cmd/compile: split slow 3 operand LEA instructions into two LEAs

go tool objdump ../bin/go | grep "\.go\:" | grep -c "LEA.*0x.*[(].*[(].*"
Before: 1012
After: 20

Updates #21735

Benchmarks thanks to [email protected]
Intel(R) Xeon(R) CPU E5-1650 v3 @ 3.50GHz
benchstat -geomean *.stdout | grep -v pkg:
name                                       old time/op    new time/op    delta
FastTest2KB-12                              131.0ns ± 0%   131.0ns ± 0%    ~     (all equal)
BaseTest2KB-12                              601.3ns ± 0%   601.3ns ± 0%    ~     (p=0.942 n=45+41)
Encoding4KBVerySparse-12                    15.38µs ± 1%   15.41µs ± 0%  +0.24%  (p=0.000 n=44+43)
Join_8-12                                    2.117s ± 2%    2.128s ± 2%  +0.51%  (p=0.007 n=48+48)
HashimotoLight-12                           1.663ms ± 1%   1.668ms ± 1%  +0.35%  (p=0.006 n=49+49)
Sha3_224_MTU-12                             4.843µs ± 0%   4.836µs ± 0%  -0.14%  (p=0.000 n=45+42)
GenSharedKeyP256-12                         73.74µs ± 0%   70.92µs ± 0%  -3.82%  (p=0.000 n=48+47)
Run/10k/1-12                                 24.81s ± 0%    24.88s ± 0%  +0.30%  (p=0.000 n=48+47)
Run/10k/16-12                                4.621s ± 2%    4.625s ± 3%    ~     (p=0.776 n=50+50)
Dnrm2MediumPosInc-12                        4.018µs ± 0%   4.019µs ± 0%    ~     (p=0.060 n=50+48)
DasumMediumUnitaryInc-12                    855.3ns ± 0%   855.0ns ± 0%  -0.03%  (p=0.000 n=47+42)
Dgeev/Circulant10-12                        40.45µs ± 1%   41.11µs ± 1%  +1.62%  (p=0.000 n=45+49)
Dgeev/Circulant100-12                       10.42ms ± 2%   10.61ms ± 1%  +1.83%  (p=0.000 n=49+48)
MulWorkspaceDense1000Hundredth-12           64.69ms ± 1%   64.63ms ± 1%    ~     (p=0.718 n=48+48)
ScaleVec10000Inc20-12                       22.31µs ± 1%   22.29µs ± 1%    ~     (p=0.424 n=50+50)
ValidateVersionTildeFail-12                 737.6ns ± 1%   736.0ns ± 1%  -0.22%  (p=0.000 n=49+49)
StripHTML-12                                2.846µs ± 0%   2.806µs ± 1%  -1.40%  (p=0.000 n=43+50)
ReaderContains-12                           6.073µs ± 0%   5.999µs ± 0%  -1.22%  (p=0.000 n=48+48)
EncodeCodecFromInternalProtobuf-12          5.817µs ± 2%   5.555µs ± 2%  -4.51%  (p=0.000 n=47+47)
TarjanSCCGnp_10_tenth-12                    7.091µs ± 5%   7.132µs ± 7%    ~     (p=0.361 n=50+50)
TarjanSCCGnp_1000_half-12                   82.25ms ± 3%   81.29ms ± 2%  -1.16%  (p=0.000 n=50+43)
AStarUndirectedmallWorld_10_2_2_2_Heur-12   15.18µs ± 8%   15.11µs ± 7%    ~     (p=0.511 n=50+49)
LouvainDirectedMultiplex-12                 20.92ms ± 1%   21.00ms ± 1%  +0.36%  (p=0.000 n=48+49)
WalkAllBreadthFirstGnp_10_tenth-12          2.974µs ± 4%   2.964µs ± 5%    ~     (p=0.504 n=50+50)
WalkAllBreadthFirstGnp_1000_tenth-12        9.733ms ± 4%   9.741ms ± 4%    ~     (p=0.774 n=48+50)
TextMovementBetweenSegments-12              432.8µs ± 0%   433.2µs ± 1%    ~     (p=0.128 n=50+50)
Growth_MultiSegment-12                      13.11ms ± 0%   13.19ms ± 1%  +0.58%  (p=0.000 n=44+46)
AddingFields/Zap.Sugar-12                   1.296µs ± 1%   1.310µs ± 2%  +1.09%  (p=0.000 n=43+43)
AddingFields/apex/log-12                    34.19µs ± 1%   34.31µs ± 1%  +0.35%  (p=0.000 n=45+45)
AddingFields/inconshreveable/log15-12       30.08µs ± 2%   30.07µs ± 2%    ~     (p=0.803 n=48+47)
AddingFields/sirupsen/logrus-12             6.683µs ± 3%   6.735µs ± 3%  +0.78%  (p=0.000 n=43+42)

[Geo mean]                                  143.5µs        143.3µs       -0.16%

Change-Id: I637203c75c837737f1febced75d5985703e51044
Reviewed-on: https://go-review.googlesource.com/114655
Run-TryBot: Martin Möhrmann <[email protected]>
TryBot-Result: Gobot Gobot <[email protected]>
Reviewed-by: David Chase <[email protected]>
Reviewed-by: Ilya Tocar <[email protected]>

Author: martisch

src/cmd/compile/internal/amd64/ssa.go

@@ -524,6 +524,7 @@ func ssaGenValue(s *gc.SSAGenState, v *ssa.Value) {
 	case ssa.OpAMD64LEAQ1, ssa.OpAMD64LEAQ2, ssa.OpAMD64LEAQ4, ssa.OpAMD64LEAQ8,
 		ssa.OpAMD64LEAL1, ssa.OpAMD64LEAL2, ssa.OpAMD64LEAL4, ssa.OpAMD64LEAL8,
 		ssa.OpAMD64LEAW1, ssa.OpAMD64LEAW2, ssa.OpAMD64LEAW4, ssa.OpAMD64LEAW8:
+		o := v.Reg()
 		r := v.Args[0].Reg()
 		i := v.Args[1].Reg()
 		p := s.Prog(v.Op.Asm())
@@ -543,9 +544,24 @@ func ssaGenValue(s *gc.SSAGenState, v *ssa.Value) {
 		p.From.Type = obj.TYPE_MEM
 		p.From.Reg = r
 		p.From.Index = i
-		gc.AddAux(&p.From, v)
 		p.To.Type = obj.TYPE_REG
-		p.To.Reg = v.Reg()
+		p.To.Reg = o
+		if v.AuxInt != 0 && v.Aux == nil {
+			// Emit an additional LEA to add the displacement instead of creating a slow 3 operand LEA.
+			switch v.Op {
+			case ssa.OpAMD64LEAQ1, ssa.OpAMD64LEAQ2, ssa.OpAMD64LEAQ4, ssa.OpAMD64LEAQ8:
+				p = s.Prog(x86.ALEAQ)
+			case ssa.OpAMD64LEAL1, ssa.OpAMD64LEAL2, ssa.OpAMD64LEAL4, ssa.OpAMD64LEAL8:
+				p = s.Prog(x86.ALEAL)
+			case ssa.OpAMD64LEAW1, ssa.OpAMD64LEAW2, ssa.OpAMD64LEAW4, ssa.OpAMD64LEAW8:
+				p = s.Prog(x86.ALEAW)
+			}
+			p.From.Type = obj.TYPE_MEM
+			p.From.Reg = o
+			p.To.Type = obj.TYPE_REG
+			p.To.Reg = o
+		}
+		gc.AddAux(&p.From, v)
 	case ssa.OpAMD64LEAQ, ssa.OpAMD64LEAL, ssa.OpAMD64LEAW:
 		p := s.Prog(v.Op.Asm())
 		p.From.Type = obj.TYPE_MEM