Merge commit 'f2cdd4a78d
' into sync_cg_clif-2022-04-22
This commit is contained in:
commit
b70b01b112
21 changed files with 189 additions and 142 deletions
|
@ -821,7 +821,8 @@ pub(crate) fn codegen_place<'tcx>(
|
|||
if cplace.layout().ty.is_box() {
|
||||
cplace = cplace
|
||||
.place_field(fx, Field::new(0)) // Box<T> -> Unique<T>
|
||||
.place_field(fx, Field::new(0)) // Unique<T> -> *const T
|
||||
.place_field(fx, Field::new(0)) // Unique<T> -> NonNull<T>
|
||||
.place_field(fx, Field::new(0)) // NonNull<T> -> *mut T
|
||||
.place_deref(fx);
|
||||
} else {
|
||||
cplace = cplace.place_deref(fx);
|
||||
|
|
|
@ -84,7 +84,7 @@ pub(crate) fn clif_int_or_float_cast(
|
|||
fx.bcx.ins().fcvt_from_uint(to_ty, from)
|
||||
}
|
||||
} else if from_ty.is_float() && to_ty.is_int() {
|
||||
if to_ty == types::I128 {
|
||||
let val = if to_ty == types::I128 {
|
||||
// _____sssf___
|
||||
// __fix sfti: f32 -> i128
|
||||
// __fix dfti: f64 -> i128
|
||||
|
@ -109,13 +109,9 @@ pub(crate) fn clif_int_or_float_cast(
|
|||
|
||||
let to_rust_ty = if to_signed { fx.tcx.types.i128 } else { fx.tcx.types.u128 };
|
||||
|
||||
return fx
|
||||
.easy_call(&name, &[CValue::by_val(from, fx.layout_of(from_rust_ty))], to_rust_ty)
|
||||
.load_scalar(fx);
|
||||
}
|
||||
|
||||
// float -> int-like
|
||||
if to_ty == types::I8 || to_ty == types::I16 {
|
||||
fx.easy_call(&name, &[CValue::by_val(from, fx.layout_of(from_rust_ty))], to_rust_ty)
|
||||
.load_scalar(fx)
|
||||
} else if to_ty == types::I8 || to_ty == types::I16 {
|
||||
// FIXME implement fcvt_to_*int_sat.i8/i16
|
||||
let val = if to_signed {
|
||||
fx.bcx.ins().fcvt_to_sint_sat(types::I32, from)
|
||||
|
@ -146,6 +142,23 @@ pub(crate) fn clif_int_or_float_cast(
|
|||
fx.bcx.ins().fcvt_to_sint_sat(to_ty, from)
|
||||
} else {
|
||||
fx.bcx.ins().fcvt_to_uint_sat(to_ty, from)
|
||||
};
|
||||
|
||||
if let Some(false) = fx.tcx.sess.opts.debugging_opts.saturating_float_casts {
|
||||
return val;
|
||||
}
|
||||
|
||||
let is_not_nan = fx.bcx.ins().fcmp(FloatCC::Equal, from, from);
|
||||
if to_ty == types::I128 {
|
||||
// FIXME(bytecodealliance/wasmtime#3963): select.i128 on fcmp eq miscompiles
|
||||
let (lsb, msb) = fx.bcx.ins().isplit(val);
|
||||
let zero = fx.bcx.ins().iconst(types::I64, 0);
|
||||
let lsb = fx.bcx.ins().select(is_not_nan, lsb, zero);
|
||||
let msb = fx.bcx.ins().select(is_not_nan, msb, zero);
|
||||
fx.bcx.ins().iconcat(lsb, msb)
|
||||
} else {
|
||||
let zero = fx.bcx.ins().iconst(to_ty, 0);
|
||||
fx.bcx.ins().select(is_not_nan, val, zero)
|
||||
}
|
||||
} else if from_ty.is_float() && to_ty.is_float() {
|
||||
// float -> float
|
||||
|
|
|
@ -48,12 +48,6 @@ pub struct BackendConfig {
|
|||
/// Can be set using `-Cllvm-args=display_cg_time=...`.
|
||||
pub display_cg_time: bool,
|
||||
|
||||
/// The register allocator to use.
|
||||
///
|
||||
/// Defaults to the value of `CG_CLIF_REGALLOC` or `backtracking` otherwise. Can be set using
|
||||
/// `-Cllvm-args=regalloc=...`.
|
||||
pub regalloc: String,
|
||||
|
||||
/// Enable the Cranelift ir verifier for all compilation passes. If not set it will only run
|
||||
/// once before passing the clif ir to Cranelift for compilation.
|
||||
///
|
||||
|
@ -80,8 +74,6 @@ impl Default for BackendConfig {
|
|||
args.split(' ').map(|arg| arg.to_string()).collect()
|
||||
},
|
||||
display_cg_time: bool_env_var("CG_CLIF_DISPLAY_CG_TIME"),
|
||||
regalloc: std::env::var("CG_CLIF_REGALLOC")
|
||||
.unwrap_or_else(|_| "backtracking".to_string()),
|
||||
enable_verifier: cfg!(debug_assertions) || bool_env_var("CG_CLIF_ENABLE_VERIFIER"),
|
||||
disable_incr_cache: bool_env_var("CG_CLIF_DISABLE_INCR_CACHE"),
|
||||
}
|
||||
|
@ -101,7 +93,6 @@ impl BackendConfig {
|
|||
match name {
|
||||
"mode" => config.codegen_mode = value.parse()?,
|
||||
"display_cg_time" => config.display_cg_time = parse_bool(name, value)?,
|
||||
"regalloc" => config.regalloc = value.to_string(),
|
||||
"enable_verifier" => config.enable_verifier = parse_bool(name, value)?,
|
||||
"disable_incr_cache" => config.disable_incr_cache = parse_bool(name, value)?,
|
||||
_ => return Err(format!("Unknown option `{}`", name)),
|
||||
|
|
|
@ -128,8 +128,16 @@ pub(crate) fn codegen_get_discriminant<'tcx>(
|
|||
let relative_discr = if niche_start == 0 {
|
||||
tag
|
||||
} else {
|
||||
// FIXME handle niche_start > i64::MAX
|
||||
fx.bcx.ins().iadd_imm(tag, -i64::try_from(niche_start).unwrap())
|
||||
let niche_start = match fx.bcx.func.dfg.value_type(tag) {
|
||||
types::I128 => {
|
||||
let lsb = fx.bcx.ins().iconst(types::I64, niche_start as u64 as i64);
|
||||
let msb =
|
||||
fx.bcx.ins().iconst(types::I64, (niche_start >> 64) as u64 as i64);
|
||||
fx.bcx.ins().iconcat(lsb, msb)
|
||||
}
|
||||
ty => fx.bcx.ins().iconst(ty, niche_start as i64),
|
||||
};
|
||||
fx.bcx.ins().isub(tag, niche_start)
|
||||
};
|
||||
let relative_max = niche_variants.end().as_u32() - niche_variants.start().as_u32();
|
||||
let is_niche = {
|
||||
|
|
|
@ -1019,39 +1019,23 @@ fn codegen_regular_intrinsic_call<'tcx>(
|
|||
ret.write_cvalue(fx, old);
|
||||
};
|
||||
|
||||
// In Rust floating point min and max don't propagate NaN. In Cranelift they do however.
|
||||
// For this reason it is necessary to use `a.is_nan() ? b : (a >= b ? b : a)` for `minnumf*`
|
||||
// and `a.is_nan() ? b : (a <= b ? b : a)` for `maxnumf*`. NaN checks are done by comparing
|
||||
// a float against itself. Only in case of NaN is it not equal to itself.
|
||||
minnumf32, (v a, v b) {
|
||||
let a_is_nan = fx.bcx.ins().fcmp(FloatCC::NotEqual, a, a);
|
||||
let a_ge_b = fx.bcx.ins().fcmp(FloatCC::GreaterThanOrEqual, a, b);
|
||||
let temp = fx.bcx.ins().select(a_ge_b, b, a);
|
||||
let val = fx.bcx.ins().select(a_is_nan, b, temp);
|
||||
let val = crate::num::codegen_float_min(fx, a, b);
|
||||
let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f32));
|
||||
ret.write_cvalue(fx, val);
|
||||
};
|
||||
minnumf64, (v a, v b) {
|
||||
let a_is_nan = fx.bcx.ins().fcmp(FloatCC::NotEqual, a, a);
|
||||
let a_ge_b = fx.bcx.ins().fcmp(FloatCC::GreaterThanOrEqual, a, b);
|
||||
let temp = fx.bcx.ins().select(a_ge_b, b, a);
|
||||
let val = fx.bcx.ins().select(a_is_nan, b, temp);
|
||||
let val = crate::num::codegen_float_min(fx, a, b);
|
||||
let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f64));
|
||||
ret.write_cvalue(fx, val);
|
||||
};
|
||||
maxnumf32, (v a, v b) {
|
||||
let a_is_nan = fx.bcx.ins().fcmp(FloatCC::NotEqual, a, a);
|
||||
let a_le_b = fx.bcx.ins().fcmp(FloatCC::LessThanOrEqual, a, b);
|
||||
let temp = fx.bcx.ins().select(a_le_b, b, a);
|
||||
let val = fx.bcx.ins().select(a_is_nan, b, temp);
|
||||
let val = crate::num::codegen_float_max(fx, a, b);
|
||||
let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f32));
|
||||
ret.write_cvalue(fx, val);
|
||||
};
|
||||
maxnumf64, (v a, v b) {
|
||||
let a_is_nan = fx.bcx.ins().fcmp(FloatCC::NotEqual, a, a);
|
||||
let a_le_b = fx.bcx.ins().fcmp(FloatCC::LessThanOrEqual, a, b);
|
||||
let temp = fx.bcx.ins().select(a_le_b, b, a);
|
||||
let val = fx.bcx.ins().select(a_is_nan, b, temp);
|
||||
let val = crate::num::codegen_float_max(fx, a, b);
|
||||
let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f64));
|
||||
ret.write_cvalue(fx, val);
|
||||
};
|
||||
|
|
|
@ -322,20 +322,21 @@ pub(super) fn codegen_simd_intrinsic_call<'tcx>(
|
|||
}
|
||||
assert_eq!(a.layout(), b.layout());
|
||||
assert_eq!(a.layout(), c.layout());
|
||||
let layout = a.layout();
|
||||
assert_eq!(a.layout(), ret.layout());
|
||||
|
||||
let (lane_count, _lane_ty) = layout.ty.simd_size_and_type(fx.tcx);
|
||||
let (ret_lane_count, ret_lane_ty) = ret.layout().ty.simd_size_and_type(fx.tcx);
|
||||
assert_eq!(lane_count, ret_lane_count);
|
||||
let ret_lane_layout = fx.layout_of(ret_lane_ty);
|
||||
let layout = a.layout();
|
||||
let (lane_count, lane_ty) = layout.ty.simd_size_and_type(fx.tcx);
|
||||
|
||||
for lane in 0..lane_count {
|
||||
let a_lane = a.value_lane(fx, lane).load_scalar(fx);
|
||||
let b_lane = b.value_lane(fx, lane).load_scalar(fx);
|
||||
let c_lane = c.value_lane(fx, lane).load_scalar(fx);
|
||||
let a_lane = a.value_lane(fx, lane);
|
||||
let b_lane = b.value_lane(fx, lane);
|
||||
let c_lane = c.value_lane(fx, lane);
|
||||
|
||||
let mul_lane = fx.bcx.ins().fmul(a_lane, b_lane);
|
||||
let res_lane = CValue::by_val(fx.bcx.ins().fadd(mul_lane, c_lane), ret_lane_layout);
|
||||
let res_lane = match lane_ty.kind() {
|
||||
ty::Float(FloatTy::F32) => fx.easy_call("fmaf", &[a_lane, b_lane, c_lane], lane_ty),
|
||||
ty::Float(FloatTy::F64) => fx.easy_call("fma", &[a_lane, b_lane, c_lane], lane_ty),
|
||||
_ => unreachable!(),
|
||||
};
|
||||
|
||||
ret.place_lane(fx, lane).write_cvalue(fx, res_lane);
|
||||
}
|
||||
|
@ -354,8 +355,8 @@ pub(super) fn codegen_simd_intrinsic_call<'tcx>(
|
|||
_ => unreachable!("{:?}", lane_ty),
|
||||
}
|
||||
match intrinsic {
|
||||
sym::simd_fmin => fx.bcx.ins().fmin(x_lane, y_lane),
|
||||
sym::simd_fmax => fx.bcx.ins().fmax(x_lane, y_lane),
|
||||
sym::simd_fmin => crate::num::codegen_float_min(fx, x_lane, y_lane),
|
||||
sym::simd_fmax => crate::num::codegen_float_max(fx, x_lane, y_lane),
|
||||
_ => unreachable!(),
|
||||
}
|
||||
});
|
||||
|
@ -495,7 +496,7 @@ pub(super) fn codegen_simd_intrinsic_call<'tcx>(
|
|||
let lt = match ty.kind() {
|
||||
ty::Int(_) => fx.bcx.ins().icmp(IntCC::SignedLessThan, a, b),
|
||||
ty::Uint(_) => fx.bcx.ins().icmp(IntCC::UnsignedLessThan, a, b),
|
||||
ty::Float(_) => fx.bcx.ins().fcmp(FloatCC::LessThan, a, b),
|
||||
ty::Float(_) => return crate::num::codegen_float_min(fx, a, b),
|
||||
_ => unreachable!(),
|
||||
};
|
||||
fx.bcx.ins().select(lt, a, b)
|
||||
|
@ -512,7 +513,7 @@ pub(super) fn codegen_simd_intrinsic_call<'tcx>(
|
|||
let gt = match ty.kind() {
|
||||
ty::Int(_) => fx.bcx.ins().icmp(IntCC::SignedGreaterThan, a, b),
|
||||
ty::Uint(_) => fx.bcx.ins().icmp(IntCC::UnsignedGreaterThan, a, b),
|
||||
ty::Float(_) => fx.bcx.ins().fcmp(FloatCC::GreaterThan, a, b),
|
||||
ty::Float(_) => return crate::num::codegen_float_max(fx, a, b),
|
||||
_ => unreachable!(),
|
||||
};
|
||||
fx.bcx.ins().select(gt, a, b)
|
||||
|
|
|
@ -256,8 +256,6 @@ fn build_isa(sess: &Session, backend_config: &BackendConfig) -> Box<dyn isa::Tar
|
|||
|
||||
flags_builder.set("enable_llvm_abi_extensions", "true").unwrap();
|
||||
|
||||
flags_builder.set("regalloc", &backend_config.regalloc).unwrap();
|
||||
|
||||
use rustc_session::config::OptLevel;
|
||||
match sess.opts.optimize {
|
||||
OptLevel::No => {
|
||||
|
|
|
@ -420,3 +420,21 @@ pub(crate) fn codegen_ptr_binop<'tcx>(
|
|||
CValue::by_val(fx.bcx.ins().bint(types::I8, res), fx.layout_of(fx.tcx.types.bool))
|
||||
}
|
||||
}
|
||||
|
||||
// In Rust floating point min and max don't propagate NaN. In Cranelift they do however.
|
||||
// For this reason it is necessary to use `a.is_nan() ? b : (a >= b ? b : a)` for `minnumf*`
|
||||
// and `a.is_nan() ? b : (a <= b ? b : a)` for `maxnumf*`. NaN checks are done by comparing
|
||||
// a float against itself. Only in case of NaN is it not equal to itself.
|
||||
pub(crate) fn codegen_float_min(fx: &mut FunctionCx<'_, '_, '_>, a: Value, b: Value) -> Value {
|
||||
let a_is_nan = fx.bcx.ins().fcmp(FloatCC::NotEqual, a, a);
|
||||
let a_ge_b = fx.bcx.ins().fcmp(FloatCC::GreaterThanOrEqual, a, b);
|
||||
let temp = fx.bcx.ins().select(a_ge_b, b, a);
|
||||
fx.bcx.ins().select(a_is_nan, b, temp)
|
||||
}
|
||||
|
||||
pub(crate) fn codegen_float_max(fx: &mut FunctionCx<'_, '_, '_>, a: Value, b: Value) -> Value {
|
||||
let a_is_nan = fx.bcx.ins().fcmp(FloatCC::NotEqual, a, a);
|
||||
let a_le_b = fx.bcx.ins().fcmp(FloatCC::LessThanOrEqual, a, b);
|
||||
let temp = fx.bcx.ins().select(a_le_b, b, a);
|
||||
fx.bcx.ins().select(a_is_nan, b, temp)
|
||||
}
|
||||
|
|
Loading…
Add table
Add a link
Reference in a new issue