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/base"
9 "cmd/compile/internal/logopt"
10 "cmd/compile/internal/types"
12 "cmd/internal/obj/s390x"
24 type deadValueChoice bool
27 leaveDeadValues deadValueChoice = false
28 removeDeadValues = true
31 // deadcode indicates whether rewrite should try to remove any values that become dead.
32 func applyRewrite(f *Func, rb blockRewriter, rv valueRewriter, deadcode deadValueChoice) {
33 // repeat rewrites until we find no more rewrites
34 pendingLines := f.cachedLineStarts // Holds statement boundaries that need to be moved to a new value/block
38 fmt.Printf("%s: rewriting for %s\n", f.pass.name, f.Name)
41 var states map[string]bool
45 for _, b := range f.Blocks {
50 b0.Succs = append([]Edge{}, b.Succs...) // make a new copy, not aliasing
52 for i, c := range b.ControlValues() {
55 b.ReplaceControl(i, c)
61 fmt.Printf("rewriting %s -> %s\n", b0.LongString(), b.LongString())
64 for j, v := range b.Values {
69 v0.Args = append([]*Value{}, v.Args...) // make a new copy, not aliasing
71 if v.Uses == 0 && v.removeable() {
72 if v.Op != OpInvalid && deadcode == removeDeadValues {
73 // Reset any values that are now unused, so that we decrement
74 // the use count of all of its arguments.
75 // Not quite a deadcode pass, because it does not handle cycles.
76 // But it should help Uses==1 rules to fire.
80 // No point rewriting values which aren't used.
84 vchange := phielimValue(v)
85 if vchange && debug > 1 {
86 fmt.Printf("rewriting %s -> %s\n", v0.LongString(), v.LongString())
89 // Eliminate copy inputs.
90 // If any copy input becomes unused, mark it
91 // as invalid and discard its argument. Repeat
92 // recursively on the discarded argument.
93 // This phase helps remove phantom "dead copy" uses
94 // of a value so that a x.Uses==1 rule condition
96 for i, a := range v.Args {
102 // If a, a copy, has a line boundary indicator, attempt to find a new value
103 // to hold it. The first candidate is the value that will replace a (aa),
104 // if it shares the same block and line and is eligible.
105 // The second option is v, which has a as an input. Because aa is earlier in
106 // the data flow, it is the better choice.
107 if a.Pos.IsStmt() == src.PosIsStmt {
108 if aa.Block == a.Block && aa.Pos.Line() == a.Pos.Line() && aa.Pos.IsStmt() != src.PosNotStmt {
109 aa.Pos = aa.Pos.WithIsStmt()
110 } else if v.Block == a.Block && v.Pos.Line() == a.Pos.Line() && v.Pos.IsStmt() != src.PosNotStmt {
111 v.Pos = v.Pos.WithIsStmt()
113 // Record the lost line and look for a new home after all rewrites are complete.
114 // TODO: it's possible (in FOR loops, in particular) for statement boundaries for the same
115 // line to appear in more than one block, but only one block is stored, so if both end
116 // up here, then one will be lost.
117 pendingLines.set(a.Pos, int32(a.Block.ID))
119 a.Pos = a.Pos.WithNotStmt()
128 if vchange && debug > 1 {
129 fmt.Printf("rewriting %s -> %s\n", v0.LongString(), v.LongString())
132 // apply rewrite function
135 // If value changed to a poor choice for a statement boundary, move the boundary
136 if v.Pos.IsStmt() == src.PosIsStmt {
137 if k := nextGoodStatementIndex(v, j, b); k != j {
138 v.Pos = v.Pos.WithNotStmt()
139 b.Values[k].Pos = b.Values[k].Pos.WithIsStmt()
144 change = change || vchange
145 if vchange && debug > 1 {
146 fmt.Printf("rewriting %s -> %s\n", v0.LongString(), v.LongString())
150 if !change && !deadChange {
154 if (iters > 1000 || debug >= 2) && change {
155 // We've done a suspiciously large number of rewrites (or we're in debug mode).
156 // As of Sep 2021, 90% of rewrites complete in 4 iterations or fewer
157 // and the maximum value encountered during make.bash is 12.
158 // Start checking for cycles. (This is too expensive to do routinely.)
159 // Note: we avoid this path for deadChange-only iterations, to fix #51639.
161 states = make(map[string]bool)
164 if _, ok := states[h]; ok {
165 // We've found a cycle.
166 // To diagnose it, set debug to 2 and start again,
167 // so that we'll print all rules applied until we complete another cycle.
168 // If debug is already >= 2, we've already done that, so it's time to crash.
171 states = make(map[string]bool)
173 f.Fatalf("rewrite cycle detected")
179 // remove clobbered values
180 for _, b := range f.Blocks {
182 for i, v := range b.Values {
184 if v.Op == OpInvalid {
185 if v.Pos.IsStmt() == src.PosIsStmt {
186 pendingLines.set(vl, int32(b.ID))
191 if v.Pos.IsStmt() != src.PosNotStmt && !notStmtBoundary(v.Op) && pendingLines.get(vl) == int32(b.ID) {
192 pendingLines.remove(vl)
193 v.Pos = v.Pos.WithIsStmt()
200 if pendingLines.get(b.Pos) == int32(b.ID) {
201 b.Pos = b.Pos.WithIsStmt()
202 pendingLines.remove(b.Pos)
208 // Common functions called from rewriting rules
210 func is64BitFloat(t *types.Type) bool {
211 return t.Size() == 8 && t.IsFloat()
214 func is32BitFloat(t *types.Type) bool {
215 return t.Size() == 4 && t.IsFloat()
218 func is64BitInt(t *types.Type) bool {
219 return t.Size() == 8 && t.IsInteger()
222 func is32BitInt(t *types.Type) bool {
223 return t.Size() == 4 && t.IsInteger()
226 func is16BitInt(t *types.Type) bool {
227 return t.Size() == 2 && t.IsInteger()
230 func is8BitInt(t *types.Type) bool {
231 return t.Size() == 1 && t.IsInteger()
234 func isPtr(t *types.Type) bool {
235 return t.IsPtrShaped()
238 func isSigned(t *types.Type) bool {
242 // mergeSym merges two symbolic offsets. There is no real merging of
243 // offsets, we just pick the non-nil one.
244 func mergeSym(x, y Sym) Sym {
251 panic(fmt.Sprintf("mergeSym with two non-nil syms %v %v", x, y))
254 func canMergeSym(x, y Sym) bool {
255 return x == nil || y == nil
258 // canMergeLoadClobber reports whether the load can be merged into target without
259 // invalidating the schedule.
260 // It also checks that the other non-load argument x is something we
261 // are ok with clobbering.
262 func canMergeLoadClobber(target, load, x *Value) bool {
263 // The register containing x is going to get clobbered.
264 // Don't merge if we still need the value of x.
265 // We don't have liveness information here, but we can
266 // approximate x dying with:
267 // 1) target is x's only use.
268 // 2) target is not in a deeper loop than x.
272 loopnest := x.Block.Func.loopnest()
273 loopnest.calculateDepths()
274 if loopnest.depth(target.Block.ID) > loopnest.depth(x.Block.ID) {
277 return canMergeLoad(target, load)
280 // canMergeLoad reports whether the load can be merged into target without
281 // invalidating the schedule.
282 func canMergeLoad(target, load *Value) bool {
283 if target.Block.ID != load.Block.ID {
284 // If the load is in a different block do not merge it.
288 // We can't merge the load into the target if the load
289 // has more than one use.
294 mem := load.MemoryArg()
296 // We need the load's memory arg to still be alive at target. That
297 // can't be the case if one of target's args depends on a memory
298 // state that is a successor of load's memory arg.
300 // For example, it would be invalid to merge load into target in
301 // the following situation because newmem has killed oldmem
302 // before target is reached:
303 // load = read ... oldmem
304 // newmem = write ... oldmem
305 // arg0 = read ... newmem
306 // target = add arg0 load
308 // If the argument comes from a different block then we can exclude
309 // it immediately because it must dominate load (which is in the
310 // same block as target).
312 for _, a := range target.Args {
313 if a != load && a.Block.ID == target.Block.ID {
314 args = append(args, a)
318 // memPreds contains memory states known to be predecessors of load's
319 // memory state. It is lazily initialized.
320 var memPreds map[*Value]bool
321 for i := 0; len(args) > 0; i++ {
324 // Give up if we have done a lot of iterations.
327 v := args[len(args)-1]
328 args = args[:len(args)-1]
329 if target.Block.ID != v.Block.ID {
330 // Since target and load are in the same block
331 // we can stop searching when we leave the block.
335 // A Phi implies we have reached the top of the block.
336 // The memory phi, if it exists, is always
337 // the first logical store in the block.
340 if v.Type.IsTuple() && v.Type.FieldType(1).IsMemory() {
341 // We could handle this situation however it is likely
345 if v.Op.SymEffect()&SymAddr != 0 {
346 // This case prevents an operation that calculates the
347 // address of a local variable from being forced to schedule
348 // before its corresponding VarDef.
354 // We don't want to combine the CMPQ with the load, because
355 // that would force the CMPQ to schedule before the VARDEF, which
356 // in turn requires the LEAQ to schedule before the VARDEF.
359 if v.Type.IsMemory() {
361 // Initialise a map containing memory states
362 // known to be predecessors of load's memory
364 memPreds = make(map[*Value]bool)
367 for i := 0; i < limit; i++ {
369 // The memory phi, if it exists, is always
370 // the first logical store in the block.
373 if m.Block.ID != target.Block.ID {
376 if !m.Type.IsMemory() {
380 if len(m.Args) == 0 {
387 // We can merge if v is a predecessor of mem.
389 // For example, we can merge load into target in the
390 // following scenario:
393 // load = read ... mem
394 // target = add x load
400 if len(v.Args) > 0 && v.Args[len(v.Args)-1] == mem {
401 // If v takes mem as an input then we know mem
402 // is valid at this point.
405 for _, a := range v.Args {
406 if target.Block.ID == a.Block.ID {
407 args = append(args, a)
415 // isSameCall reports whether sym is the same as the given named symbol.
416 func isSameCall(sym interface{}, name string) bool {
417 fn := sym.(*AuxCall).Fn
418 return fn != nil && fn.String() == name
421 // canLoadUnaligned reports if the architecture supports unaligned load operations.
422 func canLoadUnaligned(c *Config) bool {
423 return c.ctxt.Arch.Alignment == 1
426 // nlzX returns the number of leading zeros.
427 func nlz64(x int64) int { return bits.LeadingZeros64(uint64(x)) }
428 func nlz32(x int32) int { return bits.LeadingZeros32(uint32(x)) }
429 func nlz16(x int16) int { return bits.LeadingZeros16(uint16(x)) }
430 func nlz8(x int8) int { return bits.LeadingZeros8(uint8(x)) }
432 // ntzX returns the number of trailing zeros.
433 func ntz64(x int64) int { return bits.TrailingZeros64(uint64(x)) }
434 func ntz32(x int32) int { return bits.TrailingZeros32(uint32(x)) }
435 func ntz16(x int16) int { return bits.TrailingZeros16(uint16(x)) }
436 func ntz8(x int8) int { return bits.TrailingZeros8(uint8(x)) }
438 func oneBit(x int64) bool { return x&(x-1) == 0 && x != 0 }
439 func oneBit8(x int8) bool { return x&(x-1) == 0 && x != 0 }
440 func oneBit16(x int16) bool { return x&(x-1) == 0 && x != 0 }
441 func oneBit32(x int32) bool { return x&(x-1) == 0 && x != 0 }
442 func oneBit64(x int64) bool { return x&(x-1) == 0 && x != 0 }
444 // nto returns the number of trailing ones.
445 func nto(x int64) int64 {
446 return int64(ntz64(^x))
449 // logX returns logarithm of n base 2.
450 // n must be a positive power of 2 (isPowerOfTwoX returns true).
451 func log8(n int8) int64 {
452 return int64(bits.Len8(uint8(n))) - 1
454 func log16(n int16) int64 {
455 return int64(bits.Len16(uint16(n))) - 1
457 func log32(n int32) int64 {
458 return int64(bits.Len32(uint32(n))) - 1
460 func log64(n int64) int64 {
461 return int64(bits.Len64(uint64(n))) - 1
464 // log2uint32 returns logarithm in base 2 of uint32(n), with log2(0) = -1.
466 func log2uint32(n int64) int64 {
467 return int64(bits.Len32(uint32(n))) - 1
470 // isPowerOfTwoX functions report whether n is a power of 2.
471 func isPowerOfTwo8(n int8) bool {
472 return n > 0 && n&(n-1) == 0
474 func isPowerOfTwo16(n int16) bool {
475 return n > 0 && n&(n-1) == 0
477 func isPowerOfTwo32(n int32) bool {
478 return n > 0 && n&(n-1) == 0
480 func isPowerOfTwo64(n int64) bool {
481 return n > 0 && n&(n-1) == 0
484 // isUint64PowerOfTwo reports whether uint64(n) is a power of 2.
485 func isUint64PowerOfTwo(in int64) bool {
487 return n > 0 && n&(n-1) == 0
490 // isUint32PowerOfTwo reports whether uint32(n) is a power of 2.
491 func isUint32PowerOfTwo(in int64) bool {
492 n := uint64(uint32(in))
493 return n > 0 && n&(n-1) == 0
496 // is32Bit reports whether n can be represented as a signed 32 bit integer.
497 func is32Bit(n int64) bool {
498 return n == int64(int32(n))
501 // is16Bit reports whether n can be represented as a signed 16 bit integer.
502 func is16Bit(n int64) bool {
503 return n == int64(int16(n))
506 // is8Bit reports whether n can be represented as a signed 8 bit integer.
507 func is8Bit(n int64) bool {
508 return n == int64(int8(n))
511 // isU8Bit reports whether n can be represented as an unsigned 8 bit integer.
512 func isU8Bit(n int64) bool {
513 return n == int64(uint8(n))
516 // isU12Bit reports whether n can be represented as an unsigned 12 bit integer.
517 func isU12Bit(n int64) bool {
518 return 0 <= n && n < (1<<12)
521 // isU16Bit reports whether n can be represented as an unsigned 16 bit integer.
522 func isU16Bit(n int64) bool {
523 return n == int64(uint16(n))
526 // isU32Bit reports whether n can be represented as an unsigned 32 bit integer.
527 func isU32Bit(n int64) bool {
528 return n == int64(uint32(n))
531 // is20Bit reports whether n can be represented as a signed 20 bit integer.
532 func is20Bit(n int64) bool {
533 return -(1<<19) <= n && n < (1<<19)
536 // b2i translates a boolean value to 0 or 1 for assigning to auxInt.
537 func b2i(b bool) int64 {
544 // b2i32 translates a boolean value to 0 or 1.
545 func b2i32(b bool) int32 {
552 // shiftIsBounded reports whether (left/right) shift Value v is known to be bounded.
553 // A shift is bounded if it is shifting by less than the width of the shifted value.
554 func shiftIsBounded(v *Value) bool {
558 // canonLessThan returns whether x is "ordered" less than y, for purposes of normalizing
559 // generated code as much as possible.
560 func canonLessThan(x, y *Value) bool {
564 if !x.Pos.SameFileAndLine(y.Pos) {
565 return x.Pos.Before(y.Pos)
570 // truncate64Fto32F converts a float64 value to a float32 preserving the bit pattern
571 // of the mantissa. It will panic if the truncation results in lost information.
572 func truncate64Fto32F(f float64) float32 {
573 if !isExactFloat32(f) {
574 panic("truncate64Fto32F: truncation is not exact")
579 // NaN bit patterns aren't necessarily preserved across conversion
580 // instructions so we need to do the conversion manually.
581 b := math.Float64bits(f)
582 m := b & ((1 << 52) - 1) // mantissa (a.k.a. significand)
583 // | sign | exponent | mantissa |
584 r := uint32(((b >> 32) & (1 << 31)) | 0x7f800000 | (m >> (52 - 23)))
585 return math.Float32frombits(r)
588 // extend32Fto64F converts a float32 value to a float64 value preserving the bit
589 // pattern of the mantissa.
590 func extend32Fto64F(f float32) float64 {
591 if !math.IsNaN(float64(f)) {
594 // NaN bit patterns aren't necessarily preserved across conversion
595 // instructions so we need to do the conversion manually.
596 b := uint64(math.Float32bits(f))
597 // | sign | exponent | mantissa |
598 r := ((b << 32) & (1 << 63)) | (0x7ff << 52) | ((b & 0x7fffff) << (52 - 23))
599 return math.Float64frombits(r)
602 // DivisionNeedsFixUp reports whether the division needs fix-up code.
603 func DivisionNeedsFixUp(v *Value) bool {
607 // auxFrom64F encodes a float64 value so it can be stored in an AuxInt.
608 func auxFrom64F(f float64) int64 {
610 panic("can't encode a NaN in AuxInt field")
612 return int64(math.Float64bits(f))
615 // auxFrom32F encodes a float32 value so it can be stored in an AuxInt.
616 func auxFrom32F(f float32) int64 {
618 panic("can't encode a NaN in AuxInt field")
620 return int64(math.Float64bits(extend32Fto64F(f)))
623 // auxTo32F decodes a float32 from the AuxInt value provided.
624 func auxTo32F(i int64) float32 {
625 return truncate64Fto32F(math.Float64frombits(uint64(i)))
628 // auxTo64F decodes a float64 from the AuxInt value provided.
629 func auxTo64F(i int64) float64 {
630 return math.Float64frombits(uint64(i))
633 func auxIntToBool(i int64) bool {
639 func auxIntToInt8(i int64) int8 {
642 func auxIntToInt16(i int64) int16 {
645 func auxIntToInt32(i int64) int32 {
648 func auxIntToInt64(i int64) int64 {
651 func auxIntToUint8(i int64) uint8 {
654 func auxIntToFloat32(i int64) float32 {
655 return float32(math.Float64frombits(uint64(i)))
657 func auxIntToFloat64(i int64) float64 {
658 return math.Float64frombits(uint64(i))
660 func auxIntToValAndOff(i int64) ValAndOff {
663 func auxIntToArm64BitField(i int64) arm64BitField {
664 return arm64BitField(i)
666 func auxIntToInt128(x int64) int128 {
668 panic("nonzero int128 not allowed")
672 func auxIntToFlagConstant(x int64) flagConstant {
673 return flagConstant(x)
676 func auxIntToOp(cc int64) Op {
680 func boolToAuxInt(b bool) int64 {
686 func int8ToAuxInt(i int8) int64 {
689 func int16ToAuxInt(i int16) int64 {
692 func int32ToAuxInt(i int32) int64 {
695 func int64ToAuxInt(i int64) int64 {
698 func uint8ToAuxInt(i uint8) int64 {
699 return int64(int8(i))
701 func float32ToAuxInt(f float32) int64 {
702 return int64(math.Float64bits(float64(f)))
704 func float64ToAuxInt(f float64) int64 {
705 return int64(math.Float64bits(f))
707 func valAndOffToAuxInt(v ValAndOff) int64 {
710 func arm64BitFieldToAuxInt(v arm64BitField) int64 {
713 func int128ToAuxInt(x int128) int64 {
715 panic("nonzero int128 not allowed")
719 func flagConstantToAuxInt(x flagConstant) int64 {
723 func opToAuxInt(o Op) int64 {
727 // Aux is an interface to hold miscellaneous data in Blocks and Values.
732 // stringAux wraps string values for use in Aux.
733 type stringAux string
735 func (stringAux) CanBeAnSSAAux() {}
737 func auxToString(i Aux) string {
738 return string(i.(stringAux))
740 func auxToSym(i Aux) Sym {
741 // TODO: kind of a hack - allows nil interface through
745 func auxToType(i Aux) *types.Type {
746 return i.(*types.Type)
748 func auxToCall(i Aux) *AuxCall {
751 func auxToS390xCCMask(i Aux) s390x.CCMask {
752 return i.(s390x.CCMask)
754 func auxToS390xRotateParams(i Aux) s390x.RotateParams {
755 return i.(s390x.RotateParams)
758 func StringToAux(s string) Aux {
761 func symToAux(s Sym) Aux {
764 func callToAux(s *AuxCall) Aux {
767 func typeToAux(t *types.Type) Aux {
770 func s390xCCMaskToAux(c s390x.CCMask) Aux {
773 func s390xRotateParamsToAux(r s390x.RotateParams) Aux {
777 // uaddOvf reports whether unsigned a+b would overflow.
778 func uaddOvf(a, b int64) bool {
779 return uint64(a)+uint64(b) < uint64(a)
782 // loadLSymOffset simulates reading a word at an offset into a
783 // read-only symbol's runtime memory. If it would read a pointer to
784 // another symbol, that symbol is returned. Otherwise, it returns nil.
785 func loadLSymOffset(lsym *obj.LSym, offset int64) *obj.LSym {
786 if lsym.Type != objabi.SRODATA {
790 for _, r := range lsym.R {
791 if int64(r.Off) == offset && r.Type&^objabi.R_WEAK == objabi.R_ADDR && r.Add == 0 {
799 // de-virtualize an InterLECall
800 // 'sym' is the symbol for the itab.
801 func devirtLESym(v *Value, aux Aux, sym Sym, offset int64) *obj.LSym {
802 n, ok := sym.(*obj.LSym)
807 lsym := loadLSymOffset(n, offset)
808 if f := v.Block.Func; f.pass.debug > 0 {
810 f.Warnl(v.Pos, "de-virtualizing call")
812 f.Warnl(v.Pos, "couldn't de-virtualize call")
818 func devirtLECall(v *Value, sym *obj.LSym) *Value {
819 v.Op = OpStaticLECall
820 auxcall := v.Aux.(*AuxCall)
824 copy(v.Args[0:], v.Args[1:])
825 v.Args[len(v.Args)-1] = nil // aid GC
826 v.Args = v.Args[:len(v.Args)-1]
830 // isSamePtr reports whether p1 and p2 point to the same address.
831 func isSamePtr(p1, p2 *Value) bool {
840 return p1.AuxInt == p2.AuxInt && isSamePtr(p1.Args[0], p2.Args[0])
841 case OpAddr, OpLocalAddr:
842 return p1.Aux == p2.Aux
844 return p1.Args[1] == p2.Args[1] && isSamePtr(p1.Args[0], p2.Args[0])
849 func isStackPtr(v *Value) bool {
850 for v.Op == OpOffPtr || v.Op == OpAddPtr {
853 return v.Op == OpSP || v.Op == OpLocalAddr
856 // disjoint reports whether the memory region specified by [p1:p1+n1)
857 // does not overlap with [p2:p2+n2).
858 // A return value of false does not imply the regions overlap.
859 func disjoint(p1 *Value, n1 int64, p2 *Value, n2 int64) bool {
860 if n1 == 0 || n2 == 0 {
866 baseAndOffset := func(ptr *Value) (base *Value, offset int64) {
867 base, offset = ptr, 0
868 for base.Op == OpOffPtr {
869 offset += base.AuxInt
874 p1, off1 := baseAndOffset(p1)
875 p2, off2 := baseAndOffset(p2)
876 if isSamePtr(p1, p2) {
877 return !overlap(off1, n1, off2, n2)
879 // p1 and p2 are not the same, so if they are both OpAddrs then
880 // they point to different variables.
881 // If one pointer is on the stack and the other is an argument
882 // then they can't overlap.
884 case OpAddr, OpLocalAddr:
885 if p2.Op == OpAddr || p2.Op == OpLocalAddr || p2.Op == OpSP {
888 return (p2.Op == OpArg || p2.Op == OpArgIntReg) && p1.Args[0].Op == OpSP
889 case OpArg, OpArgIntReg:
890 if p2.Op == OpSP || p2.Op == OpLocalAddr {
894 return p2.Op == OpAddr || p2.Op == OpLocalAddr || p2.Op == OpArg || p2.Op == OpArgIntReg || p2.Op == OpSP
899 // moveSize returns the number of bytes an aligned MOV instruction moves.
900 func moveSize(align int64, c *Config) int64 {
902 case align%8 == 0 && c.PtrSize == 8:
912 // mergePoint finds a block among a's blocks which dominates b and is itself
913 // dominated by all of a's blocks. Returns nil if it can't find one.
914 // Might return nil even if one does exist.
915 func mergePoint(b *Block, a ...*Value) *Block {
916 // Walk backward from b looking for one of the a's blocks.
922 for _, x := range a {
927 if len(b.Preds) > 1 {
928 // Don't know which way to go back. Abort.
934 return nil // too far away
936 // At this point, r is the first value in a that we find by walking backwards.
937 // if we return anything, r will be it.
940 // Keep going, counting the other a's that we find. They must all dominate r.
943 for _, x := range a {
949 // Found all of a in a backwards walk. We can return r.
952 if len(b.Preds) > 1 {
959 return nil // too far away
962 // clobber invalidates values. Returns true.
963 // clobber is used by rewrite rules to:
965 // A) make sure the values are really dead and never used again.
966 // B) decrement use counts of the values' args.
967 func clobber(vv ...*Value) bool {
968 for _, v := range vv {
970 // Note: leave v.Block intact. The Block field is used after clobber.
975 // clobberIfDead resets v when use count is 1. Returns true.
976 // clobberIfDead is used by rewrite rules to decrement
977 // use counts of v's args when v is dead and never used.
978 func clobberIfDead(v *Value) bool {
982 // Note: leave v.Block intact. The Block field is used after clobberIfDead.
986 // noteRule is an easy way to track if a rule is matched when writing
987 // new ones. Make the rule of interest also conditional on
989 // noteRule("note to self: rule of interest matched")
991 // and that message will print when the rule matches.
992 func noteRule(s string) bool {
997 // countRule increments Func.ruleMatches[key].
998 // If Func.ruleMatches is non-nil at the end
999 // of compilation, it will be printed to stdout.
1000 // This is intended to make it easier to find which functions
1001 // which contain lots of rules matches when developing new rules.
1002 func countRule(v *Value, key string) bool {
1004 if f.ruleMatches == nil {
1005 f.ruleMatches = make(map[string]int)
1007 f.ruleMatches[key]++
1011 // warnRule generates compiler debug output with string s when
1012 // v is not in autogenerated code, cond is true and the rule has fired.
1013 func warnRule(cond bool, v *Value, s string) bool {
1014 if pos := v.Pos; pos.Line() > 1 && cond {
1015 v.Block.Func.Warnl(pos, s)
1020 // for a pseudo-op like (LessThan x), extract x.
1021 func flagArg(v *Value) *Value {
1022 if len(v.Args) != 1 || !v.Args[0].Type.IsFlags() {
1028 // arm64Negate finds the complement to an ARM64 condition code,
1029 // for example !Equal -> NotEqual or !LessThan -> GreaterEqual
1031 // For floating point, it's more subtle because NaN is unordered. We do
1032 // !LessThanF -> NotLessThanF, the latter takes care of NaNs.
1033 func arm64Negate(op Op) Op {
1035 case OpARM64LessThan:
1036 return OpARM64GreaterEqual
1037 case OpARM64LessThanU:
1038 return OpARM64GreaterEqualU
1039 case OpARM64GreaterThan:
1040 return OpARM64LessEqual
1041 case OpARM64GreaterThanU:
1042 return OpARM64LessEqualU
1043 case OpARM64LessEqual:
1044 return OpARM64GreaterThan
1045 case OpARM64LessEqualU:
1046 return OpARM64GreaterThanU
1047 case OpARM64GreaterEqual:
1048 return OpARM64LessThan
1049 case OpARM64GreaterEqualU:
1050 return OpARM64LessThanU
1052 return OpARM64NotEqual
1053 case OpARM64NotEqual:
1055 case OpARM64LessThanF:
1056 return OpARM64NotLessThanF
1057 case OpARM64NotLessThanF:
1058 return OpARM64LessThanF
1059 case OpARM64LessEqualF:
1060 return OpARM64NotLessEqualF
1061 case OpARM64NotLessEqualF:
1062 return OpARM64LessEqualF
1063 case OpARM64GreaterThanF:
1064 return OpARM64NotGreaterThanF
1065 case OpARM64NotGreaterThanF:
1066 return OpARM64GreaterThanF
1067 case OpARM64GreaterEqualF:
1068 return OpARM64NotGreaterEqualF
1069 case OpARM64NotGreaterEqualF:
1070 return OpARM64GreaterEqualF
1072 panic("unreachable")
1076 // arm64Invert evaluates (InvertFlags op), which
1077 // is the same as altering the condition codes such
1078 // that the same result would be produced if the arguments
1079 // to the flag-generating instruction were reversed, e.g.
1080 // (InvertFlags (CMP x y)) -> (CMP y x)
1081 func arm64Invert(op Op) Op {
1083 case OpARM64LessThan:
1084 return OpARM64GreaterThan
1085 case OpARM64LessThanU:
1086 return OpARM64GreaterThanU
1087 case OpARM64GreaterThan:
1088 return OpARM64LessThan
1089 case OpARM64GreaterThanU:
1090 return OpARM64LessThanU
1091 case OpARM64LessEqual:
1092 return OpARM64GreaterEqual
1093 case OpARM64LessEqualU:
1094 return OpARM64GreaterEqualU
1095 case OpARM64GreaterEqual:
1096 return OpARM64LessEqual
1097 case OpARM64GreaterEqualU:
1098 return OpARM64LessEqualU
1099 case OpARM64Equal, OpARM64NotEqual:
1101 case OpARM64LessThanF:
1102 return OpARM64GreaterThanF
1103 case OpARM64GreaterThanF:
1104 return OpARM64LessThanF
1105 case OpARM64LessEqualF:
1106 return OpARM64GreaterEqualF
1107 case OpARM64GreaterEqualF:
1108 return OpARM64LessEqualF
1109 case OpARM64NotLessThanF:
1110 return OpARM64NotGreaterThanF
1111 case OpARM64NotGreaterThanF:
1112 return OpARM64NotLessThanF
1113 case OpARM64NotLessEqualF:
1114 return OpARM64NotGreaterEqualF
1115 case OpARM64NotGreaterEqualF:
1116 return OpARM64NotLessEqualF
1118 panic("unreachable")
1122 // evaluate an ARM64 op against a flags value
1123 // that is potentially constant; return 1 for true,
1124 // -1 for false, and 0 for not constant.
1125 func ccARM64Eval(op Op, flags *Value) int {
1127 if fop == OpARM64InvertFlags {
1128 return -ccARM64Eval(op, flags.Args[0])
1130 if fop != OpARM64FlagConstant {
1133 fc := flagConstant(flags.AuxInt)
1134 b2i := func(b bool) int {
1143 case OpARM64NotEqual:
1145 case OpARM64LessThan:
1147 case OpARM64LessThanU:
1148 return b2i(fc.ult())
1149 case OpARM64GreaterThan:
1151 case OpARM64GreaterThanU:
1152 return b2i(fc.ugt())
1153 case OpARM64LessEqual:
1155 case OpARM64LessEqualU:
1156 return b2i(fc.ule())
1157 case OpARM64GreaterEqual:
1159 case OpARM64GreaterEqualU:
1160 return b2i(fc.uge())
1165 // logRule logs the use of the rule s. This will only be enabled if
1166 // rewrite rules were generated with the -log option, see _gen/rulegen.go.
1167 func logRule(s string) {
1168 if ruleFile == nil {
1169 // Open a log file to write log to. We open in append
1170 // mode because all.bash runs the compiler lots of times,
1171 // and we want the concatenation of all of those logs.
1172 // This means, of course, that users need to rm the old log
1173 // to get fresh data.
1174 // TODO: all.bash runs compilers in parallel. Need to synchronize logging somehow?
1175 w, err := os.OpenFile(filepath.Join(os.Getenv("GOROOT"), "src", "rulelog"),
1176 os.O_CREATE|os.O_WRONLY|os.O_APPEND, 0666)
1182 _, err := fmt.Fprintln(ruleFile, s)
1188 var ruleFile io.Writer
1190 func min(x, y int64) int64 {
1197 func isConstZero(v *Value) bool {
1201 case OpConst64, OpConst32, OpConst16, OpConst8, OpConstBool, OpConst32F, OpConst64F:
1202 return v.AuxInt == 0
1207 // reciprocalExact64 reports whether 1/c is exactly representable.
1208 func reciprocalExact64(c float64) bool {
1209 b := math.Float64bits(c)
1210 man := b & (1<<52 - 1)
1212 return false // not a power of 2, denormal, or NaN
1214 exp := b >> 52 & (1<<11 - 1)
1215 // exponent bias is 0x3ff. So taking the reciprocal of a number
1216 // changes the exponent to 0x7fe-exp.
1221 return false // ±inf
1223 return false // exponent is not representable
1229 // reciprocalExact32 reports whether 1/c is exactly representable.
1230 func reciprocalExact32(c float32) bool {
1231 b := math.Float32bits(c)
1232 man := b & (1<<23 - 1)
1234 return false // not a power of 2, denormal, or NaN
1236 exp := b >> 23 & (1<<8 - 1)
1237 // exponent bias is 0x7f. So taking the reciprocal of a number
1238 // changes the exponent to 0xfe-exp.
1243 return false // ±inf
1245 return false // exponent is not representable
1251 // check if an immediate can be directly encoded into an ARM's instruction.
1252 func isARMImmRot(v uint32) bool {
1253 for i := 0; i < 16; i++ {
1263 // overlap reports whether the ranges given by the given offset and
1264 // size pairs overlap.
1265 func overlap(offset1, size1, offset2, size2 int64) bool {
1266 if offset1 >= offset2 && offset2+size2 > offset1 {
1269 if offset2 >= offset1 && offset1+size1 > offset2 {
1275 func areAdjacentOffsets(off1, off2, size int64) bool {
1276 return off1+size == off2 || off1 == off2+size
1279 // check if value zeroes out upper 32-bit of 64-bit register.
1280 // depth limits recursion depth. In AMD64.rules 3 is used as limit,
1281 // because it catches same amount of cases as 4.
1282 func zeroUpper32Bits(x *Value, depth int) bool {
1284 case OpAMD64MOVLconst, OpAMD64MOVLload, OpAMD64MOVLQZX, OpAMD64MOVLloadidx1,
1285 OpAMD64MOVWload, OpAMD64MOVWloadidx1, OpAMD64MOVBload, OpAMD64MOVBloadidx1,
1286 OpAMD64MOVLloadidx4, OpAMD64ADDLload, OpAMD64SUBLload, OpAMD64ANDLload,
1287 OpAMD64ORLload, OpAMD64XORLload, OpAMD64CVTTSD2SL,
1288 OpAMD64ADDL, OpAMD64ADDLconst, OpAMD64SUBL, OpAMD64SUBLconst,
1289 OpAMD64ANDL, OpAMD64ANDLconst, OpAMD64ORL, OpAMD64ORLconst,
1290 OpAMD64XORL, OpAMD64XORLconst, OpAMD64NEGL, OpAMD64NOTL,
1291 OpAMD64SHRL, OpAMD64SHRLconst, OpAMD64SARL, OpAMD64SARLconst,
1292 OpAMD64SHLL, OpAMD64SHLLconst:
1295 return x.Type.Size() == 4
1296 case OpPhi, OpSelect0, OpSelect1:
1297 // Phis can use each-other as an arguments, instead of tracking visited values,
1298 // just limit recursion depth.
1302 for i := range x.Args {
1303 if !zeroUpper32Bits(x.Args[i], depth-1) {
1313 // zeroUpper48Bits is similar to zeroUpper32Bits, but for upper 48 bits.
1314 func zeroUpper48Bits(x *Value, depth int) bool {
1316 case OpAMD64MOVWQZX, OpAMD64MOVWload, OpAMD64MOVWloadidx1, OpAMD64MOVWloadidx2:
1319 return x.Type.Size() == 2
1320 case OpPhi, OpSelect0, OpSelect1:
1321 // Phis can use each-other as an arguments, instead of tracking visited values,
1322 // just limit recursion depth.
1326 for i := range x.Args {
1327 if !zeroUpper48Bits(x.Args[i], depth-1) {
1337 // zeroUpper56Bits is similar to zeroUpper32Bits, but for upper 56 bits.
1338 func zeroUpper56Bits(x *Value, depth int) bool {
1340 case OpAMD64MOVBQZX, OpAMD64MOVBload, OpAMD64MOVBloadidx1:
1343 return x.Type.Size() == 1
1344 case OpPhi, OpSelect0, OpSelect1:
1345 // Phis can use each-other as an arguments, instead of tracking visited values,
1346 // just limit recursion depth.
1350 for i := range x.Args {
1351 if !zeroUpper56Bits(x.Args[i], depth-1) {
1361 // isInlinableMemmove reports whether the given arch performs a Move of the given size
1362 // faster than memmove. It will only return true if replacing the memmove with a Move is
1363 // safe, either because Move will do all of its loads before any of its stores, or
1364 // because the arguments are known to be disjoint.
1365 // This is used as a check for replacing memmove with Move ops.
1366 func isInlinableMemmove(dst, src *Value, sz int64, c *Config) bool {
1367 // It is always safe to convert memmove into Move when its arguments are disjoint.
1368 // Move ops may or may not be faster for large sizes depending on how the platform
1369 // lowers them, so we only perform this optimization on platforms that we know to
1370 // have fast Move ops.
1373 return sz <= 16 || (sz < 1024 && disjoint(dst, sz, src, sz))
1374 case "386", "arm64":
1376 case "s390x", "ppc64", "ppc64le":
1377 return sz <= 8 || disjoint(dst, sz, src, sz)
1378 case "arm", "loong64", "mips", "mips64", "mipsle", "mips64le":
1383 func IsInlinableMemmove(dst, src *Value, sz int64, c *Config) bool {
1384 return isInlinableMemmove(dst, src, sz, c)
1387 // logLargeCopy logs the occurrence of a large copy.
1388 // The best place to do this is in the rewrite rules where the size of the move is easy to find.
1389 // "Large" is arbitrarily chosen to be 128 bytes; this may change.
1390 func logLargeCopy(v *Value, s int64) bool {
1394 if logopt.Enabled() {
1395 logopt.LogOpt(v.Pos, "copy", "lower", v.Block.Func.Name, fmt.Sprintf("%d bytes", s))
1399 func LogLargeCopy(funcName string, pos src.XPos, s int64) {
1403 if logopt.Enabled() {
1404 logopt.LogOpt(pos, "copy", "lower", funcName, fmt.Sprintf("%d bytes", s))
1408 // hasSmallRotate reports whether the architecture has rotate instructions
1409 // for sizes < 32-bit. This is used to decide whether to promote some rotations.
1410 func hasSmallRotate(c *Config) bool {
1412 case "amd64", "386":
1419 func newPPC64ShiftAuxInt(sh, mb, me, sz int64) int32 {
1420 if sh < 0 || sh >= sz {
1421 panic("PPC64 shift arg sh out of range")
1423 if mb < 0 || mb >= sz {
1424 panic("PPC64 shift arg mb out of range")
1426 if me < 0 || me >= sz {
1427 panic("PPC64 shift arg me out of range")
1429 return int32(sh<<16 | mb<<8 | me)
1432 func GetPPC64Shiftsh(auxint int64) int64 {
1433 return int64(int8(auxint >> 16))
1436 func GetPPC64Shiftmb(auxint int64) int64 {
1437 return int64(int8(auxint >> 8))
1440 func GetPPC64Shiftme(auxint int64) int64 {
1441 return int64(int8(auxint))
1444 // Test if this value can encoded as a mask for a rlwinm like
1445 // operation. Masks can also extend from the msb and wrap to
1446 // the lsb too. That is, the valid masks are 32 bit strings
1447 // of the form: 0..01..10..0 or 1..10..01..1 or 1...1
1448 func isPPC64WordRotateMask(v64 int64) bool {
1449 // Isolate rightmost 1 (if none 0) and add.
1452 // Likewise, for the wrapping case.
1454 vpn := (vn & -vn) + vn
1455 return (v&vp == 0 || vn&vpn == 0) && v != 0
1458 // Compress mask and shift into single value of the form
1459 // me | mb<<8 | rotate<<16 | nbits<<24 where me and mb can
1460 // be used to regenerate the input mask.
1461 func encodePPC64RotateMask(rotate, mask, nbits int64) int64 {
1462 var mb, me, mbn, men int
1464 // Determine boundaries and then decode them
1465 if mask == 0 || ^mask == 0 || rotate >= nbits {
1466 panic("Invalid PPC64 rotate mask")
1467 } else if nbits == 32 {
1468 mb = bits.LeadingZeros32(uint32(mask))
1469 me = 32 - bits.TrailingZeros32(uint32(mask))
1470 mbn = bits.LeadingZeros32(^uint32(mask))
1471 men = 32 - bits.TrailingZeros32(^uint32(mask))
1473 mb = bits.LeadingZeros64(uint64(mask))
1474 me = 64 - bits.TrailingZeros64(uint64(mask))
1475 mbn = bits.LeadingZeros64(^uint64(mask))
1476 men = 64 - bits.TrailingZeros64(^uint64(mask))
1478 // Check for a wrapping mask (e.g bits at 0 and 63)
1479 if mb == 0 && me == int(nbits) {
1480 // swap the inverted values
1484 return int64(me) | int64(mb<<8) | int64(rotate<<16) | int64(nbits<<24)
1487 // DecodePPC64RotateMask is the inverse operation of encodePPC64RotateMask. The values returned as
1488 // mb and me satisfy the POWER ISA definition of MASK(x,y) where MASK(mb,me) = mask.
1489 func DecodePPC64RotateMask(sauxint int64) (rotate, mb, me int64, mask uint64) {
1490 auxint := uint64(sauxint)
1491 rotate = int64((auxint >> 16) & 0xFF)
1492 mb = int64((auxint >> 8) & 0xFF)
1493 me = int64((auxint >> 0) & 0xFF)
1494 nbits := int64((auxint >> 24) & 0xFF)
1495 mask = ((1 << uint(nbits-mb)) - 1) ^ ((1 << uint(nbits-me)) - 1)
1500 mask = uint64(uint32(mask))
1503 // Fixup ME to match ISA definition. The second argument to MASK(..,me)
1505 me = (me - 1) & (nbits - 1)
1509 // This verifies that the mask is a set of
1510 // consecutive bits including the least
1512 func isPPC64ValidShiftMask(v int64) bool {
1513 if (v != 0) && ((v+1)&v) == 0 {
1519 func getPPC64ShiftMaskLength(v int64) int64 {
1520 return int64(bits.Len64(uint64(v)))
1523 // Decompose a shift right into an equivalent rotate/mask,
1524 // and return mask & m.
1525 func mergePPC64RShiftMask(m, s, nbits int64) int64 {
1526 smask := uint64((1<<uint(nbits))-1) >> uint(s)
1527 return m & int64(smask)
1530 // Combine (ANDconst [m] (SRWconst [s])) into (RLWINM [y]) or return 0
1531 func mergePPC64AndSrwi(m, s int64) int64 {
1532 mask := mergePPC64RShiftMask(m, s, 32)
1533 if !isPPC64WordRotateMask(mask) {
1536 return encodePPC64RotateMask((32-s)&31, mask, 32)
1539 // Test if a shift right feeding into a CLRLSLDI can be merged into RLWINM.
1540 // Return the encoded RLWINM constant, or 0 if they cannot be merged.
1541 func mergePPC64ClrlsldiSrw(sld, srw int64) int64 {
1542 mask_1 := uint64(0xFFFFFFFF >> uint(srw))
1543 // for CLRLSLDI, it's more convient to think of it as a mask left bits then rotate left.
1544 mask_2 := uint64(0xFFFFFFFFFFFFFFFF) >> uint(GetPPC64Shiftmb(int64(sld)))
1546 // Rewrite mask to apply after the final left shift.
1547 mask_3 := (mask_1 & mask_2) << uint(GetPPC64Shiftsh(sld))
1550 r_2 := GetPPC64Shiftsh(sld)
1551 r_3 := (r_1 + r_2) & 31 // This can wrap.
1553 if uint64(uint32(mask_3)) != mask_3 || mask_3 == 0 {
1556 return encodePPC64RotateMask(int64(r_3), int64(mask_3), 32)
1559 // Test if a RLWINM feeding into a CLRLSLDI can be merged into RLWINM. Return
1560 // the encoded RLWINM constant, or 0 if they cannot be merged.
1561 func mergePPC64ClrlsldiRlwinm(sld int32, rlw int64) int64 {
1562 r_1, _, _, mask_1 := DecodePPC64RotateMask(rlw)
1563 // for CLRLSLDI, it's more convient to think of it as a mask left bits then rotate left.
1564 mask_2 := uint64(0xFFFFFFFFFFFFFFFF) >> uint(GetPPC64Shiftmb(int64(sld)))
1566 // combine the masks, and adjust for the final left shift.
1567 mask_3 := (mask_1 & mask_2) << uint(GetPPC64Shiftsh(int64(sld)))
1568 r_2 := GetPPC64Shiftsh(int64(sld))
1569 r_3 := (r_1 + r_2) & 31 // This can wrap.
1571 // Verify the result is still a valid bitmask of <= 32 bits.
1572 if !isPPC64WordRotateMask(int64(mask_3)) || uint64(uint32(mask_3)) != mask_3 {
1575 return encodePPC64RotateMask(r_3, int64(mask_3), 32)
1578 // Compute the encoded RLWINM constant from combining (SLDconst [sld] (SRWconst [srw] x)),
1579 // or return 0 if they cannot be combined.
1580 func mergePPC64SldiSrw(sld, srw int64) int64 {
1581 if sld > srw || srw >= 32 {
1584 mask_r := uint32(0xFFFFFFFF) >> uint(srw)
1585 mask_l := uint32(0xFFFFFFFF) >> uint(sld)
1586 mask := (mask_r & mask_l) << uint(sld)
1587 return encodePPC64RotateMask((32-srw+sld)&31, int64(mask), 32)
1590 // Convenience function to rotate a 32 bit constant value by another constant.
1591 func rotateLeft32(v, rotate int64) int64 {
1592 return int64(bits.RotateLeft32(uint32(v), int(rotate)))
1595 func rotateRight64(v, rotate int64) int64 {
1596 return int64(bits.RotateLeft64(uint64(v), int(-rotate)))
1599 // encodes the lsb and width for arm(64) bitfield ops into the expected auxInt format.
1600 func armBFAuxInt(lsb, width int64) arm64BitField {
1601 if lsb < 0 || lsb > 63 {
1602 panic("ARM(64) bit field lsb constant out of range")
1604 if width < 1 || lsb+width > 64 {
1605 panic("ARM(64) bit field width constant out of range")
1607 return arm64BitField(width | lsb<<8)
1610 // returns the lsb part of the auxInt field of arm64 bitfield ops.
1611 func (bfc arm64BitField) getARM64BFlsb() int64 {
1612 return int64(uint64(bfc) >> 8)
1615 // returns the width part of the auxInt field of arm64 bitfield ops.
1616 func (bfc arm64BitField) getARM64BFwidth() int64 {
1617 return int64(bfc) & 0xff
1620 // checks if mask >> rshift applied at lsb is a valid arm64 bitfield op mask.
1621 func isARM64BFMask(lsb, mask, rshift int64) bool {
1622 shiftedMask := int64(uint64(mask) >> uint64(rshift))
1623 return shiftedMask != 0 && isPowerOfTwo64(shiftedMask+1) && nto(shiftedMask)+lsb < 64
1626 // returns the bitfield width of mask >> rshift for arm64 bitfield ops.
1627 func arm64BFWidth(mask, rshift int64) int64 {
1628 shiftedMask := int64(uint64(mask) >> uint64(rshift))
1629 if shiftedMask == 0 {
1630 panic("ARM64 BF mask is zero")
1632 return nto(shiftedMask)
1635 // sizeof returns the size of t in bytes.
1636 // It will panic if t is not a *types.Type.
1637 func sizeof(t interface{}) int64 {
1638 return t.(*types.Type).Size()
1641 // registerizable reports whether t is a primitive type that fits in
1642 // a register. It assumes float64 values will always fit into registers
1643 // even if that isn't strictly true.
1644 func registerizable(b *Block, typ *types.Type) bool {
1645 if typ.IsPtrShaped() || typ.IsFloat() || typ.IsBoolean() {
1648 if typ.IsInteger() {
1649 return typ.Size() <= b.Func.Config.RegSize
1654 // needRaceCleanup reports whether this call to racefuncenter/exit isn't needed.
1655 func needRaceCleanup(sym *AuxCall, v *Value) bool {
1660 if !isSameCall(sym, "runtime.racefuncenter") && !isSameCall(sym, "runtime.racefuncexit") {
1663 for _, b := range f.Blocks {
1664 for _, v := range b.Values {
1666 case OpStaticCall, OpStaticLECall:
1667 // Check for racefuncenter will encounter racefuncexit and vice versa.
1668 // Allow calls to panic*
1669 s := v.Aux.(*AuxCall).Fn.String()
1671 case "runtime.racefuncenter", "runtime.racefuncexit",
1672 "runtime.panicdivide", "runtime.panicwrap",
1673 "runtime.panicshift":
1676 // If we encountered any call, we need to keep racefunc*,
1677 // for accurate stacktraces.
1679 case OpPanicBounds, OpPanicExtend:
1680 // Note: these are panic generators that are ok (like the static calls above).
1681 case OpClosureCall, OpInterCall, OpClosureLECall, OpInterLECall:
1682 // We must keep the race functions if there are any other call types.
1687 if isSameCall(sym, "runtime.racefuncenter") {
1688 // TODO REGISTER ABI this needs to be cleaned up.
1689 // If we're removing racefuncenter, remove its argument as well.
1690 if v.Args[0].Op != OpStore {
1691 if v.Op == OpStaticLECall {
1692 // there is no store, yet.
1697 mem := v.Args[0].Args[2]
1698 v.Args[0].reset(OpCopy)
1699 v.Args[0].AddArg(mem)
1704 // symIsRO reports whether sym is a read-only global.
1705 func symIsRO(sym interface{}) bool {
1706 lsym := sym.(*obj.LSym)
1707 return lsym.Type == objabi.SRODATA && len(lsym.R) == 0
1710 // symIsROZero reports whether sym is a read-only global whose data contains all zeros.
1711 func symIsROZero(sym Sym) bool {
1712 lsym := sym.(*obj.LSym)
1713 if lsym.Type != objabi.SRODATA || len(lsym.R) != 0 {
1716 for _, b := range lsym.P {
1724 // read8 reads one byte from the read-only global sym at offset off.
1725 func read8(sym interface{}, off int64) uint8 {
1726 lsym := sym.(*obj.LSym)
1727 if off >= int64(len(lsym.P)) || off < 0 {
1728 // Invalid index into the global sym.
1729 // This can happen in dead code, so we don't want to panic.
1730 // Just return any value, it will eventually get ignored.
1737 // read16 reads two bytes from the read-only global sym at offset off.
1738 func read16(sym interface{}, off int64, byteorder binary.ByteOrder) uint16 {
1739 lsym := sym.(*obj.LSym)
1740 // lsym.P is written lazily.
1741 // Bytes requested after the end of lsym.P are 0.
1743 if 0 <= off && off < int64(len(lsym.P)) {
1746 buf := make([]byte, 2)
1748 return byteorder.Uint16(buf)
1751 // read32 reads four bytes from the read-only global sym at offset off.
1752 func read32(sym interface{}, off int64, byteorder binary.ByteOrder) uint32 {
1753 lsym := sym.(*obj.LSym)
1755 if 0 <= off && off < int64(len(lsym.P)) {
1758 buf := make([]byte, 4)
1760 return byteorder.Uint32(buf)
1763 // read64 reads eight bytes from the read-only global sym at offset off.
1764 func read64(sym interface{}, off int64, byteorder binary.ByteOrder) uint64 {
1765 lsym := sym.(*obj.LSym)
1767 if 0 <= off && off < int64(len(lsym.P)) {
1770 buf := make([]byte, 8)
1772 return byteorder.Uint64(buf)
1775 // sequentialAddresses reports true if it can prove that x + n == y
1776 func sequentialAddresses(x, y *Value, n int64) bool {
1777 if x == y && n == 0 {
1780 if x.Op == Op386ADDL && y.Op == Op386LEAL1 && y.AuxInt == n && y.Aux == nil &&
1781 (x.Args[0] == y.Args[0] && x.Args[1] == y.Args[1] ||
1782 x.Args[0] == y.Args[1] && x.Args[1] == y.Args[0]) {
1785 if x.Op == Op386LEAL1 && y.Op == Op386LEAL1 && y.AuxInt == x.AuxInt+n && x.Aux == y.Aux &&
1786 (x.Args[0] == y.Args[0] && x.Args[1] == y.Args[1] ||
1787 x.Args[0] == y.Args[1] && x.Args[1] == y.Args[0]) {
1790 if x.Op == OpAMD64ADDQ && y.Op == OpAMD64LEAQ1 && y.AuxInt == n && y.Aux == nil &&
1791 (x.Args[0] == y.Args[0] && x.Args[1] == y.Args[1] ||
1792 x.Args[0] == y.Args[1] && x.Args[1] == y.Args[0]) {
1795 if x.Op == OpAMD64LEAQ1 && y.Op == OpAMD64LEAQ1 && y.AuxInt == x.AuxInt+n && x.Aux == y.Aux &&
1796 (x.Args[0] == y.Args[0] && x.Args[1] == y.Args[1] ||
1797 x.Args[0] == y.Args[1] && x.Args[1] == y.Args[0]) {
1803 // flagConstant represents the result of a compile-time comparison.
1804 // The sense of these flags does not necessarily represent the hardware's notion
1805 // of a flags register - these are just a compile-time construct.
1806 // We happen to match the semantics to those of arm/arm64.
1807 // Note that these semantics differ from x86: the carry flag has the opposite
1808 // sense on a subtraction!
1810 // On amd64, C=1 represents a borrow, e.g. SBB on amd64 does x - y - C.
1811 // On arm64, C=0 represents a borrow, e.g. SBC on arm64 does x - y - ^C.
1812 // (because it does x + ^y + C).
1814 // See https://en.wikipedia.org/wiki/Carry_flag#Vs._borrow_flag
1815 type flagConstant uint8
1817 // N reports whether the result of an operation is negative (high bit set).
1818 func (fc flagConstant) N() bool {
1822 // Z reports whether the result of an operation is 0.
1823 func (fc flagConstant) Z() bool {
1827 // C reports whether an unsigned add overflowed (carry), or an
1828 // unsigned subtract did not underflow (borrow).
1829 func (fc flagConstant) C() bool {
1833 // V reports whether a signed operation overflowed or underflowed.
1834 func (fc flagConstant) V() bool {
1838 func (fc flagConstant) eq() bool {
1841 func (fc flagConstant) ne() bool {
1844 func (fc flagConstant) lt() bool {
1845 return fc.N() != fc.V()
1847 func (fc flagConstant) le() bool {
1848 return fc.Z() || fc.lt()
1850 func (fc flagConstant) gt() bool {
1851 return !fc.Z() && fc.ge()
1853 func (fc flagConstant) ge() bool {
1854 return fc.N() == fc.V()
1856 func (fc flagConstant) ult() bool {
1859 func (fc flagConstant) ule() bool {
1860 return fc.Z() || fc.ult()
1862 func (fc flagConstant) ugt() bool {
1863 return !fc.Z() && fc.uge()
1865 func (fc flagConstant) uge() bool {
1869 func (fc flagConstant) ltNoov() bool {
1870 return fc.lt() && !fc.V()
1872 func (fc flagConstant) leNoov() bool {
1873 return fc.le() && !fc.V()
1875 func (fc flagConstant) gtNoov() bool {
1876 return fc.gt() && !fc.V()
1878 func (fc flagConstant) geNoov() bool {
1879 return fc.ge() && !fc.V()
1882 func (fc flagConstant) String() string {
1883 return fmt.Sprintf("N=%v,Z=%v,C=%v,V=%v", fc.N(), fc.Z(), fc.C(), fc.V())
1886 type flagConstantBuilder struct {
1893 func (fcs flagConstantBuilder) encode() flagConstant {
1910 // Note: addFlags(x,y) != subFlags(x,-y) in some situations:
1911 // - the results of the C flag are different
1912 // - the results of the V flag when y==minint are different
1914 // addFlags64 returns the flags that would be set from computing x+y.
1915 func addFlags64(x, y int64) flagConstant {
1916 var fcb flagConstantBuilder
1919 fcb.C = uint64(x+y) < uint64(x)
1920 fcb.V = x >= 0 && y >= 0 && x+y < 0 || x < 0 && y < 0 && x+y >= 0
1924 // subFlags64 returns the flags that would be set from computing x-y.
1925 func subFlags64(x, y int64) flagConstant {
1926 var fcb flagConstantBuilder
1929 fcb.C = uint64(y) <= uint64(x) // This code follows the arm carry flag model.
1930 fcb.V = x >= 0 && y < 0 && x-y < 0 || x < 0 && y >= 0 && x-y >= 0
1934 // addFlags32 returns the flags that would be set from computing x+y.
1935 func addFlags32(x, y int32) flagConstant {
1936 var fcb flagConstantBuilder
1939 fcb.C = uint32(x+y) < uint32(x)
1940 fcb.V = x >= 0 && y >= 0 && x+y < 0 || x < 0 && y < 0 && x+y >= 0
1944 // subFlags32 returns the flags that would be set from computing x-y.
1945 func subFlags32(x, y int32) flagConstant {
1946 var fcb flagConstantBuilder
1949 fcb.C = uint32(y) <= uint32(x) // This code follows the arm carry flag model.
1950 fcb.V = x >= 0 && y < 0 && x-y < 0 || x < 0 && y >= 0 && x-y >= 0
1954 // logicFlags64 returns flags set to the sign/zeroness of x.
1955 // C and V are set to false.
1956 func logicFlags64(x int64) flagConstant {
1957 var fcb flagConstantBuilder
1963 // logicFlags32 returns flags set to the sign/zeroness of x.
1964 // C and V are set to false.
1965 func logicFlags32(x int32) flagConstant {
1966 var fcb flagConstantBuilder
1972 func makeJumpTableSym(b *Block) *obj.LSym {
1973 s := base.Ctxt.Lookup(fmt.Sprintf("%s.jump%d", b.Func.fe.LSym(), b.ID))
1974 s.Set(obj.AttrDuplicateOK, true)
1975 s.Set(obj.AttrLocal, true)
1979 // canRotate reports whether the architecture supports
1980 // rotates of integer registers with the given number of bits.
1981 func canRotate(c *Config, bits int64) bool {
1982 if bits > c.PtrSize*8 {
1983 // Don't rewrite to rotates bigger than the machine word.
1987 case "386", "amd64", "arm64":
1989 case "arm", "s390x", "ppc64", "ppc64le", "wasm", "loong64":
1996 // isARM64bitcon reports whether a constant can be encoded into a logical instruction.
1997 func isARM64bitcon(x uint64) bool {
1998 if x == 1<<64-1 || x == 0 {
2001 // determine the period and sign-extend a unit to 64 bits
2003 case x != x>>32|x<<32:
2006 case x != x>>16|x<<48:
2008 x = uint64(int64(int32(x)))
2009 case x != x>>8|x<<56:
2011 x = uint64(int64(int16(x)))
2012 case x != x>>4|x<<60:
2014 x = uint64(int64(int8(x)))
2016 // period is 4 or 2, always true
2017 // 0001, 0010, 0100, 1000 -- 0001 rotate
2018 // 0011, 0110, 1100, 1001 -- 0011 rotate
2019 // 0111, 1011, 1101, 1110 -- 0111 rotate
2020 // 0101, 1010 -- 01 rotate, repeat
2023 return sequenceOfOnes(x) || sequenceOfOnes(^x)
2026 // sequenceOfOnes tests whether a constant is a sequence of ones in binary, with leading and trailing zeros.
2027 func sequenceOfOnes(x uint64) bool {
2028 y := x & -x // lowest set bit of x. x is good iff x+y is a power of 2
2033 // isARM64addcon reports whether x can be encoded as the immediate value in an ADD or SUB instruction.
2034 func isARM64addcon(v int64) bool {
2035 /* uimm12 or uimm24? */
2039 if (v & 0xFFF) == 0 {