901: Set target triple and datalayout when creating the LLVM module. r=nlewycky a=nlewycky
We were giving LLVM a triple and datalayout only when producing native code from the LLVM IR. With this change, we tell LLVM as early as possible so that the entire optimization stack knows that it's safe to use target-specific constructs (including target intrinsics `@llvm.x86.sse2.ucomieq.sd`) as well as cost models (for autovectorization) and knowing the bitwidth of the registers so that we can know it's profitable to eliminate redundant extend/truncate casts.
Co-authored-by: Nick Lewycky <nick@wasmer.io>
904: Use getelementptr instruction instead of int_to_ptr and ptr_to_int. r=nlewycky a=nlewycky
The main part of this change is that we no longer turn pointers into integers to do arithmetic on them, then turn them back into pointers. Doing so is a signal to LLVM that it should not attempt to analyze the provenance of the pointers, disabling some optimizations. Using getelementptr allows us to perform arithmetic on pointers while keeping them in pointer types, which LLVM can then analyze.
Most of the textual change is a refactoring to reorder the operations. Previously the bounds checking and determining of the base and bounds were combined because you could put both into the same match, since both actions are performed differently depending on whether the memory is static or dynamic. In this case, we simply check the type twice and do two different things, with comments labelling the steps.
Co-authored-by: Nick Lewycky <nick@wasmer.io>
ptr_to_int and int_to_ptr are not treated the same as type-safe pointer operations. LLVM intentionally performs fewer optimizations to code using the former as compared with the latter. We don't need that in our code, standard pointer optimizations are safe for us.
887: Get aggressive about running cargo check over all packages with all features. r=nlewycky a=nlewycky
Fixes the one issue uncovered. The capstone disassembling support in the LLVM backend was broken. Fixed by removing it. Instead, use the `--llvm-object-file` flag to get a finished object file to disassemble with any disassembler.
Co-authored-by: Nick Lewycky <nick@wasmer.io>
Co-authored-by: Mark McCaskey <mark@wasmer.io>
Fixes the one issue uncovered. The capstone disassembling support in the LLVM backend was broken. Fixed by removing it. Instead, use the `--llvm-object-file` flag to get a finished object file to disassemble with any disassembler.
Instead of ensuring outputs are arithmetic NaNs on every function, we tag them as pending such a check, so that a sequence of computation can have a single canonicalization step at the end.
There's an extra wriggle for SIMD. The Wasm type system only indicates them as V128, so it's possible that we might do computations as F32x4Add, I8x16Add, F64x2Add in a row with no other computations in between. Thus, most SIMD functions apply pending canonicalizations to their inputs, even integer SIMD operations.
863: Rewrite Min/Max to handle all cases correctly. Fixes 545 spectest failures. r=nlewycky a=nlewycky
# Description
The llvm backend was not quite following the Wasm spec for {F32,F64,F32x4xF64x2}{Min,Max}. We used the `@llvm.minnum` and `@llvm.maxnum` intrinsics which don't handle the corner cases the same. When we tried to use `@llvm.minimum` and `@llvm.maximum` which do, we get an internal error from the x86 backend. I was hoping that crash would go away with the upgrade to LLVM 9, but it does not.
Reimplement these operations using plain LLVM instructions.
# Review
- [x] Add a short description of the the change to the CHANGELOG.md file
Co-authored-by: Nick Lewycky <nick@wasmer.io>
870: Fix unused value warning due to inkwell API change. NFC. r=nlewycky a=nlewycky
# Description
Fix unused value warning due to inkwell API change. No functionality change.
Co-authored-by: Nick Lewycky <nick@wasmer.io>
Use them to implement three features in the LLVM backend, getting a valid ELF object file, the post-optimization LLVM IR and the pre-optimization LLVM IR.
Presently they are also global to the backend which is not ideal.
Not handled here is @llvm.minnum and @llvm.maxnum which should be replaced with
@llvm.minimum and @llvm.maximum, but using those currently leads to LLVM backend
fatal errors.
These is one test failure remaining with V128 global variables.
* Fix trunc_sat. We need both the largest float that can be converted to an int
and the largest int, they are not the same number.
* Implement calling of functions that take V128 by passing in two i64's.
* Improve support for V128 in spectests. Parse binary modules with the same
features as the outer spectest. Fix compilation error involving Result in
emitted .rs file. Handle V128 in more cases when producing .rs file. Parse
the wast script with SIMD enabled.
* Adjust the WAVM spectest so that it parses with WABT and mostly passes with
wasmer. Wabt is particular about ints not having decimal places and floats
having decimal places. Wasmer does not support mutable globals or shared
memory. Tests of shuffles are disabled. Some assert_invalid tests that wabt
won't even parse are disabled.
Fix typo in function name. Use two fcmp instructions instead of unpacking the bits of the IEEE float and using integer arithmetic to determine details about its value.
A few notes:
a) the inliner doesn't help because all the calls are indirect and not even opt -O2 can figure out which functions they're actually calling.
b) aggressive instruction combining is not a super-set of the instruction combiner. Instcombine is made up of a large number (probably 10,000s) of patterns, and some particularly slow ones were taken out and moved to the aggressive instruction combiner. Aggressive instcombine *only* runs that handful of optimizations, which fired zero times on our example wasm files.
c) NewGVN is not ready for production, it has asserts that fire when building sqlite or cowsay. This is why sqlite didn't build with the llvm backend.
d) Scalar-replacement-of-aggregates (sroa) is a strict superset of promote-memory-to-registers (mem2reg), and you probably want sroa because it's usually faster. It also fires 10,000s more times than mem2reg on lua.wasm.
e) Aggressive-dead-code-elimination was only deleting as much regular dead-code-elimination, but is slower because it depends on a postdominator tree (PDT) analysis that. Other passes don't need PDT so we'll have to build it for just this one pass (as opposed to regular dominator-tree which is reused by many passes). I've replaced this with bit-tracking dead-code-elimination which deletes more code than dce/adce.
