cmd/compile: fold single-bit TST; CSET NE into UBFX on arm64

A boolean materialized from a single-bit test compiles to
TST $(1<<k); CSET NE. The bit is a bitfield extract: rewrite the
NotEqual of a one-bit TSTconst/TSTWconst to UBFX #k, #1, the value
form of the existing single-bit TBZ/TBNZ branch rules. The flags
producer remains for any other reader, and a shared bool's branch
then tests the extracted bit directly, so the TST disappears there
too.

Measured with armlint on darwin/arm64:

cmd/go: 166 TST; CSET pairs eliminated, __text -672 bytes.
gofmt: 35 pairs eliminated, __text -192 bytes.

Author: gaul

src/cmd/compile/internal/ssa/_gen/ARM64.rules

@@ -570,6 +570,12 @@
 ((Equal|NotEqual|LessThan|LessEqual|GreaterThan|GreaterEqual) (CMPconst [0]  x:(ANDconst [c] y))) && x.Uses == 1 =>
 	((Equal|NotEqual|LessThan|LessEqual|GreaterThan|GreaterEqual) (TSTconst [c] y))
 
+// A single-bit test materialized as a boolean is a bitfield extract:
+// TST $(1<<k); CSET NE is UBFX #k, #1. The flags stay for any other
+// reader; the branch form is handled by the TBZ/TBNZ rules below.
+(NotEqual (TSTconst  [c] x)) && oneBit(c) => (UBFX [armBFAuxInt(int64(ntz64(c)), 1)] x)
+(NotEqual (TSTWconst [c] x)) && oneBit(int64(uint32(c))) => (UBFX [armBFAuxInt(int64(ntz64(int64(uint32(c)))), 1)] x)
+
 ((EQ|NE|LT|LE|GT|GE) (CMPconst  [0] x:(ADDconst [c] y)) yes no) && x.Uses == 1 => ((EQ|NE|LTnoov|LEnoov|GTnoov|GEnoov) (CMNconst         [c] y) yes no)
 ((EQ|NE|LT|LE|GT|GE) (CMPWconst [0] x:(ADDconst [c] y)) yes no) && x.Uses == 1 => ((EQ|NE|LTnoov|LEnoov|GTnoov|GEnoov) (CMNWconst [int32(c)] y) yes no)
 ((EQ|NE|LT|LE|GT|GE) (CMPconst  [0] z:(ADD        x y)) yes no) && z.Uses == 1 => ((EQ|NE|LTnoov|LEnoov|GTnoov|GEnoov) (CMN                x y) yes no)

src/cmd/compile/internal/ssa/rewriteARM64.go

@@ -327,6 +327,22 @@ func ubfx16(x uint64) uint64 {
 	return ((x >> 3) & 0xfff) >> 1 & 0x3f
 }
 
+// materialize a single-bit test as ubfx instead of TST + CSET.
+func ubfx17(x uint64) bool {
+	// arm64:"UBFX [$]3, R[0-9]+, [$]1" -"TST" -"CSET"
+	return x&8 != 0
+}
+
+func ubfx18(x uint64) bool {
+	// arm64:"UBFX [$]63, R[0-9]+, [$]1" -"TST" -"CSET"
+	return x&(1<<63) != 0
+}
+
+func ubfx19(x uint32) bool {
+	// arm64:"UBFX [$]31, R[0-9]+, [$]1" -"TST" -"CSET"
+	return x&(1<<31) != 0
+}
+
 // Check that we don't emit comparisons for constant shifts.
 //
 //go:nosplit