cmd/link/internal/ld: assign temporary addresses to per-package text

If trampolines may be required, the current text addressing second
pass resets all assigned addresses, before assigning addresses and
laying down trampolines in a linear fashion. However, this approach
means that intra-package calls are to a symbol that has not yet
been assigned an address, when the symbol is ahead of the current
function.

In the case of RISC-V the JAL instruction is limited to +/-1MiB.
As such, if a call is to a symbol with no address currently assigned,
we have to assume that a trampoline will be required. During the
relocation phase we can fix up and avoid trampolines in some cases,
however this results in unused trampolines that are still present
in the binary (since removing them would change text addresses).

In order to significantly reduce the number of unused trampolines,
assign temporary addresses to functions within the same package,
based on the maximum number of trampolines that may be required by
a function. This allows for better decisions to be made regarding
the requirement for intra-package trampolines, as we reset the
addressing for a function, assign its final address and lay down
any resulting trampolines.

This results in ~2,300 unused trampolines being removed from the
Go binary and ~5,600 unused trampolines being removed from the
compile binary, on linux/riscv64.

This reapplies CL 349650, however does not pass big to assignAddress
when assigning temporary addresses, as this can result in side
effects such as section splitting.

Change-Id: Id7febdb65d962d6b1297a91294a8dc27c94d8696
Reviewed-on: https://go-review.googlesource.com/c/go/+/534760
Reviewed-by: Cherry Mui <[email protected]>
Reviewed-by: Dmitri Shuralyov <[email protected]>
Run-TryBot: Joel Sing <[email protected]>
TryBot-Result: Gopher Robot <[email protected]>
Reviewed-by: Than McIntosh <[email protected]>

Author: 4a6f656c

src/cmd/link/internal/ld/data.go

@@ -84,14 +84,15 @@ func maxSizeTrampolines(ctxt *Link, ldr *loader.Loader, s loader.Sym, isTramp bo
 		}
 	}
 
-	if ctxt.IsARM() {
+	switch {
+	case ctxt.IsARM():
 		return n * 20 // Trampolines in ARM range from 3 to 5 instructions.
-	}
-	if ctxt.IsPPC64() {
-		return n * 16 // Trampolines in PPC64 are 4 instructions.
-	}
-	if ctxt.IsARM64() {
+	case ctxt.IsARM64():
 		return n * 12 // Trampolines in ARM64 are 3 instructions.
+	case ctxt.IsPPC64():
+		return n * 16 // Trampolines in PPC64 are 4 instructions.
+	case ctxt.IsRISCV64():
+		return n * 8 // Trampolines in RISCV64 are 2 instructions.
 	}
 	panic("unreachable")
 }
@@ -118,18 +119,21 @@ func trampoline(ctxt *Link, s loader.Sym) {
 			continue // something is wrong. skip it here and we'll emit a better error later
 		}
 
-		// RISC-V is only able to reach +/-1MiB via a JAL instruction,
-		// which we can readily exceed in the same package. As such, we
-		// need to generate trampolines when the address is unknown.
-		if ldr.SymValue(rs) == 0 && !ctxt.Target.IsRISCV64() && ldr.SymType(rs) != sym.SDYNIMPORT && ldr.SymType(rs) != sym.SUNDEFEXT {
+		if ldr.SymValue(rs) == 0 && ldr.SymType(rs) != sym.SDYNIMPORT && ldr.SymType(rs) != sym.SUNDEFEXT {
+			// Symbols in the same package are laid out together.
+			// Except that if SymPkg(s) == "", it is a host object symbol
+			// which may call an external symbol via PLT.
 			if ldr.SymPkg(s) != "" && ldr.SymPkg(rs) == ldr.SymPkg(s) {
-				// Symbols in the same package are laid out together.
-				// Except that if SymPkg(s) == "", it is a host object symbol
-				// which may call an external symbol via PLT.
-				continue
+				// RISC-V is only able to reach +/-1MiB via a JAL instruction.
+				// We need to generate a trampoline when an address is
+				// currently unknown.
+				if !ctxt.Target.IsRISCV64() {
+					continue
+				}
 			}
+			// Runtime packages are laid out together.
 			if isRuntimeDepPkg(ldr.SymPkg(s)) && isRuntimeDepPkg(ldr.SymPkg(rs)) {
-				continue // runtime packages are laid out together
+				continue
 			}
 		}
 		thearch.Trampoline(ctxt, ldr, ri, rs, s)
@@ -2435,8 +2439,8 @@ func (ctxt *Link) textaddress() {
 		limit = 1
 	}
 
-	// First pass: assign addresses assuming the program is small and
-	// don't generate trampolines.
+	// First pass: assign addresses assuming the program is small and will
+	// not require trampoline generation.
 	big := false
 	for _, s := range ctxt.Textp {
 		sect, n, va = assignAddress(ctxt, sect, n, s, va, false, big)
@@ -2451,21 +2455,45 @@ func (ctxt *Link) textaddress() {
 	if big {
 		// reset addresses
 		for _, s := range ctxt.Textp {
-			if ldr.OuterSym(s) != 0 || s == text {
-				continue
-			}
-			oldv := ldr.SymValue(s)
-			for sub := s; sub != 0; sub = ldr.SubSym(sub) {
-				ldr.SetSymValue(sub, ldr.SymValue(sub)-oldv)
+			if s != text {
+				resetAddress(ctxt, s)
 			}
 		}
 		va = start
 
 		ntramps := 0
-		for _, s := range ctxt.Textp {
+		var curPkg string
+		for i, s := range ctxt.Textp {
+			// When we find the first symbol in a package, perform a
+			// single iteration that assigns temporary addresses to all
+			// of the text in the same package, using the maximum possible
+			// number of trampolines. This allows for better decisions to
+			// be made regarding reachability and the need for trampolines.
+			if symPkg := ldr.SymPkg(s); symPkg != "" && curPkg != symPkg {
+				curPkg = symPkg
+				vaTmp := va
+				for j := i; j < len(ctxt.Textp); j++ {
+					curSym := ctxt.Textp[j]
+					if symPkg := ldr.SymPkg(curSym); symPkg == "" || curPkg != symPkg {
+						break
+					}
+					// We do not pass big to assignAddress here, as this
+					// can result in side effects such as section splitting.
+					sect, n, vaTmp = assignAddress(ctxt, sect, n, curSym, vaTmp, false, false)
+					vaTmp += maxSizeTrampolines(ctxt, ldr, curSym, false)
+				}
+			}
+
+			// Reset address for current symbol.
+			if s != text {
+				resetAddress(ctxt, s)
+			}
+
+			// Assign actual address for current symbol.
 			sect, n, va = assignAddress(ctxt, sect, n, s, va, false, big)
 
-			trampoline(ctxt, s) // resolve jumps, may add trampolines if jump too far
+			// Resolve jumps, adding trampolines if they are needed.
+			trampoline(ctxt, s)
 
 			// lay down trampolines after each function
 			for ; ntramps < len(ctxt.tramps); ntramps++ {
@@ -2613,6 +2641,17 @@ func assignAddress(ctxt *Link, sect *sym.Section, n int, s loader.Sym, va uint64
 	return sect, n, va
 }
 
+func resetAddress(ctxt *Link, s loader.Sym) {
+	ldr := ctxt.loader
+	if ldr.OuterSym(s) != 0 {
+		return
+	}
+	oldv := ldr.SymValue(s)
+	for sub := s; sub != 0; sub = ldr.SubSym(sub) {
+		ldr.SetSymValue(sub, ldr.SymValue(sub)-oldv)
+	}
+}
+
 // Return whether we may need to split text sections.
 //
 // On PPC64x, when external linking, a text section should not be

src/cmd/link/internal/ld/ld_test.go

@@ -344,3 +344,71 @@ func main() {
 		})
 	}
 }
+
+func TestRISCVTrampolines(t *testing.T) {
+	testenv.MustHaveGoBuild(t)
+	t.Parallel()
+
+	tmpDir := t.TempDir()
+	tmpFile := filepath.Join(tmpDir, "x.s")
+
+	// Calling b from a or c should not use trampolines, however
+	// calling from d to a will require one.
+	buf := new(bytes.Buffer)
+	fmt.Fprintf(buf, "TEXT a(SB),$0-0\n")
+	for i := 0; i < 1<<17; i++ {
+		fmt.Fprintf(buf, "\tADD $0, X0, X0\n")
+	}
+	fmt.Fprintf(buf, "\tCALL b(SB)\n")
+	fmt.Fprintf(buf, "\tRET\n")
+	fmt.Fprintf(buf, "TEXT b(SB),$0-0\n")
+	fmt.Fprintf(buf, "\tRET\n")
+	fmt.Fprintf(buf, "TEXT c(SB),$0-0\n")
+	fmt.Fprintf(buf, "\tCALL b(SB)\n")
+	fmt.Fprintf(buf, "\tRET\n")
+	fmt.Fprintf(buf, "TEXT ·d(SB),0,$0-0\n")
+	for i := 0; i < 1<<17; i++ {
+		fmt.Fprintf(buf, "\tADD $0, X0, X0\n")
+	}
+	fmt.Fprintf(buf, "\tCALL a(SB)\n")
+	fmt.Fprintf(buf, "\tCALL c(SB)\n")
+	fmt.Fprintf(buf, "\tRET\n")
+	if err := os.WriteFile(tmpFile, buf.Bytes(), 0644); err != nil {
+		t.Fatalf("Failed to write assembly file: %v", err)
+	}
+
+	if err := os.WriteFile(filepath.Join(tmpDir, "go.mod"), []byte("module riscvtramp"), 0644); err != nil {
+		t.Fatalf("Failed to write file: %v\n", err)
+	}
+	main := `package main
+func main() {
+	d()
+}
+
+func d()
+`
+	if err := os.WriteFile(filepath.Join(tmpDir, "x.go"), []byte(main), 0644); err != nil {
+		t.Fatalf("failed to write main: %v\n", err)
+	}
+	cmd := testenv.Command(t, testenv.GoToolPath(t), "build", "-ldflags=-linkmode=internal")
+	cmd.Dir = tmpDir
+	cmd.Env = append(os.Environ(), "GOARCH=riscv64", "GOOS=linux")
+	out, err := cmd.CombinedOutput()
+	if err != nil {
+		t.Fatalf("Build failed: %v, output: %s", err, out)
+	}
+
+	// Check what trampolines exist.
+	cmd = testenv.Command(t, testenv.GoToolPath(t), "tool", "nm", filepath.Join(tmpDir, "riscvtramp"))
+	cmd.Env = append(os.Environ(), "GOARCH=riscv64", "GOOS=linux")
+	out, err = cmd.CombinedOutput()
+	if err != nil {
+		t.Fatalf("nm failure: %s\n%s\n", err, string(out))
+	}
+	if !bytes.Contains(out, []byte(" T a-tramp0")) {
+		t.Errorf("Trampoline a-tramp0 is missing")
+	}
+	if bytes.Contains(out, []byte(" T b-tramp0")) {
+		t.Errorf("Trampoline b-tramp0 exists unnecessarily")
+	}
+}