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 // truncate64Fto32F converts a float64 value to a float32 preserving the bit pattern
525 // of the mantissa. It will panic if the truncation results in lost information.
526 func truncate64Fto32F(f float64) float32 {
527 if !isExactFloat32(f) {
528 panic("truncate64Fto32F: truncation is not exact")
533 // NaN bit patterns aren't necessarily preserved across conversion
534 // instructions so we need to do the conversion manually.
535 b := math.Float64bits(f)
536 m := b & ((1 << 52) - 1) // mantissa (a.k.a. significand)
537 // | sign | exponent | mantissa |
538 r := uint32(((b >> 32) & (1 << 31)) | 0x7f800000 | (m >> (52 - 23)))
539 return math.Float32frombits(r)
542 // extend32Fto64F converts a float32 value to a float64 value preserving the bit
543 // pattern of the mantissa.
544 func extend32Fto64F(f float32) float64 {
545 if !math.IsNaN(float64(f)) {
548 // NaN bit patterns aren't necessarily preserved across conversion
549 // instructions so we need to do the conversion manually.
550 b := uint64(math.Float32bits(f))
551 // | sign | exponent | mantissa |
552 r := ((b << 32) & (1 << 63)) | (0x7ff << 52) | ((b & 0x7fffff) << (52 - 23))
553 return math.Float64frombits(r)
556 // DivisionNeedsFixUp reports whether the division needs fix-up code.
557 func DivisionNeedsFixUp(v *Value) bool {
561 // auxFrom64F encodes a float64 value so it can be stored in an AuxInt.
562 func auxFrom64F(f float64) int64 {
564 panic("can't encode a NaN in AuxInt field")
566 return int64(math.Float64bits(f))
569 // auxFrom32F encodes a float32 value so it can be stored in an AuxInt.
570 func auxFrom32F(f float32) int64 {
572 panic("can't encode a NaN in AuxInt field")
574 return int64(math.Float64bits(extend32Fto64F(f)))
577 // auxTo32F decodes a float32 from the AuxInt value provided.
578 func auxTo32F(i int64) float32 {
579 return truncate64Fto32F(math.Float64frombits(uint64(i)))
582 // auxTo64F decodes a float64 from the AuxInt value provided.
583 func auxTo64F(i int64) float64 {
584 return math.Float64frombits(uint64(i))
587 func auxIntToBool(i int64) bool {
593 func auxIntToInt8(i int64) int8 {
596 func auxIntToInt16(i int64) int16 {
599 func auxIntToInt32(i int64) int32 {
602 func auxIntToInt64(i int64) int64 {
605 func auxIntToUint8(i int64) uint8 {
608 func auxIntToFloat32(i int64) float32 {
609 return float32(math.Float64frombits(uint64(i)))
611 func auxIntToFloat64(i int64) float64 {
612 return math.Float64frombits(uint64(i))
614 func auxIntToValAndOff(i int64) ValAndOff {
617 func auxIntToArm64BitField(i int64) arm64BitField {
618 return arm64BitField(i)
620 func auxIntToInt128(x int64) int128 {
622 panic("nonzero int128 not allowed")
626 func auxIntToFlagConstant(x int64) flagConstant {
627 return flagConstant(x)
630 func auxIntToOp(cc int64) Op {
634 func boolToAuxInt(b bool) int64 {
640 func int8ToAuxInt(i int8) int64 {
643 func int16ToAuxInt(i int16) int64 {
646 func int32ToAuxInt(i int32) int64 {
649 func int64ToAuxInt(i int64) int64 {
652 func uint8ToAuxInt(i uint8) int64 {
653 return int64(int8(i))
655 func float32ToAuxInt(f float32) int64 {
656 return int64(math.Float64bits(float64(f)))
658 func float64ToAuxInt(f float64) int64 {
659 return int64(math.Float64bits(f))
661 func valAndOffToAuxInt(v ValAndOff) int64 {
664 func arm64BitFieldToAuxInt(v arm64BitField) int64 {
667 func int128ToAuxInt(x int128) int64 {
669 panic("nonzero int128 not allowed")
673 func flagConstantToAuxInt(x flagConstant) int64 {
677 func opToAuxInt(o Op) int64 {
681 // Aux is an interface to hold miscellaneous data in Blocks and Values.
686 // stringAux wraps string values for use in Aux.
687 type stringAux string
689 func (stringAux) CanBeAnSSAAux() {}
691 func auxToString(i Aux) string {
692 return string(i.(stringAux))
694 func auxToSym(i Aux) Sym {
695 // TODO: kind of a hack - allows nil interface through
699 func auxToType(i Aux) *types.Type {
700 return i.(*types.Type)
702 func auxToCall(i Aux) *AuxCall {
705 func auxToS390xCCMask(i Aux) s390x.CCMask {
706 return i.(s390x.CCMask)
708 func auxToS390xRotateParams(i Aux) s390x.RotateParams {
709 return i.(s390x.RotateParams)
712 func StringToAux(s string) Aux {
715 func symToAux(s Sym) Aux {
718 func callToAux(s *AuxCall) Aux {
721 func typeToAux(t *types.Type) Aux {
724 func s390xCCMaskToAux(c s390x.CCMask) Aux {
727 func s390xRotateParamsToAux(r s390x.RotateParams) Aux {
731 // uaddOvf reports whether unsigned a+b would overflow.
732 func uaddOvf(a, b int64) bool {
733 return uint64(a)+uint64(b) < uint64(a)
736 // de-virtualize an InterCall
737 // 'sym' is the symbol for the itab
738 func devirt(v *Value, aux Aux, sym Sym, offset int64) *AuxCall {
740 n, ok := sym.(*obj.LSym)
744 lsym := f.fe.DerefItab(n, offset)
745 if f.pass.debug > 0 {
747 f.Warnl(v.Pos, "de-virtualizing call")
749 f.Warnl(v.Pos, "couldn't de-virtualize call")
756 return StaticAuxCall(lsym, va.args, va.results)
759 // de-virtualize an InterLECall
760 // 'sym' is the symbol for the itab
761 func devirtLESym(v *Value, aux Aux, sym Sym, offset int64) *obj.LSym {
762 n, ok := sym.(*obj.LSym)
768 lsym := f.fe.DerefItab(n, offset)
769 if f.pass.debug > 0 {
771 f.Warnl(v.Pos, "de-virtualizing call")
773 f.Warnl(v.Pos, "couldn't de-virtualize call")
782 func devirtLECall(v *Value, sym *obj.LSym) *Value {
783 v.Op = OpStaticLECall
784 v.Aux.(*AuxCall).Fn = sym
789 // isSamePtr reports whether p1 and p2 point to the same address.
790 func isSamePtr(p1, p2 *Value) bool {
799 return p1.AuxInt == p2.AuxInt && isSamePtr(p1.Args[0], p2.Args[0])
800 case OpAddr, OpLocalAddr:
801 // OpAddr's 0th arg is either OpSP or OpSB, which means that it is uniquely identified by its Op.
802 // Checking for value equality only works after [z]cse has run.
803 return p1.Aux == p2.Aux && p1.Args[0].Op == p2.Args[0].Op
805 return p1.Args[1] == p2.Args[1] && isSamePtr(p1.Args[0], p2.Args[0])
810 func isStackPtr(v *Value) bool {
811 for v.Op == OpOffPtr || v.Op == OpAddPtr {
814 return v.Op == OpSP || v.Op == OpLocalAddr
817 // disjoint reports whether the memory region specified by [p1:p1+n1)
818 // does not overlap with [p2:p2+n2).
819 // A return value of false does not imply the regions overlap.
820 func disjoint(p1 *Value, n1 int64, p2 *Value, n2 int64) bool {
821 if n1 == 0 || n2 == 0 {
827 baseAndOffset := func(ptr *Value) (base *Value, offset int64) {
828 base, offset = ptr, 0
829 for base.Op == OpOffPtr {
830 offset += base.AuxInt
835 p1, off1 := baseAndOffset(p1)
836 p2, off2 := baseAndOffset(p2)
837 if isSamePtr(p1, p2) {
838 return !overlap(off1, n1, off2, n2)
840 // p1 and p2 are not the same, so if they are both OpAddrs then
841 // they point to different variables.
842 // If one pointer is on the stack and the other is an argument
843 // then they can't overlap.
845 case OpAddr, OpLocalAddr:
846 if p2.Op == OpAddr || p2.Op == OpLocalAddr || p2.Op == OpSP {
849 return p2.Op == OpArg && p1.Args[0].Op == OpSP
851 if p2.Op == OpSP || p2.Op == OpLocalAddr {
855 return p2.Op == OpAddr || p2.Op == OpLocalAddr || p2.Op == OpArg || p2.Op == OpSP
860 // moveSize returns the number of bytes an aligned MOV instruction moves
861 func moveSize(align int64, c *Config) int64 {
863 case align%8 == 0 && c.PtrSize == 8:
873 // mergePoint finds a block among a's blocks which dominates b and is itself
874 // dominated by all of a's blocks. Returns nil if it can't find one.
875 // Might return nil even if one does exist.
876 func mergePoint(b *Block, a ...*Value) *Block {
877 // Walk backward from b looking for one of the a's blocks.
883 for _, x := range a {
888 if len(b.Preds) > 1 {
889 // Don't know which way to go back. Abort.
895 return nil // too far away
897 // At this point, r is the first value in a that we find by walking backwards.
898 // if we return anything, r will be it.
901 // Keep going, counting the other a's that we find. They must all dominate r.
904 for _, x := range a {
910 // Found all of a in a backwards walk. We can return r.
913 if len(b.Preds) > 1 {
920 return nil // too far away
923 // clobber invalidates values. Returns true.
924 // clobber is used by rewrite rules to:
925 // A) make sure the values are really dead and never used again.
926 // B) decrement use counts of the values' args.
927 func clobber(vv ...*Value) bool {
928 for _, v := range vv {
930 // Note: leave v.Block intact. The Block field is used after clobber.
935 // clobberIfDead resets v when use count is 1. Returns true.
936 // clobberIfDead is used by rewrite rules to decrement
937 // use counts of v's args when v is dead and never used.
938 func clobberIfDead(v *Value) bool {
942 // Note: leave v.Block intact. The Block field is used after clobberIfDead.
946 // noteRule is an easy way to track if a rule is matched when writing
947 // new ones. Make the rule of interest also conditional on
948 // noteRule("note to self: rule of interest matched")
949 // and that message will print when the rule matches.
950 func noteRule(s string) bool {
955 // countRule increments Func.ruleMatches[key].
956 // If Func.ruleMatches is non-nil at the end
957 // of compilation, it will be printed to stdout.
958 // This is intended to make it easier to find which functions
959 // which contain lots of rules matches when developing new rules.
960 func countRule(v *Value, key string) bool {
962 if f.ruleMatches == nil {
963 f.ruleMatches = make(map[string]int)
969 // warnRule generates compiler debug output with string s when
970 // v is not in autogenerated code, cond is true and the rule has fired.
971 func warnRule(cond bool, v *Value, s string) bool {
972 if pos := v.Pos; pos.Line() > 1 && cond {
973 v.Block.Func.Warnl(pos, s)
978 // for a pseudo-op like (LessThan x), extract x
979 func flagArg(v *Value) *Value {
980 if len(v.Args) != 1 || !v.Args[0].Type.IsFlags() {
986 // arm64Negate finds the complement to an ARM64 condition code,
987 // for example Equal -> NotEqual or LessThan -> GreaterEqual
989 // TODO: add floating-point conditions
990 func arm64Negate(op Op) Op {
992 case OpARM64LessThan:
993 return OpARM64GreaterEqual
994 case OpARM64LessThanU:
995 return OpARM64GreaterEqualU
996 case OpARM64GreaterThan:
997 return OpARM64LessEqual
998 case OpARM64GreaterThanU:
999 return OpARM64LessEqualU
1000 case OpARM64LessEqual:
1001 return OpARM64GreaterThan
1002 case OpARM64LessEqualU:
1003 return OpARM64GreaterThanU
1004 case OpARM64GreaterEqual:
1005 return OpARM64LessThan
1006 case OpARM64GreaterEqualU:
1007 return OpARM64LessThanU
1009 return OpARM64NotEqual
1010 case OpARM64NotEqual:
1012 case OpARM64LessThanF:
1013 return OpARM64GreaterEqualF
1014 case OpARM64GreaterThanF:
1015 return OpARM64LessEqualF
1016 case OpARM64LessEqualF:
1017 return OpARM64GreaterThanF
1018 case OpARM64GreaterEqualF:
1019 return OpARM64LessThanF
1021 panic("unreachable")
1025 // arm64Invert evaluates (InvertFlags op), which
1026 // is the same as altering the condition codes such
1027 // that the same result would be produced if the arguments
1028 // to the flag-generating instruction were reversed, e.g.
1029 // (InvertFlags (CMP x y)) -> (CMP y x)
1031 // TODO: add floating-point conditions
1032 func arm64Invert(op Op) Op {
1034 case OpARM64LessThan:
1035 return OpARM64GreaterThan
1036 case OpARM64LessThanU:
1037 return OpARM64GreaterThanU
1038 case OpARM64GreaterThan:
1039 return OpARM64LessThan
1040 case OpARM64GreaterThanU:
1041 return OpARM64LessThanU
1042 case OpARM64LessEqual:
1043 return OpARM64GreaterEqual
1044 case OpARM64LessEqualU:
1045 return OpARM64GreaterEqualU
1046 case OpARM64GreaterEqual:
1047 return OpARM64LessEqual
1048 case OpARM64GreaterEqualU:
1049 return OpARM64LessEqualU
1050 case OpARM64Equal, OpARM64NotEqual:
1052 case OpARM64LessThanF:
1053 return OpARM64GreaterThanF
1054 case OpARM64GreaterThanF:
1055 return OpARM64LessThanF
1056 case OpARM64LessEqualF:
1057 return OpARM64GreaterEqualF
1058 case OpARM64GreaterEqualF:
1059 return OpARM64LessEqualF
1061 panic("unreachable")
1065 // evaluate an ARM64 op against a flags value
1066 // that is potentially constant; return 1 for true,
1067 // -1 for false, and 0 for not constant.
1068 func ccARM64Eval(op Op, flags *Value) int {
1070 if fop == OpARM64InvertFlags {
1071 return -ccARM64Eval(op, flags.Args[0])
1073 if fop != OpARM64FlagConstant {
1076 fc := flagConstant(flags.AuxInt)
1077 b2i := func(b bool) int {
1086 case OpARM64NotEqual:
1088 case OpARM64LessThan:
1090 case OpARM64LessThanU:
1091 return b2i(fc.ult())
1092 case OpARM64GreaterThan:
1094 case OpARM64GreaterThanU:
1095 return b2i(fc.ugt())
1096 case OpARM64LessEqual:
1098 case OpARM64LessEqualU:
1099 return b2i(fc.ule())
1100 case OpARM64GreaterEqual:
1102 case OpARM64GreaterEqualU:
1103 return b2i(fc.uge())
1108 // logRule logs the use of the rule s. This will only be enabled if
1109 // rewrite rules were generated with the -log option, see gen/rulegen.go.
1110 func logRule(s string) {
1111 if ruleFile == nil {
1112 // Open a log file to write log to. We open in append
1113 // mode because all.bash runs the compiler lots of times,
1114 // and we want the concatenation of all of those logs.
1115 // This means, of course, that users need to rm the old log
1116 // to get fresh data.
1117 // TODO: all.bash runs compilers in parallel. Need to synchronize logging somehow?
1118 w, err := os.OpenFile(filepath.Join(os.Getenv("GOROOT"), "src", "rulelog"),
1119 os.O_CREATE|os.O_WRONLY|os.O_APPEND, 0666)
1125 _, err := fmt.Fprintln(ruleFile, s)
1131 var ruleFile io.Writer
1133 func min(x, y int64) int64 {
1140 func isConstZero(v *Value) bool {
1144 case OpConst64, OpConst32, OpConst16, OpConst8, OpConstBool, OpConst32F, OpConst64F:
1145 return v.AuxInt == 0
1150 // reciprocalExact64 reports whether 1/c is exactly representable.
1151 func reciprocalExact64(c float64) bool {
1152 b := math.Float64bits(c)
1153 man := b & (1<<52 - 1)
1155 return false // not a power of 2, denormal, or NaN
1157 exp := b >> 52 & (1<<11 - 1)
1158 // exponent bias is 0x3ff. So taking the reciprocal of a number
1159 // changes the exponent to 0x7fe-exp.
1164 return false // ±inf
1166 return false // exponent is not representable
1172 // reciprocalExact32 reports whether 1/c is exactly representable.
1173 func reciprocalExact32(c float32) bool {
1174 b := math.Float32bits(c)
1175 man := b & (1<<23 - 1)
1177 return false // not a power of 2, denormal, or NaN
1179 exp := b >> 23 & (1<<8 - 1)
1180 // exponent bias is 0x7f. So taking the reciprocal of a number
1181 // changes the exponent to 0xfe-exp.
1186 return false // ±inf
1188 return false // exponent is not representable
1194 // check if an immediate can be directly encoded into an ARM's instruction
1195 func isARMImmRot(v uint32) bool {
1196 for i := 0; i < 16; i++ {
1206 // overlap reports whether the ranges given by the given offset and
1207 // size pairs overlap.
1208 func overlap(offset1, size1, offset2, size2 int64) bool {
1209 if offset1 >= offset2 && offset2+size2 > offset1 {
1212 if offset2 >= offset1 && offset1+size1 > offset2 {
1218 func areAdjacentOffsets(off1, off2, size int64) bool {
1219 return off1+size == off2 || off1 == off2+size
1222 // check if value zeroes out upper 32-bit of 64-bit register.
1223 // depth limits recursion depth. In AMD64.rules 3 is used as limit,
1224 // because it catches same amount of cases as 4.
1225 func zeroUpper32Bits(x *Value, depth int) bool {
1227 case OpAMD64MOVLconst, OpAMD64MOVLload, OpAMD64MOVLQZX, OpAMD64MOVLloadidx1,
1228 OpAMD64MOVWload, OpAMD64MOVWloadidx1, OpAMD64MOVBload, OpAMD64MOVBloadidx1,
1229 OpAMD64MOVLloadidx4, OpAMD64ADDLload, OpAMD64SUBLload, OpAMD64ANDLload,
1230 OpAMD64ORLload, OpAMD64XORLload, OpAMD64CVTTSD2SL,
1231 OpAMD64ADDL, OpAMD64ADDLconst, OpAMD64SUBL, OpAMD64SUBLconst,
1232 OpAMD64ANDL, OpAMD64ANDLconst, OpAMD64ORL, OpAMD64ORLconst,
1233 OpAMD64XORL, OpAMD64XORLconst, OpAMD64NEGL, OpAMD64NOTL,
1234 OpAMD64SHRL, OpAMD64SHRLconst, OpAMD64SARL, OpAMD64SARLconst,
1235 OpAMD64SHLL, OpAMD64SHLLconst:
1238 return x.Type.Width == 4
1239 case OpPhi, OpSelect0, OpSelect1:
1240 // Phis can use each-other as an arguments, instead of tracking visited values,
1241 // just limit recursion depth.
1245 for i := range x.Args {
1246 if !zeroUpper32Bits(x.Args[i], depth-1) {
1256 // zeroUpper48Bits is similar to zeroUpper32Bits, but for upper 48 bits
1257 func zeroUpper48Bits(x *Value, depth int) bool {
1259 case OpAMD64MOVWQZX, OpAMD64MOVWload, OpAMD64MOVWloadidx1, OpAMD64MOVWloadidx2:
1262 return x.Type.Width == 2
1263 case OpPhi, OpSelect0, OpSelect1:
1264 // Phis can use each-other as an arguments, instead of tracking visited values,
1265 // just limit recursion depth.
1269 for i := range x.Args {
1270 if !zeroUpper48Bits(x.Args[i], depth-1) {
1280 // zeroUpper56Bits is similar to zeroUpper32Bits, but for upper 56 bits
1281 func zeroUpper56Bits(x *Value, depth int) bool {
1283 case OpAMD64MOVBQZX, OpAMD64MOVBload, OpAMD64MOVBloadidx1:
1286 return x.Type.Width == 1
1287 case OpPhi, OpSelect0, OpSelect1:
1288 // Phis can use each-other as an arguments, instead of tracking visited values,
1289 // just limit recursion depth.
1293 for i := range x.Args {
1294 if !zeroUpper56Bits(x.Args[i], depth-1) {
1304 // isInlinableMemmove reports whether the given arch performs a Move of the given size
1305 // faster than memmove. It will only return true if replacing the memmove with a Move is
1306 // safe, either because Move is small or because the arguments are disjoint.
1307 // This is used as a check for replacing memmove with Move ops.
1308 func isInlinableMemmove(dst, src *Value, sz int64, c *Config) bool {
1309 // It is always safe to convert memmove into Move when its arguments are disjoint.
1310 // Move ops may or may not be faster for large sizes depending on how the platform
1311 // lowers them, so we only perform this optimization on platforms that we know to
1312 // have fast Move ops.
1315 return sz <= 16 || (sz < 1024 && disjoint(dst, sz, src, sz))
1316 case "386", "arm64":
1318 case "s390x", "ppc64", "ppc64le":
1319 return sz <= 8 || disjoint(dst, sz, src, sz)
1320 case "arm", "mips", "mips64", "mipsle", "mips64le":
1326 // logLargeCopy logs the occurrence of a large copy.
1327 // The best place to do this is in the rewrite rules where the size of the move is easy to find.
1328 // "Large" is arbitrarily chosen to be 128 bytes; this may change.
1329 func logLargeCopy(v *Value, s int64) bool {
1333 if logopt.Enabled() {
1334 logopt.LogOpt(v.Pos, "copy", "lower", v.Block.Func.Name, fmt.Sprintf("%d bytes", s))
1339 // hasSmallRotate reports whether the architecture has rotate instructions
1340 // for sizes < 32-bit. This is used to decide whether to promote some rotations.
1341 func hasSmallRotate(c *Config) bool {
1343 case "amd64", "386":
1350 func newPPC64ShiftAuxInt(sh, mb, me, sz int64) int32 {
1351 if sh < 0 || sh >= sz {
1352 panic("PPC64 shift arg sh out of range")
1354 if mb < 0 || mb >= sz {
1355 panic("PPC64 shift arg mb out of range")
1357 if me < 0 || me >= sz {
1358 panic("PPC64 shift arg me out of range")
1360 return int32(sh<<16 | mb<<8 | me)
1363 func GetPPC64Shiftsh(auxint int64) int64 {
1364 return int64(int8(auxint >> 16))
1367 func GetPPC64Shiftmb(auxint int64) int64 {
1368 return int64(int8(auxint >> 8))
1371 func GetPPC64Shiftme(auxint int64) int64 {
1372 return int64(int8(auxint))
1375 // Test if this value can encoded as a mask for a rlwinm like
1376 // operation. Masks can also extend from the msb and wrap to
1377 // the lsb too. That is, the valid masks are 32 bit strings
1378 // of the form: 0..01..10..0 or 1..10..01..1 or 1...1
1379 func isPPC64WordRotateMask(v64 int64) bool {
1380 // Isolate rightmost 1 (if none 0) and add.
1383 // Likewise, for the wrapping case.
1385 vpn := (vn & -vn) + vn
1386 return (v&vp == 0 || vn&vpn == 0) && v != 0
1389 // Compress mask and and shift into single value of the form
1390 // me | mb<<8 | rotate<<16 | nbits<<24 where me and mb can
1391 // be used to regenerate the input mask.
1392 func encodePPC64RotateMask(rotate, mask, nbits int64) int64 {
1393 var mb, me, mbn, men int
1395 // Determine boundaries and then decode them
1396 if mask == 0 || ^mask == 0 || rotate >= nbits {
1397 panic("Invalid PPC64 rotate mask")
1398 } else if nbits == 32 {
1399 mb = bits.LeadingZeros32(uint32(mask))
1400 me = 32 - bits.TrailingZeros32(uint32(mask))
1401 mbn = bits.LeadingZeros32(^uint32(mask))
1402 men = 32 - bits.TrailingZeros32(^uint32(mask))
1404 mb = bits.LeadingZeros64(uint64(mask))
1405 me = 64 - bits.TrailingZeros64(uint64(mask))
1406 mbn = bits.LeadingZeros64(^uint64(mask))
1407 men = 64 - bits.TrailingZeros64(^uint64(mask))
1409 // Check for a wrapping mask (e.g bits at 0 and 63)
1410 if mb == 0 && me == int(nbits) {
1411 // swap the inverted values
1415 return int64(me) | int64(mb<<8) | int64(rotate<<16) | int64(nbits<<24)
1418 // The inverse operation of encodePPC64RotateMask. The values returned as
1419 // mb and me satisfy the POWER ISA definition of MASK(x,y) where MASK(mb,me) = mask.
1420 func DecodePPC64RotateMask(sauxint int64) (rotate, mb, me int64, mask uint64) {
1421 auxint := uint64(sauxint)
1422 rotate = int64((auxint >> 16) & 0xFF)
1423 mb = int64((auxint >> 8) & 0xFF)
1424 me = int64((auxint >> 0) & 0xFF)
1425 nbits := int64((auxint >> 24) & 0xFF)
1426 mask = ((1 << uint(nbits-mb)) - 1) ^ ((1 << uint(nbits-me)) - 1)
1431 mask = uint64(uint32(mask))
1434 // Fixup ME to match ISA definition. The second argument to MASK(..,me)
1436 me = (me - 1) & (nbits - 1)
1440 // This verifies that the mask is a set of
1441 // consecutive bits including the least
1443 func isPPC64ValidShiftMask(v int64) bool {
1444 if (v != 0) && ((v+1)&v) == 0 {
1450 func getPPC64ShiftMaskLength(v int64) int64 {
1451 return int64(bits.Len64(uint64(v)))
1454 // Decompose a shift right into an equivalent rotate/mask,
1455 // and return mask & m.
1456 func mergePPC64RShiftMask(m, s, nbits int64) int64 {
1457 smask := uint64((1<<uint(nbits))-1) >> uint(s)
1458 return m & int64(smask)
1461 // Combine (ANDconst [m] (SRWconst [s])) into (RLWINM [y]) or return 0
1462 func mergePPC64AndSrwi(m, s int64) int64 {
1463 mask := mergePPC64RShiftMask(m, s, 32)
1464 if !isPPC64WordRotateMask(mask) {
1467 return encodePPC64RotateMask(32-s, mask, 32)
1470 // Test if a shift right feeding into a CLRLSLDI can be merged into RLWINM.
1471 // Return the encoded RLWINM constant, or 0 if they cannot be merged.
1472 func mergePPC64ClrlsldiSrw(sld, srw int64) int64 {
1473 mask_1 := uint64(0xFFFFFFFF >> uint(srw))
1474 // for CLRLSLDI, it's more convient to think of it as a mask left bits then rotate left.
1475 mask_2 := uint64(0xFFFFFFFFFFFFFFFF) >> uint(GetPPC64Shiftmb(int64(sld)))
1477 // Rewrite mask to apply after the final left shift.
1478 mask_3 := (mask_1 & mask_2) << uint(GetPPC64Shiftsh(sld))
1481 r_2 := GetPPC64Shiftsh(sld)
1482 r_3 := (r_1 + r_2) & 31 // This can wrap.
1484 if uint64(uint32(mask_3)) != mask_3 || mask_3 == 0 {
1487 return encodePPC64RotateMask(int64(r_3), int64(mask_3), 32)
1490 // Test if a RLWINM feeding into a CLRLSLDI can be merged into RLWINM. Return
1491 // the encoded RLWINM constant, or 0 if they cannot be merged.
1492 func mergePPC64ClrlsldiRlwinm(sld int32, rlw int64) int64 {
1493 r_1, _, _, mask_1 := DecodePPC64RotateMask(rlw)
1494 // for CLRLSLDI, it's more convient to think of it as a mask left bits then rotate left.
1495 mask_2 := uint64(0xFFFFFFFFFFFFFFFF) >> uint(GetPPC64Shiftmb(int64(sld)))
1497 // combine the masks, and adjust for the final left shift.
1498 mask_3 := (mask_1 & mask_2) << uint(GetPPC64Shiftsh(int64(sld)))
1499 r_2 := GetPPC64Shiftsh(int64(sld))
1500 r_3 := (r_1 + r_2) & 31 // This can wrap.
1502 // Verify the result is still a valid bitmask of <= 32 bits.
1503 if !isPPC64WordRotateMask(int64(mask_3)) || uint64(uint32(mask_3)) != mask_3 {
1506 return encodePPC64RotateMask(r_3, int64(mask_3), 32)
1509 // Compute the encoded RLWINM constant from combining (SLDconst [sld] (SRWconst [srw] x)),
1510 // or return 0 if they cannot be combined.
1511 func mergePPC64SldiSrw(sld, srw int64) int64 {
1512 if sld > srw || srw >= 32 {
1515 mask_r := uint32(0xFFFFFFFF) >> uint(srw)
1516 mask_l := uint32(0xFFFFFFFF) >> uint(sld)
1517 mask := (mask_r & mask_l) << uint(sld)
1518 return encodePPC64RotateMask((32-srw+sld)&31, int64(mask), 32)
1521 // Convenience function to rotate a 32 bit constant value by another constant.
1522 func rotateLeft32(v, rotate int64) int64 {
1523 return int64(bits.RotateLeft32(uint32(v), int(rotate)))
1526 // encodes the lsb and width for arm(64) bitfield ops into the expected auxInt format.
1527 func armBFAuxInt(lsb, width int64) arm64BitField {
1528 if lsb < 0 || lsb > 63 {
1529 panic("ARM(64) bit field lsb constant out of range")
1531 if width < 1 || width > 64 {
1532 panic("ARM(64) bit field width constant out of range")
1534 return arm64BitField(width | lsb<<8)
1537 // returns the lsb part of the auxInt field of arm64 bitfield ops.
1538 func (bfc arm64BitField) getARM64BFlsb() int64 {
1539 return int64(uint64(bfc) >> 8)
1542 // returns the width part of the auxInt field of arm64 bitfield ops.
1543 func (bfc arm64BitField) getARM64BFwidth() int64 {
1544 return int64(bfc) & 0xff
1547 // checks if mask >> rshift applied at lsb is a valid arm64 bitfield op mask.
1548 func isARM64BFMask(lsb, mask, rshift int64) bool {
1549 shiftedMask := int64(uint64(mask) >> uint64(rshift))
1550 return shiftedMask != 0 && isPowerOfTwo64(shiftedMask+1) && nto(shiftedMask)+lsb < 64
1553 // returns the bitfield width of mask >> rshift for arm64 bitfield ops
1554 func arm64BFWidth(mask, rshift int64) int64 {
1555 shiftedMask := int64(uint64(mask) >> uint64(rshift))
1556 if shiftedMask == 0 {
1557 panic("ARM64 BF mask is zero")
1559 return nto(shiftedMask)
1562 // sizeof returns the size of t in bytes.
1563 // It will panic if t is not a *types.Type.
1564 func sizeof(t interface{}) int64 {
1565 return t.(*types.Type).Size()
1568 // registerizable reports whether t is a primitive type that fits in
1569 // a register. It assumes float64 values will always fit into registers
1570 // even if that isn't strictly true.
1571 func registerizable(b *Block, typ *types.Type) bool {
1572 if typ.IsPtrShaped() || typ.IsFloat() {
1575 if typ.IsInteger() {
1576 return typ.Size() <= b.Func.Config.RegSize
1581 // needRaceCleanup reports whether this call to racefuncenter/exit isn't needed.
1582 func needRaceCleanup(sym *AuxCall, v *Value) bool {
1587 if !isSameCall(sym, "runtime.racefuncenter") && !isSameCall(sym, "runtime.racefuncenterfp") && !isSameCall(sym, "runtime.racefuncexit") {
1590 for _, b := range f.Blocks {
1591 for _, v := range b.Values {
1594 // Check for racefuncenter/racefuncenterfp will encounter racefuncexit and vice versa.
1595 // Allow calls to panic*
1596 s := v.Aux.(*AuxCall).Fn.String()
1598 case "runtime.racefuncenter", "runtime.racefuncenterfp", "runtime.racefuncexit",
1599 "runtime.panicdivide", "runtime.panicwrap",
1600 "runtime.panicshift":
1603 // If we encountered any call, we need to keep racefunc*,
1604 // for accurate stacktraces.
1606 case OpPanicBounds, OpPanicExtend:
1607 // Note: these are panic generators that are ok (like the static calls above).
1608 case OpClosureCall, OpInterCall:
1609 // We must keep the race functions if there are any other call types.
1614 if isSameCall(sym, "runtime.racefuncenter") {
1615 // If we're removing racefuncenter, remove its argument as well.
1616 if v.Args[0].Op != OpStore {
1619 mem := v.Args[0].Args[2]
1620 v.Args[0].reset(OpCopy)
1621 v.Args[0].AddArg(mem)
1626 // symIsRO reports whether sym is a read-only global.
1627 func symIsRO(sym interface{}) bool {
1628 lsym := sym.(*obj.LSym)
1629 return lsym.Type == objabi.SRODATA && len(lsym.R) == 0
1632 // symIsROZero reports whether sym is a read-only global whose data contains all zeros.
1633 func symIsROZero(sym Sym) bool {
1634 lsym := sym.(*obj.LSym)
1635 if lsym.Type != objabi.SRODATA || len(lsym.R) != 0 {
1638 for _, b := range lsym.P {
1646 // read8 reads one byte from the read-only global sym at offset off.
1647 func read8(sym interface{}, off int64) uint8 {
1648 lsym := sym.(*obj.LSym)
1649 if off >= int64(len(lsym.P)) || off < 0 {
1650 // Invalid index into the global sym.
1651 // This can happen in dead code, so we don't want to panic.
1652 // Just return any value, it will eventually get ignored.
1659 // read16 reads two bytes from the read-only global sym at offset off.
1660 func read16(sym interface{}, off int64, byteorder binary.ByteOrder) uint16 {
1661 lsym := sym.(*obj.LSym)
1662 // lsym.P is written lazily.
1663 // Bytes requested after the end of lsym.P are 0.
1665 if 0 <= off && off < int64(len(lsym.P)) {
1668 buf := make([]byte, 2)
1670 return byteorder.Uint16(buf)
1673 // read32 reads four bytes from the read-only global sym at offset off.
1674 func read32(sym interface{}, off int64, byteorder binary.ByteOrder) uint32 {
1675 lsym := sym.(*obj.LSym)
1677 if 0 <= off && off < int64(len(lsym.P)) {
1680 buf := make([]byte, 4)
1682 return byteorder.Uint32(buf)
1685 // read64 reads eight bytes from the read-only global sym at offset off.
1686 func read64(sym interface{}, off int64, byteorder binary.ByteOrder) uint64 {
1687 lsym := sym.(*obj.LSym)
1689 if 0 <= off && off < int64(len(lsym.P)) {
1692 buf := make([]byte, 8)
1694 return byteorder.Uint64(buf)
1697 // sequentialAddresses reports true if it can prove that x + n == y
1698 func sequentialAddresses(x, y *Value, n int64) bool {
1699 if x.Op == Op386ADDL && y.Op == Op386LEAL1 && y.AuxInt == n && y.Aux == nil &&
1700 (x.Args[0] == y.Args[0] && x.Args[1] == y.Args[1] ||
1701 x.Args[0] == y.Args[1] && x.Args[1] == y.Args[0]) {
1704 if x.Op == Op386LEAL1 && y.Op == Op386LEAL1 && y.AuxInt == x.AuxInt+n && x.Aux == y.Aux &&
1705 (x.Args[0] == y.Args[0] && x.Args[1] == y.Args[1] ||
1706 x.Args[0] == y.Args[1] && x.Args[1] == y.Args[0]) {
1709 if x.Op == OpAMD64ADDQ && y.Op == OpAMD64LEAQ1 && y.AuxInt == n && y.Aux == nil &&
1710 (x.Args[0] == y.Args[0] && x.Args[1] == y.Args[1] ||
1711 x.Args[0] == y.Args[1] && x.Args[1] == y.Args[0]) {
1714 if x.Op == OpAMD64LEAQ1 && y.Op == OpAMD64LEAQ1 && y.AuxInt == x.AuxInt+n && x.Aux == y.Aux &&
1715 (x.Args[0] == y.Args[0] && x.Args[1] == y.Args[1] ||
1716 x.Args[0] == y.Args[1] && x.Args[1] == y.Args[0]) {
1722 // flagConstant represents the result of a compile-time comparison.
1723 // The sense of these flags does not necessarily represent the hardware's notion
1724 // of a flags register - these are just a compile-time construct.
1725 // We happen to match the semantics to those of arm/arm64.
1726 // Note that these semantics differ from x86: the carry flag has the opposite
1727 // sense on a subtraction!
1728 // On amd64, C=1 represents a borrow, e.g. SBB on amd64 does x - y - C.
1729 // On arm64, C=0 represents a borrow, e.g. SBC on arm64 does x - y - ^C.
1730 // (because it does x + ^y + C).
1731 // See https://en.wikipedia.org/wiki/Carry_flag#Vs._borrow_flag
1732 type flagConstant uint8
1734 // N reports whether the result of an operation is negative (high bit set).
1735 func (fc flagConstant) N() bool {
1739 // Z reports whether the result of an operation is 0.
1740 func (fc flagConstant) Z() bool {
1744 // C reports whether an unsigned add overflowed (carry), or an
1745 // unsigned subtract did not underflow (borrow).
1746 func (fc flagConstant) C() bool {
1750 // V reports whether a signed operation overflowed or underflowed.
1751 func (fc flagConstant) V() bool {
1755 func (fc flagConstant) eq() bool {
1758 func (fc flagConstant) ne() bool {
1761 func (fc flagConstant) lt() bool {
1762 return fc.N() != fc.V()
1764 func (fc flagConstant) le() bool {
1765 return fc.Z() || fc.lt()
1767 func (fc flagConstant) gt() bool {
1768 return !fc.Z() && fc.ge()
1770 func (fc flagConstant) ge() bool {
1771 return fc.N() == fc.V()
1773 func (fc flagConstant) ult() bool {
1776 func (fc flagConstant) ule() bool {
1777 return fc.Z() || fc.ult()
1779 func (fc flagConstant) ugt() bool {
1780 return !fc.Z() && fc.uge()
1782 func (fc flagConstant) uge() bool {
1786 func (fc flagConstant) ltNoov() bool {
1787 return fc.lt() && !fc.V()
1789 func (fc flagConstant) leNoov() bool {
1790 return fc.le() && !fc.V()
1792 func (fc flagConstant) gtNoov() bool {
1793 return fc.gt() && !fc.V()
1795 func (fc flagConstant) geNoov() bool {
1796 return fc.ge() && !fc.V()
1799 func (fc flagConstant) String() string {
1800 return fmt.Sprintf("N=%v,Z=%v,C=%v,V=%v", fc.N(), fc.Z(), fc.C(), fc.V())
1803 type flagConstantBuilder struct {
1810 func (fcs flagConstantBuilder) encode() flagConstant {
1827 // Note: addFlags(x,y) != subFlags(x,-y) in some situations:
1828 // - the results of the C flag are different
1829 // - the results of the V flag when y==minint are different
1831 // addFlags64 returns the flags that would be set from computing x+y.
1832 func addFlags64(x, y int64) flagConstant {
1833 var fcb flagConstantBuilder
1836 fcb.C = uint64(x+y) < uint64(x)
1837 fcb.V = x >= 0 && y >= 0 && x+y < 0 || x < 0 && y < 0 && x+y >= 0
1841 // subFlags64 returns the flags that would be set from computing x-y.
1842 func subFlags64(x, y int64) flagConstant {
1843 var fcb flagConstantBuilder
1846 fcb.C = uint64(y) <= uint64(x) // This code follows the arm carry flag model.
1847 fcb.V = x >= 0 && y < 0 && x-y < 0 || x < 0 && y >= 0 && x-y >= 0
1851 // addFlags32 returns the flags that would be set from computing x+y.
1852 func addFlags32(x, y int32) flagConstant {
1853 var fcb flagConstantBuilder
1856 fcb.C = uint32(x+y) < uint32(x)
1857 fcb.V = x >= 0 && y >= 0 && x+y < 0 || x < 0 && y < 0 && x+y >= 0
1861 // subFlags32 returns the flags that would be set from computing x-y.
1862 func subFlags32(x, y int32) flagConstant {
1863 var fcb flagConstantBuilder
1866 fcb.C = uint32(y) <= uint32(x) // This code follows the arm carry flag model.
1867 fcb.V = x >= 0 && y < 0 && x-y < 0 || x < 0 && y >= 0 && x-y >= 0
1871 // logicFlags64 returns flags set to the sign/zeroness of x.
1872 // C and V are set to false.
1873 func logicFlags64(x int64) flagConstant {
1874 var fcb flagConstantBuilder
1880 // logicFlags32 returns flags set to the sign/zeroness of x.
1881 // C and V are set to false.
1882 func logicFlags32(x int32) flagConstant {
1883 var fcb flagConstantBuilder