Skip to content

Commit

Permalink
Rollup merge of rust-lang#99769 - bjorn3:sync_cg_clif-2022-07-26, r=b…
Browse files Browse the repository at this point in the history 
…jorn3

Sync rustc_codegen_cranelift

I did a large refactoring of the intrinsics module to remove the intrinsic_match macro which is not very clear to other people. This also enables rustfmt to run on this code. While I already did a sync yesterday, I am going to do another sync again to avoid potential conflicts as those will likely be painful to resolve.

r? `@ghost`

`@rustbot` label +A-codegen +A-cranelift +T-compiler
  • Loading branch information
@matthiaskrgr
matthiaskrgr authored Jul 26, 2022
2 parents 09eac4e + 30a5eb0 commit a8084f5
Show file tree
Hide file tree
Showing 3 changed files with 705 additions and 461 deletions.
145 changes: 81 additions & 64 deletions compiler/rustc_codegen_cranelift/src/intrinsics/llvm.rs
View file
Original file line number Diff line number Diff line change
Expand Up @@ -13,23 +13,20 @@ pub(crate) fn codegen_llvm_intrinsic_call<'tcx>(
ret: CPlace<'tcx>,
target: Option<BasicBlock>,
) {
intrinsic_match! {
fx, intrinsic, args,
_ => {
fx.tcx.sess.warn(&format!("unsupported llvm intrinsic {}; replacing with trap", intrinsic));
crate::trap::trap_unimplemented(fx, intrinsic);
};

match intrinsic {
// Used by `_mm_movemask_epi8` and `_mm256_movemask_epi8`
"llvm.x86.sse2.pmovmskb.128" | "llvm.x86.avx2.pmovmskb" | "llvm.x86.sse2.movmsk.pd", (c a) {
"llvm.x86.sse2.pmovmskb.128" | "llvm.x86.avx2.pmovmskb" | "llvm.x86.sse2.movmsk.pd" => {
intrinsic_args!(fx, args => (a); intrinsic);

let (lane_count, lane_ty) = a.layout().ty.simd_size_and_type(fx.tcx);
let lane_ty = fx.clif_type(lane_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 =
a.value_field(fx, mir::Field::new(lane.try_into().unwrap())).load_scalar(fx);

// cast float to int
let a_lane = match lane_ty {
Expand All @@ -49,26 +46,29 @@ pub(crate) fn codegen_llvm_intrinsic_call<'tcx>(

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 = crate::constant::mir_operand_get_const_val(fx, kind).expect("llvm.x86.sse2.cmp.* kind not const");
let flt_cc = match kind.try_to_bits(Size::from_bytes(1)).unwrap_or_else(|| panic!("kind not scalar: {:?}", kind)) {
}
"llvm.x86.sse2.cmp.ps" | "llvm.x86.sse2.cmp.pd" => {
let (x, y, kind) = match args {
[x, y, kind] => (x, y, kind),
_ => bug!("wrong number of args for intrinsic {intrinsic}"),
};
let x = codegen_operand(fx, x);
let y = codegen_operand(fx, y);
let kind = crate::constant::mir_operand_get_const_val(fx, kind)
.expect("llvm.x86.sse2.cmp.* kind not const");

let flt_cc = match kind
.try_to_bits(Size::from_bytes(1))
.unwrap_or_else(|| panic!("kind not scalar: {:?}", kind))
{
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`.");
}
7 => FloatCC::Ordered,
3 => FloatCC::Unordered,
4 => FloatCC::NotEqual,
5 => {
unimplemented!("not less than");
}
6 => {
unimplemented!("not less than or equal");
}
5 => FloatCC::UnorderedOrGreaterThanOrEqual,
6 => FloatCC::UnorderedOrGreaterThan,
kind => unreachable!("kind {:?}", kind),
};

Expand All @@ -79,50 +79,67 @@ pub(crate) fn codegen_llvm_intrinsic_call<'tcx>(
};
bool_to_zero_or_max_uint(fx, res_lane_ty, res_lane)
});
};
"llvm.x86.sse2.psrli.d", (c a, o imm8) {
let imm8 = crate::constant::mir_operand_get_const_val(fx, imm8).expect("llvm.x86.sse2.psrli.d imm8 not const");
simd_for_each_lane(fx, a, ret, &|fx, _lane_ty, _res_lane_ty, lane| {
match imm8.try_to_bits(Size::from_bytes(4)).unwrap_or_else(|| panic!("imm8 not scalar: {:?}", imm8)) {
imm8 if imm8 < 32 => fx.bcx.ins().ushr_imm(lane, i64::from(imm8 as u8)),
_ => fx.bcx.ins().iconst(types::I32, 0),
}
}
"llvm.x86.sse2.psrli.d" => {
let (a, imm8) = match args {
[a, imm8] => (a, imm8),
_ => bug!("wrong number of args for intrinsic {intrinsic}"),
};
let a = codegen_operand(fx, a);
let imm8 = crate::constant::mir_operand_get_const_val(fx, imm8)
.expect("llvm.x86.sse2.psrli.d imm8 not const");

simd_for_each_lane(fx, a, ret, &|fx, _lane_ty, _res_lane_ty, lane| match imm8
.try_to_bits(Size::from_bytes(4))
.unwrap_or_else(|| panic!("imm8 not scalar: {:?}", imm8))
{
imm8 if imm8 < 32 => fx.bcx.ins().ushr_imm(lane, i64::from(imm8 as u8)),
_ => fx.bcx.ins().iconst(types::I32, 0),
});
};
"llvm.x86.sse2.pslli.d", (c a, o imm8) {
let imm8 = crate::constant::mir_operand_get_const_val(fx, imm8).expect("llvm.x86.sse2.psrli.d imm8 not const");
simd_for_each_lane(fx, a, ret, &|fx, _lane_ty, _res_lane_ty, lane| {
match imm8.try_to_bits(Size::from_bytes(4)).unwrap_or_else(|| panic!("imm8 not scalar: {:?}", imm8)) {
imm8 if imm8 < 32 => fx.bcx.ins().ishl_imm(lane, i64::from(imm8 as u8)),
_ => fx.bcx.ins().iconst(types::I32, 0),
}
}
"llvm.x86.sse2.pslli.d" => {
let (a, imm8) = match args {
[a, imm8] => (a, imm8),
_ => bug!("wrong number of args for intrinsic {intrinsic}"),
};
let a = codegen_operand(fx, a);
let imm8 = crate::constant::mir_operand_get_const_val(fx, imm8)
.expect("llvm.x86.sse2.psrli.d imm8 not const");

simd_for_each_lane(fx, a, ret, &|fx, _lane_ty, _res_lane_ty, lane| match imm8
.try_to_bits(Size::from_bytes(4))
.unwrap_or_else(|| panic!("imm8 not scalar: {:?}", imm8))
{
imm8 if imm8 < 32 => fx.bcx.ins().ishl_imm(lane, i64::from(imm8 as u8)),
_ => fx.bcx.ins().iconst(types::I32, 0),
});
};
"llvm.x86.sse2.storeu.dq", (v mem_addr, c a) {
}
"llvm.x86.sse2.storeu.dq" => {
intrinsic_args!(fx, args => (mem_addr, a); intrinsic);
let mem_addr = mem_addr.load_scalar(fx);

// FIXME correctly handle the unalignment
let dest = CPlace::for_ptr(Pointer::new(mem_addr), a.layout());
dest.write_cvalue(fx, a);
};
"llvm.x86.addcarry.64", (v c_in, c a, c b) {
llvm_add_sub(
fx,
BinOp::Add,
ret,
c_in,
a,
b
);
};
"llvm.x86.subborrow.64", (v b_in, c a, c b) {
llvm_add_sub(
fx,
BinOp::Sub,
ret,
b_in,
a,
b
);
};
}
"llvm.x86.addcarry.64" => {
intrinsic_args!(fx, args => (c_in, a, b); intrinsic);
let c_in = c_in.load_scalar(fx);

llvm_add_sub(fx, BinOp::Add, ret, c_in, a, b);
}
"llvm.x86.subborrow.64" => {
intrinsic_args!(fx, args => (b_in, a, b); intrinsic);
let b_in = b_in.load_scalar(fx);

llvm_add_sub(fx, BinOp::Sub, ret, b_in, a, b);
}
_ => {
fx.tcx
.sess
.warn(&format!("unsupported llvm intrinsic {}; replacing with trap", intrinsic));
crate::trap::trap_unimplemented(fx, intrinsic);
}
}

let dest = target.expect("all llvm intrinsics used by stdlib should return");
Expand Down
Loading

0 comments on commit a8084f5

Please sign in to comment.