Skip to content

PoC: remove unsafe from generated calls to intrinsics #236

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
Shnatsel wants to merge 2 commits into
base: main
Choose a base branch
from
Draft

PoC: remove unsafe from generated calls to intrinsics #236

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
e81a6fc
PoC: remove unsafe from generated calls to intrinsics
Shnatsel May 27, 2026
717f943
Merge branch 'main' into safety-in-kernels
Shnatsel May 28, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2,829 changes: 1,140 additions & 1,689 deletions fearless_simd/src/generated/avx2.rs
View file
Open in desktop

Large diffs are not rendered by default.

921 changes: 548 additions & 373 deletions fearless_simd/src/generated/neon.rs
View file
Open in desktop

Large diffs are not rendered by default.

1,117 changes: 516 additions & 601 deletions fearless_simd/src/generated/sse4_2.rs
View file
Open in desktop

Large diffs are not rendered by default.

174 changes: 70 additions & 104 deletions fearless_simd_gen/src/mk_neon.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -83,14 +83,14 @@ impl Level for Neon {
OpSig::Splat => {
let expr = neon::expr(method, vec_ty, &[quote! { val }]);
let normalize_mask = integer_lane_mask_splat_arg(vec_ty);
quote! {
#method_sig {
unsafe {
#normalize_mask
#expr.simd_into(self)
}
}
}
op.simd_trait_kernel_method(
self.token(),
vec_ty,
quote! {
#normalize_mask
#expr.simd_into(self)
},
)
}
OpSig::Shift => {
let dup_type = vec_ty.cast(ScalarType::Int);
Expand All @@ -109,26 +109,14 @@ impl Level for Neon {
vec_ty,
&[quote! { a.into() }, quote! { #dup_intrinsic ( #shift ) }],
);
quote! {
#method_sig {
unsafe {
#expr.simd_into(self)
}
}
}
op.simd_trait_kernel_method(self.token(), vec_ty, quote! { #expr.simd_into(self) })
}
OpSig::Unary => {
let args = [quote! { a.into() }];

let expr = neon::expr(method, vec_ty, &args);

quote! {
#method_sig {
unsafe {
#expr.simd_into(self)
}
}
}
op.simd_trait_kernel_method(self.token(), vec_ty, quote! { #expr.simd_into(self) })
}
OpSig::LoadInterleaved {
block_size,
Expand Down Expand Up @@ -187,34 +175,34 @@ impl Level for Neon {
let id2 =
Ident::new(&format!("vcombine_{}", target_scalar_ty), Span::call_site());

quote! {
#method_sig {
unsafe {
let converted: #arch = a.into();
let low = #id1(converted.0);
let high = #id1(converted.1);
op.simd_trait_kernel_method(
self.token(),
vec_ty,
quote! {
let converted: #arch = a.into();
let low = #id1(converted.0);
let high = #id1(converted.1);

#id2(low, high).simd_into(self)
}
}
}
#id2(low, high).simd_into(self)
},
)
} else {
let arch = self.arch_ty(&target_ty);
let id1 = Ident::new(&format!("vmovl_{}", vec_scalar_ty), Span::call_site());
let id2 = Ident::new(&format!("vget_low_{}", vec_scalar_ty), Span::call_site());
let id3 =
Ident::new(&format!("vget_high_{}", vec_scalar_ty), Span::call_site());

quote! {
#method_sig {
unsafe {
let low = #id1(#id2(a.into()));
let high = #id1(#id3(a.into()));
op.simd_trait_kernel_method(
self.token(),
vec_ty,
quote! {
let low = #id1(#id2(a.into()));
let high = #id1(#id3(a.into()));

#arch(low, high).simd_into(self)
}
}
}
#arch(low, high).simd_into(self)
},
)
}
}
OpSig::Binary => {
Expand Down Expand Up @@ -263,13 +251,7 @@ impl Level for Neon {
}
};

quote! {
#method_sig {
unsafe {
#expr
}
}
}
op.simd_trait_kernel_method(self.token(), vec_ty, expr)
}
OpSig::Ternary => {
let args = match method {
Expand All @@ -291,13 +273,7 @@ impl Level for Neon {
let neg = simple_intrinsic("vneg", vec_ty);
expr = quote! { #neg(#expr) };
}
quote! {
#method_sig {
unsafe {
#expr.simd_into(self)
}
}
}
op.simd_trait_kernel_method(self.token(), vec_ty, quote! { #expr.simd_into(self) })
}
OpSig::Compare => {
let args = [quote! { a.into() }, quote! { b.into() }];
Expand All @@ -306,27 +282,23 @@ impl Level for Neon {
let scalar_bits = vec_ty.scalar_bits;
let reinterpret_str = format!("vreinterpret{opt_q}_s{scalar_bits}_u{scalar_bits}");
let reinterpret = Ident::new(&reinterpret_str, Span::call_site());
quote! {
#method_sig {
unsafe {
#reinterpret(#expr).simd_into(self)
}
}
}
op.simd_trait_kernel_method(
self.token(),
vec_ty,
quote! { #reinterpret(#expr).simd_into(self) },
)
}
OpSig::Select => {
let opt_q = neon::opt_q(vec_ty);
let scalar_bits = vec_ty.scalar_bits;
let reinterpret_str = format!("vreinterpret{opt_q}_u{scalar_bits}_s{scalar_bits}");
let reinterpret = Ident::new(&reinterpret_str, Span::call_site());
let vbsl = simple_intrinsic("vbsl", vec_ty);
quote! {
#method_sig {
unsafe {
#vbsl(#reinterpret(a.into()), b.into(), c.into()).simd_into(self)
}
}
}
op.simd_trait_kernel_method(
self.token(),
vec_ty,
quote! { #vbsl(#reinterpret(a.into()), b.into(), c.into()).simd_into(self) },
)
}
OpSig::Combine { combined_ty } => {
let combined_wrapper = combined_ty.aligned_wrapper();
Expand Down Expand Up @@ -375,28 +347,28 @@ impl Level for Neon {
OpSig::Zip { select_low } => {
let neon = if select_low { "vzip1" } else { "vzip2" };
let zip = simple_intrinsic(neon, vec_ty);
quote! {
#method_sig {
op.simd_trait_kernel_method(
self.token(),
vec_ty,
quote! {
let x = a.into();
let y = b.into();
unsafe {
#zip(x, y).simd_into(self)
}
}
}
#zip(x, y).simd_into(self)
},
)
}
OpSig::Unzip { select_even } => {
let neon = if select_even { "vuzp1" } else { "vuzp2" };
let zip = simple_intrinsic(neon, vec_ty);
quote! {
#method_sig {
op.simd_trait_kernel_method(
self.token(),
vec_ty,
quote! {
let x = a.into();
let y = b.into();
unsafe {
#zip(x, y).simd_into(self)
}
}
}
#zip(x, y).simd_into(self)
},
)
}
OpSig::Slide { granularity } => {
use SlideGranularity::*;
Expand Down Expand Up @@ -480,13 +452,11 @@ impl Level for Neon {
} else {
let to_ty = &vec_ty.reinterpret(target_ty, scalar_bits);
let neon = cvt_intrinsic("vcvt", to_ty, vec_ty);
quote! {
#method_sig {
unsafe {
#neon(a.into()).simd_into(self)
}
}
}
op.simd_trait_kernel_method(
self.token(),
vec_ty,
quote! { #neon(a.into()).simd_into(self) },
)
}
}
OpSig::Reinterpret {
Expand All @@ -497,13 +467,11 @@ impl Level for Neon {
let to_ty = vec_ty.reinterpret(target_ty, scalar_bits);
let neon = cvt_intrinsic("vreinterpret", &to_ty, vec_ty);

quote! {
#method_sig {
unsafe {
#neon(a.into()).simd_into(self)
}
}
}
op.simd_trait_kernel_method(
self.token(),
vec_ty,
quote! { #neon(a.into()).simd_into(self) },
)
} else {
quote! {}
}
Expand All @@ -522,13 +490,11 @@ impl Level for Neon {
let u32_ty = vec_ty.reinterpret(ScalarType::Unsigned, 32);
let min_max = simple_intrinsic(reduction, &u32_ty);
let reinterpret = format_ident!("vreinterpretq_u32_s{}", vec_ty.scalar_bits);
quote! {
#method_sig {
unsafe {
#min_max(#reinterpret(a.into())) #target
}
}
}
op.simd_trait_kernel_method(
self.token(),
vec_ty,
quote! { #min_max(#reinterpret(a.into())) #target },
)
}
OpSig::MaskFromBitmask => self.handle_mask_from_bitmask(method_sig, vec_ty),
OpSig::MaskToBitmask => self.handle_mask_to_bitmask(method_sig, vec_ty),
Expand Down
Loading
Loading