cmd/compile: make a copy of Phi input if it is still live

Register of Phi input is allocated to the Phi. So if the Phi
input is still live after Phi, we may need to use a spill. In
this case, copy the Phi input to a spare register to avoid a
spill.

Originally targeted the code in issue #16187, and this CL
indeed removes the spill, but doesn't seem to help on benchmark
result. It may help in general, though.

On AMD64:
name                      old time/op    new time/op    delta
BinaryTree17-12              2.79s ± 1%     2.76s ± 0%  -1.33%  (p=0.000 n=10+10)
Fannkuch11-12                3.02s ± 0%     3.14s ± 0%  +3.99%  (p=0.000 n=10+10)
FmtFprintfEmpty-12          51.2ns ± 0%    51.4ns ± 3%    ~      (p=0.368 n=8+10)
FmtFprintfString-12          145ns ± 0%     144ns ± 0%  -0.69%    (p=0.000 n=6+9)
FmtFprintfInt-12             127ns ± 1%     124ns ± 1%  -2.79%   (p=0.000 n=10+9)
FmtFprintfIntInt-12          186ns ± 0%     184ns ± 0%  -1.34%   (p=0.000 n=10+9)
FmtFprintfPrefixedInt-12     196ns ± 0%     194ns ± 0%  -0.97%    (p=0.000 n=9+9)
FmtFprintfFloat-12           293ns ± 2%     287ns ± 0%  -2.00%   (p=0.000 n=10+9)
FmtManyArgs-12               847ns ± 1%     829ns ± 0%  -2.17%   (p=0.000 n=10+7)
GobDecode-12                7.17ms ± 0%    7.18ms ± 0%    ~     (p=0.123 n=10+10)
GobEncode-12                6.08ms ± 1%    6.08ms ± 0%    ~      (p=0.497 n=10+9)
Gzip-12                      277ms ± 1%     275ms ± 1%  -0.47%   (p=0.028 n=10+9)
Gunzip-12                   39.1ms ± 2%    38.2ms ± 1%  -2.20%   (p=0.000 n=10+9)
HTTPClientServer-12         90.9µs ± 4%    87.7µs ± 2%  -3.51%   (p=0.001 n=9+10)
JSONEncode-12               17.3ms ± 1%    16.5ms ± 0%  -5.02%    (p=0.000 n=9+9)
JSONDecode-12               54.6ms ± 1%    54.1ms ± 0%  -0.99%    (p=0.000 n=9+9)
Mandelbrot200-12            4.45ms ± 0%    4.45ms ± 0%  -0.02%    (p=0.006 n=8+9)
GoParse-12                  3.44ms ± 0%    3.48ms ± 1%  +0.95%  (p=0.000 n=10+10)
RegexpMatchEasy0_32-12      84.9ns ± 0%    85.0ns ± 0%    ~       (p=0.241 n=8+8)
RegexpMatchEasy0_1K-12       867ns ± 3%     915ns ±11%  +5.55%  (p=0.037 n=10+10)
RegexpMatchEasy1_32-12      82.7ns ± 5%    83.9ns ± 4%    ~      (p=0.161 n=9+10)
RegexpMatchEasy1_1K-12       361ns ± 1%     363ns ± 0%    ~      (p=0.098 n=10+8)
RegexpMatchMedium_32-12      126ns ± 0%     126ns ± 1%    ~      (p=0.549 n=8+10)
RegexpMatchMedium_1K-12     38.8µs ± 0%    39.1µs ± 0%  +0.67%    (p=0.000 n=9+8)
RegexpMatchHard_32-12       1.95µs ± 0%    1.96µs ± 0%  +0.43%    (p=0.000 n=9+9)
RegexpMatchHard_1K-12       59.0µs ± 0%    59.1µs ± 0%  +0.27%   (p=0.000 n=10+9)
Revcomp-12                   436ms ± 1%     431ms ± 1%  -1.19%  (p=0.005 n=10+10)
Template-12                 56.7ms ± 1%    57.1ms ± 1%  +0.71%   (p=0.001 n=10+9)
TimeParse-12                 312ns ± 0%     310ns ± 0%  -0.80%   (p=0.000 n=10+9)
TimeFormat-12                336ns ± 0%     332ns ± 0%  -1.19%    (p=0.000 n=8+7)
[Geo mean]                  59.2µs         58.9µs       -0.42%

On PPC64:
name                     old time/op    new time/op    delta
BinaryTree17-2              4.67s ± 2%     4.71s ± 1%    ~     (p=0.421 n=5+5)
Fannkuch11-2                3.92s ± 1%     3.94s ± 0%  +0.46%  (p=0.032 n=5+5)
FmtFprintfEmpty-2           122ns ± 0%     120ns ± 2%  -1.80%  (p=0.016 n=4+5)
FmtFprintfString-2          305ns ± 1%     299ns ± 1%  -1.84%  (p=0.008 n=5+5)
FmtFprintfInt-2             243ns ± 0%     241ns ± 1%  -0.66%  (p=0.016 n=4+5)
FmtFprintfIntInt-2          361ns ± 1%     356ns ± 1%  -1.49%  (p=0.016 n=5+5)
FmtFprintfPrefixedInt-2     355ns ± 1%     357ns ± 1%    ~     (p=0.333 n=5+5)
FmtFprintfFloat-2           502ns ± 2%     498ns ± 1%    ~     (p=0.151 n=5+5)
FmtManyArgs-2              1.55µs ± 2%    1.59µs ± 1%  +2.52%  (p=0.008 n=5+5)
GobDecode-2                13.0ms ± 1%    13.0ms ± 1%    ~     (p=0.841 n=5+5)
GobEncode-2                11.8ms ± 1%    11.8ms ± 1%    ~     (p=0.690 n=5+5)
Gzip-2                      499ms ± 1%     503ms ± 0%    ~     (p=0.421 n=5+5)
Gunzip-2                   86.5ms ± 0%    86.4ms ± 1%    ~     (p=0.841 n=5+5)
HTTPClientServer-2         68.2µs ± 2%    69.6µs ± 3%    ~     (p=0.151 n=5+5)
JSONEncode-2               39.0ms ± 1%    37.2ms ± 1%  -4.65%  (p=0.008 n=5+5)
JSONDecode-2                122ms ± 1%     126ms ± 1%  +2.63%  (p=0.008 n=5+5)
Mandelbrot200-2            6.08ms ± 1%    5.89ms ± 1%  -3.06%  (p=0.008 n=5+5)
GoParse-2                  5.95ms ± 2%    5.98ms ± 1%    ~     (p=0.421 n=5+5)
RegexpMatchEasy0_32-2       331ns ± 1%     328ns ± 1%    ~     (p=0.056 n=5+5)
RegexpMatchEasy0_1K-2      1.45µs ± 0%    1.47µs ± 0%  +1.13%  (p=0.008 n=5+5)
RegexpMatchEasy1_32-2       359ns ± 0%     353ns ± 0%  -1.84%  (p=0.008 n=5+5)
RegexpMatchEasy1_1K-2      1.79µs ± 0%    1.81µs ± 1%  +1.16%  (p=0.008 n=5+5)
RegexpMatchMedium_32-2      420ns ± 2%     413ns ± 0%  -1.72%  (p=0.008 n=5+5)
RegexpMatchMedium_1K-2     70.2µs ± 1%    69.5µs ± 1%  -1.09%  (p=0.032 n=5+5)
RegexpMatchHard_32-2       3.87µs ± 1%    3.65µs ± 0%  -5.86%  (p=0.008 n=5+5)
RegexpMatchHard_1K-2        111µs ± 0%     105µs ± 0%  -5.49%  (p=0.016 n=5+4)
Revcomp-2                   1.00s ± 1%     1.01s ± 2%    ~     (p=0.151 n=5+5)
Template-2                  113ms ± 1%     113ms ± 2%    ~     (p=0.841 n=5+5)
TimeParse-2                 555ns ± 0%     550ns ± 1%  -0.87%  (p=0.032 n=5+5)
TimeFormat-2                736ns ± 1%     704ns ± 1%  -4.35%  (p=0.008 n=5+5)
[Geo mean]                  120µs          119µs       -0.77%

Reduce "spilled value remains" by 0.6% in cmd/go on AMD64.

Change-Id: If655df343b0f30d1a49ab1ab644f10c698b96f3e
Reviewed-on: https://go-review.googlesource.com/32442
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>

src/cmd/compile/internal/ssa/regalloc.go

@@ -869,7 +869,6 @@ func (s *regAllocState) regalloc(f *Func) {
 				m := s.values[a.ID].regs &^ phiUsed & s.allocatable
 				if m != 0 {
 					r := pickReg(m)
-					s.freeReg(r)
 					phiUsed |= regMask(1) << r
 					phiRegs = append(phiRegs, r)
 				} else {
@@ -878,7 +877,7 @@ func (s *regAllocState) regalloc(f *Func) {
 			}
 
 			// Second pass - deallocate any phi inputs which are now dead.
-			for _, v := range phis {
+			for i, v := range phis {
 				if !s.values[v.ID].needReg {
 					continue
 				}
@@ -887,6 +886,31 @@ func (s *regAllocState) regalloc(f *Func) {
 					// Input is dead beyond the phi, deallocate
 					// anywhere else it might live.
 					s.freeRegs(s.values[a.ID].regs)
+				} else {
+					// Input is still live.
+					// Try to move it around before kicking out, if there is a free register.
+					// We generate a Copy in the predecessor block and record it. It will be
+					// deleted if never used.
+					r := phiRegs[i]
+					if r == noRegister {
+						continue
+					}
+					// Pick a free register. At this point some registers used in the predecessor
+					// block may have been deallocated. Those are the ones used for Phis. Exclude
+					// them (and they are not going to be helpful anyway).
+					m := s.compatRegs(a.Type) &^ s.used &^ phiUsed
+					if m != 0 && !s.values[a.ID].rematerializeable && countRegs(s.values[a.ID].regs) == 1 {
+						r2 := pickReg(m)
+						c := p.NewValue1(a.Line, OpCopy, a.Type, s.regs[r].c)
+						s.copies[c] = false
+						if s.f.pass.debug > regDebug {
+							fmt.Printf("copy %s to %s : %s\n", a, c, s.registers[r2].Name())
+						}
+						s.setOrig(c, a)
+						s.assignReg(r2, a, c)
+						s.endRegs[p.ID] = append(s.endRegs[p.ID], endReg{r2, a, c})
+					}
+					s.freeReg(r)
 				}
 			}