Add support for Float16, Float32, Float64 and Float128 #17

zedar · 2024-04-10T21:09:00Z

No description provided.

antoyo · 2024-04-11T19:39:28Z

src/context.rs

+        #[cfg(feature = "master")]
+        let float64_type = context.new_c_type(CType::Float64);
+        #[cfg(feature = "master")]
+        let float128_type = context.new_c_type(CType::Float128);


Those types might not be available on some platforms, so it might be wrong to query these types here.

Depending on how much you want to add stuff in this PR, you have two options:

Do the same as detecting 128-bit integers support: we might want to change the GCC API to take a type as a parameter.

Only query the types in the functions like type_f16 below so that it only panics when used.

So far I query the types only in the functions().
As the next step I will try to detect whether new types are supported.

I mean, just calling new_c_type() here will give an error if the type is not supported.
Here's how I detect 128-bit support with the old way of doing things.

antoyo · 2024-04-11T19:41:00Z

src/type_.rs

@@ -132,8 +137,13 @@ impl<'gcc, 'tcx> BaseTypeMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
        self.double_type


Depending on what you decide to do above, you could return Float32/Float64 in those functions if you decide to add the support for checking if those types are supported.

Of course, we would keep returing Float and Double on platforms where those types are not supported.

antoyo · 2024-04-13T11:45:19Z

src/type_.rs

    fn type_f16(&self) -> Type<'gcc> {
-        unimplemented!("f16_f128")
+        self.context.new_c_type(CType::Float16)
+    }


I'm surprised the CI didn't complain about formatting, but it's missing an empty line between the two methods.

antoyo · 2024-04-13T11:45:26Z

src/type_.rs

+    #[cfg(feature = "master")]
+    fn type_f128(&self) -> Type<'gcc> {
+        self.context.new_c_type(CType::Float128)
+    }


antoyo · 2024-04-13T11:46:39Z

Do you want to add more stuff for this PR (like the checks if the types are supported on the target platform for at least Float32 and Float64) or do you want to merge as is (with the above formatting nitpicks fixed)?

zedar · 2024-04-14T12:19:48Z

Do you want to add more stuff for this PR (like the checks if the types are supported on the target platform for at least Float32 and Float64) or do you want to merge as is (with the above formatting nitpicks fixed)?

Yes, I plan to add checks if Float32 and Float64 are supported.

Upgrade libgccjit.version Limit new Floatxx types to master branch only apply rustfmt Make new types available only when requested Make new types available only when requested Check if Float16 and Float128 are supported by the target platform Replace Float with Float32 and Double with Float64 if target dependent type is defined Add support for Float16|32|64|128 in the builder Fix cargo fmt errors Update gccjit wrapper update hash of ligccjit

…he no-f16-f128 feautes introduced incompatibility

reorder type_kind reorder type_kind reorder type_kind fix fix fix fix

fix formatting

antoyo · 2024-06-21T14:20:39Z

src/type_.rs

+    #[cfg(feature = "master")]
+    fn type_f32(&self) -> Type<'gcc> {
+        if self.supports_f32_type {
+            return self.context.new_c_type(CType::Float32);
+        }
+        self.float_type
    }

+    #[cfg(not(feature = "master"))]
    fn type_f32(&self) -> Type<'gcc> {
        self.float_type
    }


Would it work to do the following?

Suggested change

#[cfg(feature = "master")]

fn type_f32(&self) -> Type<'gcc> {

if self.supports_f32_type {

return self.context.new_c_type(CType::Float32);

}

self.float_type

}

#[cfg(not(feature = "master"))]

fn type_f32(&self) -> Type<'gcc> {

self.float_type

}

fn type_f32(&self) -> Type<'gcc> {

#[cfg(feature = "master")]

if self.supports_f32_type {

return self.context.new_c_type(CType::Float32);

}

self.float_type

}

antoyo · 2024-06-21T14:45:10Z

src/type_.rs

+                4 => TypeKind::Float,
+                8 => TypeKind::Double,
+                16 => TypeKind::FP128,
+                _ => TypeKind::Void,


Maybe this should be unreachable!() instead?

antoyo · 2024-06-21T14:45:34Z

src/type_.rs

+            TypeKind::Float
+        } else if typ.is_compatible_with(self.double_type) {
+            TypeKind::Double
+        } else if typ.is_floating_point() {


You'll need to do this:

Suggested change

} else if typ.is_floating_point() {

}

#[cfg(feature="master")]

else if typ.is_floating_point() {

This syntax isn't supported.
Unless you can put this above the if, I guess you can revert to 2 different implementations:

#[cfg(feature="master")] if typ.is_floating_point() { match typ.get_size() { 2 => return TypeKind::Half, 4 => return TypeKind::Float, 8 => return TypeKind::Double, 16 => return TypeKind::FP128, _ => return TypeKind::Void, } }

antoyo · 2024-06-21T14:45:54Z

src/type_.rs

+        #[cfg(feature = "master")]
+        if self.supports_f32_type {
+            return self.context.new_c_type(CType::Float32);
+        }
        self.float_type
    }



Don't forget to implement type_f64 as well.

fix

antoyo · 2024-06-25T13:33:10Z

Thank you so much for your work!

zedar force-pushed the add_support_for_float16_float128 branch 2 times, most recently from 35bc8ff to 6b2a99a Compare April 11, 2024 18:56

antoyo reviewed Apr 11, 2024

View reviewed changes

zedar force-pushed the add_support_for_float16_float128 branch from 6b2a99a to f9c3617 Compare April 12, 2024 06:57

antoyo reviewed Apr 13, 2024

View reviewed changes

zedar force-pushed the add_support_for_float16_float128 branch from f9c3617 to 46eea99 Compare April 14, 2024 18:20

zedar force-pushed the add_support_for_float16_float128 branch from 46eea99 to 2743807 Compare April 22, 2024 18:23

zedar force-pushed the add_support_for_float16_float128 branch 5 times, most recently from d71f067 to 6280647 Compare May 3, 2024 04:48

zedar force-pushed the add_support_for_float16_float128 branch 4 times, most recently from 98af2e1 to 8920002 Compare May 13, 2024 06:02

zedar force-pushed the add_support_for_float16_float128 branch 4 times, most recently from d9dc4cb to 7004c96 Compare May 13, 2024 20:31

zedar force-pushed the add_support_for_float16_float128 branch from 850b944 to 6f3aa00 Compare May 24, 2024 06:35

Robert Zakrzewski and others added 7 commits June 21, 2024 16:12

Upgrade libgccjit.version

a486dbf

Temporary downgrade compiler_builtins library. From version 0.1.110 t…

fa18a18

…he no-f16-f128 feautes introduced incompatibility

Add missing types in the type_kind function

b94cb8c

reorder type_kind reorder type_kind reorder type_kind fix fix fix fix

Do not use target dependent Float32

0dad11f

fix formatting

Fix location of check for sized floating-point types

c4e7c04

Update libgccjit version with fixed is_same_type_as for vector types

55788e4

zedar force-pushed the add_support_for_float16_float128 branch from 894f00c to 55788e4 Compare June 21, 2024 14:12

antoyo reviewed Jun 21, 2024

View reviewed changes

Refactor type_f16|32|128 functions. Common type_kind()

2eaac23

fix

zedar force-pushed the add_support_for_float16_float128 branch from 0d05848 to 2eaac23 Compare June 21, 2024 15:33

Robert Zakrzewski and others added 4 commits June 21, 2024 22:10

Add support for Float64

dabf5fa

Fix clippy warnings

16e1ad7

Add comment about compiler_builtins

73db249

Change Void to unreachable

9274c63

antoyo force-pushed the add_support_for_float16_float128 branch from 55467aa to 9274c63 Compare June 25, 2024 13:00

antoyo merged commit 70d3655 into antoyo:master Jun 25, 2024
34 checks passed

tgross35 mentioned this pull request Jun 25, 2024

Support f16 and f128 rust-lang/rustc_codegen_gcc#461

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add support for Float16, Float32, Float64 and Float128 #17

Add support for Float16, Float32, Float64 and Float128 #17

zedar commented Apr 10, 2024

antoyo Apr 11, 2024

zedar Apr 12, 2024 •

edited

Loading

antoyo Apr 12, 2024

antoyo Apr 11, 2024

antoyo Apr 13, 2024

antoyo Apr 13, 2024

antoyo commented Apr 13, 2024

zedar commented Apr 14, 2024

antoyo Jun 21, 2024

antoyo Jun 21, 2024

antoyo Jun 21, 2024

antoyo Jun 21, 2024 •

edited

Loading

antoyo Jun 21, 2024

antoyo commented Jun 25, 2024

		@@ -132,8 +137,13 @@ impl<'gcc, 'tcx> BaseTypeMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
		self.double_type

Add support for Float16, Float32, Float64 and Float128 #17

Add support for Float16, Float32, Float64 and Float128 #17

Conversation

zedar commented Apr 10, 2024

Choose a reason for hiding this comment

zedar Apr 12, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

antoyo commented Apr 13, 2024

zedar commented Apr 14, 2024

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

antoyo Jun 21, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

antoyo commented Jun 25, 2024

zedar Apr 12, 2024 •

edited

Loading

antoyo Jun 21, 2024 •

edited

Loading