diff --git a/include/verilated.cpp b/include/verilated.cpp
index 74ceae09d..33cf917d4 100644
--- a/include/verilated.cpp
+++ b/include/verilated.cpp
@@ -3061,7 +3061,9 @@ void VerilatedContext::statsPrintSummary() VL_MT_UNSAFE {
         = vl_timescaled_double((cputime != 0.0) ? (simtimeInUnits / cputime) : 0, "%0.3f %s");
     VL_PRINTF("- Verilator: %s at %s; walltime %0.3f s; speed %s/s\n", endwhy.c_str(),
               simtime.c_str(), walltime, simtimePerf.c_str());
-    const double modelMB = VlOs::memPeakUsageBytes() / 1024.0 / 1024.0;
+    uint64_t memPeak, memCurrent;
+    VlOs::memUsageBytes(memPeak /*ref*/, memCurrent /*ref*/);
+    const double modelMB = memPeak / 1024.0 / 1024.0;
     VL_PRINTF("- Verilator: cpu %0.3f s on %u threads; alloced %0.0f MB\n", cputime,
               threadsInModels(), modelMB);
 }
diff --git a/include/verilatedos.h b/include/verilatedos.h
index 085bb2544..6b289292f 100644
--- a/include/verilatedos.h
+++ b/include/verilatedos.h
@@ -637,7 +637,7 @@ extern std::string getenvStr(const std::string& envvar,
 extern uint16_t getcpu() VL_MT_SAFE;
 
 /// Return memory usage in bytes, or 0 if unknown
-extern uint64_t memPeakUsageBytes() VL_MT_SAFE;
+extern void memUsageBytes(uint64_t& peakr, uint64_t& currentr) VL_MT_SAFE;
 
 // Internal: Record CPU time, starting point on construction, and current delta from that
 class DeltaCpuTime final {
diff --git a/include/verilatedos_c.h b/include/verilatedos_c.h
index 6ecdf7dba..2b4619816 100644
--- a/include/verilatedos_c.h
+++ b/include/verilatedos_c.h
@@ -25,6 +25,9 @@
 
 #include "verilatedos.h"
 
+#include <fstream>
+#include <sstream>
+
 // clang-format off
 #if defined(_WIN32) || defined(__MINGW32__)
 # include <windows.h>   // LONG for bcrypt.h on MINGW
@@ -104,28 +107,40 @@ uint16_t getcpu() VL_MT_SAFE {
 //=========================================================================
 // VlOs::memPeakUsageBytes implementation
 
-uint64_t memPeakUsageBytes() VL_MT_SAFE {
+void memUsageBytes(uint64_t& peakr, uint64_t& currentr) VL_MT_SAFE {
+    peakr = 0;
+    currentr = 0;
 #if defined(_WIN32) || defined(__MINGW32__)
     const HANDLE process = GetCurrentProcess();
     PROCESS_MEMORY_COUNTERS pmc;
     if (GetProcessMemoryInfo(process, &pmc, sizeof(pmc))) {
         // The best we can do using simple Windows APIs is to get the size of the working set.
-        return pmc.WorkingSetSize;
+        peakr = pmc.PeakWorkingSetSize;
+        currentr = pmc.WorkingSetSize;
     }
-    return 0;
 #else
     // Highly unportable. Sorry
-    const char* const statmFilename = "/proc/self/statm";
-    FILE* const fp = fopen(statmFilename, "r");
-    if (!fp) return 0;
-    uint64_t size, resident, share, text, lib, data, dt;  // All in pages
-    const int items = fscanf(
-        fp, "%" SCNu64 " %" SCNu64 " %" SCNu64 " %" SCNu64 " %" SCNu64 " %" SCNu64 " %" SCNu64,
-        &size, &resident, &share, &text, &lib, &data, &dt);
-    fclose(fp);
-    if (VL_UNCOVERABLE(7 != items)) return 0;
-    // Return the vm size, not the current active set size (/proc/self/status VmRSS + VmSwap)
-    return (text + data) * getpagesize();
+    std::ifstream is{"/proc/self/status"};
+    if (!is) return;
+    std::string line;
+    uint64_t vmPeak = 0;
+    uint64_t vmRss = 0;
+    uint64_t vmSwap = 0;
+    std::string field;
+    while (std::getline(is, line)) {
+        if (line.rfind("VmPeak:", 0) == 0) {
+            std::stringstream ss{line};
+            ss >> field >> vmPeak;
+        } else if (line.rfind("VmRSS:", 0) == 0) {
+            std::stringstream ss{line};
+            ss >> field >> vmRss;
+        } else if (line.rfind("VmSwap:", 0) == 0) {
+            std::stringstream ss{line};
+            ss >> field >> vmSwap;
+        }
+    }
+    peakr = vmPeak * 1024;
+    currentr = (vmRss + vmSwap) * 1024;
 #endif
 }
 
diff --git a/src/V3Stats.cpp b/src/V3Stats.cpp
index 394ba1de2..cee37d85a 100644
--- a/src/V3Stats.cpp
+++ b/src/V3Stats.cpp
@@ -179,7 +179,4 @@ void V3Stats::statsStageAll(AstNetlist* nodep, const std::string& stage, bool fa
     StatsVisitor{nodep, stage, fastOnly};
 }
 
-void V3Stats::statsFinalAll(AstNetlist* nodep) {
-    statsStageAll(nodep, "Final all");
-    statsStageAll(nodep, "Final fast", true);
-}
+void V3Stats::statsFinalAll(AstNetlist* nodep) { statsStageAll(nodep, "Final"); }
diff --git a/src/V3StatsReport.cpp b/src/V3StatsReport.cpp
index 2c5dadb80..3efe8e947 100644
--- a/src/V3StatsReport.cpp
+++ b/src/V3StatsReport.cpp
@@ -215,8 +215,10 @@ void V3Stats::statsStage(const string& name) {
     V3Stats::addStatPerf("Stage, Elapsed time (sec), " + digitName, wallTimeDelta);
     V3Stats::addStatPerf("Stage, Elapsed time (sec), TOTAL", wallTimeDelta);
 
-    const double memory = VlOs::memPeakUsageBytes() / 1024.0 / 1024.0;
-    V3Stats::addStatPerf("Stage, Memory (MB), " + digitName, memory);
+    uint64_t memPeak, memCurrent;
+    VlOs::memUsageBytes(memPeak /*ref*/, memCurrent /*ref*/);
+    V3Stats::addStatPerf("Stage, Memory current (MB), " + digitName, memCurrent / 1024.0 / 1024.0);
+    V3Stats::addStatPerf("Stage, Memory peak (MB), " + digitName, memPeak / 1024.0 / 1024.0);
 }
 
 void V3Stats::infoHeader(std::ofstream& os, const string& prefix) {
@@ -266,7 +268,9 @@ void V3Stats::summaryReport() {
               << ", cvt=" << walltimeCvt << ", bld=" << walltimeBuild << "); cpu " << cputime
               << " s on " << std::max(v3Global.opt.verilateJobs(), v3Global.opt.buildJobs())
               << " threads";
-    const double memory = VlOs::memPeakUsageBytes() / 1024.0 / 1024.0;
+    uint64_t memPeak, memCurrent;
+    VlOs::memUsageBytes(memPeak /*ref*/, memCurrent /*ref*/);
+    const double memory = memPeak / 1024.0 / 1024.0;
     if (VL_UNCOVERABLE(memory != 0.0)) std::cout << "; alloced " << memory << " MB";
     std::cout << "\n";
 }
diff --git a/src/Verilator.cpp b/src/Verilator.cpp
index f712412b8..61cc3b9b5 100644
--- a/src/Verilator.cpp
+++ b/src/Verilator.cpp
@@ -768,7 +768,8 @@ static void verilate(const string& argString) {
 
     V3Os::filesystemFlushBuildDir(v3Global.opt.makeDir());
     if (v3Global.opt.hierTop()) V3Os::filesystemFlushBuildDir(v3Global.opt.hierTopDataDir());
-    if (v3Global.opt.stats()) V3Stats::statsStage("wrote");
+    if (v3Global.opt.stats()) V3Stats::statsStageAll(v3Global.rootp(), "WroteAll");
+    if (v3Global.opt.stats()) V3Stats::statsStageAll(v3Global.rootp(), "WroteFast");
 
     // Final writing shouldn't throw warnings, but...
     V3Error::abortIfWarnings();
diff --git a/test_regress/t/t_display_merge.py b/test_regress/t/t_display_merge.py
index 03f439c26..21d0b3b95 100755
--- a/test_regress/t/t_display_merge.py
+++ b/test_regress/t/t_display_merge.py
@@ -9,7 +9,7 @@
 
 import vltest_bootstrap
 
-test.scenarios('simulator')
+test.scenarios('simulator_st')
 
 test.compile(verilator_flags2=["--stats"])