// zerorange clears the stack in the given range.
func zerorange(p *obj.Prog, frame int64, lo int64, hi int64) *obj.Prog {
cnt := hi - lo
if cnt == 0 {
return p
}
// Adjust the frame to account for LR.
frame += gc.Ctxt.FixedFrameSize()
offset := frame + lo
reg := int16(s390x.REGSP)
// If the offset cannot fit in a 12-bit unsigned displacement then we
// need to create a copy of the stack pointer that we can adjust.
// We also need to do this if we are going to loop.
if offset < 0 || offset > 4096-clearLoopCutoff || cnt > clearLoopCutoff {
p = appendpp(p, s390x.AADD, obj.TYPE_CONST, 0, offset, obj.TYPE_REG, s390x.REGRT1, 0)
p.Reg = int16(s390x.REGSP)
reg = s390x.REGRT1
offset = 0
}
// Generate a loop of large clears.
if cnt > clearLoopCutoff {
n := cnt - (cnt % 256)
end := int16(s390x.REGRT2)
p = appendpp(p, s390x.AADD, obj.TYPE_CONST, 0, offset+n, obj.TYPE_REG, end, 0)
p.Reg = reg
p = appendpp(p, s390x.AXC, obj.TYPE_MEM, reg, offset, obj.TYPE_MEM, reg, offset)
p.From3 = new(obj.Addr)
p.From3.Type = obj.TYPE_CONST
p.From3.Offset = 256
pl := p
p = appendpp(p, s390x.AADD, obj.TYPE_CONST, 0, 256, obj.TYPE_REG, reg, 0)
p = appendpp(p, s390x.ACMP, obj.TYPE_REG, reg, 0, obj.TYPE_REG, end, 0)
p = appendpp(p, s390x.ABNE, obj.TYPE_NONE, 0, 0, obj.TYPE_BRANCH, 0, 0)
gc.Patch(p, pl)
cnt -= n
}
// Generate remaining clear instructions without a loop.
for cnt > 0 {
n := cnt
// Can clear at most 256 bytes per instruction.
if n > 256 {
n = 256
}
switch n {
// Handle very small clears with move instructions.
case 8, 4, 2, 1:
ins := s390x.AMOVB
switch n {
case 8:
ins = s390x.AMOVD
case 4:
ins = s390x.AMOVW
case 2:
ins = s390x.AMOVH
}
p = appendpp(p, ins, obj.TYPE_CONST, 0, 0, obj.TYPE_MEM, reg, offset)
// Handle clears that would require multiple move instructions with XC.
default:
p = appendpp(p, s390x.AXC, obj.TYPE_MEM, reg, offset, obj.TYPE_MEM, reg, offset)
p.From3 = new(obj.Addr)
p.From3.Type = obj.TYPE_CONST
p.From3.Offset = n
}
cnt -= n
offset += n
}
return p
}
