diff --git a/Changes b/Changes
index 2ccef80e3..bcb162f96 100644
--- a/Changes
+++ b/Changes
@@ -14,6 +14,8 @@ The contributors that suggested a given feature are shown in []. Thanks!
 
 ****  Report interface ports connected to wrong interface, bug1294. [Todd Strader]
 
+****  When tracing, use scalars on single bit arrays to appease vcddiff.
+
 ****  Fix parsing "output signed" in V2K port list, msg2540. [James Jung]
 
 ****  Fix parsing error on bad missing #, bug1308. [Dan Kirkham]
diff --git a/include/verilated.h b/include/verilated.h
index 49b270a99..c9ebf8977 100644
--- a/include/verilated.h
+++ b/include/verilated.h
@@ -186,9 +186,7 @@ public:
     // METHODS
     /// Check that the current thread ID is the same as the construction thread ID
     void check() VL_MT_UNSAFE_ONE {
-        // Memoize results in local thread, to prevent slow get_id() call
-        VL_THREAD_LOCAL bool t_okThread = (m_threadid == VL_THREAD_ID());
-        if (!VL_LIKELY(t_okThread)) {
+        if (!VL_LIKELY(m_threadid == VL_THREAD_ID())) {
             fatal_different();
         }
     }
diff --git a/include/verilated_vcd_c.cpp b/include/verilated_vcd_c.cpp
index 8fb2701d4..9f3f5c72e 100644
--- a/include/verilated_vcd_c.cpp
+++ b/include/verilated_vcd_c.cpp
@@ -343,7 +343,7 @@ void VerilatedVcd::bufferResize(vluint64_t minsize) {
 	memcpy(m_wrBufp, oldbufp, m_writep - oldbufp);
         m_writep = m_wrBufp + (m_writep - oldbufp);
 	m_wrFlushp = m_wrBufp + m_wrChunkSize * 6;
-	delete oldbufp; oldbufp=NULL;
+	delete [] oldbufp; oldbufp=NULL;
     }
 }
 
diff --git a/src/V3EmitC.cpp b/src/V3EmitC.cpp
index 7f37f52cb..158911a21 100644
--- a/src/V3EmitC.cpp
+++ b/src/V3EmitC.cpp
@@ -2405,14 +2405,20 @@ class EmitCTrace : EmitCStmts {
 	string full = ((m_funcp->funcType() == AstCFuncType::TRACE_FULL
 			|| m_funcp->funcType() == AstCFuncType::TRACE_FULL_SUB)
 		       ? "full":"chg");
+        bool emitWidth = false;
 	if (nodep->dtypep()->basicp()->isDouble()) {
 	    puts("vcdp->"+full+"Double");
 	} else if (nodep->isWide() || emitTraceIsScBv(nodep) || emitTraceIsScBigUint(nodep)) {
 	    puts("vcdp->"+full+"Array");
+            emitWidth = true;
 	} else if (nodep->isQuad()) {
 	    puts("vcdp->"+full+"Quad ");
-	} else if (nodep->declp()->bitRange().ranged()) {
+            emitWidth = true;
+        } else if (nodep->declp()->bitRange().ranged()
+                   // 1 element smaller to use Bit dump
+                   && nodep->declp()->bitRange().elements() != 1) {
 	    puts("vcdp->"+full+"Bus  ");
+            emitWidth = true;
 	} else {
 	    puts("vcdp->"+full+"Bit  ");
 	}
@@ -2420,8 +2426,7 @@ class EmitCTrace : EmitCStmts {
 			    + ((arrayindex<0) ? 0 : (arrayindex*nodep->declp()->widthWords()))));
 	puts(",");
 	emitTraceValue(nodep, arrayindex);
-	if (!nodep->dtypep()->basicp()->isDouble()  // When float/double no longer have widths this can go
-	    && (nodep->declp()->bitRange().ranged() || emitTraceIsScBv(nodep) || emitTraceIsScBigUint(nodep))) {
+        if (emitWidth) {
 	    puts(","+cvtToStr(nodep->declp()->widthMin()));
 	}
 	puts(");\n");
diff --git a/test_regress/t/t_vlcov_merge.pl b/test_regress/t/t_vlcov_merge.pl
index 6a0e57121..72de4372c 100755
--- a/test_regress/t/t_vlcov_merge.pl
+++ b/test_regress/t/t_vlcov_merge.pl
@@ -19,6 +19,9 @@ run(cmd => ["../bin/verilator_coverage",
 
 # Older clib's didn't properly sort maps, but the coverage data doesn't
 # really care about ordering. So avoid false failures by sorting.
+# Set LC_ALL as suggested in the sort manpage to avoid sort order
+# changes from the locale.
+$ENV{LC_ALL} = "C";
 run(cmd => ["sort",
             "$Self->{obj_dir}/coverage.dat",
             "> $Self->{obj_dir}/coverage-sort.dat",
diff --git a/test_regress/t/t_vpi_get.cpp b/test_regress/t/t_vpi_get.cpp
index 4cd41d87a..066deaa87 100644
--- a/test_regress/t/t_vpi_get.cpp
+++ b/test_regress/t/t_vpi_get.cpp
@@ -156,20 +156,27 @@ struct params {
       unsigned int  scalar;
       int  type;
     } attributes, children;
-} values[] = {
-    {"onebit", {1, vpiNoDirection, 1, vpiReg}, {0, 0, 0, 0}},
-    {"twoone", {2, vpiNoDirection, 0, vpiReg}, {0, 0, 0, 0}},
-    {"onetwo", {2, vpiNoDirection, 0, TestSimulator::is_verilator() ? vpiReg : vpiMemory}, {0, 0, 0, 0}},
-    {"fourthreetwoone", {2, vpiNoDirection, 0, vpiMemory}, {2, vpiNoDirection, 0, vpiMemoryWord}},
-    {"clk", {1, vpiInput, 1, vpiPort}, {0, 0, 0, 0}},
-    {"testin", {16, vpiInput, 0, vpiPort}, {0, 0, 0, 0}},
-    {"testout", {24, vpiOutput, 0, vpiPort}, {0, 0, 0, 0}},
-    {"sub.subin", {1, vpiInput, 1, vpiPort}, {0, 0, 0, 0}},
-    {"sub.subout", {1, vpiOutput, 1, vpiPort}, {0, 0, 0, 0}},
-    {NULL, {0, 0, 0, 0}, {0, 0, 0, 0}}
 };
 
 int mon_check_props() {
+    // This table needs to be function-static.
+    // This avoids calling is_verilator() below at global-static init time.
+    // When global-static led to a race between the is_verilator call below, and
+    // the code that sets up the VerilatedAssertOneThread() check in
+    // verilated_vpi.cc, it was causing the check to falsely fail
+    // (due to m_threadid within the check not being initted yet.)
+    static struct params values[] = {
+        {"onebit", {1, vpiNoDirection, 1, vpiReg}, {0, 0, 0, 0}},
+        {"twoone", {2, vpiNoDirection, 0, vpiReg}, {0, 0, 0, 0}},
+        {"onetwo", {2, vpiNoDirection, 0, TestSimulator::is_verilator() ? vpiReg : vpiMemory}, {0, 0, 0, 0}},
+        {"fourthreetwoone", {2, vpiNoDirection, 0, vpiMemory}, {2, vpiNoDirection, 0, vpiMemoryWord}},
+        {"clk", {1, vpiInput, 1, vpiPort}, {0, 0, 0, 0}},
+        {"testin", {16, vpiInput, 0, vpiPort}, {0, 0, 0, 0}},
+        {"testout", {24, vpiOutput, 0, vpiPort}, {0, 0, 0, 0}},
+        {"sub.subin", {1, vpiInput, 1, vpiPort}, {0, 0, 0, 0}},
+        {"sub.subout", {1, vpiOutput, 1, vpiPort}, {0, 0, 0, 0}},
+        {NULL, {0, 0, 0, 0}, {0, 0, 0, 0}}
+    };
     struct params* value = values;
     while (value->signal) {
       TestVpiHandle h = VPI_HANDLE(value->signal);