Enable more CLIF tests on AArch64
The tests for the SIMD floating-point maximum and minimum operations require particular care because the handling of the NaN values is non-deterministic and may vary between platforms. There is no way to match several NaN values in a test, so the solution is to extract the non-deterministic test cases into a separate file that is subsequently replicated for every backend under test, with adjustments made to the expected results. Copyright (c) 2021, Arm Limited.
This commit is contained in:
@@ -13,7 +13,7 @@ use crate::ir::Inst as IRInst;
|
||||
use crate::ir::{Opcode, Type};
|
||||
use crate::machinst::lower::*;
|
||||
use crate::machinst::*;
|
||||
use crate::CodegenResult;
|
||||
use crate::{CodegenError, CodegenResult};
|
||||
|
||||
use crate::isa::aarch64::inst::*;
|
||||
use crate::isa::aarch64::AArch64Backend;
|
||||
@@ -1103,50 +1103,96 @@ pub(crate) fn lower_vector_compare<C: LowerCtx<I = Inst>>(
|
||||
_ => false,
|
||||
};
|
||||
let size = VectorSize::from_ty(ty);
|
||||
// 'Less than' operations are implemented by swapping
|
||||
// the order of operands and using the 'greater than'
|
||||
// instructions.
|
||||
// 'Not equal' is implemented with 'equal' and inverting
|
||||
// the result.
|
||||
let (alu_op, swap) = match (is_float, cond) {
|
||||
(false, Cond::Eq) => (VecALUOp::Cmeq, false),
|
||||
(false, Cond::Ne) => (VecALUOp::Cmeq, false),
|
||||
(false, Cond::Ge) => (VecALUOp::Cmge, false),
|
||||
(false, Cond::Gt) => (VecALUOp::Cmgt, false),
|
||||
(false, Cond::Le) => (VecALUOp::Cmge, true),
|
||||
(false, Cond::Lt) => (VecALUOp::Cmgt, true),
|
||||
(false, Cond::Hs) => (VecALUOp::Cmhs, false),
|
||||
(false, Cond::Hi) => (VecALUOp::Cmhi, false),
|
||||
(false, Cond::Ls) => (VecALUOp::Cmhs, true),
|
||||
(false, Cond::Lo) => (VecALUOp::Cmhi, true),
|
||||
(true, Cond::Eq) => (VecALUOp::Fcmeq, false),
|
||||
(true, Cond::Ne) => (VecALUOp::Fcmeq, false),
|
||||
(true, Cond::Mi) => (VecALUOp::Fcmgt, true),
|
||||
(true, Cond::Ls) => (VecALUOp::Fcmge, true),
|
||||
(true, Cond::Ge) => (VecALUOp::Fcmge, false),
|
||||
(true, Cond::Gt) => (VecALUOp::Fcmgt, false),
|
||||
_ => unreachable!(),
|
||||
};
|
||||
|
||||
if swap {
|
||||
std::mem::swap(&mut rn, &mut rm);
|
||||
}
|
||||
if is_float && (cond == Cond::Vc || cond == Cond::Vs) {
|
||||
let tmp = ctx.alloc_tmp(ty).only_reg().unwrap();
|
||||
|
||||
ctx.emit(Inst::VecRRR {
|
||||
alu_op,
|
||||
rd,
|
||||
rn,
|
||||
rm,
|
||||
size,
|
||||
});
|
||||
|
||||
if cond == Cond::Ne {
|
||||
ctx.emit(Inst::VecMisc {
|
||||
op: VecMisc2::Not,
|
||||
ctx.emit(Inst::VecRRR {
|
||||
alu_op: VecALUOp::Fcmeq,
|
||||
rd,
|
||||
rn: rd.to_reg(),
|
||||
rn,
|
||||
rm: rn,
|
||||
size,
|
||||
});
|
||||
ctx.emit(Inst::VecRRR {
|
||||
alu_op: VecALUOp::Fcmeq,
|
||||
rd: tmp,
|
||||
rn: rm,
|
||||
rm,
|
||||
size,
|
||||
});
|
||||
ctx.emit(Inst::VecRRR {
|
||||
alu_op: VecALUOp::And,
|
||||
rd,
|
||||
rn: rd.to_reg(),
|
||||
rm: tmp.to_reg(),
|
||||
size,
|
||||
});
|
||||
|
||||
if cond == Cond::Vs {
|
||||
ctx.emit(Inst::VecMisc {
|
||||
op: VecMisc2::Not,
|
||||
rd,
|
||||
rn: rd.to_reg(),
|
||||
size,
|
||||
});
|
||||
}
|
||||
} else {
|
||||
// 'Less than' operations are implemented by swapping
|
||||
// the order of operands and using the 'greater than'
|
||||
// instructions.
|
||||
// 'Not equal' is implemented with 'equal' and inverting
|
||||
// the result.
|
||||
let (alu_op, swap) = match (is_float, cond) {
|
||||
(false, Cond::Eq) => (VecALUOp::Cmeq, false),
|
||||
(false, Cond::Ne) => (VecALUOp::Cmeq, false),
|
||||
(false, Cond::Ge) => (VecALUOp::Cmge, false),
|
||||
(false, Cond::Gt) => (VecALUOp::Cmgt, false),
|
||||
(false, Cond::Le) => (VecALUOp::Cmge, true),
|
||||
(false, Cond::Lt) => (VecALUOp::Cmgt, true),
|
||||
(false, Cond::Hs) => (VecALUOp::Cmhs, false),
|
||||
(false, Cond::Hi) => (VecALUOp::Cmhi, false),
|
||||
(false, Cond::Ls) => (VecALUOp::Cmhs, true),
|
||||
(false, Cond::Lo) => (VecALUOp::Cmhi, true),
|
||||
(true, Cond::Eq) => (VecALUOp::Fcmeq, false),
|
||||
(true, Cond::Ne) => (VecALUOp::Fcmeq, false),
|
||||
(true, Cond::Mi) => (VecALUOp::Fcmgt, true),
|
||||
(true, Cond::Ls) => (VecALUOp::Fcmge, true),
|
||||
(true, Cond::Ge) => (VecALUOp::Fcmge, false),
|
||||
(true, Cond::Gt) => (VecALUOp::Fcmgt, false),
|
||||
_ => {
|
||||
return Err(CodegenError::Unsupported(format!(
|
||||
"Unsupported {} SIMD vector comparison: {:?}",
|
||||
if is_float {
|
||||
"floating-point"
|
||||
} else {
|
||||
"integer"
|
||||
},
|
||||
cond
|
||||
)))
|
||||
}
|
||||
};
|
||||
|
||||
if swap {
|
||||
std::mem::swap(&mut rn, &mut rm);
|
||||
}
|
||||
|
||||
ctx.emit(Inst::VecRRR {
|
||||
alu_op,
|
||||
rd,
|
||||
rn,
|
||||
rm,
|
||||
size,
|
||||
});
|
||||
|
||||
if cond == Cond::Ne {
|
||||
ctx.emit(Inst::VecMisc {
|
||||
op: VecMisc2::Not,
|
||||
rd,
|
||||
rn: rd.to_reg(),
|
||||
size,
|
||||
});
|
||||
}
|
||||
}
|
||||
|
||||
Ok(())
|
||||
|
||||
@@ -1803,23 +1803,30 @@ pub(crate) fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
|
||||
}
|
||||
|
||||
Opcode::Bint => {
|
||||
let ty = ty.unwrap();
|
||||
|
||||
if ty.is_vector() {
|
||||
return Err(CodegenError::Unsupported(format!(
|
||||
"Bint: Unsupported type: {:?}",
|
||||
ty
|
||||
)));
|
||||
}
|
||||
|
||||
// Booleans are stored as all-zeroes (0) or all-ones (-1). We AND
|
||||
// out the LSB to give a 0 / 1-valued integer result.
|
||||
let rn = put_input_in_reg(ctx, inputs[0], NarrowValueMode::None);
|
||||
let rd = get_output_reg(ctx, outputs[0]).only_reg().unwrap();
|
||||
let output_bits = ty_bits(ctx.output_ty(insn, 0));
|
||||
let input = put_input_in_regs(ctx, inputs[0]);
|
||||
let output = get_output_reg(ctx, outputs[0]);
|
||||
|
||||
let (imm_ty, alu_op) = if output_bits > 32 {
|
||||
(I64, ALUOp::And64)
|
||||
} else {
|
||||
(I32, ALUOp::And32)
|
||||
};
|
||||
ctx.emit(Inst::AluRRImmLogic {
|
||||
alu_op,
|
||||
rd,
|
||||
rn,
|
||||
imml: ImmLogic::maybe_from_u64(1, imm_ty).unwrap(),
|
||||
alu_op: ALUOp::And32,
|
||||
rd: output.regs()[0],
|
||||
rn: input.regs()[0],
|
||||
imml: ImmLogic::maybe_from_u64(1, I32).unwrap(),
|
||||
});
|
||||
|
||||
if ty_bits(ty) > 64 {
|
||||
lower_constant_u64(ctx, output.regs()[1], 0);
|
||||
}
|
||||
}
|
||||
|
||||
Opcode::Bitcast => {
|
||||
@@ -2240,7 +2247,9 @@ pub(crate) fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
|
||||
}
|
||||
}
|
||||
|
||||
Opcode::VallTrue if ctx.input_ty(insn, 0) == I64X2 => {
|
||||
Opcode::VallTrue if ty_bits(ctx.input_ty(insn, 0).lane_type()) == 64 => {
|
||||
debug_assert!(ctx.input_ty(insn, 0).is_vector());
|
||||
|
||||
let rd = get_output_reg(ctx, outputs[0]).only_reg().unwrap();
|
||||
let rm = put_input_in_reg(ctx, inputs[0], NarrowValueMode::None);
|
||||
let tmp = ctx.alloc_tmp(I64X2).only_reg().unwrap();
|
||||
|
||||
Reference in New Issue
Block a user