[dev.simd] cmd/compile: fix SIMD const rematerialization condition

This CL fixes a condition for the previous fix CL 704056.

Change-Id: I1f1f8c6f72870403cb3dff14755c43385dc0c933
Reviewed-on: https://go-review.googlesource.com/c/go/+/705499
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>

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

@@ -2576,22 +2576,25 @@ func (e *edgeState) processDest(loc Location, vid ID, splice **Value, pos src.XP
 			e.s.f.Fatalf("can't find source for %s->%s: %s\n", e.p, e.b, v.LongString())
 		}
 		if dstReg {
-			// Handle incompatible registers.
+			// We want to rematerialize v into a register that is incompatible with v's op's register mask.
+			// Instead of setting the wrong register for the rematerialized v, we should find the right register
+			// for it and emit an additional copy to move to the desired register.
 			// For #70451.
-			if e.s.regspec(v).outputs[0].regs&regMask(1<<register(loc.(*Register).num)) == 0 && c != nil {
+			if e.s.regspec(v).outputs[0].regs&regMask(1<<register(loc.(*Register).num)) == 0 {
 				_, srcReg := src.(*Register)
-				if !srcReg {
+				if srcReg {
+					// It exists in a valid register already, so just copy it to the desired register
+					// If src is a Register, c must have already been set.
+					x = e.p.NewValue1(pos, OpCopy, c.Type, c)
+				} else {
 					// We need a tmp register
 					x = v.copyInto(e.p)
 					r := e.findRegFor(x.Type)
 					e.erase(r)
-					// Rematerialize to a tmp register
+					// Rematerialize to the tmp register
 					e.set(r, vid, x, false, pos)
 					// Copy from tmp to the desired register
 					x = e.p.NewValue1(pos, OpCopy, x.Type, x)
-				} else {
-					// It exist in a valid register already, so just copy it to the desired register
-					x = e.p.NewValue1(pos, OpCopy, c.Type, c)
 				}
 			} else {
 				x = v.copyInto(e.p)

test/simd/bug2.go

@@ -24,3 +24,46 @@ func PackComplex(b bool) {
 		}
 	}
 }
+
+func PackComplex2(x0 uint16, src [][4]float32, b, b2 bool) {
+	var out [][4]byte
+	if b2 {
+		for y := range x0 {
+			row := out[:x0]
+			for x := range row {
+				px := &src[y]
+				if b {
+					var indices [4]uint32
+					fu := simd.LoadFloat32x4(px).AsUint32x4()
+					fu.ShiftAllRight(0).Store(nil)
+					entry := simd.LoadUint32x4(&[4]uint32{
+						toSrgbTable[indices[0]],
+					})
+					var res [4]uint32
+					entry.ShiftAllRight(19).Store(nil)
+					row[x] = [4]uint8{
+						uint8(res[0]),
+						uint8(res[1]),
+						uint8(res[2]),
+					}
+				} else {
+					row[x] = [4]uint8{
+						float32ToSrgb8(0),
+						float32ToSrgb8(1),
+						float32ToSrgb8(2),
+					}
+				}
+			}
+			out = out[len(out):]
+		}
+	}
+}
+
+var toSrgbTable = [4]uint32{}
+
+func float32ToSrgb8(f float32) uint8 {
+	f = min(0, f)
+	fu := uint32(f)
+	entry := toSrgbTable[fu]
+	return uint8(entry * fu)
+}