cmd/compile: load some live values into registers before loop

If we're about to enter a loop, load values which are live
and will soon be used in the loop into registers.

name                     old time/op    new time/op    delta
BinaryTree17-8              2.80s ± 4%     2.62s ± 2%   -6.43%          (p=0.008 n=5+5)
Fannkuch11-8                2.45s ± 2%     2.14s ± 1%  -12.43%          (p=0.008 n=5+5)
FmtFprintfEmpty-8          49.0ns ± 1%    48.4ns ± 1%   -1.35%          (p=0.032 n=5+5)
FmtFprintfString-8          160ns ± 1%     153ns ± 0%   -4.63%          (p=0.008 n=5+5)
FmtFprintfInt-8             152ns ± 0%     150ns ± 0%   -1.57%          (p=0.000 n=5+4)
FmtFprintfIntInt-8          252ns ± 2%     244ns ± 1%   -3.02%          (p=0.008 n=5+5)
FmtFprintfPrefixedInt-8     223ns ± 0%     223ns ± 0%     ~     (all samples are equal)
FmtFprintfFloat-8           293ns ± 2%     291ns ± 2%     ~             (p=0.389 n=5+5)
FmtManyArgs-8               956ns ± 0%     936ns ± 0%   -2.05%          (p=0.008 n=5+5)
GobDecode-8                7.18ms ± 0%    7.11ms ± 0%   -1.02%          (p=0.008 n=5+5)
GobEncode-8                6.12ms ± 3%    6.07ms ± 1%     ~             (p=0.690 n=5+5)
Gzip-8                      284ms ± 1%     284ms ± 0%     ~             (p=1.000 n=5+5)
Gunzip-8                   40.8ms ± 1%    40.6ms ± 1%     ~             (p=0.310 n=5+5)
HTTPClientServer-8         69.8µs ± 1%    72.2µs ± 4%     ~             (p=0.056 n=5+5)
JSONEncode-8               16.1ms ± 2%    16.2ms ± 1%     ~             (p=0.151 n=5+5)
JSONDecode-8               54.9ms ± 0%    57.0ms ± 1%   +3.79%          (p=0.008 n=5+5)
Mandelbrot200-8            4.35ms ± 0%    4.39ms ± 0%   +0.85%          (p=0.008 n=5+5)
GoParse-8                  3.56ms ± 1%    3.42ms ± 1%   -4.03%          (p=0.008 n=5+5)
RegexpMatchEasy0_32-8      75.6ns ± 1%    75.0ns ± 0%   -0.83%          (p=0.016 n=5+4)
RegexpMatchEasy0_1K-8       250ns ± 0%     252ns ± 1%   +0.80%          (p=0.016 n=4+5)
RegexpMatchEasy1_32-8      75.0ns ± 0%    75.4ns ± 2%     ~             (p=0.206 n=5+5)
RegexpMatchEasy1_1K-8       401ns ± 0%     398ns ± 1%     ~             (p=0.056 n=5+5)
RegexpMatchMedium_32-8      119ns ± 0%     118ns ± 0%   -0.84%          (p=0.008 n=5+5)
RegexpMatchMedium_1K-8     36.6µs ± 0%    36.9µs ± 0%   +0.91%          (p=0.008 n=5+5)
RegexpMatchHard_32-8       1.95µs ± 1%    1.92µs ± 0%   -1.23%          (p=0.032 n=5+5)
RegexpMatchHard_1K-8       58.3µs ± 1%    58.1µs ± 1%     ~             (p=0.548 n=5+5)
Revcomp-8                   425ms ± 1%     389ms ± 1%   -8.39%          (p=0.008 n=5+5)
Template-8                 65.5ms ± 1%    63.6ms ± 1%   -2.86%          (p=0.008 n=5+5)
TimeParse-8                 363ns ± 0%     354ns ± 1%   -2.59%          (p=0.008 n=5+5)
TimeFormat-8                363ns ± 0%     364ns ± 1%     ~             (p=0.159 n=5+5)

Fixes #14511

Change-Id: I1b79d2545271fa90d5b04712cc25573bdc94f2ce
Reviewed-on: https://go-review.googlesource.com/20151
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
diff --git a/src/cmd/compile/internal/gc/ssa.go b/src/cmd/compile/internal/gc/ssa.go
index e5d72ce..f83b88d 100644
--- a/src/cmd/compile/internal/gc/ssa.go
+++ b/src/cmd/compile/internal/gc/ssa.go
@@ -5081,8 +5081,10 @@
 			return x86.AMOVL
 		case 8:
 			return x86.AMOVQ
+		case 16:
+			return x86.AMOVUPS // int128s are in SSE registers
 		default:
-			panic("bad int register width")
+			panic(fmt.Sprintf("bad int register width %d:%s", t.Size(), t))
 		}
 	}
 	panic("bad register type")