T0b1/wasmtime
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

x64: port fcmp to ISLE (#3967)

Browse Source 
* x64: port scalar `fcmp` to ISLE

Implement the CLIF lowering for the `fcmp` to ISLE. This adds a new
type-matcher, `ty_scalar_float`, for detecting uses of `F32` and `F64`.

* isle: rename `vec128` to `ty_vec12`

This refactoring changes the name of the `vec128` matcher function to
follow the `ty_*` convention of the other type matchers. It also makes
the helper an inline function call.

* x64: port vector `fcmp` to ISLE
This commit is contained in:
Andrew Brown
2022-03-29 15:41:49 -07:00
committed by GitHub
parent 819b61b661
commit 5d8dd648d7
13 changed files with 746 additions and 400 deletions
Download Patch File Download Diff File Expand all files Collapse all files

99
cranelift/codegen/src/isa/x64/lower.rs
View File

@@ -930,104 +930,7 @@ fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
}
Opcode::Fcmp => {
let cond_code = ctx.data(insn).fp_cond_code().unwrap();
let input_ty = ctx.input_ty(insn, 0);
if !input_ty.is_vector() {
// Unordered is returned by setting ZF, PF, CF <- 111
// Greater than by ZF, PF, CF <- 000
// Less than by ZF, PF, CF <- 001
// Equal by ZF, PF, CF <- 100
//
// Checking the result of comiss is somewhat annoying because you don't have setcc
// instructions that explicitly check simultaneously for the condition (i.e. eq, le,
// gt, etc) *and* orderedness.
//
// So that might mean we need more than one setcc check and then a logical "and" or
// "or" to determine both, in some cases. However knowing that if the parity bit is
// set, then the result was considered unordered and knowing that if the parity bit is
// set, then both the ZF and CF flag bits must also be set we can get away with using
// one setcc for most condition codes.
let dst = get_output_reg(ctx, outputs[0]).only_reg().unwrap();
match emit_fcmp(ctx, insn, cond_code, FcmpSpec::Normal) {
FcmpCondResult::Condition(cc) => {
ctx.emit(Inst::setcc(cc, dst));
}
FcmpCondResult::AndConditions(cc1, cc2) => {
let tmp = ctx.alloc_tmp(types::I32).only_reg().unwrap();
ctx.emit(Inst::setcc(cc1, tmp));
ctx.emit(Inst::setcc(cc2, dst));
ctx.emit(Inst::alu_rmi_r(
OperandSize::Size32,
AluRmiROpcode::And,
RegMemImm::reg(tmp.to_reg()),
dst,
));
}
FcmpCondResult::OrConditions(cc1, cc2) => {
let tmp = ctx.alloc_tmp(types::I32).only_reg().unwrap();
ctx.emit(Inst::setcc(cc1, tmp));
ctx.emit(Inst::setcc(cc2, dst));
ctx.emit(Inst::alu_rmi_r(
OperandSize::Size32,
AluRmiROpcode::Or,
RegMemImm::reg(tmp.to_reg()),
dst,
));
}
FcmpCondResult::InvertedEqualOrConditions(_, _) => unreachable!(),
}
} else {
let op = match input_ty {
types::F32X4 => SseOpcode::Cmpps,
types::F64X2 => SseOpcode::Cmppd,
_ => panic!("Bad input type to fcmp: {}", input_ty),
};
// Since some packed comparisons are not available, some of the condition codes
// must be inverted, with a corresponding `flip` of the operands.
let (imm, flip) = match cond_code {
FloatCC::GreaterThan => (FcmpImm::LessThan, true),
FloatCC::GreaterThanOrEqual => (FcmpImm::LessThanOrEqual, true),
FloatCC::UnorderedOrLessThan => (FcmpImm::UnorderedOrGreaterThan, true),
FloatCC::UnorderedOrLessThanOrEqual => {
(FcmpImm::UnorderedOrGreaterThanOrEqual, true)
}
FloatCC::OrderedNotEqual | FloatCC::UnorderedOrEqual => {
panic!("unsupported float condition code: {}", cond_code)
}
_ => (FcmpImm::from(cond_code), false),
};
// Determine the operands of the comparison, possibly by flipping them.
let (lhs, rhs) = if flip {
(
put_input_in_reg(ctx, inputs[1]),
input_to_reg_mem(ctx, inputs[0]),
)
} else {
(
put_input_in_reg(ctx, inputs[0]),
input_to_reg_mem(ctx, inputs[1]),
)
};
// Move the `lhs` to the same register as `dst`; this may not emit an actual move
// but ensures that the registers are the same to match x86's read-write operand
// encoding.
let dst = get_output_reg(ctx, outputs[0]).only_reg().unwrap();
ctx.emit(Inst::gen_move(dst, lhs, input_ty));
// Emit the comparison.
ctx.emit(Inst::xmm_rm_r_imm(
op,
rhs,
dst,
imm.encode(),
OperandSize::Size32,
));
}
implemented_in_isle(ctx);
}
Opcode::FallthroughReturn | Opcode::Return => {