Rollup merge of rust-lang#95961 - RalfJung:gather-scatter, r=workingjubilee

implement SIMD gather/scatter via vector getelementptr

Fixes rust-lang/portable-simd#271

However, I don't *really* know what I am doing here... Cc `@workingjubilee` `@calebzulawski`

I didn't do anything for cranelift -- `@bjorn3` not sure if it's okay for that backend to temporarily break. I'm happy to cherry-pick a patch that adds cranelift support. :)

Author: RalfJung

compiler/rustc_codegen_llvm/src/intrinsic.rs

@@ -1839,6 +1839,27 @@ unsupported {} from `{}` with element `{}` of size `{}` to `{}`"#,
         simd_neg: Int => neg, Float => fneg;
     }
 
+    if name == sym::simd_arith_offset {
+        // This also checks that the first operand is a ptr type.
+        let pointee = in_elem.builtin_deref(true).unwrap_or_else(|| {
+            span_bug!(span, "must be called with a vector of pointer types as first argument")
+        });
+        let layout = bx.layout_of(pointee.ty);
+        let ptrs = args[0].immediate();
+        // The second argument must be a ptr-sized integer.
+        // (We don't care about the signedness, this is wrapping anyway.)
+        let (_offsets_len, offsets_elem) = arg_tys[1].simd_size_and_type(bx.tcx());
+        if !matches!(offsets_elem.kind(), ty::Int(ty::IntTy::Isize) | ty::Uint(ty::UintTy::Usize)) {
+            span_bug!(
+                span,
+                "must be called with a vector of pointer-sized integers as second argument"
+            );
+        }
+        let offsets = args[1].immediate();
+
+        return Ok(bx.gep(bx.backend_type(layout), ptrs, &[offsets]));
+    }
+
     if name == sym::simd_saturating_add || name == sym::simd_saturating_sub {
         let lhs = args[0].immediate();
         let rhs = args[1].immediate();

compiler/rustc_span/src/symbol.rs

@@ -1247,6 +1247,7 @@ symbols! {
         simd,
         simd_add,
         simd_and,
+        simd_arith_offset,
         simd_as,
         simd_bitmask,
         simd_cast,

compiler/rustc_typeck/src/check/intrinsic.rs

@@ -437,6 +437,7 @@ pub fn check_platform_intrinsic_type(tcx: TyCtxt<'_>, it: &hir::ForeignItem<'_>)
         | sym::simd_fpow
         | sym::simd_saturating_add
         | sym::simd_saturating_sub => (1, vec![param(0), param(0)], param(0)),
+        sym::simd_arith_offset => (2, vec![param(0), param(1)], param(0)),
         sym::simd_neg
         | sym::simd_fsqrt
         | sym::simd_fsin

library/portable-simd/crates/core_simd/src/intrinsics.rs

@@ -61,6 +61,10 @@ extern "platform-intrinsic" {
     /// xor
     pub(crate) fn simd_xor<T>(x: T, y: T) -> T;
 
+    /// getelementptr (without inbounds)
+    #[cfg(not(bootstrap))]
+    pub(crate) fn simd_arith_offset<T, U>(ptrs: T, offsets: U) -> T;
+
     /// fptoui/fptosi/uitofp/sitofp
     /// casting floats to integers is truncating, so it is safe to convert values like e.g. 1.5
     /// but the truncated value must fit in the target type or the result is poison.

library/portable-simd/crates/core_simd/src/vector/ptr.rs

@@ -1,5 +1,8 @@
 //! Private implementation details of public gather/scatter APIs.
+#[cfg(not(bootstrap))]
+use crate::simd::intrinsics;
 use crate::simd::{LaneCount, Simd, SupportedLaneCount};
+#[cfg(bootstrap)]
 use core::mem;
 
 /// A vector of *const T.
@@ -21,12 +24,16 @@ where
     #[inline]
     #[must_use]
     pub fn wrapping_add(self, addend: Simd<usize, LANES>) -> Self {
+        #[cfg(bootstrap)]
         // Safety: converting pointers to usize and vice-versa is safe
         // (even if using that pointer is not)
         unsafe {
             let x: Simd<usize, LANES> = mem::transmute_copy(&self);
             mem::transmute_copy(&{ x + (addend * Simd::splat(mem::size_of::<T>())) })
         }
+        #[cfg(not(bootstrap))]
+        // Safety: this intrinsic doesn't have a precondition
+        unsafe { intrinsics::simd_arith_offset(self, addend) }
     }
 }
 
@@ -49,11 +56,15 @@ where
     #[inline]
     #[must_use]
     pub fn wrapping_add(self, addend: Simd<usize, LANES>) -> Self {
+        #[cfg(bootstrap)]
         // Safety: converting pointers to usize and vice-versa is safe
         // (even if using that pointer is not)
         unsafe {
             let x: Simd<usize, LANES> = mem::transmute_copy(&self);
             mem::transmute_copy(&{ x + (addend * Simd::splat(mem::size_of::<T>())) })
         }
+        #[cfg(not(bootstrap))]
+        // Safety: this intrinsic doesn't have a precondition
+        unsafe { intrinsics::simd_arith_offset(self, addend) }
     }
 }

src/test/codegen/simd_arith_offset.rs

@@ -0,0 +1,26 @@
+// compile-flags: -C no-prepopulate-passes
+// only-64bit (because the LLVM type of i64 for usize shows up)
+//
+
+#![crate_type = "lib"]
+#![feature(repr_simd, platform_intrinsics)]
+
+extern "platform-intrinsic" {
+    pub(crate) fn simd_arith_offset<T, U>(ptrs: T, offsets: U) -> T;
+}
+
+/// A vector of *const T.
+#[derive(Debug, Copy, Clone)]
+#[repr(simd)]
+pub struct SimdConstPtr<T, const LANES: usize>([*const T; LANES]);
+
+#[derive(Debug, Copy, Clone)]
+#[repr(simd)]
+pub struct Simd<T, const LANES: usize>([T; LANES]);
+
+// CHECK-LABEL: smoke
+#[no_mangle]
+pub fn smoke(ptrs: SimdConstPtr<u8, 8>, offsets: Simd<usize, 8>) -> SimdConstPtr<u8, 8> {
+    // CHECK: getelementptr i8, <8 x i8*> %_3, <8 x i64> %_4
+    unsafe { simd_arith_offset(ptrs, offsets) }
+}