esp32s3: fix register-window corruption under interrupt load

Remove the C3 bluetooth hook addresses from esp32s3.ld (on the S3 they
point into the ROM md5/crc thunk table, and being bare assignments they
also shadowed the blob's own definitions), keep the interrupt frame clear
of the 16-byte windowed-ABI save area below SP, and make tinygo_swapTask
hold INTLEVEL across the stack switch while keeping the running frame's
WINDOWSTART bit set.

Signed-off-by: deadprogram <ron@hybridgroup.com>

.github/workflows/build-macos.yml

@@ -40,7 +40,7 @@ jobs:
       - name: Install Go
         uses: actions/setup-go@v6
         with:
-          go-version: '1.27.0-rc.2'
+          go-version: '1.27.0'
           cache: true
       - name: Restore LLVM source cache
         uses: actions/cache/restore@v5
@@ -131,7 +131,7 @@ jobs:
       - name: Install Go
         uses: actions/setup-go@v6
         with:
-          go-version: '1.27.0-rc.2'
+          go-version: '1.27.0'
           cache: true
       - name: Build TinyGo (LLVM ${{ matrix.version }})
         run: go install -tags=llvm${{ matrix.version }}

.github/workflows/linux.yml

@@ -23,7 +23,7 @@ jobs:
       - name: Install Go
         uses: actions/setup-go@v6
         with:
-          go-version: '1.27.0-rc.2'
+          go-version: '1.27.0'
           cache: true
       - name: Run go mod tidy
         run: go mod tidy
@@ -36,7 +36,7 @@ jobs:
     # statically linked binary.
     runs-on: ubuntu-latest
     container:
-      image: golang:1.27rc2-alpine
+      image: golang:1.27-alpine
     outputs:
       version: ${{ steps.version.outputs.version }}
     steps:
@@ -160,7 +160,7 @@ jobs:
       - name: Install Go
         uses: actions/setup-go@v6
         with:
-          go-version: '1.27.0-rc.2'
+          go-version: '1.27.0'
           cache: true
       - name: Install wasmtime
         uses: bytecodealliance/actions/wasmtime/setup@v1
@@ -196,6 +196,10 @@ jobs:
           cat /proc/cpuinfo
           sudo apt-get update
           sudo apt-get install --no-install-recommends \
+              libgcrypt20 \
+              libpixman-1-0 \
+              libsdl2-2.0-0 \
+              libslirp0 \
               qemu-system-arm \
               qemu-system-riscv32 \
               qemu-user \
@@ -204,7 +208,7 @@ jobs:
       - name: Install Go
         uses: actions/setup-go@v6
         with:
-          go-version: '1.27.0-rc.2'
+          go-version: '1.27.0'
           cache: true
       - name: Install Node.js
         uses: actions/setup-node@v6
@@ -216,6 +220,14 @@ jobs:
           version: "29.0.1"
       - name: Setup `wasm-tools`
         uses: bytecodealliance/actions/wasm-tools/setup@v1
+      - name: Install Espressif QEMU
+        run: |
+          release=esp-develop-9.2.2-20260417
+          archive=qemu-xtensa-softmmu-${release//-/_}-x86_64-linux-gnu.tar.xz
+          curl -fL --retry 3 -o "$RUNNER_TEMP/$archive" \
+            "https://github.com/espressif/qemu/releases/download/$release/$archive"
+          tar -xJf "$RUNNER_TEMP/$archive" -C "$RUNNER_TEMP"
+          echo "$RUNNER_TEMP/qemu/bin" >> "$GITHUB_PATH"
       - name: Restore LLVM source cache
         uses: actions/cache/restore@v5
         id: cache-llvm-source
@@ -323,7 +335,7 @@ jobs:
       - name: Install Go
         uses: actions/setup-go@v6
         with:
-          go-version: '1.27.0-rc.2'
+          go-version: '1.27.0'
           cache: true
       - name: Restore LLVM source cache
         uses: actions/cache/restore@v5

.github/workflows/windows.yml

@@ -34,7 +34,7 @@ jobs:
       - name: Install Go
         uses: actions/setup-go@v6
         with:
-          go-version: '1.27.0-rc.2'
+          go-version: '1.27.0'
           cache: true
       - name: Restore cached LLVM source
         uses: actions/cache/restore@v5
@@ -129,7 +129,7 @@ jobs:
       - name: Install Go
         uses: actions/setup-go@v6
         with:
-          go-version: '1.27.0-rc.2'
+          go-version: '1.27.0'
           cache: true
       - name: Download TinyGo build
         uses: actions/download-artifact@v8
@@ -150,7 +150,7 @@ jobs:
       - name: Install Go
         uses: actions/setup-go@v6
         with:
-          go-version: '1.27.0-rc.2'
+          go-version: '1.27.0'
           cache: true
       - name: Download TinyGo build
         uses: actions/download-artifact@v8
@@ -176,7 +176,7 @@ jobs:
       - name: Install Go
         uses: actions/setup-go@v6
         with:
-          go-version: '1.27.0-rc.2'
+          go-version: '1.27.0'
           cache: true
       - name: Download TinyGo build
         uses: actions/download-artifact@v8

Dockerfile

@@ -1,5 +1,5 @@
 # tinygo-llvm stage obtains the llvm source for TinyGo
-FROM golang:1.27rc2 AS tinygo-llvm
+FROM golang:1.27 AS tinygo-llvm
 
 RUN apt-get update && \
     apt-get install -y apt-utils make cmake clang-17 ninja-build && \
@@ -33,7 +33,7 @@ RUN cd /tinygo/ && \
 
 # tinygo-compiler copies the compiler build over to a base Go container (without
 # all the build tools etc).
-FROM golang:1.27rc2 AS tinygo-compiler
+FROM golang:1.27 AS tinygo-compiler
 
 # Copy tinygo build.
 COPY --from=tinygo-compiler-build /tinygo/build/release/tinygo /tinygo

GNUmakefile

@@ -473,9 +473,15 @@ TEST_PACKAGES_NONWASM = \
 #     (just like wasm).
 #   * picolibc math functions apparently are less precise, the math package
 #     fails on baremetal.
+#   * Since Go 1.27 the crypto tests below go through cryptotest.TestHash, which
+#     calls cryptotest.BoundarySlices. These targets report GOOS=linux, so they
+#     build boundary.go (//go:build linux || darwin) rather than
+#     boundary_compat.go, and that needs a working syscall.Mmap/syscall.Mprotect
+#     which we don't have. See #5593.
 TEST_PACKAGES_BAREMETAL = $(filter-out $(TEST_PACKAGES_NONBAREMETAL), $(TEST_PACKAGES_FAST))
 TEST_PACKAGES_NONBAREMETAL = \
 	$(TEST_PACKAGES_NONWASM) \
+	$(TEST_PACKAGES_NOBOUNDARYSLICES) \
 	math \
 	$(nil)
 
@@ -484,6 +490,19 @@ TEST_PACKAGES_NOWASI = \
 	crypto/ecdsa \
 	$(nil)
 
+# wasip1 reports GOOS=wasip1 and so gets the boundary_compat.go fallback, but
+# wasip2 reports GOOS=linux and hits the same BoundarySlices problem as
+# baremetal. On wasip2 syscall.Mmap returns ENOSYS and t.Fatalf cannot Goexit,
+# so the test falls through and panics with "slice out of range".
+TEST_PACKAGES_FAST_WASIP2 = $(filter-out $(TEST_PACKAGES_NOBOUNDARYSLICES), $(TEST_PACKAGES_FAST_WASI))
+
+TEST_PACKAGES_NOBOUNDARYSLICES = \
+	crypto/md5 \
+	crypto/sha1 \
+	crypto/sha256 \
+	crypto/sha512 \
+	$(nil)
+
 # Report platforms on which each standard library package is known to pass tests
 report-stdlib-tests-pass:
 	$(eval jointmp := $(shell echo /tmp/join.$$$$))
@@ -510,7 +529,7 @@ TEST_PACKAGES_HOST := $(TEST_PACKAGES_FAST) $(TEST_PACKAGES_WINDOWS)
 TEST_IOFS := false
 endif
 
-TEST_SKIP_FLAG := -skip='TestExtraMethods|TestParseAndBytesRoundTrip/P256/Generic|TestAsValidation'
+TEST_SKIP_FLAG := -skip='TestExtraMethods|TestParseAndBytesRoundTrip/P256/Generic|TestAsValidation|TestUnmarshalNestingLimitSlice|TestUnmarshalNestingLimitStruct'
 TEST_ADDITIONAL_FLAGS ?=
 
 # Test known-working standard library packages.
@@ -519,6 +538,8 @@ TEST_ADDITIONAL_FLAGS ?=
 tinygo-test:
 	@# TestExtraMethods: used by many crypto packages and uses reflect.Type.Method which is not implemented.
 	@# TestParseAndBytesRoundTrip/P256/Generic: needs Goexit to run defers on wasm.
+	@# TestUnmarshalNestingLimit{Slice,Struct}: encoding/asn1 nesting limit added in
+	@# https://github.com/golang/go/commit/6a6d115f9a7422b2fa081ba6f567eefb4a099462
 	$(TINYGO) test $(TEST_ADDITIONAL_FLAGS) $(TEST_SKIP_FLAG) $(filter-out encoding/xml,$(TEST_PACKAGES_HOST)) $(TEST_PACKAGES_SLOW)
 ifeq ($(TEST_ENCODING_XML),true)
 	$(TINYGO) test $(TEST_ADDITIONAL_FLAGS) $(TEST_SKIP_FLAG) -stack-size=16MB encoding/xml
@@ -549,7 +570,7 @@ tinygo-test-wasip1-fast:
 tinygo-test-wasip2-slow:
 	$(TINYGO) test -target=wasip2 $(TEST_SKIP_FLAG) $(TEST_PACKAGES_SLOW)
 tinygo-test-wasip2-fast:
-	$(TINYGO) test -target=wasip2 $(TEST_SKIP_FLAG) $(TEST_PACKAGES_FAST_WASI) ./tests/runtime_wasi
+	$(TINYGO) test -target=wasip2 $(TEST_SKIP_FLAG) $(TEST_PACKAGES_FAST_WASIP2) ./tests/runtime_wasi
 
 tinygo-test-wasip2-sum-slow:
 	TINYGO=$(TINYGO) \

builder/build.go

@@ -24,6 +24,7 @@ import (
 	"sort"
 	"strconv"
 	"strings"
+	"sync"
 
 	"github.com/gofrs/flock"
 	"github.com/tinygo-org/tinygo/compileopts"
@@ -269,6 +270,7 @@ func Build(pkgName, outpath, tmpdir string, config *compileopts.Config) (BuildRe
 	// Create the *ssa.Program. This does not yet build the entire SSA of the
 	// program so it's pretty fast and doesn't need to be parallelized.
 	program := lprogram.LoadSSA()
+	buildProgram := sync.OnceFunc(program.Build)
 
 	// Add jobs to compile each package.
 	// Packages that have a cache hit will not be compiled again.
@@ -398,8 +400,13 @@ func Build(pkgName, outpath, tmpdir string, config *compileopts.Config) (BuildRe
 					return nil
 				}
 
-				// Compile AST to IR. The compiler.CompilePackage function will
-				// build the SSA as needed.
+				// SSA package builds may run concurrently, but the resulting
+				// functions cannot be inspected until all builds have finished:
+				// generic instances and wrappers can be shared across packages.
+				// Build the whole program once before compiling any package.
+				buildProgram()
+
+				// Compile AST to IR.
 				mod, errs := compiler.CompilePackage(pkg.ImportPath, pkg, program.Package(pkg.Pkg), machine, compilerConfig, config.DumpSSA())
 				defer mod.Context().Dispose()
 				defer mod.Dispose()

builder/esp.go

@@ -111,9 +111,6 @@ func makeESPFirmwareImage(infile, outfile, format string) error {
 		// Compute the size of the RAM portion of the image (everything the ROM
 		// bootloader loads, up to and including the appended SHA256 hash).
 		ramImageSize := 0
-		if makeImage {
-			ramImageSize += 4096
-		}
 		ramImageSize += 24 // image header (8) + trailer fields (16)
 		for _, seg := range segments {
 			ramImageSize += 8 + len(seg.data) // segment header + data (4-aligned)
@@ -284,6 +281,9 @@ func makeESPFirmwareImage(infile, outfile, format string) error {
 		dromSize := 0
 		if len(dromSegs) > 0 {
 			targetImageOffset := int(dromFlashAddr - flashBase)
+			if makeImage {
+				targetImageOffset = int(dromFlashAddr)
+			}
 			if outf.Len() > targetImageOffset {
 				return fmt.Errorf("ESP32: RAM segments too large (%d bytes), overlap DROM at flash 0x%x", outf.Len(), dromFlashAddr)
 			}
@@ -304,6 +304,9 @@ func makeESPFirmwareImage(infile, outfile, format string) error {
 			}
 			iromFlashAddr := dromFlashAddr + uint32(dromPages)*pageSize
 			targetImageOffset := int(iromFlashAddr - flashBase)
+			if makeImage {
+				targetImageOffset = int(iromFlashAddr)
+			}
 			if outf.Len() > targetImageOffset {
 				return fmt.Errorf("ESP32: DROM too large, overlaps IROM at flash 0x%x", iromFlashAddr)
 			}

builder/sizes_test.go

@@ -42,9 +42,9 @@ func TestBinarySize(t *testing.T) {
 	// This is a small number of very diverse targets that we want to test.
 	tests := []sizeTest{
 		// microcontrollers
-		{"hifive1b", "examples/echo", 4313, 323, 0, 2260},
-		{"microbit", "examples/serial", 2838, 382, 8, 2256},
-		{"wioterminal", "examples/pininterrupt", 8027, 1665, 132, 7488},
+		{"hifive1b", "examples/echo", 4321, 323, 0, 2268},
+		{"microbit", "examples/serial", 2842, 382, 8, 2264},
+		{"wioterminal", "examples/pininterrupt", 8039, 1665, 132, 7496},
 
 		// TODO: also check wasm. Right now this is difficult, because
 		// wasm binaries are run through wasm-opt and therefore the

compiler/compiler.go

@@ -296,6 +296,9 @@ func Sizes(machine llvm.TargetMachine) types.Sizes {
 }
 
 // CompilePackage compiles a single package to a LLVM module.
+//
+// The SSA package must already be built. When packages are compiled
+// concurrently, the entire SSA program must be built before compilation starts.
 func CompilePackage(moduleName string, pkg *loader.Package, ssaPkg *ssa.Package, machine llvm.TargetMachine, config *Config, dumpSSA bool) (llvm.Module, []error) {
 	c := newCompilerContext(moduleName, machine, config, dumpSSA)
 	defer c.dispose()
@@ -306,9 +309,6 @@ func CompilePackage(moduleName string, pkg *loader.Package, ssaPkg *ssa.Package,
 	c.runtimePkg = ssaPkg.Prog.ImportedPackage("runtime").Pkg
 	c.program = ssaPkg.Prog
 
-	// Convert AST to SSA.
-	ssaPkg.Build()
-
 	// Assign names to function-local named types before compiling the
 	// package, so that types declared in different functions (or in
 	// different instantiations of a generic function) do not collide.

compiler/compiler_test.go

@@ -115,7 +115,7 @@ func TestCompiler(t *testing.T) {
 
 			// Update test if needed. Do not check the result.
 			if *flagUpdate {
-				err := os.WriteFile(outPath, []byte(mod.String()), 0666)
+				err := os.WriteFile(outPath, []byte(normalizeIR(mod.String())), 0666)
 				if err != nil {
 					t.Error("failed to write updated output file:", err)
 				}
@@ -174,9 +174,8 @@ func TestOptimizedLargeAggregateABI(t *testing.T) {
 	}
 }
 
-// normalizeIR canonicalizes LLVM IR so a single golden file keeps matching
-// across LLVM versions. Golden files are written against LLVM <21; newer LLVM
-// prints some attributes differently.
+// normalizeIR canonicalizes LLVM-version-specific IR spellings for comparison
+// and when regenerating golden files.
 func normalizeIR(s string) string {
 	// Golden files are written using the pre-LLVM21 'nocapture' spelling,
 	// which LLVM printed before any co-occurring attribute such as
@@ -414,5 +413,7 @@ func testCompilePackage(t *testing.T, options *compileopts.Options, file string)
 	// Compile AST to IR.
 	program := lprogram.LoadSSA()
 	pkg := lprogram.MainPkg()
-	return CompilePackage(file, pkg, program.Package(pkg.Pkg), machine, compilerConfig, false)
+	ssaPkg := program.Package(pkg.Pkg)
+	ssaPkg.Build()
+	return CompilePackage(file, pkg, ssaPkg, machine, compilerConfig, false)
 }

compiler/testdata/channel.ll

@@ -19,33 +19,33 @@ define hidden void @main.chanIntSend(ptr dereferenceable_or_null(36) %ch, ptr %c
 entry:
   %chan.op = alloca %runtime.channelOp, align 8
   %chan.value = alloca i32, align 4
-  call void @llvm.lifetime.start.p0(i64 4, ptr nonnull %chan.value)
+  call void @llvm.lifetime.start.p0(ptr nonnull %chan.value)
   store i32 3, ptr %chan.value, align 4
-  call void @llvm.lifetime.start.p0(i64 16, ptr nonnull %chan.op)
+  call void @llvm.lifetime.start.p0(ptr nonnull %chan.op)
   call void @runtime.chanSend(ptr %ch, ptr nonnull %chan.value, ptr nonnull %chan.op, ptr undef) #3
-  call void @llvm.lifetime.end.p0(i64 16, ptr nonnull %chan.op)
-  call void @llvm.lifetime.end.p0(i64 4, ptr nonnull %chan.value)
+  call void @llvm.lifetime.end.p0(ptr nonnull %chan.op)
+  call void @llvm.lifetime.end.p0(ptr nonnull %chan.value)
   ret void
 }
 
 ; Function Attrs: nocallback nofree nosync nounwind willreturn memory(argmem: readwrite)
-declare void @llvm.lifetime.start.p0(i64 immarg, ptr nocapture) #2
+declare void @llvm.lifetime.start.p0(ptr nocapture) #2
 
 declare void @runtime.chanSend(ptr dereferenceable_or_null(36), ptr, ptr dereferenceable_or_null(16), ptr) #0
 
 ; Function Attrs: nocallback nofree nosync nounwind willreturn memory(argmem: readwrite)
-declare void @llvm.lifetime.end.p0(i64 immarg, ptr nocapture) #2
+declare void @llvm.lifetime.end.p0(ptr nocapture) #2
 
 ; Function Attrs: nounwind
 define hidden void @main.chanIntRecv(ptr dereferenceable_or_null(36) %ch, ptr %context) unnamed_addr #1 {
 entry:
   %chan.op = alloca %runtime.channelOp, align 8
   %chan.value = alloca i32, align 4
-  call void @llvm.lifetime.start.p0(i64 4, ptr nonnull %chan.value)
-  call void @llvm.lifetime.start.p0(i64 16, ptr nonnull %chan.op)
+  call void @llvm.lifetime.start.p0(ptr nonnull %chan.value)
+  call void @llvm.lifetime.start.p0(ptr nonnull %chan.op)
   %0 = call i1 @runtime.chanRecv(ptr %ch, ptr nonnull %chan.value, ptr nonnull %chan.op, ptr undef) #3
-  call void @llvm.lifetime.end.p0(i64 4, ptr nonnull %chan.value)
-  call void @llvm.lifetime.end.p0(i64 16, ptr nonnull %chan.op)
+  call void @llvm.lifetime.end.p0(ptr nonnull %chan.value)
+  call void @llvm.lifetime.end.p0(ptr nonnull %chan.op)
   ret void
 }
 
@@ -55,9 +55,9 @@ declare i1 @runtime.chanRecv(ptr dereferenceable_or_null(36), ptr, ptr dereferen
 define hidden void @main.chanZeroSend(ptr dereferenceable_or_null(36) %ch, ptr %context) unnamed_addr #1 {
 entry:
   %chan.op = alloca %runtime.channelOp, align 8
-  call void @llvm.lifetime.start.p0(i64 16, ptr nonnull %chan.op)
+  call void @llvm.lifetime.start.p0(ptr nonnull %chan.op)
   call void @runtime.chanSend(ptr %ch, ptr null, ptr nonnull %chan.op, ptr undef) #3
-  call void @llvm.lifetime.end.p0(i64 16, ptr nonnull %chan.op)
+  call void @llvm.lifetime.end.p0(ptr nonnull %chan.op)
   ret void
 }
 
@@ -65,9 +65,9 @@ entry:
 define hidden void @main.chanZeroRecv(ptr dereferenceable_or_null(36) %ch, ptr %context) unnamed_addr #1 {
 entry:
   %chan.op = alloca %runtime.channelOp, align 8
-  call void @llvm.lifetime.start.p0(i64 16, ptr nonnull %chan.op)
+  call void @llvm.lifetime.start.p0(ptr nonnull %chan.op)
   %0 = call i1 @runtime.chanRecv(ptr %ch, ptr null, ptr nonnull %chan.op, ptr undef) #3
-  call void @llvm.lifetime.end.p0(i64 16, ptr nonnull %chan.op)
+  call void @llvm.lifetime.end.p0(ptr nonnull %chan.op)
   ret void
 }
 
@@ -77,7 +77,7 @@ entry:
   %select.states.alloca = alloca [2 x %runtime.chanSelectState], align 8
   %select.send.value = alloca i32, align 4
   store i32 1, ptr %select.send.value, align 4
-  call void @llvm.lifetime.start.p0(i64 16, ptr nonnull %select.states.alloca)
+  call void @llvm.lifetime.start.p0(ptr nonnull %select.states.alloca)
   store ptr %ch1, ptr %select.states.alloca, align 4
   %select.states.alloca.repack1 = getelementptr inbounds nuw i8, ptr %select.states.alloca, i32 4
   store ptr %select.send.value, ptr %select.states.alloca.repack1, align 4
@@ -86,7 +86,7 @@ entry:
   %.repack3 = getelementptr inbounds nuw i8, ptr %select.states.alloca, i32 12
   store ptr null, ptr %.repack3, align 4
   %select.result = call { i32, i1 } @runtime.chanSelect(ptr undef, ptr nonnull %select.states.alloca, i32 2, i32 2, ptr null, i32 0, i32 0, ptr undef) #3
-  call void @llvm.lifetime.end.p0(i64 16, ptr nonnull %select.states.alloca)
+  call void @llvm.lifetime.end.p0(ptr nonnull %select.states.alloca)
   %1 = extractvalue { i32, i1 } %select.result, 0
   %2 = icmp eq i32 %1, 0
   br i1 %2, label %select.done, label %select.next

compiler/testdata/large.ll

@@ -15,7 +15,7 @@ target triple = "wasm32-unknown-wasi"
 @"reflect/types.type:basic:string" = linkonce_odr constant { i8, ptr } { i8 81, ptr @"reflect/types.type:pointer:basic:string" }, align 4
 @"reflect/types.type:pointer:basic:string" = linkonce_odr constant { i8, i16, ptr } { i8 -43, i16 0, ptr @"reflect/types.type:basic:string" }, align 4
 
-declare void @runtime.trackPointer(ptr readonly captures(none), ptr, ptr) #0
+declare void @runtime.trackPointer(ptr nocapture readonly, ptr, ptr) #0
 
 ; Function Attrs: nounwind
 define hidden void @main.init(ptr %context) unnamed_addr #1 {
@@ -31,7 +31,7 @@ entry:
 }
 
 ; Function Attrs: nocallback nofree nounwind willreturn memory(argmem: readwrite)
-declare void @llvm.memcpy.p0.p0.i32(ptr noalias writeonly captures(none), ptr noalias readonly captures(none), i32, i1 immarg) #2
+declare void @llvm.memcpy.p0.p0.i32(ptr noalias nocapture writeonly, ptr noalias nocapture readonly, i32, i1 immarg) #2
 
 ; Function Attrs: nounwind
 define hidden i8 @"(main.largeReceiver).readLargeValue"(ptr readonly dereferenceable_or_null(1025) %receiver, ptr readonly dereferenceable_or_null(1025) %value, ptr %context) unnamed_addr #1 {
@@ -344,7 +344,7 @@ if.then:                                          ; preds = %typeassert.next
 declare i1 @runtime.typeAssert(ptr, ptr dereferenceable_or_null(1), ptr) #0
 
 ; Function Attrs: nocallback nofree nounwind willreturn memory(argmem: write)
-declare void @llvm.memset.p0.i32(ptr writeonly captures(none), i8, i32, i1 immarg) #7
+declare void @llvm.memset.p0.i32(ptr nocapture writeonly, i8, i32, i1 immarg) #7
 
 ; Function Attrs: nounwind
 define hidden i8 @main.useLargeMap(ptr readonly dereferenceable_or_null(1025) %key, ptr readonly dereferenceable_or_null(1025) %value, ptr %context) unnamed_addr #1 {
@@ -423,12 +423,12 @@ if.then:                                          ; preds = %entry
 }
 
 ; Function Attrs: nocallback nofree nosync nounwind willreturn memory(argmem: readwrite)
-declare void @llvm.lifetime.start.p0(ptr captures(none)) #8
+declare void @llvm.lifetime.start.p0(ptr nocapture) #8
 
 declare void @runtime.chanSend(ptr dereferenceable_or_null(36), ptr, ptr dereferenceable_or_null(16), ptr) #0
 
 ; Function Attrs: nocallback nofree nosync nounwind willreturn memory(argmem: readwrite)
-declare void @llvm.lifetime.end.p0(ptr captures(none)) #8
+declare void @llvm.lifetime.end.p0(ptr nocapture) #8
 
 declare i1 @runtime.chanRecv(ptr dereferenceable_or_null(36), ptr, ptr dereferenceable_or_null(16), ptr) #0
 

compiler/testdata/zeromap.ll

@@ -21,23 +21,23 @@ entry:
   %0 = insertvalue %main.hasPadding zeroinitializer, i1 %s.b1, 0
   %1 = insertvalue %main.hasPadding %0, i32 %s.i, 1
   %2 = insertvalue %main.hasPadding %1, i1 %s.b2, 2
-  call void @llvm.lifetime.start.p0(i64 4, ptr nonnull %hashmap.value)
-  call void @llvm.lifetime.start.p0(i64 12, ptr nonnull %hashmap.key)
+  call void @llvm.lifetime.start.p0(ptr nonnull %hashmap.value)
+  call void @llvm.lifetime.start.p0(ptr nonnull %hashmap.key)
   store %main.hasPadding %2, ptr %hashmap.key, align 4
   %3 = call i1 @runtime.hashmapGenericGet(ptr %m, ptr nonnull %hashmap.key, ptr nonnull %hashmap.value, i32 4, ptr undef) #4
-  call void @llvm.lifetime.end.p0(i64 12, ptr nonnull %hashmap.key)
+  call void @llvm.lifetime.end.p0(ptr nonnull %hashmap.key)
   %4 = load i32, ptr %hashmap.value, align 4
-  call void @llvm.lifetime.end.p0(i64 4, ptr nonnull %hashmap.value)
+  call void @llvm.lifetime.end.p0(ptr nonnull %hashmap.value)
   ret i32 %4
 }
 
 ; Function Attrs: nocallback nofree nosync nounwind willreturn memory(argmem: readwrite)
-declare void @llvm.lifetime.start.p0(i64 immarg, ptr nocapture) #3
+declare void @llvm.lifetime.start.p0(ptr nocapture) #3
 
 declare i1 @runtime.hashmapGenericGet(ptr dereferenceable_or_null(48), ptr nocapture, ptr nocapture, i32, ptr) #0
 
 ; Function Attrs: nocallback nofree nosync nounwind willreturn memory(argmem: readwrite)
-declare void @llvm.lifetime.end.p0(i64 immarg, ptr nocapture) #3
+declare void @llvm.lifetime.end.p0(ptr nocapture) #3
 
 ; Function Attrs: noinline nounwind
 define hidden void @main.testZeroSet(ptr dereferenceable_or_null(48) %m, i1 %s.b1, i32 %s.i, i1 %s.b2, ptr %context) unnamed_addr #2 {
@@ -47,13 +47,13 @@ entry:
   %0 = insertvalue %main.hasPadding zeroinitializer, i1 %s.b1, 0
   %1 = insertvalue %main.hasPadding %0, i32 %s.i, 1
   %2 = insertvalue %main.hasPadding %1, i1 %s.b2, 2
-  call void @llvm.lifetime.start.p0(i64 4, ptr nonnull %hashmap.value)
+  call void @llvm.lifetime.start.p0(ptr nonnull %hashmap.value)
   store i32 5, ptr %hashmap.value, align 4
-  call void @llvm.lifetime.start.p0(i64 12, ptr nonnull %hashmap.key)
+  call void @llvm.lifetime.start.p0(ptr nonnull %hashmap.key)
   store %main.hasPadding %2, ptr %hashmap.key, align 4
   call void @runtime.hashmapGenericSet(ptr %m, ptr nonnull %hashmap.key, ptr nonnull %hashmap.value, ptr undef) #4
-  call void @llvm.lifetime.end.p0(i64 12, ptr nonnull %hashmap.key)
-  call void @llvm.lifetime.end.p0(i64 4, ptr nonnull %hashmap.value)
+  call void @llvm.lifetime.end.p0(ptr nonnull %hashmap.key)
+  call void @llvm.lifetime.end.p0(ptr nonnull %hashmap.value)
   ret void
 }
 
@@ -64,17 +64,17 @@ define hidden i32 @main.testZeroArrayGet(ptr dereferenceable_or_null(48) %m, [2
 entry:
   %hashmap.key = alloca [2 x %main.hasPadding], align 8
   %hashmap.value = alloca i32, align 4
-  call void @llvm.lifetime.start.p0(i64 4, ptr nonnull %hashmap.value)
-  call void @llvm.lifetime.start.p0(i64 24, ptr nonnull %hashmap.key)
+  call void @llvm.lifetime.start.p0(ptr nonnull %hashmap.value)
+  call void @llvm.lifetime.start.p0(ptr nonnull %hashmap.key)
   %s.elt = extractvalue [2 x %main.hasPadding] %s, 0
   store %main.hasPadding %s.elt, ptr %hashmap.key, align 4
   %hashmap.key.repack1 = getelementptr inbounds nuw i8, ptr %hashmap.key, i32 12
   %s.elt2 = extractvalue [2 x %main.hasPadding] %s, 1
   store %main.hasPadding %s.elt2, ptr %hashmap.key.repack1, align 4
   %0 = call i1 @runtime.hashmapGenericGet(ptr %m, ptr nonnull %hashmap.key, ptr nonnull %hashmap.value, i32 4, ptr undef) #4
-  call void @llvm.lifetime.end.p0(i64 24, ptr nonnull %hashmap.key)
+  call void @llvm.lifetime.end.p0(ptr nonnull %hashmap.key)
   %1 = load i32, ptr %hashmap.value, align 4
-  call void @llvm.lifetime.end.p0(i64 4, ptr nonnull %hashmap.value)
+  call void @llvm.lifetime.end.p0(ptr nonnull %hashmap.value)
   ret i32 %1
 }
 
@@ -83,17 +83,17 @@ define hidden void @main.testZeroArraySet(ptr dereferenceable_or_null(48) %m, [2
 entry:
   %hashmap.key = alloca [2 x %main.hasPadding], align 8
   %hashmap.value = alloca i32, align 4
-  call void @llvm.lifetime.start.p0(i64 4, ptr nonnull %hashmap.value)
+  call void @llvm.lifetime.start.p0(ptr nonnull %hashmap.value)
   store i32 5, ptr %hashmap.value, align 4
-  call void @llvm.lifetime.start.p0(i64 24, ptr nonnull %hashmap.key)
+  call void @llvm.lifetime.start.p0(ptr nonnull %hashmap.key)
   %s.elt = extractvalue [2 x %main.hasPadding] %s, 0
   store %main.hasPadding %s.elt, ptr %hashmap.key, align 4
   %hashmap.key.repack1 = getelementptr inbounds nuw i8, ptr %hashmap.key, i32 12
   %s.elt2 = extractvalue [2 x %main.hasPadding] %s, 1
   store %main.hasPadding %s.elt2, ptr %hashmap.key.repack1, align 4
   call void @runtime.hashmapGenericSet(ptr %m, ptr nonnull %hashmap.key, ptr nonnull %hashmap.value, ptr undef) #4
-  call void @llvm.lifetime.end.p0(i64 24, ptr nonnull %hashmap.key)
-  call void @llvm.lifetime.end.p0(i64 4, ptr nonnull %hashmap.value)
+  call void @llvm.lifetime.end.p0(ptr nonnull %hashmap.key)
+  call void @llvm.lifetime.end.p0(ptr nonnull %hashmap.value)
   ret void
 }
 

main.go

@@ -177,8 +177,15 @@ func Build(pkgName, outpath string, config *compileopts.Config) error {
 				// Pick a default output path based on the main directory.
 				outpath = filepath.Base(result.MainDir) + config.DefaultBinaryExtension()
 			}
+		} else if fi, statErr := os.Stat(outpath); statErr == nil && fi.IsDir() {
+			var name string
+			if strings.HasSuffix(pkgName, ".go") {
+				name = filepath.Base(pkgName[:len(pkgName)-3]) + config.DefaultBinaryExtension()
+			} else {
+				name = filepath.Base(result.MainDir) + config.DefaultBinaryExtension()
+			}
+			outpath = filepath.Join(outpath, name)
 		}
-
 		if err := os.Rename(result.Binary, outpath); err != nil {
 			// Moving failed. Do a file copy.
 			inf, err := os.Open(result.Binary)

main_test.go

@@ -261,6 +261,21 @@ func TestTimerStopResetRace(t *testing.T) {
 	runTest("timer_stop_reset_race.go", optionsFromTarget("", sema), t, nil, nil)
 }
 
+func TestESP32QEMU(t *testing.T) {
+	t.Parallel()
+
+	options := optionsFromTarget("esp32-qemu", sema)
+	emuCheck(t, options)
+	machines, err := exec.Command("qemu-system-xtensa", "-machine", "help").Output()
+	if err != nil {
+		t.Fatal("failed to list qemu-system-xtensa machines:", err)
+	}
+	if !regexp.MustCompile(`(?m)^esp32\s`).Match(machines) {
+		t.Skip("qemu-system-xtensa does not support the ESP32 machine")
+	}
+	runTest("print.go", options, t, nil, nil)
+}
+
 func runPlatTests(options compileopts.Options, tests []string, t *testing.T) {
 	emuCheck(t, options)
 
@@ -539,6 +554,9 @@ func runTestWithConfig(name string, t *testing.T, options compileopts.Options, c
 	if config.EmulatorName() == "simavr" {
 		actual = cleanSimAVRTestOutput(actual)
 	}
+	if config.EmulatorName() == "qemu-system-xtensa" {
+		actual = cleanESP32QEMUOutput(actual)
+	}
 	if name == "testing.go" {
 		// Strip actual time.
 		re := regexp.MustCompile(`\([0-9]\.[0-9][0-9]s\)`)
@@ -564,6 +582,18 @@ func runTestWithConfig(name string, t *testing.T, options compileopts.Options, c
 	}
 }
 
+func cleanESP32QEMUOutput(output []byte) []byte {
+	entryLine := bytes.Index(output, []byte("\nentry "))
+	if entryLine < 0 {
+		return output
+	}
+	entryLineEnd := bytes.IndexByte(output[entryLine+1:], '\n')
+	if entryLineEnd < 0 {
+		return output
+	}
+	return output[entryLine+1+entryLineEnd+1:]
+}
+
 func cleanSimAVRTestOutput(output []byte) []byte {
 	output = bytes.ReplaceAll(output, []byte{0x1b, '[', '3', '2', 'm'}, nil)
 	output = bytes.ReplaceAll(output, []byte{0x1b, '[', '0', 'm'}, nil)

src/internal/task/task_stack_esp32.S

@@ -68,14 +68,20 @@ tinygo_swapTask:
 
     // After the recursive spill returns, the physical register file still
     // has WindowStart bits set for the spill helper frames.
-    // We will clear WindowStart completely (to 0) right before the retw.n
-    // below, after the stack switch is done.  This prevents stale overflow
-    // when the new goroutine's calls rotate back into these panes.
+    // We will drop every WindowStart bit except this window's own
+    // (WINDOWSTART = 1 << WINDOWBASE) right before the retw.n below, after the
+    // stack switch is done.  This prevents stale overflow when the new
+    // goroutine's calls rotate back into these panes.
 
-    // Restore interrupts.
+    // Reload the saved PS, but do NOT restore it yet: everything from the stack
+    // switch below until WindowStart is made consistent with the new stack must
+    // stay masked.  Once sp points at the new task's stack, the memory below sp
+    // is that task's save area -- the a0-a3 the retw.n is about to reload
+    // through a window underflow.  An interrupt landing in that gap builds its
+    // frame there and overwrites them, and a garbage a0 makes the next retw an
+    // Illegal Instruction.  The PS restore below is therefore deferred to just
+    // before the retw.n, once the window state matches the new stack.
     l32i a4, sp, 4          // reload saved PS
-    wsr.ps a4
-    rsync
 
     // At this point, the following is true:
     //     WindowStart == 1 << WindowBase
@@ -97,16 +103,24 @@ tinygo_swapTask:
     // register also stores the parent register window.
     l32i.n a0, sp, 0
 
-    // Clear ALL WindowStart bits.  With all windows spilled to the stack,
-    // we must ensure no stale WS bits remain: the retw.n below will trigger
-    // underflow4 to load the new goroutine's registers from the new stack
-    // (which sets the appropriate WS bit via rfwu).  Any stale WS bits
-    // (from spill helpers or the old goroutine) would cause spurious
-    // overflows of garbage register values into memory.
-    movi a5, 0
+    // Drop every WindowStart bit except this window's.  Stale bits (from the
+    // spill helpers or the old goroutine) would overflow garbage registers into
+    // memory.  This window's own bit has to stay set: zeroing WindowStart marks
+    // the running frame dead, so an interrupt arriving here returns through a
+    // retw that underflows into a save area nothing ever wrote.  Keeping it
+    // costs nothing -- WindowStart[WindowBase - n] is still clear, so the
+    // retw.n below underflows as intended.
+    rsr  a6, WINDOWBASE
+    movi a5, 1
+    ssl  a6
+    sll  a5, a5             // a5 = 1 << WindowBase
     wsr  a5, WINDOWSTART
     rsync
 
+    // Now that the new task's window state is consistent, take interrupts back.
+    wsr.ps a4
+    rsync
+
     // Return into the new stack. This instruction will trigger a window
     // underflow, reloading the saved registers from the stack.
     retw.n

src/machine/machine_atsamd21_usb.go

@@ -51,7 +51,7 @@ func (dev *USBDevice) Configure(config UARTConfig) {
 	sam.USB_DEVICE.CTRLB.SetBits(sam.USB_DEVICE_CTRLB_SPDCONF_FS << sam.USB_DEVICE_CTRLB_SPDCONF_Pos)
 
 	// attach
-	sam.USB_DEVICE.CTRLB.ClearBits(sam.USB_DEVICE_CTRLB_DETACH)
+	dev.Attach()
 
 	// enable interrupt for end of reset
 	sam.USB_DEVICE.INTENSET.SetBits(sam.USB_DEVICE_INTENSET_EORST)
@@ -68,6 +68,20 @@ func (dev *USBDevice) Configure(config UARTConfig) {
 	dev.initcomplete = true
 }
 
+// Attach connects the device to the USB bus, allowing the host to detect and
+// enumerate it. It can be used together with Detach to delay enumeration
+// until the USB configuration (device identifiers, classes, ...) is complete.
+func (dev *USBDevice) Attach() {
+	sam.USB_DEVICE.CTRLB.ClearBits(sam.USB_DEVICE_CTRLB_DETACH)
+}
+
+// Detach disconnects the device from the USB bus. To the host this appears
+// as if the device was unplugged. A subsequent Attach makes the host
+// enumerate the device again.
+func (dev *USBDevice) Detach() {
+	sam.USB_DEVICE.CTRLB.SetBits(sam.USB_DEVICE_CTRLB_DETACH)
+}
+
 func handlePadCalibration() {
 	// Load Pad Calibration data from non-volatile memory
 	// This requires registers that are not included in the SVD file.

src/machine/machine_atsamd51_usb.go

@@ -51,7 +51,7 @@ func (dev *USBDevice) Configure(config UARTConfig) {
 	sam.USB_DEVICE.CTRLB.SetBits(sam.USB_DEVICE_CTRLB_SPDCONF_FS << sam.USB_DEVICE_CTRLB_SPDCONF_Pos)
 
 	// attach
-	sam.USB_DEVICE.CTRLB.ClearBits(sam.USB_DEVICE_CTRLB_DETACH)
+	dev.Attach()
 
 	// enable interrupt for end of reset
 	sam.USB_DEVICE.INTENSET.SetBits(sam.USB_DEVICE_INTENSET_EORST)
@@ -71,6 +71,20 @@ func (dev *USBDevice) Configure(config UARTConfig) {
 	dev.initcomplete = true
 }
 
+// Attach connects the device to the USB bus, allowing the host to detect and
+// enumerate it. It can be used together with Detach to delay enumeration
+// until the USB configuration (device identifiers, classes, ...) is complete.
+func (dev *USBDevice) Attach() {
+	sam.USB_DEVICE.CTRLB.ClearBits(sam.USB_DEVICE_CTRLB_DETACH)
+}
+
+// Detach disconnects the device from the USB bus. To the host this appears
+// as if the device was unplugged. A subsequent Attach makes the host
+// enumerate the device again.
+func (dev *USBDevice) Detach() {
+	sam.USB_DEVICE.CTRLB.SetBits(sam.USB_DEVICE_CTRLB_DETACH)
+}
+
 func handlePadCalibration() {
 	// Load Pad Calibration data from non-volatile memory
 	// This requires registers that are not included in the SVD file.

src/machine/machine_esp32.go

@@ -112,6 +112,11 @@ func (p Pin) configure(config PinConfig, signal uint32) {
 	// Configure the pad with the given IO mux configuration.
 	p.mux().Set(muxConfig)
 
+	// Internal pull resistors for pins with RTC function ignore
+	// the IO_MUX_GPIO0_FUN_WPU and IO_MUX_GPIO0_FUN_WPD bits set
+	// above and are instead controlled by the RTC_IO registers.
+	p.configureRTCPull(config.Mode)
+
 	switch config.Mode {
 	case PinOutput:
 		// Set the 'output enable' bit.
@@ -293,6 +298,60 @@ func (p Pin) mux() *volatile.Register32 {
 	}
 }
 
+// configureRTCPull applies the pullup/pulldown setting to a pin.
+// This mirrors ESP-IDF:
+// https://github.com/espressif/esp-idf/blob/08e0d30/components/esp_driver_gpio/src/gpio.c#L283
+// https://github.com/espressif/esp-idf/blob/08e0d30/components/esp_hal_gpio/esp32/rtc_io_periph.c#L57
+// https://github.com/espressif/esp-idf/blob/08e0d30/components/esp_hal_gpio/esp32/include/hal/rtc_io_ll.h#L175
+func (p Pin) configureRTCPull(mode PinMode) {
+	var rue, rde uint32
+	switch mode {
+	case PinInputPullup:
+		rue = 1
+	case PinInputPulldown:
+		rde = 1
+	}
+
+	switch p {
+	case 0:
+		esp.RTC_IO.SetTOUCH_PAD1_RUE(rue)
+		esp.RTC_IO.SetTOUCH_PAD1_RDE(rde)
+	case 2:
+		esp.RTC_IO.SetTOUCH_PAD2_RUE(rue)
+		esp.RTC_IO.SetTOUCH_PAD2_RDE(rde)
+	case 4:
+		esp.RTC_IO.SetTOUCH_PAD0_RUE(rue)
+		esp.RTC_IO.SetTOUCH_PAD0_RDE(rde)
+	case 12:
+		esp.RTC_IO.SetTOUCH_PAD5_RUE(rue)
+		esp.RTC_IO.SetTOUCH_PAD5_RDE(rde)
+	case 13:
+		esp.RTC_IO.SetTOUCH_PAD4_RUE(rue)
+		esp.RTC_IO.SetTOUCH_PAD4_RDE(rde)
+	case 14:
+		esp.RTC_IO.SetTOUCH_PAD6_RUE(rue)
+		esp.RTC_IO.SetTOUCH_PAD6_RDE(rde)
+	case 15:
+		esp.RTC_IO.SetTOUCH_PAD3_RUE(rue)
+		esp.RTC_IO.SetTOUCH_PAD3_RDE(rde)
+	case 25:
+		esp.RTC_IO.SetPAD_DAC1_PDAC1_RUE(rue)
+		esp.RTC_IO.SetPAD_DAC1_PDAC1_RDE(rde)
+	case 26:
+		esp.RTC_IO.SetPAD_DAC2_PDAC2_RUE(rue)
+		esp.RTC_IO.SetPAD_DAC2_PDAC2_RDE(rde)
+	case 27:
+		esp.RTC_IO.SetTOUCH_PAD7_RUE(rue)
+		esp.RTC_IO.SetTOUCH_PAD7_RDE(rde)
+	case 32:
+		esp.RTC_IO.SetXTAL_32K_PAD_X32P_RUE(rue)
+		esp.RTC_IO.SetXTAL_32K_PAD_X32P_RDE(rde)
+	case 33:
+		esp.RTC_IO.SetXTAL_32K_PAD_X32N_RUE(rue)
+		esp.RTC_IO.SetXTAL_32K_PAD_X32N_RDE(rde)
+	}
+}
+
 const maxPin = 40
 
 // cpuInterruptFromPin selects an edge-triggered CPU interrupt line for GPIO.

src/machine/machine_esp32_i2c.go

@@ -242,7 +242,6 @@ func (i2c *I2C) transmit(addr uint16, cmd []i2cCommand, timeoutMS int) error {
 	timeoutNS := int64(timeoutMS) * 1000000
 	needAddress := true
 	needRestart := false
-	readLast := false
 	var readTo []byte
 	for cmdIdx, reg := 0, &i2c.Bus.COMD0; cmdIdx < len(cmd); {
 		c := &cmd[cmdIdx]
@@ -310,7 +309,6 @@ func (i2c *I2C) transmit(addr uint16, cmd []i2cCommand, timeoutMS int) error {
 			}
 
 			if split {
-				readLast = true
 				reg.Set(i2cCMD_READLAST | 1)
 				reg = nextAddress(reg)
 				readTo = c.data[c.head : c.head+bytes+1] // read bytes + 1 last byte
@@ -342,7 +340,7 @@ func (i2c *I2C) transmit(addr uint16, cmd []i2cCommand, timeoutMS int) error {
 				}
 			}
 			switch {
-			case mask&esp.I2C_INT_STATUS_ACK_ERR_INT_ST_Msk != 0 && !readLast:
+			case mask&esp.I2C_INT_STATUS_ACK_ERR_INT_ST_Msk != 0:
 				return errI2CAckExpected
 			case mask&esp.I2C_INT_STATUS_TIME_OUT_INT_ST_Msk != 0:
 				// timeout leaves the bus in an undefined state, reset

src/machine/machine_esp32c3_usb.go

@@ -67,6 +67,11 @@ func (dev *USBDevice) SetStallEPOut(ep uint32)   {}
 func (dev *USBDevice) ClearStallEPIn(ep uint32)  {}
 func (dev *USBDevice) ClearStallEPOut(ep uint32) {}
 
+// Attach and Detach are no-ops: the USB Serial/JTAG controller has no
+// software-controlled soft-connect, it is always attached to the bus.
+func (dev *USBDevice) Attach() {}
+func (dev *USBDevice) Detach() {}
+
 // initUSB is intentionally empty — the interp phase evaluates init()
 // functions at compile time and cannot access hardware registers.
 // Actual hardware setup is deferred to the first Configure() call.

src/machine/machine_esp32c6_usb.go

@@ -66,6 +66,11 @@ func (dev *USBDevice) SetStallEPOut(ep uint32)   {}
 func (dev *USBDevice) ClearStallEPIn(ep uint32)  {}
 func (dev *USBDevice) ClearStallEPOut(ep uint32) {}
 
+// Attach and Detach are no-ops: the USB Serial/JTAG controller has no
+// software-controlled soft-connect, it is always attached to the bus.
+func (dev *USBDevice) Attach() {}
+func (dev *USBDevice) Detach() {}
+
 // initUSB is intentionally empty — the interp phase evaluates init()
 // functions at compile time and cannot access hardware registers.
 // Actual hardware setup is deferred to the first Configure() call.

src/machine/machine_esp32xx_usb.go

@@ -67,6 +67,11 @@ func (dev *USBDevice) SetStallEPOut(ep uint32)   {}
 func (dev *USBDevice) ClearStallEPIn(ep uint32)  {}
 func (dev *USBDevice) ClearStallEPOut(ep uint32) {}
 
+// Attach and Detach are no-ops: the USB Serial/JTAG controller has no
+// software-controlled soft-connect, it is always attached to the bus.
+func (dev *USBDevice) Attach() {}
+func (dev *USBDevice) Detach() {}
+
 // initUSB is intentionally empty — the interp phase evaluates init()
 // functions at compile time and cannot access hardware registers.
 // Actual hardware setup is deferred to the first Configure() call.

src/machine/machine_nrf52840_usb.go

@@ -22,6 +22,11 @@ var (
 	epinen      uint32
 	epouten     uint32
 	easyDMABusy volatile.Register8
+
+	// usbDetached keeps the device detached from the bus after Detach: the
+	// USB IRQ handler re-enables the DP pull-up on every power-ready event,
+	// which would otherwise silently undo a Detach.
+	usbDetached bool
 )
 
 // enterCriticalSection is used to protect access to easyDMA - only one thing
@@ -89,6 +94,23 @@ func (dev *USBDevice) Configure(config UARTConfig) {
 	dev.initcomplete = true
 }
 
+// Attach connects the device to the USB bus by enabling the DP pull-up,
+// allowing the host to detect and enumerate it. It can be used together with
+// Detach to delay enumeration until the USB configuration (device
+// identifiers, classes, ...) is complete.
+func (dev *USBDevice) Attach() {
+	usbDetached = false
+	nrf.USBD.USBPULLUP.Set(1)
+}
+
+// Detach disconnects the device from the USB bus by disabling the DP pull-up.
+// To the host this appears as if the device was unplugged. A subsequent
+// Attach makes the host enumerate the device again.
+func (dev *USBDevice) Detach() {
+	usbDetached = true
+	nrf.USBD.USBPULLUP.Set(0)
+}
+
 func handleUSBIRQ(interrupt.Interrupt) {
 	if nrf.USBD.EVENTS_SOF.Get() == 1 {
 		nrf.USBD.EVENTS_SOF.Set(0)
@@ -103,7 +125,9 @@ func handleUSBIRQ(interrupt.Interrupt) {
 
 			// Configure control endpoint
 			initEndpoint(0, usb.ENDPOINT_TYPE_CONTROL)
-			nrf.USBD.USBPULLUP.Set(1)
+			if !usbDetached {
+				nrf.USBD.USBPULLUP.Set(1)
+			}
 
 			usbConfiguration = 0
 		}

src/machine/machine_rp2040_usb.go

@@ -43,9 +43,24 @@ func (dev *USBDevice) Configure(config UARTConfig) {
 		rp.USBCTRL_REGS_INTE_SETUP_REQ)
 
 	// Present full speed device by enabling pull up on DP
+	dev.Attach()
+}
+
+// Attach connects the device to the USB bus by enabling the DP pull-up,
+// allowing the host to detect and enumerate it. It can be used together with
+// Detach to delay enumeration until the USB configuration (device
+// identifiers, classes, ...) is complete.
+func (dev *USBDevice) Attach() {
 	rp.USBCTRL_REGS.SIE_CTRL.SetBits(rp.USBCTRL_REGS_SIE_CTRL_PULLUP_EN)
 }
 
+// Detach disconnects the device from the USB bus by disabling the DP pull-up.
+// To the host this appears as if the device was unplugged. A subsequent
+// Attach makes the host enumerate the device again.
+func (dev *USBDevice) Detach() {
+	rp.USBCTRL_REGS.SIE_CTRL.ClearBits(rp.USBCTRL_REGS_SIE_CTRL_PULLUP_EN)
+}
+
 func handleUSBIRQ(intr interrupt.Interrupt) {
 	status := rp.USBCTRL_REGS.INTS.Get()
 

src/machine/machine_rp2350_usb.go

@@ -43,12 +43,27 @@ func (dev *USBDevice) Configure(config UARTConfig) {
 		rp.USB_INTE_SETUP_REQ)
 
 	// Present full speed device by enabling pull up on DP
-	rp.USB.SIE_CTRL.SetBits(rp.USB_SIE_CTRL_PULLUP_EN)
+	dev.Attach()
 
 	// 12.7.2 Disable phy isolation
 	rp.USB.SetMAIN_CTRL_PHY_ISO(0x0)
 }
 
+// Attach connects the device to the USB bus by enabling the DP pull-up,
+// allowing the host to detect and enumerate it. It can be used together with
+// Detach to delay enumeration until the USB configuration (device
+// identifiers, classes, ...) is complete.
+func (dev *USBDevice) Attach() {
+	rp.USB.SIE_CTRL.SetBits(rp.USB_SIE_CTRL_PULLUP_EN)
+}
+
+// Detach disconnects the device from the USB bus by disabling the DP pull-up.
+// To the host this appears as if the device was unplugged. A subsequent
+// Attach makes the host enumerate the device again.
+func (dev *USBDevice) Detach() {
+	rp.USB.SIE_CTRL.ClearBits(rp.USB_SIE_CTRL_PULLUP_EN)
+}
+
 func handleUSBIRQ(intr interrupt.Interrupt) {
 	status := rp.USB.INTS.Get()
 

src/machine/machine_stm32_otgfs_usb.go

@@ -254,7 +254,7 @@ func (dev *USBDevice) Configure(config UARTConfig) {
 	otgPower.PCGCCTL.Set(0)
 
 	// Soft-disconnect now (after CSRST reset DCTL to its default connected state).
-	otgDevice.DCTL.SetBits(dctlSDIS)
+	dev.Detach()
 
 	// ---- 7. Configure data FIFOs --------------------------------------------
 
@@ -311,11 +311,26 @@ func (dev *USBDevice) Configure(config UARTConfig) {
 
 	// ---- 12. Connect to host (clear soft-disconnect) -----------------------
 
-	otgDevice.DCTL.ClearBits(dctlSDIS)
+	dev.Attach()
 
 	dev.initcomplete = true
 }
 
+// Attach connects the device to the USB bus by releasing soft disconnect,
+// allowing the host to detect and enumerate it. It can be used together with
+// Detach to delay enumeration until the USB configuration (device
+// identifiers, classes, ...) is complete.
+func (dev *USBDevice) Attach() {
+	otgDevice.DCTL.ClearBits(dctlSDIS)
+}
+
+// Detach disconnects the device from the USB bus by asserting soft
+// disconnect. To the host this appears as if the device was unplugged. A
+// subsequent Attach makes the host enumerate the device again.
+func (dev *USBDevice) Detach() {
+	otgDevice.DCTL.SetBits(dctlSDIS)
+}
+
 // handleUSBIRQ is the OTG FS interrupt handler, dispatching on GINTSTS bits.
 func handleUSBIRQ(intr interrupt.Interrupt) {
 	status := stm32.OTG_FS_GLOBAL.GINTSTS.Get() &

src/machine/machine_stm32h7_usb.go

@@ -206,7 +206,7 @@ func (dev *USBDevice) Configure(config UARTConfig) {
 
 	// Stay soft-disconnected until configuration is complete; CSRST left
 	// DCTL at its default "connected" state.
-	usbOTG.DCTL.SetBits(DCTL_SDIS)
+	dev.Detach()
 
 	// 5. Force device mode now that the core is out of reset. The mode
 	// change takes effect only after up to 25 ms (RM0433); poll GINTSTS.CMOD
@@ -265,9 +265,24 @@ func (dev *USBDevice) Configure(config UARTConfig) {
 	dev.initcomplete = true
 
 	// Release soft-disconnect: pulls D+ high, making device visible to host.
+	dev.Attach()
+}
+
+// Attach connects the device to the USB bus by releasing soft disconnect,
+// allowing the host to detect and enumerate it. It can be used together with
+// Detach to delay enumeration until the USB configuration (device
+// identifiers, classes, ...) is complete.
+func (dev *USBDevice) Attach() {
 	usbOTG.DCTL.ClearBits(DCTL_SDIS)
 }
 
+// Detach disconnects the device from the USB bus by asserting soft
+// disconnect. To the host this appears as if the device was unplugged. A
+// subsequent Attach makes the host enumerate the device again.
+func (dev *USBDevice) Detach() {
+	usbOTG.DCTL.SetBits(DCTL_SDIS)
+}
+
 func initEndpoint(ep, config uint32) {
 	if ep == 0 {
 		// Control endpoint

src/os/file_unix.go

@@ -163,6 +163,17 @@ func (f *File) Truncate(size int64) (err error) {
 	return Truncate(f.name, size)
 }
 
+// Chown changes the numeric uid and gid of the named file.
+// A uid or gid of -1 means to not change that value.
+// If there is an error, it will be of type *PathError.
+func (f *File) Chown(uid, gid int) error {
+	if f.handle == nil {
+		return ErrClosed
+	}
+
+	return Chown(f.name, uid, gid)
+}
+
 func (f *File) chmod(mode FileMode) error {
 	if f.handle == nil {
 		return ErrClosed

src/os/getpagesize.go

@@ -0,0 +1,6 @@
+package os
+
+import "syscall"
+
+// Getpagesize returns the underlying system's memory page size.
+func Getpagesize() int { return syscall.Getpagesize() }

src/os/types_anyos.go

@@ -6,11 +6,6 @@
 
 package os
 
-import "syscall"
-
-// Getpagesize returns the underlying system's memory page size.
-func Getpagesize() int { return syscall.Getpagesize() }
-
 func (fs *fileStat) Name() string { return fs.name }
 func (fs *fileStat) IsDir() bool  { return fs.Mode().IsDir() }
 

src/runtime/gc_blocks.go

@@ -56,6 +56,7 @@ var (
 	endBlock      gcBlock        // the block just past the end of the available space
 	gcTotalAlloc  uint64         // total number of bytes allocated
 	gcMallocs     uint64         // total number of allocations
+	gcNumGC       uint32         // total number of completed collection cycles
 	gcLock        task.PMutex    // lock to avoid race conditions on multicore systems
 )
 
@@ -510,7 +511,7 @@ func realloc(ptr unsafe.Pointer, size uintptr) unsafe.Pointer {
 	lastBlock := firstBlock.findHead()
 
 	// Calculate the size of the original allocation body.
-	oldSize := uintptr(lastBlock-firstBlock)*blocksPerStateByte + (bytesPerBlock - unsafe.Sizeof(objHeader{}))
+	oldSize := uintptr(lastBlock-firstBlock)*bytesPerBlock + (bytesPerBlock - unsafe.Sizeof(objHeader{}))
 
 	if size <= oldSize {
 		// The requested size is less than the old size.
@@ -610,6 +611,11 @@ func runGC() (freeBytes uintptr) {
 		dumpHeap()
 	}
 
+	// The cycle is complete. Counted here rather than in GC() so that
+	// collections triggered by an allocation are counted too. Every caller
+	// holds gcLock, the same lock ReadMemStats reads it under.
+	gcNumGC++
+
 	return
 }
 
@@ -850,6 +856,9 @@ func ReadMemStats(m *MemStats) {
 	// Record the total allocated bytes.
 	m.TotalAlloc = gcTotalAlloc
 
+	// Record the number of completed collection cycles.
+	m.NumGC = gcNumGC
+
 	gcLock.Unlock()
 }
 

src/runtime/gc_boehm.go

@@ -129,6 +129,7 @@ func ReadMemStats(m *MemStats) {
 	m.Mallocs = 0 // not provided by bdwgc
 	m.Frees = 0   // not provided by bdwgc
 	m.Sys = uint64(gcMemStats.obtained_from_os_bytes)
+	m.NumGC = uint32(gcMemStats.gc_no)
 
 	gcLock.Unlock()
 }

src/runtime/gc_leaking.go

@@ -47,8 +47,14 @@ func alloc(size uintptr, layout unsafe.Pointer) unsafe.Pointer {
 	gcTotalAlloc += uint64(size)
 	gcMallocs++
 	heapptr += size
-	for heapptr >= heapEnd {
+	if heapptr < addr {
+		// The allocation size overflowed the heap pointer.
+		runtimePanic("out of memory")
+	}
+	for heapptr > heapEnd {
 		// Try to increase the heap and check again.
+		// Use > instead of >= because an allocation that exactly
+		// ends at heapEnd is still within heap boundaries.
 		if growHeap() {
 			continue
 		}
@@ -103,6 +109,7 @@ func ReadMemStats(m *MemStats) {
 	m.HeapAlloc = gcTotalAlloc
 	m.HeapObjects = gcMallocs
 	m.Alloc = m.HeapAlloc
+	m.NumGC = 0 // this GC never collects, so no cycle ever completes
 
 	gcLock.Unlock()
 }

src/runtime/gc_stack_cores.go

@@ -43,6 +43,12 @@ func gcMarkReachable() {
 		gcPauseCore(i)
 	}
 
+	// Busy-wait until all the other cores are ready.
+	for gcScanState.Load() != numCPU {
+		spinLoopWait()
+	}
+	gcScanState.Store(0)
+
 	// Scan the stack(s) of the current core.
 	scanCurrentStack()
 	if !task.OnSystemStack() {
@@ -53,13 +59,6 @@ func gcMarkReachable() {
 	// Scan globals.
 	findGlobals(markRoots)
 
-	// Busy-wait until all the other cores are ready. They certainly should be,
-	// after the scanning we did above.
-	for gcScanState.Load() != numCPU {
-		spinLoopWait()
-	}
-	gcScanState.Store(0)
-
 	// Signal each core in turn that they can scan the stack.
 	for i := uint32(0); i < numCPU; i++ {
 		if i == core {

src/runtime/hashmap.go

@@ -807,13 +807,13 @@ func hashmapInterfaceHash(itf interface{}, seed uintptr) uint32 {
 	case reflectlite.Array:
 		var hash uint32
 		for i := 0; i < x.Len(); i++ {
-			hash ^= hashmapInterfaceHash(valueInterfaceUnsafe(x.Index(i)), seed)
+			hash = (hash * 31) ^ hashmapInterfaceHash(valueInterfaceUnsafe(x.Index(i)), seed)
 		}
 		return hash
 	case reflectlite.Struct:
 		var hash uint32
 		for i := 0; i < x.NumField(); i++ {
-			hash ^= hashmapInterfaceHash(valueInterfaceUnsafe(x.Field(i)), seed)
+			hash = (hash * 31) ^ hashmapInterfaceHash(valueInterfaceUnsafe(x.Field(i)), seed)
 		}
 		return hash
 	default:

src/runtime/mstats.go

@@ -82,4 +82,9 @@ type MemStats struct {
 
 	// GCSys is bytes of memory in garbage collection metadata.
 	GCSys uint64
+
+	// NumGC is the number of completed GC cycles.
+	//
+	// The leaking collector never collects, so it always reports 0.
+	NumGC uint32
 }

src/runtime/runtime_esp32.go

@@ -78,12 +78,6 @@ var _ebss [0]byte
 //go:extern _vector_table
 var _vector_table [0]uintptr
 
-func abort() {
-	for {
-		device.Asm("waiti 0")
-	}
-}
-
 // interruptInit installs the Xtensa vector table by writing its address
 // to the VECBASE special register and ensures all CPU interrupts are
 // initially disabled.

src/runtime/runtime_esp32_abort.go

@@ -0,0 +1,11 @@
+//go:build esp32 && !qemu
+
+package runtime
+
+import "device"
+
+func abort() {
+	for {
+		device.Asm("waiti 0")
+	}
+}

src/runtime/runtime_esp32_qemu.go

@@ -0,0 +1,28 @@
+//go:build esp32 && qemu
+
+package runtime
+
+import "device"
+
+func exit(code int) {
+	qemuExit(code)
+}
+
+func abort() {
+	qemuExit(1)
+}
+
+func qemuExit(code int) {
+	// QEMU semihosting expects a2 = SYS_exit (1) and a3 = the exit code.
+	// AsmFull cannot declare these clobbers, so no Go code may run afterward.
+	device.AsmFull(
+		"mov a3, {code}\n"+
+			"movi a2, 1\n"+
+			"simcall",
+		map[string]interface{}{"code": code},
+	)
+	// This loop ensures the clobbered registers are never used again.
+	for {
+		device.Asm("waiti 0")
+	}
+}

src/runtime/runtime_esp32xx.go

@@ -55,10 +55,6 @@ func ticksToNanoseconds(ticks timeUnit) int64 {
 	return int64(ticks) * 25
 }
 
-func exit(code int) {
-	abort()
-}
-
 func putchar(c byte) {
 	machine.Serial.WriteByte(c)
 }

src/runtime/runtime_esp32xx_exit.go

@@ -0,0 +1,7 @@
+//go:build (esp32 && !qemu) || esp32c3 || esp32c6
+
+package runtime
+
+func exit(code int) {
+	abort()
+}

src/syscall/statfs_stubs.go

@@ -0,0 +1,39 @@
+//go:build baremetal || js || wasip1 || wasip2 || wasm_unknown || nintendoswitch
+
+// Copyright 2009 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package syscall
+
+// Statfs_t and Fsid have been copied from the Go source tree
+// (syscall/ztypes_linux_amd64.go).
+
+type Statfs_t struct {
+	Type    int64
+	Bsize   int64
+	Blocks  uint64
+	Bfree   uint64
+	Bavail  uint64
+	Files   uint64
+	Ffree   uint64
+	Fsid    Fsid
+	Namelen int64
+	Frsize  int64
+	Flags   int64
+	Spare   [4]int64
+}
+
+type Fsid struct {
+	X__val [2]int32
+}
+
+// This is a stub, it is not functional.
+func Statfs(path string, buf *Statfs_t) (err error) {
+	return ENOSYS
+}
+
+// This is a stub, it is not functional.
+func Fstatfs(fd int, buf *Statfs_t) (err error) {
+	return ENOSYS
+}

targets/esp32-qemu.json

@@ -0,0 +1,5 @@
+{
+	"inherits": ["esp32-generic"],
+	"build-tags": ["qemu"],
+	"emulator": "qemu-system-xtensa -machine esp32 -nographic -semihosting -drive file={img},if=mtd,format=raw"
+}

targets/esp32.json

@@ -25,6 +25,5 @@
 	],
 	"binary-format": "esp32",
 	"flash-method": "esp32flash",
-	"emulator": "qemu-system-xtensa -machine esp32 -nographic -drive file={img},if=mtd,format=raw",
 	"gdb": ["xtensa-esp32-elf-gdb"]
 }

targets/esp32s3-interrupts.S

@@ -299,13 +299,37 @@ _handle_kernel_exc:
 
     .global _handle_level1
 _handle_level1:
-    // --- allocate 96-byte exception frame on the interrupted stack ---
+    // --- EXCCAUSE 5 (alloca / MOVSP) is not a fault, it is a request ---
+    //
+    // MOVSP raises this whenever it moves the stack pointer while the caller's
+    // register window is still live in the register file: the hardware is
+    // asking us to spill that window, not reporting an error.  LLVM emits MOVSP
+    // for any call that needs a 7th argument on the stack (the windowed ABI
+    // passes six in registers), so it shows up in ordinary C and Go code alike.
+    //
+    // Without this, such a call takes a fatal user exception.  The check goes
+    // first, before any state is touched, so a0 is still in EXCSAVE1 where
+    // _xt_alloca_exc wants it and every other cause falls through to the
+    // handler below completely unchanged.
+    rsr     a0, EXCCAUSE
+    bnei    a0, 5, 1f
+    j       _xt_alloca_exc
+1:
+    rsr     a0, EXCSAVE1        // restore a0 clobbered by the EXCCAUSE read
+
+    // --- allocate the exception frame on the interrupted stack ---
     // Layout (offsets from a1 after adjustment):
     //   0: a0   4: a1(orig)  8: a2  12: a3  16: a4  20: a5
     //  24: a6  28: a7  32: a8  36: a9  40: a10 44: a11
     //  48: a12 52: a13 56: a14 60: a15
-    //  64: SAR 68: EPC1  72: PS
-    addi    a0, a1, -96         // a0 = new frame pointer
+    //  64: SAR 68: EPC1  72: PS  76: WINDOWBASE  80: WINDOWSTART
+    //
+    // 128, not the 84 bytes the layout above needs: the 16 bytes below the
+    // interrupted SP are the windowed-ABI caller save area, written by window
+    // overflow and read back by underflow.  Clobbering them corrupts the
+    // interrupted code's a0, which surfaces later as an Illegal Instruction on
+    // its retw.  ESP-IDF reserves the same gap (the 0x20 in XT_STK_FRMSZ).
+    addi    a0, a1, -128        // a0 = new frame pointer
     s32i    a1, a0, 4           // save original a1 (SP)
     mov     a1, a0              // a1 = frame pointer
 
@@ -339,6 +363,15 @@ _handle_level1:
     // level-1 interrupts.
     rsr     a2, PS
     s32i    a2, a1, 72          // save PS (with EXCM=1 set by hardware)
+
+    // Window state as it was at the fault.  Reading it from the C handler
+    // instead describes the handler: getting there costs a callx4, and the
+    // printfs rotate and spill windows of their own.
+    rsr     a3, WINDOWBASE
+    s32i    a3, a1, 76
+    rsr     a3, WINDOWSTART
+    s32i    a3, a1, 80
+
     movi    a3, ~0x1F           // mask: clear INTLEVEL (bits 0-3) + EXCM (bit 4)
     and     a2, a2, a3
     movi    a3, 1               // INTLEVEL = 1
@@ -412,6 +445,52 @@ _handle_level1:
 
     rfe
 
+// -----------------------------------------------------------------------
+// Alloca (MOVSP) exception handler.
+//
+// Rotates back to the window that executed MOVSP, fixes up PS.OWB to match
+// the new WINDOWBASE, and falls into the window underflow handler that
+// corresponds to the caller's call size. The underflow handler spills the
+// window and returns to the interrupted MOVSP via rfwu, which then completes.
+//
+// All of the interruptee's registers are intact except a0, which is in
+// EXCSAVE1. PS.EXCM is set by the hardware, so this cannot be interrupted.
+// The registers of the base-save area are free scratch here, because taking
+// this exception means they have already been spilled and the underflow
+// handler will restore them.
+// -----------------------------------------------------------------------
+    .balign 4
+    .global _xt_alloca_exc
+_xt_alloca_exc:
+    rsr     a0, WINDOWBASE      // grab WINDOWBASE before rotw changes it
+    rotw    -1                  // WINDOWBASE goes to a4, new a0-a3 are scratch
+    rsr     a2, PS
+    extui   a3, a2, 8, 4        // a3 = PS.OWB  (shift 8, 4 bits)
+    xor     a3, a3, a4          // bits that changed from old to current WB
+    rsr     a4, EXCSAVE1        // restore the interruptee's a0 (now in a4)
+    slli    a3, a3, 8
+    xor     a2, a2, a3          // flip those bits in PS.OWB
+    wsr     a2, PS              // PS.OWB now matches the new WINDOWBASE
+    rsync
+
+    // Dispatch on the call size encoded in the interruptee's return address:
+    // bits 31:30 of a0 are 00/01 for call4, 10 for call8, 11 for call12.
+    //
+    // The canonical sequence branches straight to the underflow vectors with
+    // _bbci.l, but that has only an 8-bit displacement and this handler lives
+    // outside the vector table, well out of reach. Inverting each test and
+    // putting the vector target on a full-range `j` is the same dispatch
+    // without the range limit.
+    _bbsi.l a4, 31, 1f
+    j       _window_underflow4
+1:
+    rotw    -1                  // interruptee's a0 moves to a8
+    _bbsi.l a8, 30, 2f
+    j       _window_underflow8
+2:
+    rotw    -1
+    j       _window_underflow12
+
 // -----------------------------------------------------------------------
 // Default weak espradio_user_exception: infinite loop halt.
 // Overridden by the strong definition in espradio's isr.c when linked.

targets/esp32s3.json

@@ -25,6 +25,5 @@
 	],
 	"binary-format": "esp32s3",
 	"flash-method": "esp32jtag",
-	"emulator": "qemu-system-xtensa -machine esp32 -nographic -drive file={img},if=mtd,format=raw",
 	"gdb": ["xtensa-esp32-elf-gdb"]
 }

transform/transform_test.go

@@ -76,7 +76,7 @@ func testTransform(t *testing.T, pathPrefix string, transform func(mod llvm.Modu
 	actual = actual[strings.Index(actual, "\ntarget datalayout = ")+1:]
 
 	if *update {
-		err := os.WriteFile(pathPrefix+".out.ll", []byte(actual), 0666)
+		err := os.WriteFile(pathPrefix+".out.ll", []byte(normalizeIR(actual)), 0666)
 		if err != nil {
 			t.Error("failed to write out new output:", err)
 		}
@@ -100,28 +100,8 @@ func testTransform(t *testing.T, pathPrefix string, transform func(mod llvm.Modu
 // equal. That means, only relevant lines are compared (excluding comments
 // etc.).
 func fuzzyEqualIR(s1, s2 string) bool {
-	// Golden files are written using the pre-LLVM21 'nocapture' spelling,
-	// which LLVM printed before any co-occurring attribute such as
-	// 'readonly' (e.g. "ptr nocapture readonly"). LLVM 21+ prints the
-	// equivalent 'captures(none)' instead, and after such attributes (e.g.
-	// "ptr readonly captures(none)"). Normalize both name and position back
-	// to the old spelling to keep a single golden file working across LLVM
-	// versions.
-	s1 = normalizeCapturesAttr(s1)
-	s2 = normalizeCapturesAttr(s2)
-
-	// LLVM 21+ also added an explicit 'nocreateundeforpoison' attribute to
-	// certain intrinsic declarations (e.g. llvm.umin) that were implicitly
-	// assumed not to create undef/poison before. It's unrelated to the
-	// escape-analysis behavior under test, so ignore it for comparison.
-	s1 = strings.ReplaceAll(s1, "nocreateundeforpoison ", "")
-	s2 = strings.ReplaceAll(s2, "nocreateundeforpoison ", "")
-
-	// LLVM 22 dropped the (redundant) i64 size argument from
-	// llvm.lifetime.start/end. Normalize away that argument so golden files
-	// written against the two-argument form still match.
-	s1 = lifetimeSizeArgRe.ReplaceAllString(s1, "$1")
-	s2 = lifetimeSizeArgRe.ReplaceAllString(s2, "$1")
+	s1 = normalizeIR(s1)
+	s2 = normalizeIR(s2)
 
 	lines1 := filterIrrelevantIRLines(strings.Split(s1, "\n"))
 	lines2 := filterIrrelevantIRLines(strings.Split(s2, "\n"))
@@ -138,6 +118,29 @@ func fuzzyEqualIR(s1, s2 string) bool {
 	return true
 }
 
+func normalizeIR(s string) string {
+	// Golden files are written using the pre-LLVM21 'nocapture' spelling,
+	// which LLVM printed before any co-occurring attribute such as
+	// 'readonly' (e.g. "ptr nocapture readonly"). LLVM 21+ prints the
+	// equivalent 'captures(none)' instead, and after such attributes (e.g.
+	// "ptr readonly captures(none)"). Normalize both name and position back
+	// to the old spelling.
+	s = normalizeCapturesAttr(s)
+
+	// LLVM 21+ also added an explicit 'nocreateundeforpoison' attribute to
+	// certain intrinsic declarations (e.g. llvm.umin) that were implicitly
+	// assumed not to create undef/poison before. It's unrelated to the
+	// escape-analysis behavior under test, so ignore it for comparison.
+	s = strings.ReplaceAll(s, "nocreateundeforpoison ", "")
+
+	// LLVM 22 dropped the (redundant) i64 size argument from
+	// llvm.lifetime.start/end. Normalize away that argument so golden files
+	// written against the two-argument form still match.
+	s = lifetimeSizeArgRe.ReplaceAllString(s, "$1")
+
+	return s
+}
+
 // capturesNoneAttrRe matches a co-occurring attribute directly followed by
 // 'captures(none)', which is how LLVM 21+ orders these two attributes when
 // printing IR (the pre-LLVM21 'nocapture' attribute printed the other way
@@ -221,7 +224,9 @@ func compileGoFileForTesting(t *testing.T, filename string) llvm.Module {
 	// Compile AST to IR.
 	program := lprogram.LoadSSA()
 	pkg := lprogram.MainPkg()
-	mod, errs := compiler.CompilePackage(filename, pkg, program.Package(pkg.Pkg), machine, compilerConfig, false)
+	ssaPkg := program.Package(pkg.Pkg)
+	ssaPkg.Build()
+	mod, errs := compiler.CompilePackage(filename, pkg, ssaPkg, machine, compilerConfig, false)
 	if errs != nil {
 		for _, err := range errs {
 			t.Error(err)