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

Cranelift AArch64: Improve the handling of callee-saved registers

Browse Source 
SIMD & FP registers are now saved and restored in pairs, similarly
to general-purpose registers. Also, only the bottom 64 bits of the
registers are saved and restored (in case of non-Baldrdash ABIs),
which is the requirement from the Procedure Call Standard for the
Arm 64-bit Architecture.

As for the callee-saved general-purpose registers, if a procedure
needs to save and restore an odd number of them, it no longer uses
store and load pair instructions for the last register.

Copyright (c) 2021, Arm Limited.
This commit is contained in:
Anton Kirilov
2021-02-10 17:42:59 +00:00
parent 8387bc0d76
commit 7248abd591
6 changed files with 747 additions and 74 deletions
Download Patch File Download Diff File Expand all files Collapse all files

302
cranelift/codegen/src/isa/aarch64/abi.rs
View File

@@ -135,12 +135,28 @@ impl Into<AMode> for StackAMode {
// Returns the size of stack space needed to store the
// `int_reg` and `vec_reg`.
fn saved_reg_stack_size(
call_conv: isa::CallConv,
int_reg: &[Writable<RealReg>],
vec_reg: &[Writable<RealReg>],
) -> (usize, usize) {
// Round up to multiple of 2, to keep 16-byte stack alignment.
let int_save_bytes = (int_reg.len() + (int_reg.len() & 1)) * 8;
let vec_save_bytes = vec_reg.len() * 16;
// The Baldrdash ABIs require saving and restoring the whole 16-byte
// SIMD & FP registers, so the necessary stack space is always a
// multiple of the mandatory 16-byte stack alignment. However, the
// Procedure Call Standard for the Arm 64-bit Architecture (AAPCS64,
// including several related ABIs such as the one used by Windows)
// mandates saving only the bottom 8 bytes of the vector registers,
// so in that case we round up the number of registers to ensure proper
// stack alignment (similarly to the situation with `int_reg`).
let vec_reg_size = if call_conv.extends_baldrdash() { 16 } else { 8 };
let vec_save_padding = if call_conv.extends_baldrdash() {
0
} else {
vec_reg.len() & 1
};
let vec_save_bytes = (vec_reg.len() + vec_save_padding) * vec_reg_size;
(int_save_bytes, vec_save_bytes)
}
@@ -591,7 +607,8 @@ impl ABIMachineSpec for AArch64MachineDeps {
let mut insts = SmallVec::new();
let (clobbered_int, clobbered_vec) = get_regs_saved_in_prologue(call_conv, clobbers);
let (int_save_bytes, vec_save_bytes) = saved_reg_stack_size(&clobbered_int, &clobbered_vec);
let (int_save_bytes, vec_save_bytes) =
saved_reg_stack_size(call_conv, &clobbered_int, &clobbered_vec);
let total_save_bytes = int_save_bytes + vec_save_bytes;
let clobber_size = total_save_bytes as i32;
@@ -620,59 +637,170 @@ impl ABIMachineSpec for AArch64MachineDeps {
// `frame_offset` tracks offset above start-of-clobbers for unwind-info
// purposes.
let mut clobber_offset = clobber_size as u32;
for reg_pair in clobbered_int.chunks(2) {
let (r1, r2) = if reg_pair.len() == 2 {
// .to_reg().to_reg(): Writable<RealReg> --> RealReg --> Reg
(reg_pair[0].to_reg().to_reg(), reg_pair[1].to_reg().to_reg())
} else {
(reg_pair[0].to_reg().to_reg(), zero_reg())
};
let clobber_offset_change = 16;
let iter = clobbered_int.chunks_exact(2);
debug_assert!(r1.get_class() == RegClass::I64);
debug_assert!(r2.get_class() == RegClass::I64);
if let [rd] = iter.remainder() {
let rd = rd.to_reg().to_reg();
// stp r1, r2, [sp, #-16]!
insts.push(Inst::StoreP64 {
rt: r1,
rt2: r2,
mem: PairAMode::PreIndexed(
debug_assert_eq!(rd.get_class(), RegClass::I64);
// str rd, [sp, #-16]!
insts.push(Inst::Store64 {
rd,
mem: AMode::PreIndexed(
writable_stack_reg(),
SImm7Scaled::maybe_from_i64(-16, types::I64).unwrap(),
SImm9::maybe_from_i64(-clobber_offset_change).unwrap(),
),
flags: MemFlags::trusted(),
});
if flags.unwind_info() {
clobber_offset -= 8;
if r2 != zero_reg() {
insts.push(Inst::Unwind {
inst: UnwindInst::SaveReg {
clobber_offset,
reg: r2.to_real_reg(),
},
});
}
clobber_offset -= 8;
clobber_offset -= clobber_offset_change as u32;
insts.push(Inst::Unwind {
inst: UnwindInst::SaveReg {
clobber_offset,
reg: r1.to_real_reg(),
reg: rd.to_real_reg(),
},
});
}
}
for reg in clobbered_vec.iter() {
insts.push(Inst::FpuStore128 {
rd: reg.to_reg().to_reg(),
mem: AMode::PreIndexed(writable_stack_reg(), SImm9::maybe_from_i64(-16).unwrap()),
let mut iter = iter.rev();
while let Some([rt, rt2]) = iter.next() {
// .to_reg().to_reg(): Writable<RealReg> --> RealReg --> Reg
let rt = rt.to_reg().to_reg();
let rt2 = rt2.to_reg().to_reg();
debug_assert!(rt.get_class() == RegClass::I64);
debug_assert!(rt2.get_class() == RegClass::I64);
// stp rt, rt2, [sp, #-16]!
insts.push(Inst::StoreP64 {
rt,
rt2,
mem: PairAMode::PreIndexed(
writable_stack_reg(),
SImm7Scaled::maybe_from_i64(-clobber_offset_change, types::I64).unwrap(),
),
flags: MemFlags::trusted(),
});
if flags.unwind_info() {
clobber_offset -= 16;
clobber_offset -= clobber_offset_change as u32;
insts.push(Inst::Unwind {
inst: UnwindInst::SaveReg {
clobber_offset,
reg: reg.to_reg(),
reg: rt.to_real_reg(),
},
});
insts.push(Inst::Unwind {
inst: UnwindInst::SaveReg {
clobber_offset: clobber_offset + (clobber_offset_change / 2) as u32,
reg: rt2.to_real_reg(),
},
});
}
}
let store_vec_reg = |rd| {
if call_conv.extends_baldrdash() {
Inst::FpuStore128 {
rd,
mem: AMode::PreIndexed(
writable_stack_reg(),
SImm9::maybe_from_i64(-clobber_offset_change).unwrap(),
),
flags: MemFlags::trusted(),
}
} else {
Inst::FpuStore64 {
rd,
mem: AMode::PreIndexed(
writable_stack_reg(),
SImm9::maybe_from_i64(-clobber_offset_change).unwrap(),
),
flags: MemFlags::trusted(),
}
}
};
let iter = clobbered_vec.chunks_exact(2);
if let [rd] = iter.remainder() {
let rd = rd.to_reg().to_reg();
debug_assert_eq!(rd.get_class(), RegClass::V128);
insts.push(store_vec_reg(rd));
if flags.unwind_info() {
clobber_offset -= clobber_offset_change as u32;
insts.push(Inst::Unwind {
inst: UnwindInst::SaveReg {
clobber_offset,
reg: rd.to_real_reg(),
},
});
}
}
let store_vec_reg_pair = |rt, rt2| {
if call_conv.extends_baldrdash() {
let clobber_offset_change = 32;
(
Inst::FpuStoreP128 {
rt,
rt2,
mem: PairAMode::PreIndexed(
writable_stack_reg(),
SImm7Scaled::maybe_from_i64(-clobber_offset_change, I8X16).unwrap(),
),
flags: MemFlags::trusted(),
},
clobber_offset_change as u32,
)
} else {
let clobber_offset_change = 16;
(
Inst::FpuStoreP64 {
rt,
rt2,
mem: PairAMode::PreIndexed(
writable_stack_reg(),
SImm7Scaled::maybe_from_i64(-clobber_offset_change, F64).unwrap(),
),
flags: MemFlags::trusted(),
},
clobber_offset_change as u32,
)
}
};
let mut iter = iter.rev();
while let Some([rt, rt2]) = iter.next() {
let rt = rt.to_reg().to_reg();
let rt2 = rt2.to_reg().to_reg();
debug_assert_eq!(rt.get_class(), RegClass::V128);
debug_assert_eq!(rt2.get_class(), RegClass::V128);
let (inst, clobber_offset_change) = store_vec_reg_pair(rt, rt2);
insts.push(inst);
if flags.unwind_info() {
clobber_offset -= clobber_offset_change;
insts.push(Inst::Unwind {
inst: UnwindInst::SaveReg {
clobber_offset,
reg: rt.to_real_reg(),
},
});
insts.push(Inst::Unwind {
inst: UnwindInst::SaveReg {
clobber_offset: clobber_offset + clobber_offset_change / 2,
reg: rt2.to_real_reg(),
},
});
}
@@ -700,31 +828,83 @@ impl ABIMachineSpec for AArch64MachineDeps {
insts.extend(Self::gen_sp_reg_adjust(fixed_frame_storage_size as i32));
}
for reg in clobbered_vec.iter().rev() {
insts.push(Inst::FpuLoad128 {
rd: Writable::from_reg(reg.to_reg().to_reg()),
mem: AMode::PostIndexed(writable_stack_reg(), SImm9::maybe_from_i64(16).unwrap()),
flags: MemFlags::trusted(),
});
let load_vec_reg = |rd| {
if call_conv.extends_baldrdash() {
Inst::FpuLoad128 {
rd,
mem: AMode::PostIndexed(
writable_stack_reg(),
SImm9::maybe_from_i64(16).unwrap(),
),
flags: MemFlags::trusted(),
}
} else {
Inst::FpuLoad64 {
rd,
mem: AMode::PostIndexed(
writable_stack_reg(),
SImm9::maybe_from_i64(16).unwrap(),
),
flags: MemFlags::trusted(),
}
}
};
let load_vec_reg_pair = |rt, rt2| {
if call_conv.extends_baldrdash() {
Inst::FpuLoadP128 {
rt,
rt2,
mem: PairAMode::PostIndexed(
writable_stack_reg(),
SImm7Scaled::maybe_from_i64(32, I8X16).unwrap(),
),
flags: MemFlags::trusted(),
}
} else {
Inst::FpuLoadP64 {
rt,
rt2,
mem: PairAMode::PostIndexed(
writable_stack_reg(),
SImm7Scaled::maybe_from_i64(16, F64).unwrap(),
),
flags: MemFlags::trusted(),
}
}
};
let mut iter = clobbered_vec.chunks_exact(2);
while let Some([rt, rt2]) = iter.next() {
let rt = rt.map(|r| r.to_reg());
let rt2 = rt2.map(|r| r.to_reg());
debug_assert_eq!(rt.to_reg().get_class(), RegClass::V128);
debug_assert_eq!(rt2.to_reg().get_class(), RegClass::V128);
insts.push(load_vec_reg_pair(rt, rt2));
}
for reg_pair in clobbered_int.chunks(2).rev() {
let (r1, r2) = if reg_pair.len() == 2 {
(
reg_pair[0].map(|r| r.to_reg()),
reg_pair[1].map(|r| r.to_reg()),
)
} else {
(reg_pair[0].map(|r| r.to_reg()), writable_zero_reg())
};
debug_assert!(iter.remainder().len() <= 1);
debug_assert!(r1.to_reg().get_class() == RegClass::I64);
debug_assert!(r2.to_reg().get_class() == RegClass::I64);
if let [rd] = iter.remainder() {
let rd = rd.map(|r| r.to_reg());
// ldp r1, r2, [sp], #16
debug_assert_eq!(rd.to_reg().get_class(), RegClass::V128);
insts.push(load_vec_reg(rd));
}
let mut iter = clobbered_int.chunks_exact(2);
while let Some([rt, rt2]) = iter.next() {
let rt = rt.map(|r| r.to_reg());
let rt2 = rt2.map(|r| r.to_reg());
debug_assert_eq!(rt.to_reg().get_class(), RegClass::I64);
debug_assert_eq!(rt2.to_reg().get_class(), RegClass::I64);
// ldp rt, rt2, [sp], #16
insts.push(Inst::LoadP64 {
rt: r1,
rt2: r2,
rt,
rt2,
mem: PairAMode::PostIndexed(
writable_stack_reg(),
SImm7Scaled::maybe_from_i64(16, I64).unwrap(),
@@ -733,6 +913,20 @@ impl ABIMachineSpec for AArch64MachineDeps {
});
}
debug_assert!(iter.remainder().len() <= 1);
if let [rd] = iter.remainder() {
let rd = rd.map(|r| r.to_reg());
debug_assert_eq!(rd.to_reg().get_class(), RegClass::I64);
// ldr rd, [sp], #16
insts.push(Inst::ULoad64 {
rd,
mem: AMode::PostIndexed(writable_stack_reg(), SImm9::maybe_from_i64(16).unwrap()),
flags: MemFlags::trusted(),
});
}
// If this is Baldrdash-2020, restore the callee (i.e., our) TLS
// register. We may have allocated it for something else and clobbered
// it, but the ABI expects us to leave the TLS register unchanged.