Support 64-bit shifts on ARM.

This code generate different versions of assembly depending on the
amount to shift.

Variable Amount: 6 cycles (5 if last shift can be folded)
Constant 1     : 2 cycles
Constant 2-31  : 3 cycles (2 if last shift can foldable)
Constant 32    : 1 cycle  (depends on the register allocator)
Constant 33-64 : 2 cycles

This should speed up softfpu on arm a bit.

git-svn-id: trunk@21686 -
This commit is contained in:
masta 2012-06-23 20:36:27 +00:00
parent 3566956389
commit ca70207bc0

View File

@ -42,6 +42,10 @@ interface
procedure second_float;override;
end;
tarmshlshrnode = class(tcgshlshrnode)
procedure second_64bit;override;
function first_shlshr64bitint: tnode; override;
end;
implementation
@ -350,9 +354,142 @@ implementation
end;
end;
function tarmshlshrnode.first_shlshr64bitint: tnode;
begin
result := nil;
end;
procedure tarmshlshrnode.second_64bit;
var
hreg64hi,hreg64lo,shiftreg:Tregister;
v : TConstExprInt;
l1,l2,l3:Tasmlabel;
so: tshifterop;
procedure emit_instr(p: tai);
begin
current_asmdata.CurrAsmList.concat(p);
end;
{Reg1 gets shifted and moved into reg2, and is set to zero afterwards}
procedure shift_more_than_32(reg1, reg2: TRegister; shiftval: Byte ; sm: TShiftMode);
begin
shifterop_reset(so); so.shiftimm:=shiftval - 32; so.shiftmode:=sm;
emit_instr(taicpu.op_reg_reg_shifterop(A_MOV, reg2, reg1, so));
emit_instr(taicpu.op_reg_const(A_MOV, reg1, 0));
end;
procedure shift_less_than_32(reg1, reg2: TRegister; shiftval: Byte; shiftright: boolean);
begin
shifterop_reset(so); so.shiftimm:=shiftval;
if shiftright then so.shiftmode:=SM_LSR else so.shiftmode:=SM_LSL;
emit_instr(taicpu.op_reg_reg_shifterop(A_MOV, reg1, reg1, so));
if shiftright then so.shiftmode:=SM_LSL else so.shiftmode:=SM_LSR;
so.shiftimm:=32-shiftval;
emit_instr(taicpu.op_reg_reg_reg_shifterop(A_ORR, reg1, reg1, reg2, so));
if shiftright then so.shiftmode:=SM_LSR else so.shiftmode:=SM_LSL;
so.shiftimm:=shiftval;
emit_instr(taicpu.op_reg_reg_shifterop(A_MOV, reg2, reg2, so));
end;
procedure shift_by_variable(reg1, reg2, shiftval: TRegister; shiftright: boolean);
var
shiftval2:TRegister;
begin
shifterop_reset(so);
shiftval2:=cg.getintregister(current_asmdata.CurrAsmList,OS_INT);
{Do we shift more than 32 bits?}
emit_instr(setoppostfix(taicpu.op_reg_reg_const(A_RSB, shiftval2, shiftval, 32), PF_S));
{This part cares for 32 bits and more}
emit_instr(setcondition(taicpu.op_reg_reg_const(A_SUB, shiftval2, shiftval, 32), C_MI));
if shiftright then so.shiftmode:=SM_LSR else so.shiftmode:=SM_LSL;
so.rs:=shiftval2;
emit_instr(setcondition(taicpu.op_reg_reg_shifterop(A_MOV, reg2, reg1, so), C_MI));
{Less than 32 bits}
so.rs:=shiftval;
emit_instr(setcondition(taicpu.op_reg_reg_shifterop(A_MOV, reg2, reg2, so), C_PL));
if shiftright then so.shiftmode:=SM_LSL else so.shiftmode:=SM_LSR;
so.rs:=shiftval2;
emit_instr(setcondition(taicpu.op_reg_reg_reg_shifterop(A_ORR, reg2, reg2, reg1, so), C_PL));
{Final adjustments}
if shiftright then so.shiftmode:=SM_LSR else so.shiftmode:=SM_LSL;
so.rs:=shiftval;
emit_instr(taicpu.op_reg_reg_shifterop(A_MOV, reg1, reg1, so));
end;
begin
location_reset(location,LOC_REGISTER,def_cgsize(resultdef));
{ load left operator in a register }
hlcg.location_force_reg(current_asmdata.CurrAsmList,left.location,left.resultdef,resultdef,false);
hreg64hi:=left.location.register64.reghi;
hreg64lo:=left.location.register64.reglo;
location.register64.reghi:=hreg64hi;
location.register64.reglo:=hreg64lo;
{ shifting by a constant directly coded: }
if (right.nodetype=ordconstn) then
begin
v:=Tordconstnode(right).value and 63;
{Single bit shift}
if v = 1 then
if nodetype=shln then
begin
{Shift left by one by 2 simple 32bit additions}
emit_instr(setoppostfix(taicpu.op_reg_reg_reg(A_ADD, hreg64lo, hreg64lo, hreg64lo), PF_S));
emit_instr(taicpu.op_reg_reg_reg(A_ADC, hreg64hi, hreg64hi, hreg64hi));
end
else
begin
{Shift right by first shifting hi by one and then using RRX (rotate right extended), which rotates through the carry}
shifterop_reset(so); so.shiftmode:=SM_LSR; so.shiftimm:=1;
emit_instr(setoppostfix(taicpu.op_reg_reg_shifterop(A_MOV, hreg64hi, hreg64hi, so), PF_S));
so.shiftmode:=SM_RRX; so.shiftimm:=0; {RRX does NOT have a shift amount}
emit_instr(taicpu.op_reg_reg_shifterop(A_MOV, hreg64lo, hreg64lo, so));
end
{A 32bit shift just replaces a register and clears the other}
else if v = 32 then
begin
if nodetype=shln then
emit_instr(taicpu.op_reg_const(A_MOV, hreg64hi, 0))
else
emit_instr(taicpu.op_reg_const(A_MOV, hreg64lo, 0));
location.register64.reghi:=hreg64lo;
location.register64.reglo:=hreg64hi;
end
{Shift LESS than 32}
else if v < 32 then
if nodetype=shln then
shift_less_than_32(hreg64hi, hreg64lo, v.uvalue, false)
else
shift_less_than_32(hreg64lo, hreg64hi, v.uvalue, true)
{More than 32}
else
if nodetype=shln then
shift_more_than_32(hreg64lo, hreg64hi, v.uvalue, SM_LSL)
else
shift_more_than_32(hreg64hi, hreg64lo, v.uvalue, SM_LSR)
end
else
begin
{ force right operators in a register }
hlcg.location_force_reg(current_asmdata.CurrAsmList,right.location,right.resultdef,resultdef,false);
if nodetype = shln then
shift_by_variable(hreg64lo,hreg64hi,right.location.register, false)
else
shift_by_variable(hreg64hi,hreg64lo,right.location.register, true);
end;
end;
begin
cmoddivnode:=tarmmoddivnode;
cnotnode:=tarmnotnode;
cunaryminusnode:=tarmunaryminusnode;
cshlshrnode:=tarmshlshrnode;
end.