over the limit, but wouldn't be if we considered
the already-shrunk instructions in the current IG. In fact, the more instructions that
shrink in an IG, the larger our target offset estimate becomes, because the source instruction
address is precise (since we know precisely how many bytes of instructions we've emitted),
but the target offset estimate is fixed.

Due to the different number of instructions in an IG between DEBUG and non-DEBUG builds,
the estimate we use for a target offset for any particular branch, given enough instruction
shrinkage and overflowing IG (such that there are a different number of IG between DEBUG and
non-DEBUG) is different. This leads to different instruction encoding between the flavors.

The solution:

The solution is simple: track instruction shrinkage as we emit instructions, and use the
latest, most accurate shrinkage adjustment when estimating branches. This gives no difference
between DEBUG and non-DEBUG because it removes the indirect dependence on IG instruction
count in recomputing `emitOffsAdj`.

There are asm diffs (with just a Checked x64 JIT compared to baseline Checked JIT) due to some
branches now being SHORT where previously they were LONG.

* Add assert
      7012cd7f
    • B
      Remove duplicate method name string initialization (#49963) · 143103b0
      由 Bruce Forstall 提交于 
      Both `compInit` and `compCompile` were initializing the method/class/full
name strings. The call to `eeGetMethodFullName` is not cheap, and shows
up high on a SuperPMI replay trace.

Remove the duplicate initialization in `compCompile`.

Speeds up SuperPMI replay and asm diffs.
      143103b0
    • B
      Add more loop alignment output (#49967) · c32c0587
      由 Bruce Forstall 提交于 
      Fix potential Checked/Release issue.

Remove default arg to emitLoopAlign to make calls more explicit.
      c32c0587
    • D
      Add arch parameter to artifact name property (#50004) · 0e4abeed
      由 Drew Scoggins 提交于
      0e4abeed
    • B
      Fix COUNT_LOOPS build (#49965) · b8c4a398
      由 Bruce Forstall 提交于
      b8c4a398
    • J
      Disable arm64 tvOS/iOS simulator builds until ICU issue is addressed (#50003) · 57393cf0
      由 Jo Shields 提交于
      57393cf0
    • J
      Update minimum recommended cmake version for Windows (#49961) · a938d2a5
      由 Jan Kotas 提交于
      a938d2a5
  3. 22 3月, 2021 14 次提交
  5. 21 3月, 2021 7 次提交
  7. 20 3月, 2021 9 次提交