fix(lint): calibrate register-clobber to the Go ABI and add legacy SSE moves

Assisted-by: Qwen 3.8 Max Preview

lint/liveness.go

@@ -4,7 +4,6 @@
 package lint
 
 import (
-	"fmt"
 	"sort"
 	"strings"
 
@@ -248,7 +247,7 @@ func instrEffect(in *ast.Instr, a arch.Arch) regEffect {
 	}
 
 	compare := isCompare(mnem)
-	dstIdx := dstIndex(in, a)
+	dstIdx := dstIndex(in)
 
 	for i, op := range in.Operands {
 		r := gprName(op, a)
@@ -281,13 +280,11 @@ func instrEffect(in *ast.Instr, a arch.Arch) regEffect {
 	return eff
 }
 
-// dstIndex returns the operand index of the destination register: last for the
-// Plan 9 (amd64) spelling, first for arm64/riscv64/loong64.
-func dstIndex(in *ast.Instr, a arch.Arch) int {
-	if a == arch.AMD64 {
-		return len(in.Operands) - 1
-	}
-	return 0
+// dstIndex returns the operand index of the destination register: in Plan 9
+// notation the destination is the last operand on every architecture Go
+// supports (amd64, arm64, riscv64 and loong64 alike).
+func dstIndex(in *ast.Instr) int {
+	return len(in.Operands) - 1
 }
 
 // isCompare reports whether the mnemonic only reads its operands (setting flags).
@@ -372,41 +369,37 @@ func sameSet(a, b map[string]bool) bool {
 	return true
 }
 
-// calleeSavedGPRs returns the general-purpose registers an assembly function
-// must preserve for its caller, using the register names the assembler accepts
-// for each architecture.
-func calleeSavedGPRs(a arch.Arch) map[string]bool {
+// goFixedGPRs returns the general-purpose registers the Go ABI designates as
+// fixed across calls — the ones hand-written assembly must not permanently
+// clobber.  This follows cmd/compile/abi-internal.md, not the platform ABI:
+// Go's stack-based ABI0 (which hand-written assembly uses) has no System V
+// style callee-saved registers, so clobbering the argument and scratch
+// registers (amd64 BX, R12, R13, R15, …) is legal.
+//
+// Two groups are returned.  always holds registers whose loss is never safe.
+// runtime holds registers that survive an ABI0 leaf only because the
+// transition machinery restores them (on amd64 the g pointer is reloaded
+// from TLS): clobbering them is safe exactly in NOSPLIT functions that make
+// no calls, which is how the runtime's own assembly uses them.
+func goFixedGPRs(a arch.Arch) (always, runtime map[string]bool) {
 	switch a {
 	case arch.AMD64:
-		return gprSet("BX", "BP", "R12", "R13", "R14", "R15")
+		// BP maintains the frame chain; R14 holds the current goroutine.
+		// R15 is scratch except in dynamically linked binaries, so it is not
+		// flagged.
+		return gprSet("BP"), gprSet("R14")
 	case arch.ARM64:
-		names := []string{"R29", "R30"} // FP, LR
-		for i := 19; i <= 28; i++ {
-			names = append(names, fmt.Sprintf("R%d", i))
-		}
-		return gprSet(names...)
+		// R18 is reserved for the OS on some platforms, R28 holds the current
+		// goroutine, R29 is the frame pointer.
+		return gprSet("R18", "R28", "R29"), nil
 	case arch.RISCV:
-		// RA (X1) and the S registers (X8, X9, X18–X27) are callee-saved.
-		names := []string{"X1", "RA", "X8", "X9", "S0", "S1", "FP"}
-		for i := 18; i <= 27; i++ {
-			names = append(names, fmt.Sprintf("X%d", i))
-		}
-		for i := 2; i <= 11; i++ {
-			names = append(names, fmt.Sprintf("S%d", i))
-		}
-		return gprSet(names...)
+		// X27 holds the current goroutine.
+		return gprSet("X27"), nil
 	case arch.LOONG64:
-		// RA (R1), FP (R22) and S0–S8 (R23–R31) are callee-saved.
-		names := []string{"R1", "RA", "R22", "FP"}
-		for i := 23; i <= 31; i++ {
-			names = append(names, fmt.Sprintf("R%d", i))
-		}
-		for i := 0; i <= 8; i++ {
-			names = append(names, fmt.Sprintf("S%d", i))
-		}
-		return gprSet(names...)
+		// R22 holds the current goroutine.
+		return gprSet("R22"), nil
 	}
-	return nil
+	return nil, nil
 }
 
 func gprSet(names ...string) map[string]bool {
@@ -417,15 +410,16 @@ func gprSet(names ...string) map[string]bool {
 	return m
 }
 
-// clobberedCalleeSaved returns the callee-saved registers a function writes
-// without also saving and restoring them — i.e. registers whose caller-owned
-// value is lost across the call.  It walks the blocks of the liveness analysis
-// (so the control-flow graph is what supplies the instruction set) and
-// aggregates each instruction's register effects.
-func clobberedCalleeSaved(l *liveness, a arch.Arch) []string {
-	callee := calleeSavedGPRs(a)
-	if len(callee) == 0 {
-		return nil
+// clobberedGoFixed returns the Go-ABI-fixed registers a function writes
+// without also saving and restoring them.  The first result lists registers
+// whose loss is never safe; the second lists the goroutine-pointer class,
+// whose loss is reported only when reachesRuntime is true (a non-NOSPLIT
+// function, or one that makes calls — the ABI0 transition machinery restores
+// the g pointer only on such paths).
+func clobberedGoFixed(l *liveness, a arch.Arch, reachesRuntime bool) (always, runtime []string) {
+	alwaysSet, runtimeSet := goFixedGPRs(a)
+	if len(alwaysSet) == 0 && len(runtimeSet) == 0 {
+		return nil, nil
 	}
 	def := map[string]bool{}
 	saved := map[string]bool{}
@@ -444,12 +438,19 @@ func clobberedCalleeSaved(l *liveness, a arch.Arch) []string {
 			}
 		}
 	}
-	var out []string
-	for r := range callee {
-		if def[r] && !(saved[r] && restored[r]) {
-			out = append(out, r)
+	clobbered := func(set map[string]bool) []string {
+		var out []string
+		for r := range set {
+			if def[r] && !(saved[r] && restored[r]) {
+				out = append(out, r)
+			}
 		}
+		sort.Strings(out)
+		return out
 	}
-	sort.Strings(out)
-	return out
+	always = clobbered(alwaysSet)
+	if reachesRuntime {
+		runtime = clobbered(runtimeSet)
+	}
+	return always, runtime
 }