diff --git a/passes.cc b/passes.cc
index 2a53dfe3..bb23bd0a 100644
--- a/passes.cc
+++ b/passes.cc
@@ -4,8 +4,6 @@
 #include <map>
 #include <tbb/parallel_do.h>
 #include <tbb/parallel_for_each.h>
-#include <tbb/parallel_scan.h>
-#include <tbb/partitioner.h>
 #include <unordered_set>
 
 template <typename E>
@@ -344,35 +342,40 @@ void compute_section_sizes(Context<E> &ctx) {
     if (osec->members.empty())
       return;
 
-    struct T {
-      i64 offset;
-      i64 align;
-    };
+    std::vector<std::span<InputSection<E> *>> slices =
+      split(osec->members, 10000);
 
-    T sum = tbb::parallel_scan(
-      tbb::blocked_range<i64>(0, osec->members.size(), 10000),
-      T{0, 1},
-      [&](const tbb::blocked_range<i64> &r, T sum, bool is_final) {
-        for (i64 i = r.begin(); i < r.end(); i++) {
-          InputSection<E> &isec = *osec->members[i];
-          if (is_final)
-            isec.offset = sum.offset;
+    std::vector<i64> size(slices.size());
+    std::vector<i64> alignments(slices.size());
 
-          sum.offset = align_to(sum.offset, isec.shdr.sh_addralign) +
-            isec.shdr.sh_size;
-          sum.align = std::max<i64>(sum.align, isec.shdr.sh_addralign);
-        }
-        return sum;
-      },
-      [](T lhs, T rhs) {
-        i64 offset = align_to(lhs.offset, rhs.align) + rhs.offset;
-        i64 align = std::max(lhs.align, rhs.align);
-        return T{offset, align};
-      },
-      tbb::simple_partitioner());
+    tbb::parallel_for((i64)0, (i64)slices.size(), [&](i64 i) {
+      i64 off = 0;
+      i64 align = 1;
 
-    osec->shdr.sh_size = sum.offset;
-    osec->shdr.sh_addralign = sum.align;
+      for (InputSection<E> *isec : slices[i]) {
+        off = align_to(off, isec->shdr.sh_addralign);
+        isec->offset = off;
+        off += isec->shdr.sh_size;
+        align = std::max<i64>(align, isec->shdr.sh_addralign);
+      }
+
+      size[i] = off;
+      alignments[i] = align;
+    });
+
+    i64 align = *std::max_element(alignments.begin(), alignments.end());
+
+    std::vector<i64> start(slices.size());
+    for (i64 i = 1; i < slices.size(); i++)
+      start[i] = align_to(start[i - 1] + size[i - 1], align);
+
+    tbb::parallel_for((i64)1, (i64)slices.size(), [&](i64 i) {
+      for (InputSection<E> *isec : slices[i])
+        isec->offset += start[i];
+    });
+
+    osec->shdr.sh_size = start.back() + size.back();
+    osec->shdr.sh_addralign = align;
   });
 }
 
diff --git a/perf.cc b/perf.cc
index c194af83..d6428884 100644
--- a/perf.cc
+++ b/perf.cc
@@ -90,8 +90,7 @@ static void print_rec(TimerRecord &rec, i64 indent) {
 
 template <typename E>
 void Timer<E>::print(Context<E> &ctx) {
-  tbb::concurrent_vector<std::unique_ptr<TimerRecord>> &records =
-    ctx.timer_records;
+  tbb::concurrent_vector<std::unique_ptr<TimerRecord>> &records = ctx.timer_records;
 
   for (i64 i = records.size() - 1; i >= 0; i--)
     records[i]->stop();
diff --git a/test/dt_init.sh b/test/dt_init.sh
index 057fbce8..0a2c7f06 100755
--- a/test/dt_init.sh
+++ b/test/dt_init.sh
@@ -26,9 +26,9 @@ grep -Pq '0000000000233000\s+0 FUNC    GLOBAL HIDDEN    \d+ _fini$' $t/log
 clang -fuse-ld=`pwd`/../mold -o $t/exe $t/a.o -Wl,-init,init -Wl,-fini,fini
 readelf -a $t/exe > $t/log
 
-grep -Pq '\(INIT\)\s+0x233117' $t/log
-grep -Pq '\(FINI\)\s+0x233118' $t/log
-grep -Pq '0000000000233117\s+0 NOTYPE  GLOBAL DEFAULT   \d+ init$' $t/log
-grep -Pq '0000000000233118\s+0 NOTYPE  GLOBAL DEFAULT   \d+ fini$' $t/log
+grep -Pq '\(INIT\)\s+0x233119' $t/log
+grep -Pq '\(FINI\)\s+0x23311a' $t/log
+grep -Pq '0000000000233119\s+0 NOTYPE  GLOBAL DEFAULT   \d+ init$' $t/log
+grep -Pq '000000000023311a\s+0 NOTYPE  GLOBAL DEFAULT   \d+ fini$' $t/log
 
 echo OK