Implement > and < intrinsics for vector registers

This commit is contained in:
Kovid Goyal 2024-02-01 08:19:34 +05:30
parent 82b7b4fcce
commit d60dacbd09
No known key found for this signature in database
GPG key ID: 06BC317B515ACE7C
3 changed files with 86 additions and 23 deletions

View file

@ -360,14 +360,12 @@ func (f *Function) Comment(x ...any) {
f.Instructions = append(f.Instructions, space_join("\t//", x...))
}
func shrn8b(r Register) int {
n, err := strconv.Atoi(r.Name[1:])
if err != nil {
panic(err)
}
n = 0x8400 + (n * 0x21)
return 0x0f0c0000 + n
func shrn8b_immediate4(a, b Register) uint32 {
return (0x0f0c84 << 8) | (a.ARMId()<<5 | b.ARMId())
}
func encode_cmgt16b(a, b, dest Register) (ans uint32) {
return 0x271<<21 | a.ARMId()<<16 | 0xd<<10 | b.ARMId()<<5 | dest.ARMId()
}
func (f *Function) CountBytesToFirstMatchDestructive(vec, ans Register) {
@ -377,7 +375,7 @@ func (f *Function) CountBytesToFirstMatchDestructive(vec, ans Register) {
// See https://community.arm.com/arm-community-blogs/b/infrastructure-solutions-blog/posts/porting-x86-vector-bitmask-optimizations-to-arm-neon
f.Comment("Go assembler doesn't support the shrn instruction, below we have: shrn.8b", vec, vec, "#4")
f.Comment("It is shifting right by four bits in every 16 bit word and truncating to 8 bits storing the result in the lower 64 bits of", vec)
f.instr("WORD", fmt.Sprintf("$0x%x", shrn8b(vec)))
f.instr("WORD", fmt.Sprintf("$0x%x", shrn8b_immediate4(vec, vec)))
f.instr("FMOVD", "F"+vec.Name[1:], ans)
f.AddTrailingComment("Extract the lower 64 bits from", vec, "and put them into", ans)
f.instr("RBIT", ans, ans)
@ -647,28 +645,53 @@ func (f *Function) SetRegsiterTo(self Register, val any) {
}
}
func (f *Function) CmpEqEpi8(a, b, ans Register) {
func (r Register) ARMId() uint32 {
num, err := strconv.Atoi(r.Name[1:])
if err != nil {
panic(err)
}
return uint32(num)
}
func (f *Function) cmp(a, b, ans Register, op, c_rep string) {
if a.Size != b.Size || a.Size != ans.Size {
panic("Can only compare registers of equal sizes")
}
if f.ISA.Goarch == ARM64 {
f.instr("VCMEQ", a.ARMFullWidth(), b.ARMFullWidth(), ans.ARMFullWidth())
if op == "EQ" {
f.instr("VCMEQ", a.ARMFullWidth(), b.ARMFullWidth(), ans.ARMFullWidth())
} else {
f.instr("WORD", fmt.Sprintf("$0x%x", encode_cmgt16b(a, b, ans)))
}
} else {
op := `PCMP` + op + "B"
if f.ISA.Bits == 128 {
switch ans.Name {
case a.Name:
f.instr("PCMPEQB", b, ans)
f.instr(op, b, ans)
case b.Name:
f.instr("PCMPEQB", a, ans)
f.instr(op, a, ans)
default:
f.CopyRegister(a, ans)
f.instr("PCMPEQB", b, ans)
f.instr(op, b, ans)
}
} else {
f.instr("VPCMPEQB", a, b, ans)
f.instr("V"+op, a, b, ans)
}
}
f.AddTrailingComment(ans, "= 0xff on every byte where", a.Name+"[n] == "+b.Name+"[n] and zero elsewhere")
f.AddTrailingComment(ans, "= 0xff on every byte where", a.Name+"[n]", c_rep, b.Name+"[n] and zero elsewhere")
}
func (f *Function) CmpGtEpi8(a, b, ans Register) {
f.cmp(a, b, ans, "GT", ">")
}
func (f *Function) CmpLtEpi8(a, b, ans Register) {
f.cmp(b, a, ans, "GT", "<")
}
func (f *Function) CmpEqEpi8(a, b, ans Register) {
f.cmp(a, b, ans, "EQ", "==")
}
func (f *Function) Set1Epi8FromParam(function_parameter string, vec Register) {
@ -1030,6 +1053,17 @@ func (s *State) test_cmpeq_epi8() {
f.store_vec_in_param(a, "ans")
}
func (s *State) test_cmplt_epi8() {
f := s.NewFunction("test_cmplt_epi8_asm", "Test byte comparison of two vectors", []FunctionParam{{"a", ByteSlice}, {"b", ByteSlice}, {"ans", ByteSlice}}, nil)
if !s.ISA.HasSIMD {
return
}
a := f.load_vec_from_param("a")
b := f.load_vec_from_param("b")
f.CmpLtEpi8(a, b, a)
f.store_vec_in_param(a, "ans")
}
func (s *State) test_or() {
f := s.NewFunction("test_or_asm", "Test OR of two vectors", []FunctionParam{{"a", ByteSlice}, {"b", ByteSlice}, {"ans", ByteSlice}}, nil)
if !s.ISA.HasSIMD {
@ -1175,6 +1209,7 @@ func (s *State) Generate() {
s.test_load()
s.test_set1_epi8()
s.test_cmpeq_epi8()
s.test_cmplt_epi8()
s.test_or()
s.test_jump_if_zero()
s.test_count_to_match()