Skip to content

Commit

Permalink
Consistently use the most significant bit of vector masks
Browse files Browse the repository at this point in the history 
This improves the codegen for vector `select`, `gather`, `scatter` and
boolean reduction intrinsics and fixes rust-lang/portable-simd#316.

The current behavior of most mask operations during llvm codegen is to
truncate the mask vector to <N x i1>, telling llvm to use the least
significat bit. The exception is the `simd_bitmask` intrinsics, which
already used the most signifiant bit.

Since sse/avx instructions are defined to use the most significant bit,
truncating means that llvm has to insert a left shift to move the bit
into the most significant position, before the mask can actually be
used.

Similarly on aarch64, mask operations like blend work bit by bit,
repeating the least significant bit across the whole lane involves
shifting it into the sign position and then comparing against zero.

By shifting before truncating to <N x i1>, we tell llvm that we only
consider the most significant bit, removing the need for additional
shift instructions in the assembly.
  • Loading branch information
@jhorstmann
jhorstmann committed Jan 26, 2025
1 parent c2270be commit 3779b8e
Show file tree
Hide file tree
Showing 13 changed files with 280 additions and 172 deletions.
197 changes: 104 additions & 93 deletions compiler/rustc_codegen_llvm/src/intrinsic.rs
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1182,6 +1182,60 @@ fn generic_simd_intrinsic<'ll, 'tcx>(
}};
}

/// Returns the bitwidth of the `$ty` argument if it is an `Int` type.
macro_rules! require_int_ty {
($ty: expr, $diag: expr) => {
match $ty {
ty::Int(i) => i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits()),
_ => {
return_error!($diag);
}
}
};
}

/// Returns the bitwidth of the `$ty` argument if it is an `Int` or `Uint` type.
macro_rules! require_int_or_uint_ty {
($ty: expr, $diag: expr) => {
match $ty {
ty::Int(i) => i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits()),
ty::Uint(i) => {
i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits())
}
_ => {
return_error!($diag);
}
}
};
}

/// Converts a vector mask, where each element has a bit width equal to the data elements it is used with,
/// down to an i1 based mask that can be used by llvm intrinsics.
///
/// The rust simd semantics are that each element should either consist of all ones or all zeroes,
/// but this information is not available to llvm. Truncating the vector effectively uses the lowest bit,
/// but codegen for several targets is better if we consider the highest bit by shifting.
///
/// For x86 SSE/AVX targets this is beneficial since most instructions with mask parameters only consider the highest bit.
/// So even though on llvm level we have an additional shift, in the final assembly there is no shift or truncate and
/// instead the mask can be used as is.
///
/// For aarch64 and other targets there is a benefit because a mask from the sign bit can be more
/// efficiently converted to an all ones / all zeroes mask by comparing whether each element is negative.
fn vector_mask_to_bitmask<'a, 'll, 'tcx>(
bx: &mut Builder<'a, 'll, 'tcx>,
i_xn: &'ll Value,
in_elem_bitwidth: u64,
in_len: u64,
) -> &'ll Value {
// Shift the MSB to the right by "in_elem_bitwidth - 1" into the first bit position.
let shift_idx = bx.cx.const_int(bx.type_ix(in_elem_bitwidth), (in_elem_bitwidth - 1) as _);
let shift_indices = vec![shift_idx; in_len as _];
let i_xn_msb = bx.lshr(i_xn, bx.const_vector(shift_indices.as_slice()));
// Truncate vector to an <i1 x N>
bx.trunc(i_xn_msb, bx.type_vector(bx.type_i1(), in_len))
}

let tcx = bx.tcx();
let sig = tcx.normalize_erasing_late_bound_regions(bx.typing_env(), callee_ty.fn_sig(tcx));
let arg_tys = sig.inputs();
Expand Down Expand Up @@ -1433,14 +1487,13 @@ fn generic_simd_intrinsic<'ll, 'tcx>(
m_len,
v_len
});
match m_elem_ty.kind() {
ty::Int(_) => {}
_ => return_error!(InvalidMonomorphization::MaskType { span, name, ty: m_elem_ty }),
}
// truncate the mask to a vector of i1s
let i1 = bx.type_i1();
let i1xn = bx.type_vector(i1, m_len as u64);
let m_i1s = bx.trunc(args[0].immediate(), i1xn);
let in_elem_bitwidth =
require_int_ty!(m_elem_ty.kind(), InvalidMonomorphization::MaskType {
span,
name,
ty: m_elem_ty
});
let m_i1s = vector_mask_to_bitmask(bx, args[0].immediate(), in_elem_bitwidth, m_len);
return Ok(bx.select(m_i1s, args[1].immediate(), args[2].immediate()));
}

Expand All @@ -1457,33 +1510,15 @@ fn generic_simd_intrinsic<'ll, 'tcx>(
let expected_bytes = in_len.div_ceil(8);

// Integer vector <i{in_bitwidth} x in_len>:
let (i_xn, in_elem_bitwidth) = match in_elem.kind() {
ty::Int(i) => (
args[0].immediate(),
i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits()),
),
ty::Uint(i) => (
args[0].immediate(),
i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits()),
),
_ => return_error!(InvalidMonomorphization::VectorArgument {
let in_elem_bitwidth =
require_int_or_uint_ty!(in_elem.kind(), InvalidMonomorphization::VectorArgument {
span,
name,
in_ty,
in_elem
}),
};
});

// LLVM doesn't always know the inputs are `0` or `!0`, so we shift here so it optimizes to
// `pmovmskb` and similar on x86.
let shift_indices =
vec![
bx.cx.const_int(bx.type_ix(in_elem_bitwidth), (in_elem_bitwidth - 1) as _);
in_len as _
];
let i_xn_msb = bx.lshr(i_xn, bx.const_vector(shift_indices.as_slice()));
// Truncate vector to an <i1 x N>
let i1xn = bx.trunc(i_xn_msb, bx.type_vector(bx.type_i1(), in_len));
let i1xn = vector_mask_to_bitmask(bx, args[0].immediate(), in_elem_bitwidth, in_len);
// Bitcast <i1 x N> to iN:
let i_ = bx.bitcast(i1xn, bx.type_ix(in_len));

Expand Down Expand Up @@ -1704,28 +1739,21 @@ fn generic_simd_intrinsic<'ll, 'tcx>(
}
);

match element_ty2.kind() {
ty::Int(_) => (),
_ => {
return_error!(InvalidMonomorphization::ThirdArgElementType {
span,
name,
expected_element: element_ty2,
third_arg: arg_tys[2]
});
}
}
let mask_elem_bitwidth =
require_int_ty!(element_ty2.kind(), InvalidMonomorphization::ThirdArgElementType {
span,
name,
expected_element: element_ty2,
third_arg: arg_tys[2]
});

// Alignment of T, must be a constant integer value:
let alignment_ty = bx.type_i32();
let alignment = bx.const_i32(bx.align_of(in_elem).bytes() as i32);

// Truncate the mask vector to a vector of i1s:
let (mask, mask_ty) = {
let i1 = bx.type_i1();
let i1xn = bx.type_vector(i1, in_len);
(bx.trunc(args[2].immediate(), i1xn), i1xn)
};
let mask = vector_mask_to_bitmask(bx, args[2].immediate(), mask_elem_bitwidth, in_len);
let mask_ty = bx.type_vector(bx.type_i1(), in_len);

// Type of the vector of pointers:
let llvm_pointer_vec_ty = llvm_vector_ty(bx, element_ty1, in_len);
Expand Down Expand Up @@ -1810,27 +1838,21 @@ fn generic_simd_intrinsic<'ll, 'tcx>(
}
);

require!(
matches!(mask_elem.kind(), ty::Int(_)),
InvalidMonomorphization::ThirdArgElementType {
let m_elem_bitwidth =
require_int_ty!(mask_elem.kind(), InvalidMonomorphization::ThirdArgElementType {
span,
name,
expected_element: values_elem,
third_arg: mask_ty,
}
);
});

let mask = vector_mask_to_bitmask(bx, args[0].immediate(), m_elem_bitwidth, mask_len);
let mask_ty = bx.type_vector(bx.type_i1(), mask_len);

// Alignment of T, must be a constant integer value:
let alignment_ty = bx.type_i32();
let alignment = bx.const_i32(bx.align_of(values_elem).bytes() as i32);

// Truncate the mask vector to a vector of i1s:
let (mask, mask_ty) = {
let i1 = bx.type_i1();
let i1xn = bx.type_vector(i1, mask_len);
(bx.trunc(args[0].immediate(), i1xn), i1xn)
};

let llvm_pointer = bx.type_ptr();

// Type of the vector of elements:
Expand Down Expand Up @@ -1901,27 +1923,21 @@ fn generic_simd_intrinsic<'ll, 'tcx>(
}
);

require!(
matches!(mask_elem.kind(), ty::Int(_)),
InvalidMonomorphization::ThirdArgElementType {
let m_elem_bitwidth =
require_int_ty!(mask_elem.kind(), InvalidMonomorphization::ThirdArgElementType {
span,
name,
expected_element: values_elem,
third_arg: mask_ty,
}
);
});

let mask = vector_mask_to_bitmask(bx, args[0].immediate(), m_elem_bitwidth, mask_len);
let mask_ty = bx.type_vector(bx.type_i1(), mask_len);

// Alignment of T, must be a constant integer value:
let alignment_ty = bx.type_i32();
let alignment = bx.const_i32(bx.align_of(values_elem).bytes() as i32);

// Truncate the mask vector to a vector of i1s:
let (mask, mask_ty) = {
let i1 = bx.type_i1();
let i1xn = bx.type_vector(i1, in_len);
(bx.trunc(args[0].immediate(), i1xn), i1xn)
};

let ret_t = bx.type_void();

let llvm_pointer = bx.type_ptr();
Expand Down Expand Up @@ -1995,28 +2011,21 @@ fn generic_simd_intrinsic<'ll, 'tcx>(
);

// The element type of the third argument must be a signed integer type of any width:
match element_ty2.kind() {
ty::Int(_) => (),
_ => {
return_error!(InvalidMonomorphization::ThirdArgElementType {
span,
name,
expected_element: element_ty2,
third_arg: arg_tys[2]
});
}
}
let mask_elem_bitwidth =
require_int_ty!(element_ty2.kind(), InvalidMonomorphization::ThirdArgElementType {
span,
name,
expected_element: element_ty2,
third_arg: arg_tys[2]
});

// Alignment of T, must be a constant integer value:
let alignment_ty = bx.type_i32();
let alignment = bx.const_i32(bx.align_of(in_elem).bytes() as i32);

// Truncate the mask vector to a vector of i1s:
let (mask, mask_ty) = {
let i1 = bx.type_i1();
let i1xn = bx.type_vector(i1, in_len);
(bx.trunc(args[2].immediate(), i1xn), i1xn)
};
let mask = vector_mask_to_bitmask(bx, args[2].immediate(), mask_elem_bitwidth, in_len);
let mask_ty = bx.type_vector(bx.type_i1(), in_len);

let ret_t = bx.type_void();

Expand Down Expand Up @@ -2164,8 +2173,13 @@ fn generic_simd_intrinsic<'ll, 'tcx>(
});
args[0].immediate()
} else {
match in_elem.kind() {
ty::Int(_) | ty::Uint(_) => {}
let bitwidth = match in_elem.kind() {
ty::Int(i) => {
i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits())
}
ty::Uint(i) => {
i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits())
}
_ => return_error!(InvalidMonomorphization::UnsupportedSymbol {
span,
name,
Expand All @@ -2174,12 +2188,9 @@ fn generic_simd_intrinsic<'ll, 'tcx>(
in_elem,
ret_ty
}),
}
};

// boolean reductions operate on vectors of i1s:
let i1 = bx.type_i1();
let i1xn = bx.type_vector(i1, in_len as u64);
bx.trunc(args[0].immediate(), i1xn)
vector_mask_to_bitmask(bx, args[0].immediate(), bitwidth, in_len as _)
};
return match in_elem.kind() {
ty::Int(_) | ty::Uint(_) => {
Expand Down
4 changes: 2 additions & 2 deletions tests/assembly/simd-intrinsic-gather.rs
View file
Original file line number Diff line number Diff line change
Expand Up @@ -35,8 +35,8 @@ pub unsafe extern "C" fn gather_f64x4(mask: m64x4, ptrs: pf64x4) -> f64x4 {
// FIXME: This should also get checked to generate a gather instruction for avx2.
// Currently llvm scalarizes this code, see https://github.com/llvm/llvm-project/issues/59789
//
// x86-avx512: vpsllq ymm0, ymm0, 63
// x86-avx512-NEXT: vpmovq2m k1, ymm0
// x86-avx512-NOT: vpsllq
// x86-avx512: vpmovq2m k1, ymm0
// x86-avx512-NEXT: vpxor xmm0, xmm0, xmm0
// x86-avx512-NEXT: vgatherqpd ymm0 {k1}, {{(ymmword)|(qword)}} ptr [1*ymm1]
simd_gather(f64x4([0_f64, 0_f64, 0_f64, 0_f64]), ptrs, mask)
Expand Down
27 changes: 13 additions & 14 deletions tests/assembly/simd-intrinsic-mask-load.rs
View file
Original file line number Diff line number Diff line change
Expand Up @@ -47,9 +47,9 @@ extern "rust-intrinsic" {
pub unsafe extern "C" fn load_i8x16(mask: m8x16, pointer: *const i8) -> i8x16 {
// Since avx2 supports no masked loads for bytes, the code tests each individual bit
// and jumps to code that inserts individual bytes.
// x86-avx2: vpsllw xmm0, xmm0, 7
// x86-avx2-NEXT: vpmovmskb eax, xmm0
// x86-avx2-NEXT: vpxor xmm0, xmm0
// x86-avx2-NOT: vpsllw
// x86-avx2-DAG: vpmovmskb eax
// x86-avx2-DAG: vpxor
// x86-avx2-NEXT: test al, 1
// x86-avx2-NEXT: jne
// x86-avx2-NEXT: test al, 2
Expand All @@ -58,32 +58,31 @@ pub unsafe extern "C" fn load_i8x16(mask: m8x16, pointer: *const i8) -> i8x16 {
// x86-avx2-NEXT: vmovd xmm0, [[REG]]
// x86-avx2-DAG: vpinsrb xmm0, xmm0, byte ptr [rdi + 1], 1
//
// x86-avx512: vpsllw xmm0, xmm0, 7
// x86-avx512-NEXT: vpmovb2m k1, xmm0
// x86-avx512-NOT: vpsllw
// x86-avx512: vpmovb2m k1, xmm0
// x86-avx512-NEXT: vmovdqu8 xmm0 {k1} {z}, xmmword ptr [rdi]
simd_masked_load(mask, pointer, i8x16([0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]))
}

// CHECK-LABEL: load_f32x8
#[no_mangle]
pub unsafe extern "C" fn load_f32x8(mask: m32x8, pointer: *const f32) -> f32x8 {
// x86-avx2: vpslld ymm0, ymm0, 31
// x86-avx2-NEXT: vmaskmovps ymm0, ymm0, ymmword ptr [rdi]
// x86-avx2-NOT: vpslld
// x86-avx2: vmaskmovps ymm0, ymm0, ymmword ptr [rdi]
//
// x86-avx512: vpslld ymm0, ymm0, 31
// x86-avx512-NEXT: vpmovd2m k1, ymm0
// x86-avx512-NOT: vpslld
// x86-avx512: vpmovd2m k1, ymm0
// x86-avx512-NEXT: vmovups ymm0 {k1} {z}, ymmword ptr [rdi]
simd_masked_load(mask, pointer, f32x8([0_f32, 0_f32, 0_f32, 0_f32, 0_f32, 0_f32, 0_f32, 0_f32]))
}

// CHECK-LABEL: load_f64x4
#[no_mangle]
pub unsafe extern "C" fn load_f64x4(mask: m64x4, pointer: *const f64) -> f64x4 {
// x86-avx2: vpsllq ymm0, ymm0, 63
// x86-avx2-NEXT: vmaskmovpd ymm0, ymm0, ymmword ptr [rdi]
// x86-avx2-NOT: vpsllq
// x86-avx2: vmaskmovpd ymm0, ymm0, ymmword ptr [rdi]
//
// x86-avx512: vpsllq ymm0, ymm0, 63
// x86-avx512-NEXT: vpmovq2m k1, ymm0
// x86-avx512-NEXT: vmovupd ymm0 {k1} {z}, ymmword ptr [rdi]
// x86-avx512-NOT: vpsllq
// x86-avx512: vpmovq2m k1, ymm0
simd_masked_load(mask, pointer, f64x4([0_f64, 0_f64, 0_f64, 0_f64]))
}
Loading

0 comments on commit 3779b8e

Please sign in to comment.