Implement some float simd intrinsics
This commit is contained in:
bjorn3
parent
8691b8b8b6
commit
69526d464f
4 changed files with 187 additions and 64 deletions
|
|
@ -1,5 +1,6 @@
|
|||
#![feature(core_intrinsics)]
|
||||
|
||||
use std::arch::x86_64::*;
|
||||
use std::io::Write;
|
||||
use std::intrinsics;
|
||||
|
||||
|
|
@ -52,8 +53,6 @@ fn main() {
|
|||
|
||||
#[target_feature(enable = "sse2")]
|
||||
unsafe fn test_simd() {
|
||||
use std::arch::x86_64::*;
|
||||
|
||||
let x = _mm_setzero_si128();
|
||||
let y = _mm_set1_epi16(7);
|
||||
let or = _mm_or_si128(x, y);
|
||||
|
|
@ -67,6 +66,8 @@ unsafe fn test_simd() {
|
|||
test_mm_slli_si128();
|
||||
test_mm_movemask_epi8();
|
||||
test_mm256_movemask_epi8();
|
||||
test_mm_add_epi8();
|
||||
test_mm_add_pd();
|
||||
|
||||
let mask1 = _mm_movemask_epi8(dbg!(_mm_setr_epi8(255u8 as i8, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0)));
|
||||
assert_eq!(mask1, 1);
|
||||
|
|
@ -74,8 +75,6 @@ unsafe fn test_simd() {
|
|||
|
||||
#[target_feature(enable = "sse2")]
|
||||
unsafe fn test_mm_slli_si128() {
|
||||
use std::arch::x86_64::*;
|
||||
|
||||
#[rustfmt::skip]
|
||||
let a = _mm_setr_epi8(
|
||||
1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16,
|
||||
|
|
@ -116,8 +115,6 @@ unsafe fn test_mm_slli_si128() {
|
|||
|
||||
#[target_feature(enable = "sse2")]
|
||||
unsafe fn test_mm_movemask_epi8() {
|
||||
use std::arch::x86_64::*;
|
||||
|
||||
#[rustfmt::skip]
|
||||
let a = _mm_setr_epi8(
|
||||
0b1000_0000u8 as i8, 0b0, 0b1000_0000u8 as i8, 0b01,
|
||||
|
|
@ -131,20 +128,48 @@ unsafe fn test_mm_movemask_epi8() {
|
|||
|
||||
#[target_feature(enable = "avx2")]
|
||||
unsafe fn test_mm256_movemask_epi8() {
|
||||
use std::arch::x86_64::*;
|
||||
|
||||
let a = _mm256_set1_epi8(-1);
|
||||
let r = _mm256_movemask_epi8(a);
|
||||
let e = -1;
|
||||
assert_eq!(r, e);
|
||||
}
|
||||
|
||||
#[target_feature(enable = "sse2")]
|
||||
unsafe fn test_mm_add_epi8() {
|
||||
let a = _mm_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
|
||||
#[rustfmt::skip]
|
||||
let b = _mm_setr_epi8(
|
||||
16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31,
|
||||
);
|
||||
let r = _mm_add_epi8(a, b);
|
||||
#[rustfmt::skip]
|
||||
let e = _mm_setr_epi8(
|
||||
16, 18, 20, 22, 24, 26, 28, 30, 32, 34, 36, 38, 40, 42, 44, 46,
|
||||
);
|
||||
assert_eq_m128i(r, e);
|
||||
}
|
||||
|
||||
#[target_feature(enable = "sse2")]
|
||||
unsafe fn test_mm_add_pd() {
|
||||
let a = _mm_setr_pd(1.0, 2.0);
|
||||
let b = _mm_setr_pd(5.0, 10.0);
|
||||
let r = _mm_add_pd(a, b);
|
||||
assert_eq_m128d(r, _mm_setr_pd(6.0, 12.0));
|
||||
}
|
||||
|
||||
fn assert_eq_m128i(x: std::arch::x86_64::__m128i, y: std::arch::x86_64::__m128i) {
|
||||
unsafe {
|
||||
assert_eq!(std::mem::transmute::<_, [u8; 16]>(x), std::mem::transmute::<_, [u8; 16]>(x));
|
||||
}
|
||||
}
|
||||
|
||||
#[target_feature(enable = "sse2")]
|
||||
pub unsafe fn assert_eq_m128d(a: __m128d, b: __m128d) {
|
||||
if _mm_movemask_pd(_mm_cmpeq_pd(a, b)) != 0b11 {
|
||||
panic!("{:?} != {:?}", a, b);
|
||||
}
|
||||
}
|
||||
|
||||
#[derive(PartialEq)]
|
||||
enum LoopState {
|
||||
Continue(()),
|
||||
|
|
|
|||
|
|
@ -88,7 +88,7 @@ pub fn trans_constant<'a, 'tcx: 'a>(
|
|||
}
|
||||
|
||||
pub fn force_eval_const<'a, 'tcx: 'a>(
|
||||
fx: &mut FunctionCx<'a, 'tcx, impl Backend>,
|
||||
fx: &FunctionCx<'a, 'tcx, impl Backend>,
|
||||
const_: &'tcx Const,
|
||||
) -> &'tcx Const<'tcx> {
|
||||
match const_.val {
|
||||
|
|
@ -422,3 +422,32 @@ impl<'mir, 'tcx> Machine<'mir, 'tcx> for TransPlaceInterpreter {
|
|||
Ok(())
|
||||
}
|
||||
}
|
||||
|
||||
pub fn mir_operand_get_const_val<'tcx>(
|
||||
fx: &FunctionCx<'_, 'tcx, impl Backend>,
|
||||
operand: &Operand<'tcx>,
|
||||
) -> Result<&'tcx Const<'tcx>, String> {
|
||||
let place = match operand {
|
||||
Operand::Copy(place) => place,
|
||||
Operand::Constant(const_) => return Ok(force_eval_const(fx, const_.literal)),
|
||||
_ => return Err(format!("{:?}", operand)),
|
||||
};
|
||||
|
||||
assert!(place.projection.is_none());
|
||||
let static_ = match &place.base {
|
||||
PlaceBase::Static(static_) => {
|
||||
static_
|
||||
}
|
||||
PlaceBase::Local(_) => return Err("local".to_string()),
|
||||
};
|
||||
|
||||
Ok(match &static_.kind {
|
||||
StaticKind::Static(_) => unimplemented!(),
|
||||
StaticKind::Promoted(promoted) => {
|
||||
fx.tcx.const_eval(ParamEnv::reveal_all().and(GlobalId {
|
||||
instance: fx.instance,
|
||||
promoted: Some(*promoted),
|
||||
})).unwrap()
|
||||
}
|
||||
})
|
||||
}
|
||||
|
|
|
|||
|
|
@ -144,7 +144,7 @@ pub fn lane_type_and_count<'tcx>(
|
|||
(lane_layout, lane_count)
|
||||
}
|
||||
|
||||
fn simd_for_each_lane<'tcx, B: Backend>(
|
||||
pub fn simd_for_each_lane<'tcx, B: Backend>(
|
||||
fx: &mut FunctionCx<'_, 'tcx, B>,
|
||||
intrinsic: &str,
|
||||
x: CValue<'tcx>,
|
||||
|
|
@ -170,23 +170,37 @@ fn simd_for_each_lane<'tcx, B: Backend>(
|
|||
}
|
||||
}
|
||||
|
||||
fn bool_to_zero_or_max_uint<'tcx>(
|
||||
pub fn bool_to_zero_or_max_uint<'tcx>(
|
||||
fx: &mut FunctionCx<'_, 'tcx, impl Backend>,
|
||||
layout: TyLayout<'tcx>,
|
||||
val: Value,
|
||||
) -> CValue<'tcx> {
|
||||
let ty = fx.clif_type(layout.ty).unwrap();
|
||||
|
||||
let zero = fx.bcx.ins().iconst(ty, 0);
|
||||
let max = fx.bcx.ins().iconst(ty, (u64::max_value() >> (64 - ty.bits())) as i64);
|
||||
let res = crate::common::codegen_select(&mut fx.bcx, val, max, zero);
|
||||
let int_ty = match ty {
|
||||
types::F32 => types::I32,
|
||||
types::F64 => types::I64,
|
||||
ty => ty,
|
||||
};
|
||||
|
||||
let zero = fx.bcx.ins().iconst(int_ty, 0);
|
||||
let max = fx.bcx.ins().iconst(int_ty, (u64::max_value() >> (64 - int_ty.bits())) as i64);
|
||||
let mut res = crate::common::codegen_select(&mut fx.bcx, val, max, zero);
|
||||
|
||||
if ty.is_float() {
|
||||
res = fx.bcx.ins().bitcast(ty, res);
|
||||
}
|
||||
|
||||
CValue::by_val(res, layout)
|
||||
}
|
||||
|
||||
macro_rules! simd_cmp {
|
||||
($fx:expr, $intrinsic:expr, $cc:ident($x:ident, $y:ident) -> $ret:ident) => {
|
||||
simd_for_each_lane($fx, $intrinsic, $x, $y, $ret, |fx, _lane_layout, res_lane_layout, x_lane, y_lane| {
|
||||
let res_lane = fx.bcx.ins().icmp(IntCC::$cc, x_lane, y_lane);
|
||||
simd_for_each_lane($fx, $intrinsic, $x, $y, $ret, |fx, lane_layout, res_lane_layout, x_lane, y_lane| {
|
||||
let res_lane = match lane_layout.ty.sty {
|
||||
ty::Uint(_) | ty::Int(_) => fx.bcx.ins().icmp(IntCC::$cc, x_lane, y_lane),
|
||||
_ => unreachable!("{:?}", lane_layout.ty),
|
||||
};
|
||||
bool_to_zero_or_max_uint(fx, res_lane_layout, res_lane)
|
||||
});
|
||||
};
|
||||
|
|
@ -203,10 +217,13 @@ macro_rules! simd_cmp {
|
|||
|
||||
}
|
||||
|
||||
macro_rules! simd_binop {
|
||||
macro_rules! simd_int_binop {
|
||||
($fx:expr, $intrinsic:expr, $op:ident($x:ident, $y:ident) -> $ret:ident) => {
|
||||
simd_for_each_lane($fx, $intrinsic, $x, $y, $ret, |fx, _lane_layout, ret_lane_layout, x_lane, y_lane| {
|
||||
let res_lane = fx.bcx.ins().$op(x_lane, y_lane);
|
||||
simd_for_each_lane($fx, $intrinsic, $x, $y, $ret, |fx, lane_layout, ret_lane_layout, x_lane, y_lane| {
|
||||
let res_lane = match lane_layout.ty.sty {
|
||||
ty::Uint(_) | ty::Int(_) => fx.bcx.ins().$op(x_lane, y_lane),
|
||||
_ => unreachable!("{:?}", lane_layout.ty),
|
||||
};
|
||||
CValue::by_val(res_lane, ret_lane_layout)
|
||||
});
|
||||
};
|
||||
|
|
@ -222,6 +239,42 @@ macro_rules! simd_binop {
|
|||
};
|
||||
}
|
||||
|
||||
macro_rules! simd_int_flt_binop {
|
||||
($fx:expr, $intrinsic:expr, $op:ident|$op_f:ident($x:ident, $y:ident) -> $ret:ident) => {
|
||||
simd_for_each_lane($fx, $intrinsic, $x, $y, $ret, |fx, lane_layout, ret_lane_layout, x_lane, y_lane| {
|
||||
let res_lane = match lane_layout.ty.sty {
|
||||
ty::Uint(_) | ty::Int(_) => fx.bcx.ins().$op(x_lane, y_lane),
|
||||
ty::Float(_) => fx.bcx.ins().$op_f(x_lane, y_lane),
|
||||
_ => unreachable!("{:?}", lane_layout.ty),
|
||||
};
|
||||
CValue::by_val(res_lane, ret_lane_layout)
|
||||
});
|
||||
};
|
||||
($fx:expr, $intrinsic:expr, $op_u:ident|$op_s:ident|$op_f:ident($x:ident, $y:ident) -> $ret:ident) => {
|
||||
simd_for_each_lane($fx, $intrinsic, $x, $y, $ret, |fx, lane_layout, ret_lane_layout, x_lane, y_lane| {
|
||||
let res_lane = match lane_layout.ty.sty {
|
||||
ty::Uint(_) => fx.bcx.ins().$op_u(x_lane, y_lane),
|
||||
ty::Int(_) => fx.bcx.ins().$op_s(x_lane, y_lane),
|
||||
ty::Float(_) => fx.bcx.ins().$op_f(x_lane, y_lane),
|
||||
_ => unreachable!("{:?}", lane_layout.ty),
|
||||
};
|
||||
CValue::by_val(res_lane, ret_lane_layout)
|
||||
});
|
||||
};
|
||||
}
|
||||
|
||||
macro_rules! simd_flt_binop {
|
||||
($fx:expr, $intrinsic:expr, $op:ident($x:ident, $y:ident) -> $ret:ident) => {
|
||||
simd_for_each_lane($fx, $intrinsic, $x, $y, $ret, |fx, lane_layout, ret_lane_layout, x_lane, y_lane| {
|
||||
let res_lane = match lane_layout.ty.sty {
|
||||
ty::Float(_) => fx.bcx.ins().$op(x_lane, y_lane),
|
||||
_ => unreachable!("{:?}", lane_layout.ty),
|
||||
};
|
||||
CValue::by_val(res_lane, ret_lane_layout)
|
||||
});
|
||||
}
|
||||
}
|
||||
|
||||
pub fn codegen_intrinsic_call<'a, 'tcx: 'a>(
|
||||
fx: &mut FunctionCx<'a, 'tcx, impl Backend>,
|
||||
def_id: DefId,
|
||||
|
|
@ -840,30 +893,7 @@ pub fn codegen_intrinsic_call<'a, 'tcx: 'a>(
|
|||
|
||||
let indexes = {
|
||||
use rustc::mir::interpret::*;
|
||||
let idx_place = match idx {
|
||||
Operand::Copy(idx_place) => {
|
||||
idx_place
|
||||
}
|
||||
_ => panic!("simd_shuffle* idx is not Operand::Copy, but {:?}", idx),
|
||||
};
|
||||
|
||||
assert!(idx_place.projection.is_none());
|
||||
let static_ = match &idx_place.base {
|
||||
PlaceBase::Static(static_) => {
|
||||
static_
|
||||
}
|
||||
PlaceBase::Local(_) => panic!("simd_shuffle* idx is not constant, but a local"),
|
||||
};
|
||||
|
||||
let idx_const = match &static_.kind {
|
||||
StaticKind::Static(_) => unimplemented!(),
|
||||
StaticKind::Promoted(promoted) => {
|
||||
fx.tcx.const_eval(ParamEnv::reveal_all().and(GlobalId {
|
||||
instance: fx.instance,
|
||||
promoted: Some(*promoted),
|
||||
})).unwrap()
|
||||
}
|
||||
};
|
||||
let idx_const = crate::constant::mir_operand_get_const_val(fx, idx).expect("simd_shuffle* idx not const");
|
||||
|
||||
let idx_bytes = match idx_const.val {
|
||||
ConstValue::ByRef { align: _, offset, alloc } => {
|
||||
|
|
@ -900,41 +930,38 @@ pub fn codegen_intrinsic_call<'a, 'tcx: 'a>(
|
|||
};
|
||||
|
||||
simd_add, (c x, c y) {
|
||||
simd_binop!(fx, intrinsic, iadd(x, y) -> ret);
|
||||
simd_int_flt_binop!(fx, intrinsic, iadd|fadd(x, y) -> ret);
|
||||
};
|
||||
simd_sub, (c x, c y) {
|
||||
simd_binop!(fx, intrinsic, isub(x, y) -> ret);
|
||||
simd_int_flt_binop!(fx, intrinsic, isub|fsub(x, y) -> ret);
|
||||
};
|
||||
simd_mul, (c x, c y) {
|
||||
simd_binop!(fx, intrinsic, imul(x, y) -> ret);
|
||||
simd_int_flt_binop!(fx, intrinsic, imul|fmul(x, y) -> ret);
|
||||
};
|
||||
simd_div, (c x, c y) {
|
||||
simd_binop!(fx, intrinsic, udiv|sdiv(x, y) -> ret);
|
||||
};
|
||||
simd_rem, (c x, c y) {
|
||||
simd_binop!(fx, intrinsic, urem|srem(x, y) -> ret);
|
||||
simd_int_flt_binop!(fx, intrinsic, udiv|sdiv|fdiv(x, y) -> ret);
|
||||
};
|
||||
simd_shl, (c x, c y) {
|
||||
simd_binop!(fx, intrinsic, ishl(x, y) -> ret);
|
||||
simd_int_binop!(fx, intrinsic, ishl(x, y) -> ret);
|
||||
};
|
||||
simd_shr, (c x, c y) {
|
||||
simd_binop!(fx, intrinsic, ushr|sshr(x, y) -> ret);
|
||||
simd_int_binop!(fx, intrinsic, ushr|sshr(x, y) -> ret);
|
||||
};
|
||||
simd_and, (c x, c y) {
|
||||
simd_binop!(fx, intrinsic, band(x, y) -> ret);
|
||||
simd_int_binop!(fx, intrinsic, band(x, y) -> ret);
|
||||
};
|
||||
simd_or, (c x, c y) {
|
||||
simd_binop!(fx, intrinsic, bor(x, y) -> ret);
|
||||
simd_int_binop!(fx, intrinsic, bor(x, y) -> ret);
|
||||
};
|
||||
simd_xor, (c x, c y) {
|
||||
simd_binop!(fx, intrinsic, bxor(x, y) -> ret);
|
||||
simd_int_binop!(fx, intrinsic, bxor(x, y) -> ret);
|
||||
};
|
||||
|
||||
simd_fmin, (c x, c y) {
|
||||
simd_binop!(fx, intrinsic, fmin(x, y) -> ret);
|
||||
simd_flt_binop!(fx, intrinsic, fmin(x, y) -> ret);
|
||||
};
|
||||
simd_fmax, (c x, c y) {
|
||||
simd_binop!(fx, intrinsic, fmax(x, y) -> ret);
|
||||
simd_flt_binop!(fx, intrinsic, fmax(x, y) -> ret);
|
||||
};
|
||||
}
|
||||
|
||||
|
|
|
|||
|
|
@ -1,4 +1,5 @@
|
|||
use crate::prelude::*;
|
||||
use crate::intrinsics::*;
|
||||
|
||||
use rustc::ty::subst::SubstsRef;
|
||||
|
||||
|
|
@ -26,7 +27,7 @@ pub fn codegen_llvm_intrinsic_call<'a, 'tcx: 'a>(
|
|||
}
|
||||
};
|
||||
|
||||
crate::intrinsics::intrinsic_match! {
|
||||
intrinsic_match! {
|
||||
fx, intrinsic, substs, args,
|
||||
_ => {
|
||||
fx.tcx.sess.warn(&format!("unsupported llvm intrinsic {}; replacing with trap", intrinsic));
|
||||
|
|
@ -34,17 +35,28 @@ pub fn codegen_llvm_intrinsic_call<'a, 'tcx: 'a>(
|
|||
};
|
||||
|
||||
// Used by `_mm_movemask_epi8` and `_mm256_movemask_epi8`
|
||||
llvm.x86.sse2.pmovmskb.128 | llvm.x86.avx2.pmovmskb, (c a) {
|
||||
let (lane_layout, lane_count) = crate::intrinsics::lane_type_and_count(fx, a.layout(), intrinsic);
|
||||
assert_eq!(lane_layout.ty.sty, fx.tcx.types.i8.sty);
|
||||
assert!(lane_count == 16 || lane_count == 32);
|
||||
llvm.x86.sse2.pmovmskb.128 | llvm.x86.avx2.pmovmskb | llvm.x86.sse2.movmsk.pd, (c a) {
|
||||
let (lane_layout, lane_count) = lane_type_and_count(fx, a.layout(), intrinsic);
|
||||
let lane_ty = fx.clif_type(lane_layout.ty).unwrap();
|
||||
assert!(lane_count <= 32);
|
||||
|
||||
let mut res = fx.bcx.ins().iconst(types::I32, 0);
|
||||
|
||||
for lane in (0..lane_count).rev() {
|
||||
let a_lane = a.value_field(fx, mir::Field::new(lane.try_into().unwrap())).load_scalar(fx);
|
||||
let a_lane_sign = fx.bcx.ins().ushr_imm(a_lane, 7); // extract sign bit of 8bit int
|
||||
let a_lane_sign = fx.bcx.ins().uextend(types::I32, a_lane_sign);
|
||||
|
||||
// cast float to int
|
||||
let a_lane = match lane_ty {
|
||||
types::F32 => fx.bcx.ins().bitcast(types::I32, a_lane),
|
||||
types::F64 => fx.bcx.ins().bitcast(types::I64, a_lane),
|
||||
_ => a_lane,
|
||||
};
|
||||
|
||||
// extract sign bit of an int
|
||||
let a_lane_sign = fx.bcx.ins().ushr_imm(a_lane, i64::from(lane_ty.bits() - 1));
|
||||
|
||||
// shift sign bit into result
|
||||
let a_lane_sign = clif_intcast(fx, a_lane_sign, types::I32, false);
|
||||
res = fx.bcx.ins().ishl_imm(res, 1);
|
||||
res = fx.bcx.ins().bor(res, a_lane_sign);
|
||||
}
|
||||
|
|
@ -52,6 +64,36 @@ pub fn codegen_llvm_intrinsic_call<'a, 'tcx: 'a>(
|
|||
let res = CValue::by_val(res, fx.layout_of(fx.tcx.types.i32));
|
||||
ret.write_cvalue(fx, res);
|
||||
};
|
||||
llvm.x86.sse2.cmp.ps | llvm.x86.sse2.cmp.pd, (c x, c y, o kind) {
|
||||
let kind_const = crate::constant::mir_operand_get_const_val(fx, kind).expect("llvm.x86.sse2.cmp.* kind not const");
|
||||
let flt_cc = match kind_const.val.try_to_bits(Size::from_bytes(1)).expect(&format!("kind not scalar: {:?}", kind_const)) {
|
||||
0 => FloatCC::Equal,
|
||||
1 => FloatCC::LessThan,
|
||||
2 => FloatCC::LessThanOrEqual,
|
||||
7 => {
|
||||
unimplemented!("Compares corresponding elements in `a` and `b` to see if neither is `NaN`.");
|
||||
}
|
||||
3 => {
|
||||
unimplemented!("Compares corresponding elements in `a` and `b` to see if either is `NaN`.");
|
||||
}
|
||||
4 => FloatCC::NotEqual,
|
||||
5 => {
|
||||
unimplemented!("not less than");
|
||||
}
|
||||
6 => {
|
||||
unimplemented!("not less than or equal");
|
||||
}
|
||||
kind => unreachable!("kind {:?}", kind),
|
||||
};
|
||||
|
||||
simd_for_each_lane(fx, intrinsic, x, y, ret, |fx, lane_layout, res_lane_layout, x_lane, y_lane| {
|
||||
let res_lane = match lane_layout.ty.sty {
|
||||
ty::Float(_) => fx.bcx.ins().fcmp(flt_cc, x_lane, y_lane),
|
||||
_ => unreachable!("{:?}", lane_layout.ty),
|
||||
};
|
||||
bool_to_zero_or_max_uint(fx, res_lane_layout, res_lane)
|
||||
});
|
||||
};
|
||||
}
|
||||
|
||||
if let Some((_, dest)) = destination {
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue