1 // Copyright 2015 The Go Authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style
3 // license that can be found in the LICENSE file.
8 "cmd/compile/internal/logopt"
9 "cmd/compile/internal/types"
11 "cmd/internal/obj/s390x"
23 type deadValueChoice bool
26 leaveDeadValues deadValueChoice = false
27 removeDeadValues = true
30 // deadcode indicates that rewrite should try to remove any values that become dead.
31 func applyRewrite(f *Func, rb blockRewriter, rv valueRewriter, deadcode deadValueChoice) {
32 // repeat rewrites until we find no more rewrites
33 pendingLines := f.cachedLineStarts // Holds statement boundaries that need to be moved to a new value/block
37 fmt.Printf("%s: rewriting for %s\n", f.pass.name, f.Name)
41 for _, b := range f.Blocks {
46 b0.Succs = append([]Edge{}, b.Succs...) // make a new copy, not aliasing
48 for i, c := range b.ControlValues() {
51 b.ReplaceControl(i, c)
57 fmt.Printf("rewriting %s -> %s\n", b0.LongString(), b.LongString())
60 for j, v := range b.Values {
65 v0.Args = append([]*Value{}, v.Args...) // make a new copy, not aliasing
67 if v.Uses == 0 && v.removeable() {
68 if v.Op != OpInvalid && deadcode == removeDeadValues {
69 // Reset any values that are now unused, so that we decrement
70 // the use count of all of its arguments.
71 // Not quite a deadcode pass, because it does not handle cycles.
72 // But it should help Uses==1 rules to fire.
76 // No point rewriting values which aren't used.
80 vchange := phielimValue(v)
81 if vchange && debug > 1 {
82 fmt.Printf("rewriting %s -> %s\n", v0.LongString(), v.LongString())
85 // Eliminate copy inputs.
86 // If any copy input becomes unused, mark it
87 // as invalid and discard its argument. Repeat
88 // recursively on the discarded argument.
89 // This phase helps remove phantom "dead copy" uses
90 // of a value so that a x.Uses==1 rule condition
92 for i, a := range v.Args {
98 // If a, a copy, has a line boundary indicator, attempt to find a new value
99 // to hold it. The first candidate is the value that will replace a (aa),
100 // if it shares the same block and line and is eligible.
101 // The second option is v, which has a as an input. Because aa is earlier in
102 // the data flow, it is the better choice.
103 if a.Pos.IsStmt() == src.PosIsStmt {
104 if aa.Block == a.Block && aa.Pos.Line() == a.Pos.Line() && aa.Pos.IsStmt() != src.PosNotStmt {
105 aa.Pos = aa.Pos.WithIsStmt()
106 } else if v.Block == a.Block && v.Pos.Line() == a.Pos.Line() && v.Pos.IsStmt() != src.PosNotStmt {
107 v.Pos = v.Pos.WithIsStmt()
109 // Record the lost line and look for a new home after all rewrites are complete.
110 // TODO: it's possible (in FOR loops, in particular) for statement boundaries for the same
111 // line to appear in more than one block, but only one block is stored, so if both end
112 // up here, then one will be lost.
113 pendingLines.set(a.Pos, int32(a.Block.ID))
115 a.Pos = a.Pos.WithNotStmt()
124 if vchange && debug > 1 {
125 fmt.Printf("rewriting %s -> %s\n", v0.LongString(), v.LongString())
128 // apply rewrite function
131 // If value changed to a poor choice for a statement boundary, move the boundary
132 if v.Pos.IsStmt() == src.PosIsStmt {
133 if k := nextGoodStatementIndex(v, j, b); k != j {
134 v.Pos = v.Pos.WithNotStmt()
135 b.Values[k].Pos = b.Values[k].Pos.WithIsStmt()
140 change = change || vchange
141 if vchange && debug > 1 {
142 fmt.Printf("rewriting %s -> %s\n", v0.LongString(), v.LongString())
150 // remove clobbered values
151 for _, b := range f.Blocks {
153 for i, v := range b.Values {
155 if v.Op == OpInvalid {
156 if v.Pos.IsStmt() == src.PosIsStmt {
157 pendingLines.set(vl, int32(b.ID))
162 if v.Pos.IsStmt() != src.PosNotStmt && pendingLines.get(vl) == int32(b.ID) {
163 pendingLines.remove(vl)
164 v.Pos = v.Pos.WithIsStmt()
171 if pendingLines.get(b.Pos) == int32(b.ID) {
172 b.Pos = b.Pos.WithIsStmt()
173 pendingLines.remove(b.Pos)
179 // Common functions called from rewriting rules
181 func is64BitFloat(t *types.Type) bool {
182 return t.Size() == 8 && t.IsFloat()
185 func is32BitFloat(t *types.Type) bool {
186 return t.Size() == 4 && t.IsFloat()
189 func is64BitInt(t *types.Type) bool {
190 return t.Size() == 8 && t.IsInteger()
193 func is32BitInt(t *types.Type) bool {
194 return t.Size() == 4 && t.IsInteger()
197 func is16BitInt(t *types.Type) bool {
198 return t.Size() == 2 && t.IsInteger()
201 func is8BitInt(t *types.Type) bool {
202 return t.Size() == 1 && t.IsInteger()
205 func isPtr(t *types.Type) bool {
206 return t.IsPtrShaped()
209 func isSigned(t *types.Type) bool {
213 // mergeSym merges two symbolic offsets. There is no real merging of
214 // offsets, we just pick the non-nil one.
215 func mergeSym(x, y Sym) Sym {
222 panic(fmt.Sprintf("mergeSym with two non-nil syms %v %v", x, y))
225 func canMergeSym(x, y Sym) bool {
226 return x == nil || y == nil
229 // canMergeLoadClobber reports whether the load can be merged into target without
230 // invalidating the schedule.
231 // It also checks that the other non-load argument x is something we
232 // are ok with clobbering.
233 func canMergeLoadClobber(target, load, x *Value) bool {
234 // The register containing x is going to get clobbered.
235 // Don't merge if we still need the value of x.
236 // We don't have liveness information here, but we can
237 // approximate x dying with:
238 // 1) target is x's only use.
239 // 2) target is not in a deeper loop than x.
243 loopnest := x.Block.Func.loopnest()
244 loopnest.calculateDepths()
245 if loopnest.depth(target.Block.ID) > loopnest.depth(x.Block.ID) {
248 return canMergeLoad(target, load)
251 // canMergeLoad reports whether the load can be merged into target without
252 // invalidating the schedule.
253 func canMergeLoad(target, load *Value) bool {
254 if target.Block.ID != load.Block.ID {
255 // If the load is in a different block do not merge it.
259 // We can't merge the load into the target if the load
260 // has more than one use.
265 mem := load.MemoryArg()
267 // We need the load's memory arg to still be alive at target. That
268 // can't be the case if one of target's args depends on a memory
269 // state that is a successor of load's memory arg.
271 // For example, it would be invalid to merge load into target in
272 // the following situation because newmem has killed oldmem
273 // before target is reached:
274 // load = read ... oldmem
275 // newmem = write ... oldmem
276 // arg0 = read ... newmem
277 // target = add arg0 load
279 // If the argument comes from a different block then we can exclude
280 // it immediately because it must dominate load (which is in the
281 // same block as target).
283 for _, a := range target.Args {
284 if a != load && a.Block.ID == target.Block.ID {
285 args = append(args, a)
289 // memPreds contains memory states known to be predecessors of load's
290 // memory state. It is lazily initialized.
291 var memPreds map[*Value]bool
292 for i := 0; len(args) > 0; i++ {
295 // Give up if we have done a lot of iterations.
298 v := args[len(args)-1]
299 args = args[:len(args)-1]
300 if target.Block.ID != v.Block.ID {
301 // Since target and load are in the same block
302 // we can stop searching when we leave the block.
306 // A Phi implies we have reached the top of the block.
307 // The memory phi, if it exists, is always
308 // the first logical store in the block.
311 if v.Type.IsTuple() && v.Type.FieldType(1).IsMemory() {
312 // We could handle this situation however it is likely
316 if v.Op.SymEffect()&SymAddr != 0 {
317 // This case prevents an operation that calculates the
318 // address of a local variable from being forced to schedule
319 // before its corresponding VarDef.
325 // We don't want to combine the CMPQ with the load, because
326 // that would force the CMPQ to schedule before the VARDEF, which
327 // in turn requires the LEAQ to schedule before the VARDEF.
330 if v.Type.IsMemory() {
332 // Initialise a map containing memory states
333 // known to be predecessors of load's memory
335 memPreds = make(map[*Value]bool)
338 for i := 0; i < limit; i++ {
340 // The memory phi, if it exists, is always
341 // the first logical store in the block.
344 if m.Block.ID != target.Block.ID {
347 if !m.Type.IsMemory() {
351 if len(m.Args) == 0 {
358 // We can merge if v is a predecessor of mem.
360 // For example, we can merge load into target in the
361 // following scenario:
364 // load = read ... mem
365 // target = add x load
371 if len(v.Args) > 0 && v.Args[len(v.Args)-1] == mem {
372 // If v takes mem as an input then we know mem
373 // is valid at this point.
376 for _, a := range v.Args {
377 if target.Block.ID == a.Block.ID {
378 args = append(args, a)
386 // isSameCall reports whether sym is the same as the given named symbol
387 func isSameCall(sym interface{}, name string) bool {
388 fn := sym.(*AuxCall).Fn
389 return fn != nil && fn.String() == name
392 // nlz returns the number of leading zeros.
393 func nlz64(x int64) int { return bits.LeadingZeros64(uint64(x)) }
394 func nlz32(x int32) int { return bits.LeadingZeros32(uint32(x)) }
395 func nlz16(x int16) int { return bits.LeadingZeros16(uint16(x)) }
396 func nlz8(x int8) int { return bits.LeadingZeros8(uint8(x)) }
398 // ntzX returns the number of trailing zeros.
399 func ntz64(x int64) int { return bits.TrailingZeros64(uint64(x)) }
400 func ntz32(x int32) int { return bits.TrailingZeros32(uint32(x)) }
401 func ntz16(x int16) int { return bits.TrailingZeros16(uint16(x)) }
402 func ntz8(x int8) int { return bits.TrailingZeros8(uint8(x)) }
404 func oneBit(x int64) bool { return x&(x-1) == 0 && x != 0 }
405 func oneBit8(x int8) bool { return x&(x-1) == 0 && x != 0 }
406 func oneBit16(x int16) bool { return x&(x-1) == 0 && x != 0 }
407 func oneBit32(x int32) bool { return x&(x-1) == 0 && x != 0 }
408 func oneBit64(x int64) bool { return x&(x-1) == 0 && x != 0 }
410 // nto returns the number of trailing ones.
411 func nto(x int64) int64 {
412 return int64(ntz64(^x))
415 // logX returns logarithm of n base 2.
416 // n must be a positive power of 2 (isPowerOfTwoX returns true).
417 func log8(n int8) int64 {
418 return int64(bits.Len8(uint8(n))) - 1
420 func log16(n int16) int64 {
421 return int64(bits.Len16(uint16(n))) - 1
423 func log32(n int32) int64 {
424 return int64(bits.Len32(uint32(n))) - 1
426 func log64(n int64) int64 {
427 return int64(bits.Len64(uint64(n))) - 1
430 // log2uint32 returns logarithm in base 2 of uint32(n), with log2(0) = -1.
432 func log2uint32(n int64) int64 {
433 return int64(bits.Len32(uint32(n))) - 1
436 // isPowerOfTwo functions report whether n is a power of 2.
437 func isPowerOfTwo8(n int8) bool {
438 return n > 0 && n&(n-1) == 0
440 func isPowerOfTwo16(n int16) bool {
441 return n > 0 && n&(n-1) == 0
443 func isPowerOfTwo32(n int32) bool {
444 return n > 0 && n&(n-1) == 0
446 func isPowerOfTwo64(n int64) bool {
447 return n > 0 && n&(n-1) == 0
450 // isUint64PowerOfTwo reports whether uint64(n) is a power of 2.
451 func isUint64PowerOfTwo(in int64) bool {
453 return n > 0 && n&(n-1) == 0
456 // isUint32PowerOfTwo reports whether uint32(n) is a power of 2.
457 func isUint32PowerOfTwo(in int64) bool {
458 n := uint64(uint32(in))
459 return n > 0 && n&(n-1) == 0
462 // is32Bit reports whether n can be represented as a signed 32 bit integer.
463 func is32Bit(n int64) bool {
464 return n == int64(int32(n))
467 // is16Bit reports whether n can be represented as a signed 16 bit integer.
468 func is16Bit(n int64) bool {
469 return n == int64(int16(n))
472 // is8Bit reports whether n can be represented as a signed 8 bit integer.
473 func is8Bit(n int64) bool {
474 return n == int64(int8(n))
477 // isU8Bit reports whether n can be represented as an unsigned 8 bit integer.
478 func isU8Bit(n int64) bool {
479 return n == int64(uint8(n))
482 // isU12Bit reports whether n can be represented as an unsigned 12 bit integer.
483 func isU12Bit(n int64) bool {
484 return 0 <= n && n < (1<<12)
487 // isU16Bit reports whether n can be represented as an unsigned 16 bit integer.
488 func isU16Bit(n int64) bool {
489 return n == int64(uint16(n))
492 // isU32Bit reports whether n can be represented as an unsigned 32 bit integer.
493 func isU32Bit(n int64) bool {
494 return n == int64(uint32(n))
497 // is20Bit reports whether n can be represented as a signed 20 bit integer.
498 func is20Bit(n int64) bool {
499 return -(1<<19) <= n && n < (1<<19)
502 // b2i translates a boolean value to 0 or 1 for assigning to auxInt.
503 func b2i(b bool) int64 {
510 // b2i32 translates a boolean value to 0 or 1.
511 func b2i32(b bool) int32 {
518 // shiftIsBounded reports whether (left/right) shift Value v is known to be bounded.
519 // A shift is bounded if it is shifting by less than the width of the shifted value.
520 func shiftIsBounded(v *Value) bool {
524 // canonLessThan returns whether x is "ordered" less than y, for purposes of normalizing
525 // generated code as much as possible.
526 func canonLessThan(x, y *Value) bool {
530 if !x.Pos.SameFileAndLine(y.Pos) {
531 return x.Pos.Before(y.Pos)
536 // truncate64Fto32F converts a float64 value to a float32 preserving the bit pattern
537 // of the mantissa. It will panic if the truncation results in lost information.
538 func truncate64Fto32F(f float64) float32 {
539 if !isExactFloat32(f) {
540 panic("truncate64Fto32F: truncation is not exact")
545 // NaN bit patterns aren't necessarily preserved across conversion
546 // instructions so we need to do the conversion manually.
547 b := math.Float64bits(f)
548 m := b & ((1 << 52) - 1) // mantissa (a.k.a. significand)
549 // | sign | exponent | mantissa |
550 r := uint32(((b >> 32) & (1 << 31)) | 0x7f800000 | (m >> (52 - 23)))
551 return math.Float32frombits(r)
554 // extend32Fto64F converts a float32 value to a float64 value preserving the bit
555 // pattern of the mantissa.
556 func extend32Fto64F(f float32) float64 {
557 if !math.IsNaN(float64(f)) {
560 // NaN bit patterns aren't necessarily preserved across conversion
561 // instructions so we need to do the conversion manually.
562 b := uint64(math.Float32bits(f))
563 // | sign | exponent | mantissa |
564 r := ((b << 32) & (1 << 63)) | (0x7ff << 52) | ((b & 0x7fffff) << (52 - 23))
565 return math.Float64frombits(r)
568 // DivisionNeedsFixUp reports whether the division needs fix-up code.
569 func DivisionNeedsFixUp(v *Value) bool {
573 // auxFrom64F encodes a float64 value so it can be stored in an AuxInt.
574 func auxFrom64F(f float64) int64 {
576 panic("can't encode a NaN in AuxInt field")
578 return int64(math.Float64bits(f))
581 // auxFrom32F encodes a float32 value so it can be stored in an AuxInt.
582 func auxFrom32F(f float32) int64 {
584 panic("can't encode a NaN in AuxInt field")
586 return int64(math.Float64bits(extend32Fto64F(f)))
589 // auxTo32F decodes a float32 from the AuxInt value provided.
590 func auxTo32F(i int64) float32 {
591 return truncate64Fto32F(math.Float64frombits(uint64(i)))
594 // auxTo64F decodes a float64 from the AuxInt value provided.
595 func auxTo64F(i int64) float64 {
596 return math.Float64frombits(uint64(i))
599 func auxIntToBool(i int64) bool {
605 func auxIntToInt8(i int64) int8 {
608 func auxIntToInt16(i int64) int16 {
611 func auxIntToInt32(i int64) int32 {
614 func auxIntToInt64(i int64) int64 {
617 func auxIntToUint8(i int64) uint8 {
620 func auxIntToFloat32(i int64) float32 {
621 return float32(math.Float64frombits(uint64(i)))
623 func auxIntToFloat64(i int64) float64 {
624 return math.Float64frombits(uint64(i))
626 func auxIntToValAndOff(i int64) ValAndOff {
629 func auxIntToArm64BitField(i int64) arm64BitField {
630 return arm64BitField(i)
632 func auxIntToInt128(x int64) int128 {
634 panic("nonzero int128 not allowed")
638 func auxIntToFlagConstant(x int64) flagConstant {
639 return flagConstant(x)
642 func auxIntToOp(cc int64) Op {
646 func boolToAuxInt(b bool) int64 {
652 func int8ToAuxInt(i int8) int64 {
655 func int16ToAuxInt(i int16) int64 {
658 func int32ToAuxInt(i int32) int64 {
661 func int64ToAuxInt(i int64) int64 {
664 func uint8ToAuxInt(i uint8) int64 {
665 return int64(int8(i))
667 func float32ToAuxInt(f float32) int64 {
668 return int64(math.Float64bits(float64(f)))
670 func float64ToAuxInt(f float64) int64 {
671 return int64(math.Float64bits(f))
673 func valAndOffToAuxInt(v ValAndOff) int64 {
676 func arm64BitFieldToAuxInt(v arm64BitField) int64 {
679 func int128ToAuxInt(x int128) int64 {
681 panic("nonzero int128 not allowed")
685 func flagConstantToAuxInt(x flagConstant) int64 {
689 func opToAuxInt(o Op) int64 {
693 // Aux is an interface to hold miscellaneous data in Blocks and Values.
698 // stringAux wraps string values for use in Aux.
699 type stringAux string
701 func (stringAux) CanBeAnSSAAux() {}
703 func auxToString(i Aux) string {
704 return string(i.(stringAux))
706 func auxToSym(i Aux) Sym {
707 // TODO: kind of a hack - allows nil interface through
711 func auxToType(i Aux) *types.Type {
712 return i.(*types.Type)
714 func auxToCall(i Aux) *AuxCall {
717 func auxToS390xCCMask(i Aux) s390x.CCMask {
718 return i.(s390x.CCMask)
720 func auxToS390xRotateParams(i Aux) s390x.RotateParams {
721 return i.(s390x.RotateParams)
724 func StringToAux(s string) Aux {
727 func symToAux(s Sym) Aux {
730 func callToAux(s *AuxCall) Aux {
733 func typeToAux(t *types.Type) Aux {
736 func s390xCCMaskToAux(c s390x.CCMask) Aux {
739 func s390xRotateParamsToAux(r s390x.RotateParams) Aux {
743 // uaddOvf reports whether unsigned a+b would overflow.
744 func uaddOvf(a, b int64) bool {
745 return uint64(a)+uint64(b) < uint64(a)
748 // de-virtualize an InterCall
749 // 'sym' is the symbol for the itab
750 func devirt(v *Value, aux Aux, sym Sym, offset int64) *AuxCall {
752 n, ok := sym.(*obj.LSym)
756 lsym := f.fe.DerefItab(n, offset)
757 if f.pass.debug > 0 {
759 f.Warnl(v.Pos, "de-virtualizing call")
761 f.Warnl(v.Pos, "couldn't de-virtualize call")
768 return StaticAuxCall(lsym, va.args, va.results)
771 // de-virtualize an InterLECall
772 // 'sym' is the symbol for the itab
773 func devirtLESym(v *Value, aux Aux, sym Sym, offset int64) *obj.LSym {
774 n, ok := sym.(*obj.LSym)
780 lsym := f.fe.DerefItab(n, offset)
781 if f.pass.debug > 0 {
783 f.Warnl(v.Pos, "de-virtualizing call")
785 f.Warnl(v.Pos, "couldn't de-virtualize call")
794 func devirtLECall(v *Value, sym *obj.LSym) *Value {
795 v.Op = OpStaticLECall
796 v.Aux.(*AuxCall).Fn = sym
801 // isSamePtr reports whether p1 and p2 point to the same address.
802 func isSamePtr(p1, p2 *Value) bool {
811 return p1.AuxInt == p2.AuxInt && isSamePtr(p1.Args[0], p2.Args[0])
812 case OpAddr, OpLocalAddr:
813 // OpAddr's 0th arg is either OpSP or OpSB, which means that it is uniquely identified by its Op.
814 // Checking for value equality only works after [z]cse has run.
815 return p1.Aux == p2.Aux && p1.Args[0].Op == p2.Args[0].Op
817 return p1.Args[1] == p2.Args[1] && isSamePtr(p1.Args[0], p2.Args[0])
822 func isStackPtr(v *Value) bool {
823 for v.Op == OpOffPtr || v.Op == OpAddPtr {
826 return v.Op == OpSP || v.Op == OpLocalAddr
829 // disjoint reports whether the memory region specified by [p1:p1+n1)
830 // does not overlap with [p2:p2+n2).
831 // A return value of false does not imply the regions overlap.
832 func disjoint(p1 *Value, n1 int64, p2 *Value, n2 int64) bool {
833 if n1 == 0 || n2 == 0 {
839 baseAndOffset := func(ptr *Value) (base *Value, offset int64) {
840 base, offset = ptr, 0
841 for base.Op == OpOffPtr {
842 offset += base.AuxInt
847 p1, off1 := baseAndOffset(p1)
848 p2, off2 := baseAndOffset(p2)
849 if isSamePtr(p1, p2) {
850 return !overlap(off1, n1, off2, n2)
852 // p1 and p2 are not the same, so if they are both OpAddrs then
853 // they point to different variables.
854 // If one pointer is on the stack and the other is an argument
855 // then they can't overlap.
857 case OpAddr, OpLocalAddr:
858 if p2.Op == OpAddr || p2.Op == OpLocalAddr || p2.Op == OpSP {
861 return p2.Op == OpArg && p1.Args[0].Op == OpSP
863 if p2.Op == OpSP || p2.Op == OpLocalAddr {
867 return p2.Op == OpAddr || p2.Op == OpLocalAddr || p2.Op == OpArg || p2.Op == OpSP
872 // moveSize returns the number of bytes an aligned MOV instruction moves
873 func moveSize(align int64, c *Config) int64 {
875 case align%8 == 0 && c.PtrSize == 8:
885 // mergePoint finds a block among a's blocks which dominates b and is itself
886 // dominated by all of a's blocks. Returns nil if it can't find one.
887 // Might return nil even if one does exist.
888 func mergePoint(b *Block, a ...*Value) *Block {
889 // Walk backward from b looking for one of the a's blocks.
895 for _, x := range a {
900 if len(b.Preds) > 1 {
901 // Don't know which way to go back. Abort.
907 return nil // too far away
909 // At this point, r is the first value in a that we find by walking backwards.
910 // if we return anything, r will be it.
913 // Keep going, counting the other a's that we find. They must all dominate r.
916 for _, x := range a {
922 // Found all of a in a backwards walk. We can return r.
925 if len(b.Preds) > 1 {
932 return nil // too far away
935 // clobber invalidates values. Returns true.
936 // clobber is used by rewrite rules to:
937 // A) make sure the values are really dead and never used again.
938 // B) decrement use counts of the values' args.
939 func clobber(vv ...*Value) bool {
940 for _, v := range vv {
942 // Note: leave v.Block intact. The Block field is used after clobber.
947 // clobberIfDead resets v when use count is 1. Returns true.
948 // clobberIfDead is used by rewrite rules to decrement
949 // use counts of v's args when v is dead and never used.
950 func clobberIfDead(v *Value) bool {
954 // Note: leave v.Block intact. The Block field is used after clobberIfDead.
958 // noteRule is an easy way to track if a rule is matched when writing
959 // new ones. Make the rule of interest also conditional on
960 // noteRule("note to self: rule of interest matched")
961 // and that message will print when the rule matches.
962 func noteRule(s string) bool {
967 // countRule increments Func.ruleMatches[key].
968 // If Func.ruleMatches is non-nil at the end
969 // of compilation, it will be printed to stdout.
970 // This is intended to make it easier to find which functions
971 // which contain lots of rules matches when developing new rules.
972 func countRule(v *Value, key string) bool {
974 if f.ruleMatches == nil {
975 f.ruleMatches = make(map[string]int)
981 // warnRule generates compiler debug output with string s when
982 // v is not in autogenerated code, cond is true and the rule has fired.
983 func warnRule(cond bool, v *Value, s string) bool {
984 if pos := v.Pos; pos.Line() > 1 && cond {
985 v.Block.Func.Warnl(pos, s)
990 // for a pseudo-op like (LessThan x), extract x
991 func flagArg(v *Value) *Value {
992 if len(v.Args) != 1 || !v.Args[0].Type.IsFlags() {
998 // arm64Negate finds the complement to an ARM64 condition code,
999 // for example Equal -> NotEqual or LessThan -> GreaterEqual
1001 // TODO: add floating-point conditions
1002 func arm64Negate(op Op) Op {
1004 case OpARM64LessThan:
1005 return OpARM64GreaterEqual
1006 case OpARM64LessThanU:
1007 return OpARM64GreaterEqualU
1008 case OpARM64GreaterThan:
1009 return OpARM64LessEqual
1010 case OpARM64GreaterThanU:
1011 return OpARM64LessEqualU
1012 case OpARM64LessEqual:
1013 return OpARM64GreaterThan
1014 case OpARM64LessEqualU:
1015 return OpARM64GreaterThanU
1016 case OpARM64GreaterEqual:
1017 return OpARM64LessThan
1018 case OpARM64GreaterEqualU:
1019 return OpARM64LessThanU
1021 return OpARM64NotEqual
1022 case OpARM64NotEqual:
1024 case OpARM64LessThanF:
1025 return OpARM64GreaterEqualF
1026 case OpARM64GreaterThanF:
1027 return OpARM64LessEqualF
1028 case OpARM64LessEqualF:
1029 return OpARM64GreaterThanF
1030 case OpARM64GreaterEqualF:
1031 return OpARM64LessThanF
1033 panic("unreachable")
1037 // arm64Invert evaluates (InvertFlags op), which
1038 // is the same as altering the condition codes such
1039 // that the same result would be produced if the arguments
1040 // to the flag-generating instruction were reversed, e.g.
1041 // (InvertFlags (CMP x y)) -> (CMP y x)
1043 // TODO: add floating-point conditions
1044 func arm64Invert(op Op) Op {
1046 case OpARM64LessThan:
1047 return OpARM64GreaterThan
1048 case OpARM64LessThanU:
1049 return OpARM64GreaterThanU
1050 case OpARM64GreaterThan:
1051 return OpARM64LessThan
1052 case OpARM64GreaterThanU:
1053 return OpARM64LessThanU
1054 case OpARM64LessEqual:
1055 return OpARM64GreaterEqual
1056 case OpARM64LessEqualU:
1057 return OpARM64GreaterEqualU
1058 case OpARM64GreaterEqual:
1059 return OpARM64LessEqual
1060 case OpARM64GreaterEqualU:
1061 return OpARM64LessEqualU
1062 case OpARM64Equal, OpARM64NotEqual:
1064 case OpARM64LessThanF:
1065 return OpARM64GreaterThanF
1066 case OpARM64GreaterThanF:
1067 return OpARM64LessThanF
1068 case OpARM64LessEqualF:
1069 return OpARM64GreaterEqualF
1070 case OpARM64GreaterEqualF:
1071 return OpARM64LessEqualF
1073 panic("unreachable")
1077 // evaluate an ARM64 op against a flags value
1078 // that is potentially constant; return 1 for true,
1079 // -1 for false, and 0 for not constant.
1080 func ccARM64Eval(op Op, flags *Value) int {
1082 if fop == OpARM64InvertFlags {
1083 return -ccARM64Eval(op, flags.Args[0])
1085 if fop != OpARM64FlagConstant {
1088 fc := flagConstant(flags.AuxInt)
1089 b2i := func(b bool) int {
1098 case OpARM64NotEqual:
1100 case OpARM64LessThan:
1102 case OpARM64LessThanU:
1103 return b2i(fc.ult())
1104 case OpARM64GreaterThan:
1106 case OpARM64GreaterThanU:
1107 return b2i(fc.ugt())
1108 case OpARM64LessEqual:
1110 case OpARM64LessEqualU:
1111 return b2i(fc.ule())
1112 case OpARM64GreaterEqual:
1114 case OpARM64GreaterEqualU:
1115 return b2i(fc.uge())
1120 // logRule logs the use of the rule s. This will only be enabled if
1121 // rewrite rules were generated with the -log option, see gen/rulegen.go.
1122 func logRule(s string) {
1123 if ruleFile == nil {
1124 // Open a log file to write log to. We open in append
1125 // mode because all.bash runs the compiler lots of times,
1126 // and we want the concatenation of all of those logs.
1127 // This means, of course, that users need to rm the old log
1128 // to get fresh data.
1129 // TODO: all.bash runs compilers in parallel. Need to synchronize logging somehow?
1130 w, err := os.OpenFile(filepath.Join(os.Getenv("GOROOT"), "src", "rulelog"),
1131 os.O_CREATE|os.O_WRONLY|os.O_APPEND, 0666)
1137 _, err := fmt.Fprintln(ruleFile, s)
1143 var ruleFile io.Writer
1145 func min(x, y int64) int64 {
1152 func isConstZero(v *Value) bool {
1156 case OpConst64, OpConst32, OpConst16, OpConst8, OpConstBool, OpConst32F, OpConst64F:
1157 return v.AuxInt == 0
1162 // reciprocalExact64 reports whether 1/c is exactly representable.
1163 func reciprocalExact64(c float64) bool {
1164 b := math.Float64bits(c)
1165 man := b & (1<<52 - 1)
1167 return false // not a power of 2, denormal, or NaN
1169 exp := b >> 52 & (1<<11 - 1)
1170 // exponent bias is 0x3ff. So taking the reciprocal of a number
1171 // changes the exponent to 0x7fe-exp.
1176 return false // ±inf
1178 return false // exponent is not representable
1184 // reciprocalExact32 reports whether 1/c is exactly representable.
1185 func reciprocalExact32(c float32) bool {
1186 b := math.Float32bits(c)
1187 man := b & (1<<23 - 1)
1189 return false // not a power of 2, denormal, or NaN
1191 exp := b >> 23 & (1<<8 - 1)
1192 // exponent bias is 0x7f. So taking the reciprocal of a number
1193 // changes the exponent to 0xfe-exp.
1198 return false // ±inf
1200 return false // exponent is not representable
1206 // check if an immediate can be directly encoded into an ARM's instruction
1207 func isARMImmRot(v uint32) bool {
1208 for i := 0; i < 16; i++ {
1218 // overlap reports whether the ranges given by the given offset and
1219 // size pairs overlap.
1220 func overlap(offset1, size1, offset2, size2 int64) bool {
1221 if offset1 >= offset2 && offset2+size2 > offset1 {
1224 if offset2 >= offset1 && offset1+size1 > offset2 {
1230 func areAdjacentOffsets(off1, off2, size int64) bool {
1231 return off1+size == off2 || off1 == off2+size
1234 // check if value zeroes out upper 32-bit of 64-bit register.
1235 // depth limits recursion depth. In AMD64.rules 3 is used as limit,
1236 // because it catches same amount of cases as 4.
1237 func zeroUpper32Bits(x *Value, depth int) bool {
1239 case OpAMD64MOVLconst, OpAMD64MOVLload, OpAMD64MOVLQZX, OpAMD64MOVLloadidx1,
1240 OpAMD64MOVWload, OpAMD64MOVWloadidx1, OpAMD64MOVBload, OpAMD64MOVBloadidx1,
1241 OpAMD64MOVLloadidx4, OpAMD64ADDLload, OpAMD64SUBLload, OpAMD64ANDLload,
1242 OpAMD64ORLload, OpAMD64XORLload, OpAMD64CVTTSD2SL,
1243 OpAMD64ADDL, OpAMD64ADDLconst, OpAMD64SUBL, OpAMD64SUBLconst,
1244 OpAMD64ANDL, OpAMD64ANDLconst, OpAMD64ORL, OpAMD64ORLconst,
1245 OpAMD64XORL, OpAMD64XORLconst, OpAMD64NEGL, OpAMD64NOTL,
1246 OpAMD64SHRL, OpAMD64SHRLconst, OpAMD64SARL, OpAMD64SARLconst,
1247 OpAMD64SHLL, OpAMD64SHLLconst:
1250 return x.Type.Width == 4
1251 case OpPhi, OpSelect0, OpSelect1:
1252 // Phis can use each-other as an arguments, instead of tracking visited values,
1253 // just limit recursion depth.
1257 for i := range x.Args {
1258 if !zeroUpper32Bits(x.Args[i], depth-1) {
1268 // zeroUpper48Bits is similar to zeroUpper32Bits, but for upper 48 bits
1269 func zeroUpper48Bits(x *Value, depth int) bool {
1271 case OpAMD64MOVWQZX, OpAMD64MOVWload, OpAMD64MOVWloadidx1, OpAMD64MOVWloadidx2:
1274 return x.Type.Width == 2
1275 case OpPhi, OpSelect0, OpSelect1:
1276 // Phis can use each-other as an arguments, instead of tracking visited values,
1277 // just limit recursion depth.
1281 for i := range x.Args {
1282 if !zeroUpper48Bits(x.Args[i], depth-1) {
1292 // zeroUpper56Bits is similar to zeroUpper32Bits, but for upper 56 bits
1293 func zeroUpper56Bits(x *Value, depth int) bool {
1295 case OpAMD64MOVBQZX, OpAMD64MOVBload, OpAMD64MOVBloadidx1:
1298 return x.Type.Width == 1
1299 case OpPhi, OpSelect0, OpSelect1:
1300 // Phis can use each-other as an arguments, instead of tracking visited values,
1301 // just limit recursion depth.
1305 for i := range x.Args {
1306 if !zeroUpper56Bits(x.Args[i], depth-1) {
1316 // isInlinableMemmove reports whether the given arch performs a Move of the given size
1317 // faster than memmove. It will only return true if replacing the memmove with a Move is
1318 // safe, either because Move is small or because the arguments are disjoint.
1319 // This is used as a check for replacing memmove with Move ops.
1320 func isInlinableMemmove(dst, src *Value, sz int64, c *Config) bool {
1321 // It is always safe to convert memmove into Move when its arguments are disjoint.
1322 // Move ops may or may not be faster for large sizes depending on how the platform
1323 // lowers them, so we only perform this optimization on platforms that we know to
1324 // have fast Move ops.
1327 return sz <= 16 || (sz < 1024 && disjoint(dst, sz, src, sz))
1328 case "386", "arm64":
1330 case "s390x", "ppc64", "ppc64le":
1331 return sz <= 8 || disjoint(dst, sz, src, sz)
1332 case "arm", "mips", "mips64", "mipsle", "mips64le":
1338 // logLargeCopy logs the occurrence of a large copy.
1339 // The best place to do this is in the rewrite rules where the size of the move is easy to find.
1340 // "Large" is arbitrarily chosen to be 128 bytes; this may change.
1341 func logLargeCopy(v *Value, s int64) bool {
1345 if logopt.Enabled() {
1346 logopt.LogOpt(v.Pos, "copy", "lower", v.Block.Func.Name, fmt.Sprintf("%d bytes", s))
1351 // hasSmallRotate reports whether the architecture has rotate instructions
1352 // for sizes < 32-bit. This is used to decide whether to promote some rotations.
1353 func hasSmallRotate(c *Config) bool {
1355 case "amd64", "386":
1362 func newPPC64ShiftAuxInt(sh, mb, me, sz int64) int32 {
1363 if sh < 0 || sh >= sz {
1364 panic("PPC64 shift arg sh out of range")
1366 if mb < 0 || mb >= sz {
1367 panic("PPC64 shift arg mb out of range")
1369 if me < 0 || me >= sz {
1370 panic("PPC64 shift arg me out of range")
1372 return int32(sh<<16 | mb<<8 | me)
1375 func GetPPC64Shiftsh(auxint int64) int64 {
1376 return int64(int8(auxint >> 16))
1379 func GetPPC64Shiftmb(auxint int64) int64 {
1380 return int64(int8(auxint >> 8))
1383 func GetPPC64Shiftme(auxint int64) int64 {
1384 return int64(int8(auxint))
1387 // Test if this value can encoded as a mask for a rlwinm like
1388 // operation. Masks can also extend from the msb and wrap to
1389 // the lsb too. That is, the valid masks are 32 bit strings
1390 // of the form: 0..01..10..0 or 1..10..01..1 or 1...1
1391 func isPPC64WordRotateMask(v64 int64) bool {
1392 // Isolate rightmost 1 (if none 0) and add.
1395 // Likewise, for the wrapping case.
1397 vpn := (vn & -vn) + vn
1398 return (v&vp == 0 || vn&vpn == 0) && v != 0
1401 // Compress mask and and shift into single value of the form
1402 // me | mb<<8 | rotate<<16 | nbits<<24 where me and mb can
1403 // be used to regenerate the input mask.
1404 func encodePPC64RotateMask(rotate, mask, nbits int64) int64 {
1405 var mb, me, mbn, men int
1407 // Determine boundaries and then decode them
1408 if mask == 0 || ^mask == 0 || rotate >= nbits {
1409 panic("Invalid PPC64 rotate mask")
1410 } else if nbits == 32 {
1411 mb = bits.LeadingZeros32(uint32(mask))
1412 me = 32 - bits.TrailingZeros32(uint32(mask))
1413 mbn = bits.LeadingZeros32(^uint32(mask))
1414 men = 32 - bits.TrailingZeros32(^uint32(mask))
1416 mb = bits.LeadingZeros64(uint64(mask))
1417 me = 64 - bits.TrailingZeros64(uint64(mask))
1418 mbn = bits.LeadingZeros64(^uint64(mask))
1419 men = 64 - bits.TrailingZeros64(^uint64(mask))
1421 // Check for a wrapping mask (e.g bits at 0 and 63)
1422 if mb == 0 && me == int(nbits) {
1423 // swap the inverted values
1427 return int64(me) | int64(mb<<8) | int64(rotate<<16) | int64(nbits<<24)
1430 // The inverse operation of encodePPC64RotateMask. The values returned as
1431 // mb and me satisfy the POWER ISA definition of MASK(x,y) where MASK(mb,me) = mask.
1432 func DecodePPC64RotateMask(sauxint int64) (rotate, mb, me int64, mask uint64) {
1433 auxint := uint64(sauxint)
1434 rotate = int64((auxint >> 16) & 0xFF)
1435 mb = int64((auxint >> 8) & 0xFF)
1436 me = int64((auxint >> 0) & 0xFF)
1437 nbits := int64((auxint >> 24) & 0xFF)
1438 mask = ((1 << uint(nbits-mb)) - 1) ^ ((1 << uint(nbits-me)) - 1)
1443 mask = uint64(uint32(mask))
1446 // Fixup ME to match ISA definition. The second argument to MASK(..,me)
1448 me = (me - 1) & (nbits - 1)
1452 // This verifies that the mask is a set of
1453 // consecutive bits including the least
1455 func isPPC64ValidShiftMask(v int64) bool {
1456 if (v != 0) && ((v+1)&v) == 0 {
1462 func getPPC64ShiftMaskLength(v int64) int64 {
1463 return int64(bits.Len64(uint64(v)))
1466 // Decompose a shift right into an equivalent rotate/mask,
1467 // and return mask & m.
1468 func mergePPC64RShiftMask(m, s, nbits int64) int64 {
1469 smask := uint64((1<<uint(nbits))-1) >> uint(s)
1470 return m & int64(smask)
1473 // Combine (ANDconst [m] (SRWconst [s])) into (RLWINM [y]) or return 0
1474 func mergePPC64AndSrwi(m, s int64) int64 {
1475 mask := mergePPC64RShiftMask(m, s, 32)
1476 if !isPPC64WordRotateMask(mask) {
1479 return encodePPC64RotateMask(32-s, mask, 32)
1482 // Test if a shift right feeding into a CLRLSLDI can be merged into RLWINM.
1483 // Return the encoded RLWINM constant, or 0 if they cannot be merged.
1484 func mergePPC64ClrlsldiSrw(sld, srw int64) int64 {
1485 mask_1 := uint64(0xFFFFFFFF >> uint(srw))
1486 // for CLRLSLDI, it's more convient to think of it as a mask left bits then rotate left.
1487 mask_2 := uint64(0xFFFFFFFFFFFFFFFF) >> uint(GetPPC64Shiftmb(int64(sld)))
1489 // Rewrite mask to apply after the final left shift.
1490 mask_3 := (mask_1 & mask_2) << uint(GetPPC64Shiftsh(sld))
1493 r_2 := GetPPC64Shiftsh(sld)
1494 r_3 := (r_1 + r_2) & 31 // This can wrap.
1496 if uint64(uint32(mask_3)) != mask_3 || mask_3 == 0 {
1499 return encodePPC64RotateMask(int64(r_3), int64(mask_3), 32)
1502 // Test if a RLWINM feeding into a CLRLSLDI can be merged into RLWINM. Return
1503 // the encoded RLWINM constant, or 0 if they cannot be merged.
1504 func mergePPC64ClrlsldiRlwinm(sld int32, rlw int64) int64 {
1505 r_1, _, _, mask_1 := DecodePPC64RotateMask(rlw)
1506 // for CLRLSLDI, it's more convient to think of it as a mask left bits then rotate left.
1507 mask_2 := uint64(0xFFFFFFFFFFFFFFFF) >> uint(GetPPC64Shiftmb(int64(sld)))
1509 // combine the masks, and adjust for the final left shift.
1510 mask_3 := (mask_1 & mask_2) << uint(GetPPC64Shiftsh(int64(sld)))
1511 r_2 := GetPPC64Shiftsh(int64(sld))
1512 r_3 := (r_1 + r_2) & 31 // This can wrap.
1514 // Verify the result is still a valid bitmask of <= 32 bits.
1515 if !isPPC64WordRotateMask(int64(mask_3)) || uint64(uint32(mask_3)) != mask_3 {
1518 return encodePPC64RotateMask(r_3, int64(mask_3), 32)
1521 // Compute the encoded RLWINM constant from combining (SLDconst [sld] (SRWconst [srw] x)),
1522 // or return 0 if they cannot be combined.
1523 func mergePPC64SldiSrw(sld, srw int64) int64 {
1524 if sld > srw || srw >= 32 {
1527 mask_r := uint32(0xFFFFFFFF) >> uint(srw)
1528 mask_l := uint32(0xFFFFFFFF) >> uint(sld)
1529 mask := (mask_r & mask_l) << uint(sld)
1530 return encodePPC64RotateMask((32-srw+sld)&31, int64(mask), 32)
1533 // Convenience function to rotate a 32 bit constant value by another constant.
1534 func rotateLeft32(v, rotate int64) int64 {
1535 return int64(bits.RotateLeft32(uint32(v), int(rotate)))
1538 // encodes the lsb and width for arm(64) bitfield ops into the expected auxInt format.
1539 func armBFAuxInt(lsb, width int64) arm64BitField {
1540 if lsb < 0 || lsb > 63 {
1541 panic("ARM(64) bit field lsb constant out of range")
1543 if width < 1 || width > 64 {
1544 panic("ARM(64) bit field width constant out of range")
1546 return arm64BitField(width | lsb<<8)
1549 // returns the lsb part of the auxInt field of arm64 bitfield ops.
1550 func (bfc arm64BitField) getARM64BFlsb() int64 {
1551 return int64(uint64(bfc) >> 8)
1554 // returns the width part of the auxInt field of arm64 bitfield ops.
1555 func (bfc arm64BitField) getARM64BFwidth() int64 {
1556 return int64(bfc) & 0xff
1559 // checks if mask >> rshift applied at lsb is a valid arm64 bitfield op mask.
1560 func isARM64BFMask(lsb, mask, rshift int64) bool {
1561 shiftedMask := int64(uint64(mask) >> uint64(rshift))
1562 return shiftedMask != 0 && isPowerOfTwo64(shiftedMask+1) && nto(shiftedMask)+lsb < 64
1565 // returns the bitfield width of mask >> rshift for arm64 bitfield ops
1566 func arm64BFWidth(mask, rshift int64) int64 {
1567 shiftedMask := int64(uint64(mask) >> uint64(rshift))
1568 if shiftedMask == 0 {
1569 panic("ARM64 BF mask is zero")
1571 return nto(shiftedMask)
1574 // sizeof returns the size of t in bytes.
1575 // It will panic if t is not a *types.Type.
1576 func sizeof(t interface{}) int64 {
1577 return t.(*types.Type).Size()
1580 // registerizable reports whether t is a primitive type that fits in
1581 // a register. It assumes float64 values will always fit into registers
1582 // even if that isn't strictly true.
1583 func registerizable(b *Block, typ *types.Type) bool {
1584 if typ.IsPtrShaped() || typ.IsFloat() {
1587 if typ.IsInteger() {
1588 return typ.Size() <= b.Func.Config.RegSize
1593 // needRaceCleanup reports whether this call to racefuncenter/exit isn't needed.
1594 func needRaceCleanup(sym *AuxCall, v *Value) bool {
1599 if !isSameCall(sym, "runtime.racefuncenter") && !isSameCall(sym, "runtime.racefuncenterfp") && !isSameCall(sym, "runtime.racefuncexit") {
1602 for _, b := range f.Blocks {
1603 for _, v := range b.Values {
1606 // Check for racefuncenter/racefuncenterfp will encounter racefuncexit and vice versa.
1607 // Allow calls to panic*
1608 s := v.Aux.(*AuxCall).Fn.String()
1610 case "runtime.racefuncenter", "runtime.racefuncenterfp", "runtime.racefuncexit",
1611 "runtime.panicdivide", "runtime.panicwrap",
1612 "runtime.panicshift":
1615 // If we encountered any call, we need to keep racefunc*,
1616 // for accurate stacktraces.
1618 case OpPanicBounds, OpPanicExtend:
1619 // Note: these are panic generators that are ok (like the static calls above).
1620 case OpClosureCall, OpInterCall:
1621 // We must keep the race functions if there are any other call types.
1626 if isSameCall(sym, "runtime.racefuncenter") {
1627 // If we're removing racefuncenter, remove its argument as well.
1628 if v.Args[0].Op != OpStore {
1631 mem := v.Args[0].Args[2]
1632 v.Args[0].reset(OpCopy)
1633 v.Args[0].AddArg(mem)
1638 // symIsRO reports whether sym is a read-only global.
1639 func symIsRO(sym interface{}) bool {
1640 lsym := sym.(*obj.LSym)
1641 return lsym.Type == objabi.SRODATA && len(lsym.R) == 0
1644 // symIsROZero reports whether sym is a read-only global whose data contains all zeros.
1645 func symIsROZero(sym Sym) bool {
1646 lsym := sym.(*obj.LSym)
1647 if lsym.Type != objabi.SRODATA || len(lsym.R) != 0 {
1650 for _, b := range lsym.P {
1658 // read8 reads one byte from the read-only global sym at offset off.
1659 func read8(sym interface{}, off int64) uint8 {
1660 lsym := sym.(*obj.LSym)
1661 if off >= int64(len(lsym.P)) || off < 0 {
1662 // Invalid index into the global sym.
1663 // This can happen in dead code, so we don't want to panic.
1664 // Just return any value, it will eventually get ignored.
1671 // read16 reads two bytes from the read-only global sym at offset off.
1672 func read16(sym interface{}, off int64, byteorder binary.ByteOrder) uint16 {
1673 lsym := sym.(*obj.LSym)
1674 // lsym.P is written lazily.
1675 // Bytes requested after the end of lsym.P are 0.
1677 if 0 <= off && off < int64(len(lsym.P)) {
1680 buf := make([]byte, 2)
1682 return byteorder.Uint16(buf)
1685 // read32 reads four bytes from the read-only global sym at offset off.
1686 func read32(sym interface{}, off int64, byteorder binary.ByteOrder) uint32 {
1687 lsym := sym.(*obj.LSym)
1689 if 0 <= off && off < int64(len(lsym.P)) {
1692 buf := make([]byte, 4)
1694 return byteorder.Uint32(buf)
1697 // read64 reads eight bytes from the read-only global sym at offset off.
1698 func read64(sym interface{}, off int64, byteorder binary.ByteOrder) uint64 {
1699 lsym := sym.(*obj.LSym)
1701 if 0 <= off && off < int64(len(lsym.P)) {
1704 buf := make([]byte, 8)
1706 return byteorder.Uint64(buf)
1709 // sequentialAddresses reports true if it can prove that x + n == y
1710 func sequentialAddresses(x, y *Value, n int64) bool {
1711 if x.Op == Op386ADDL && y.Op == Op386LEAL1 && y.AuxInt == n && y.Aux == nil &&
1712 (x.Args[0] == y.Args[0] && x.Args[1] == y.Args[1] ||
1713 x.Args[0] == y.Args[1] && x.Args[1] == y.Args[0]) {
1716 if x.Op == Op386LEAL1 && y.Op == Op386LEAL1 && y.AuxInt == x.AuxInt+n && x.Aux == y.Aux &&
1717 (x.Args[0] == y.Args[0] && x.Args[1] == y.Args[1] ||
1718 x.Args[0] == y.Args[1] && x.Args[1] == y.Args[0]) {
1721 if x.Op == OpAMD64ADDQ && y.Op == OpAMD64LEAQ1 && y.AuxInt == n && y.Aux == nil &&
1722 (x.Args[0] == y.Args[0] && x.Args[1] == y.Args[1] ||
1723 x.Args[0] == y.Args[1] && x.Args[1] == y.Args[0]) {
1726 if x.Op == OpAMD64LEAQ1 && y.Op == OpAMD64LEAQ1 && y.AuxInt == x.AuxInt+n && x.Aux == y.Aux &&
1727 (x.Args[0] == y.Args[0] && x.Args[1] == y.Args[1] ||
1728 x.Args[0] == y.Args[1] && x.Args[1] == y.Args[0]) {
1734 // flagConstant represents the result of a compile-time comparison.
1735 // The sense of these flags does not necessarily represent the hardware's notion
1736 // of a flags register - these are just a compile-time construct.
1737 // We happen to match the semantics to those of arm/arm64.
1738 // Note that these semantics differ from x86: the carry flag has the opposite
1739 // sense on a subtraction!
1740 // On amd64, C=1 represents a borrow, e.g. SBB on amd64 does x - y - C.
1741 // On arm64, C=0 represents a borrow, e.g. SBC on arm64 does x - y - ^C.
1742 // (because it does x + ^y + C).
1743 // See https://en.wikipedia.org/wiki/Carry_flag#Vs._borrow_flag
1744 type flagConstant uint8
1746 // N reports whether the result of an operation is negative (high bit set).
1747 func (fc flagConstant) N() bool {
1751 // Z reports whether the result of an operation is 0.
1752 func (fc flagConstant) Z() bool {
1756 // C reports whether an unsigned add overflowed (carry), or an
1757 // unsigned subtract did not underflow (borrow).
1758 func (fc flagConstant) C() bool {
1762 // V reports whether a signed operation overflowed or underflowed.
1763 func (fc flagConstant) V() bool {
1767 func (fc flagConstant) eq() bool {
1770 func (fc flagConstant) ne() bool {
1773 func (fc flagConstant) lt() bool {
1774 return fc.N() != fc.V()
1776 func (fc flagConstant) le() bool {
1777 return fc.Z() || fc.lt()
1779 func (fc flagConstant) gt() bool {
1780 return !fc.Z() && fc.ge()
1782 func (fc flagConstant) ge() bool {
1783 return fc.N() == fc.V()
1785 func (fc flagConstant) ult() bool {
1788 func (fc flagConstant) ule() bool {
1789 return fc.Z() || fc.ult()
1791 func (fc flagConstant) ugt() bool {
1792 return !fc.Z() && fc.uge()
1794 func (fc flagConstant) uge() bool {
1798 func (fc flagConstant) ltNoov() bool {
1799 return fc.lt() && !fc.V()
1801 func (fc flagConstant) leNoov() bool {
1802 return fc.le() && !fc.V()
1804 func (fc flagConstant) gtNoov() bool {
1805 return fc.gt() && !fc.V()
1807 func (fc flagConstant) geNoov() bool {
1808 return fc.ge() && !fc.V()
1811 func (fc flagConstant) String() string {
1812 return fmt.Sprintf("N=%v,Z=%v,C=%v,V=%v", fc.N(), fc.Z(), fc.C(), fc.V())
1815 type flagConstantBuilder struct {
1822 func (fcs flagConstantBuilder) encode() flagConstant {
1839 // Note: addFlags(x,y) != subFlags(x,-y) in some situations:
1840 // - the results of the C flag are different
1841 // - the results of the V flag when y==minint are different
1843 // addFlags64 returns the flags that would be set from computing x+y.
1844 func addFlags64(x, y int64) flagConstant {
1845 var fcb flagConstantBuilder
1848 fcb.C = uint64(x+y) < uint64(x)
1849 fcb.V = x >= 0 && y >= 0 && x+y < 0 || x < 0 && y < 0 && x+y >= 0
1853 // subFlags64 returns the flags that would be set from computing x-y.
1854 func subFlags64(x, y int64) flagConstant {
1855 var fcb flagConstantBuilder
1858 fcb.C = uint64(y) <= uint64(x) // This code follows the arm carry flag model.
1859 fcb.V = x >= 0 && y < 0 && x-y < 0 || x < 0 && y >= 0 && x-y >= 0
1863 // addFlags32 returns the flags that would be set from computing x+y.
1864 func addFlags32(x, y int32) flagConstant {
1865 var fcb flagConstantBuilder
1868 fcb.C = uint32(x+y) < uint32(x)
1869 fcb.V = x >= 0 && y >= 0 && x+y < 0 || x < 0 && y < 0 && x+y >= 0
1873 // subFlags32 returns the flags that would be set from computing x-y.
1874 func subFlags32(x, y int32) flagConstant {
1875 var fcb flagConstantBuilder
1878 fcb.C = uint32(y) <= uint32(x) // This code follows the arm carry flag model.
1879 fcb.V = x >= 0 && y < 0 && x-y < 0 || x < 0 && y >= 0 && x-y >= 0
1883 // logicFlags64 returns flags set to the sign/zeroness of x.
1884 // C and V are set to false.
1885 func logicFlags64(x int64) flagConstant {
1886 var fcb flagConstantBuilder
1892 // logicFlags32 returns flags set to the sign/zeroness of x.
1893 // C and V are set to false.
1894 func logicFlags32(x int32) flagConstant {
1895 var fcb flagConstantBuilder