460 builtin functions #511
Conversation
Implement the mux builtin
Codecov Report
@@ Coverage Diff @@
## master #511 +/- ##
==========================================
+ Coverage 93.04% 93.15% +0.11%
==========================================
Files 40 41 +1
Lines 14592 14838 +246
==========================================
+ Hits 13577 13823 +246
Misses 1015 1015
Continue to review full report at Codecov.
|
If one of the types is generics (because it was builtin) it will not be cast.
Added the variadics to the index members to allow the resolver to correctly address them / modify them once a generic type needs to be replaced
|
In one of the commits, I added the possibility to have a "transformation" function that will spawn out new (replacement) ast. The idea is to run it in the preprocessor phase. I'm not sure how useful it will be though, it turned out I did not really need it yet. |
src/builtins.rs
Outdated
| transformation: |it| it, | ||
| code: |generator, params, location| { | ||
| //Generate an access from the first param | ||
| if let &[k, ..] = params { |
There was a problem hiding this comment.
I think you can deconstruct this in one go:
if let &[k, params..] = params {
There was a problem hiding this comment.
this is not yet stable
There was a problem hiding this comment.
i changed it to if let (&[k], params) = params.split_at(1) {
| let arr_storage = self.llvm.builder.build_alloca(arr, ""); | ||
| for (i, ele) in generated_params.iter().enumerate() { | ||
| let ele_ptr = self.llvm.load_array_element( | ||
| arr_storage, | ||
| &[ | ||
| self.llvm.context.i32_type().const_zero(), | ||
| self.llvm.context.i32_type().const_int(i as u64, true), | ||
| ], | ||
| "", | ||
| )?; | ||
| self.llvm.builder.build_store(ele_ptr, *ele); | ||
| } |
There was a problem hiding this comment.
llvm supports constant-arrays. I think runtime performance would be better
@X = global i32 17
@Y = global i32 42
@Z = global [2 x i32*] [ i32* @X, i32* @Y ]https://llvm.org/docs/LangRef.html#global-variable-and-function-addresses
There was a problem hiding this comment.
This does not work on non-constant variables (expressions that need to be resolved) it was the initial implementation and I had to scrape it
There was a problem hiding this comment.
why do you think so? did you try?
the inkwell type's are a bit opinionated. I just tried and at least I get the IR for the test builtin_function_call_mux generated:
15 15 │ %load_c = load i32, i32* %c, align 4
16 16 │ %load_d = load i32, i32* %d, align 4
17 17 │ %load_e = load i32, i32* %e, align 4
18 18 │ %1 = alloca [4 x i32], align 4
- 19 │- %2 = getelementptr inbounds [4 x i32], [4 x i32]* %1, i32 0, i32 0
- 20 │- store i32 %load_b, i32* %2, align 4
- 21 │- %3 = getelementptr inbounds [4 x i32], [4 x i32]* %1, i32 0, i32 1
- 22 │- store i32 %load_c, i32* %3, align 4
- 23 │- %4 = getelementptr inbounds [4 x i32], [4 x i32]* %1, i32 0, i32 2
- 24 │- store i32 %load_d, i32* %4, align 4
- 25 │- %5 = getelementptr inbounds [4 x i32], [4 x i32]* %1, i32 0, i32 3
- 26 │- store i32 %load_e, i32* %5, align 4
- 27 │- %6 = getelementptr inbounds [4 x i32], [4 x i32]* %1, i32 0, i32 3
- 28 │- %7 = load i32, i32* %6, align 4
- 29 │- store i32 %7, i32* %a, align 4
+ 19 │+ store [4 x i32] [i32 %load_b, i32 %load_c, i32 %load_d, i32 %load_e], [4 x i32]* %1, align 4
+ 20 │+ %2 = getelementptr inbounds [4 x i32], [4 x i32]* %1, i32 0, i32 3
+ 21 │+ %3 = load i32, i32* %2, align 4
+ 22 │+ store i32 %3, i32* %a, align 4
30 23 │ ret void
31 24 │ }It just works for int_values here, so i think you just need a helper method, where you pass the generated variadic parameters and it creates you an int_array or a ptr_array or a float_array etc. (depending on the variadic values you passed it --> see that all variadic parameters need to be the same type-nature anyway)
I think its a cleaner implementation and faster at runtime (and much easier for the optimizer)
here is my change:
index 9487bb518..893e72506 100644
--- a/src/codegen/generators/expression_generator.rs
+++ b/src/codegen/generators/expression_generator.rs
@@ -777,22 +777,19 @@ impl<'ink, 'b> ExpressionCodeGenerator<'ink, 'b> {
} else {
self.llvm_index.get_associated_type(type_name)
}?;
- let size = generated_params.len();
- let size_param = self.llvm.i32_type().const_int(size as u64, true);
- let arr = Llvm::get_array_type(llvm_type, size as u32);
- let arr_storage = self.llvm.builder.build_alloca(arr, "");
- for (i, ele) in generated_params.iter().enumerate() {
- let ele_ptr = self.llvm.load_array_element(
- arr_storage,
- &[
- self.llvm.context.i32_type().const_zero(),
- self.llvm.context.i32_type().const_int(i as u64, true),
- ],
- "",
- )?;
- self.llvm.builder.build_store(ele_ptr, *ele);
- }
- Ok(vec![size_param.into(), arr_storage.into()])
+ // generate the variadic values
+ let size_value = self.llvm.i32_type().const_int(generated_params.len() as u64, false);
+ let arr_storage = self.llvm.builder.build_array_alloca(llvm_type, size_value, "");
+ // these 2 lines need a helper function (it now assumes intValues)
+ let values: Vec<IntValue> = generated_params
+ .iter()
+ .map(|it| it.into_int_value())
+ .collect();
+ self.llvm.builder.build_store(arr_storage, llvm_type.into_int_type().const_array(values.as_slice()));
+
+ let size_param = self.llvm.builder.build_alloca(self.llvm.i32_type(), "");
+ self.llvm.builder.build_store(size_param, size_value);
+ Ok(vec![size_param.into(), arr_storage.into()])
} else {
Ok(generated_params)
}There was a problem hiding this comment.
There should be a correctness test for it. It will segfault or at least my implementation segfaults.
llialso refused to run it.
I didn't check your version yet maybe it's different I'll check tomorrow
There was a problem hiding this comment.
hmm strange. in the given example [ i32* @X, i32* @Y ] X and Y are also not const
ghaith
force-pushed
the
460-builtin-functions
branch
from
13c9eff
to
d796dfd
Compare
|
@riederm I would like to introduce an extra built in property here that allows us to override the resolver for a specific built in, this allows us to derive correct return types (oscat is having issue with cases like |
The MUX, SEL and MOVE functions will have their own annotations This allows more control on the return types of the functions
The return type for builtins is now correctly annotated, the special annotation is not yet used, will be needed for EXPT
| let arr_storage = self.llvm.builder.build_alloca(arr, ""); | ||
| for (i, ele) in generated_params.iter().enumerate() { | ||
| let ele_ptr = self.llvm.load_array_element( | ||
| arr_storage, | ||
| &[ | ||
| self.llvm.context.i32_type().const_zero(), | ||
| self.llvm.context.i32_type().const_int(i as u64, true), | ||
| ], | ||
| "", | ||
| )?; | ||
| self.llvm.builder.build_store(ele_ptr, *ele); | ||
| } |
There was a problem hiding this comment.
hmm strange. in the given example [ i32* @X, i32* @Y ] X and Y are also not const
More support for the built-in functions
Added a {sized} pragma for the varargs which when used will make sure the functions receive a sized argument and a pointer to the varargs, it then passes the size into the call.
This allows externally implemented functions to receive a pointer to the parameter data and loop through it.
Added MUX and SEL as built-in functions.