diff --git a/src/faster/R2Stretcher.cpp b/src/faster/R2Stretcher.cpp
index 4159c24..52737a4 100644
--- a/src/faster/R2Stretcher.cpp
+++ b/src/faster/R2Stretcher.cpp
@@ -993,6 +993,8 @@ R2Stretcher::study(const float *const *input, size_t samples, bool final)
         mixdown = input[0];
     }
 
+    std::vector<float> optionalFoldBuffer;
+    
     while (consumed < samples) {
 
 	size_t writable = inbuf.getWriteSpace();
@@ -1042,15 +1044,18 @@ R2Stretcher::study(const float *const *input, size_t samples, bool final)
 
                 // Note that we can't do this in-place.  Pity
 
-                float *tmp = (float *)alloca
-                    (std::max(m_fftSize, m_aWindowSize) * sizeof(float));
+                // This doesn't have to be allocation-free, as we're
+                // in offline-mode study
+
+                optionalFoldBuffer.resize(std::max(m_fftSize, m_aWindowSize));
 
                 if (m_aWindowSize > m_fftSize) {
                     m_afilter->cut(cd.accumulator);
                 }
 
-                cutShiftAndFold(tmp, m_fftSize, cd.accumulator, m_awindow);
-                v_copy(cd.accumulator, tmp, m_fftSize);
+                cutShiftAndFold(optionalFoldBuffer.data(), m_fftSize,
+                                cd.accumulator, m_awindow);
+                v_copy(cd.accumulator, optionalFoldBuffer.data(), m_fftSize);
             }
 
             m_studyFFT->forwardMagnitude(cd.accumulator, cd.fltbuf);
@@ -1058,8 +1063,6 @@ R2Stretcher::study(const float *const *input, size_t samples, bool final)
             float df = m_phaseResetAudioCurve->processFloat(cd.fltbuf, m_increment);
             m_phaseResetDf.push_back(df);
 
-//            cout << m_phaseResetDf.size() << " [" << final << "] -> " << df << " \t: ";
-
             df = m_silentAudioCurve->processFloat(cd.fltbuf, m_increment);
             bool silent = (df > 0.f);
             if (silent) {
@@ -1067,8 +1070,6 @@ R2Stretcher::study(const float *const *input, size_t samples, bool final)
             }
             m_silence.push_back(silent);
 
-//            cout << df << endl;
-
             // We have augmented the input by m_aWindowSize/2 so that
             // the first chunk is centred on the first audio sample.
             // We want to ensure that m_inputDuration contains the
@@ -1080,7 +1081,7 @@ R2Stretcher::study(const float *const *input, size_t samples, bool final)
             inbuf.skip(m_increment);
 	}
     }
-
+    
     if (final) {
         int rs = inbuf.getReadSpace();
         m_inputDuration += rs;
diff --git a/src/faster/StretcherProcess.cpp b/src/faster/StretcherProcess.cpp
index 36fe27b..74d09e7 100644
--- a/src/faster/StretcherProcess.cpp
+++ b/src/faster/StretcherProcess.cpp
@@ -236,6 +236,9 @@ R2Stretcher::consumeChannel(size_t c,
 
         inbuf.write(cd.resamplebuf, toWrite);
         cd.inCount += samples;
+
+        m_log.log(2, "consumeChannel: wrote to inbuf from resamplebuf, inCount now", toWrite, cd.inCount);
+        
         return samples;
 
     } else {
@@ -249,6 +252,9 @@ R2Stretcher::consumeChannel(size_t c,
 
         inbuf.write(input, toWrite);
         cd.inCount += toWrite;
+
+        m_log.log(2, "consumeChannel: wrote to inbuf from input, inCount now", toWrite, cd.inCount);
+        
         return toWrite;
     }
 }
@@ -340,7 +346,7 @@ R2Stretcher::processOneChunk()
             return false;
         }
         ChannelData &cd = *m_channelData[c];
-        m_log.log(3, "read space and draining", cd.inbuf->getReadSpace(), cd.draining);
+        m_log.log(2, "read space and draining", cd.inbuf->getReadSpace(), cd.draining);
         if (!cd.draining) {
             size_t ready = cd.inbuf->getReadSpace();
             assert(ready >= m_aWindowSize || cd.inputSize >= 0);
diff --git a/src/test/TestStretcher.cpp b/src/test/TestStretcher.cpp
index fc72d99..a2352c4 100644
--- a/src/test/TestStretcher.cpp
+++ b/src/test/TestStretcher.cpp
@@ -1516,4 +1516,32 @@ BOOST_AUTO_TEST_CASE(with_resets_2x_5up_realtime_faster)
     with_resets(RubberBandStretcher::OptionProcessRealTime | RubberBandStretcher::OptionEngineFaster, 2.0, 1.5);
 }
 
+BOOST_AUTO_TEST_CASE(long_blocksize_with_smoothing)
+{
+    // Test added because the smoothing option was calling alloca in a
+    // loop, which could run us out of stack. This test is best used
+    // with a small stack artificially enforced via e.g. ulimit -s 32
+    
+    int n = 10000;
+    float freq = 440.f;
+    int rate = 44100;
+    RubberBandStretcher stretcher
+        (rate, 1,
+         RubberBandStretcher::OptionEngineFaster |
+         RubberBandStretcher::OptionProcessOffline |
+         RubberBandStretcher::OptionSmoothingOn);
+    
+    vector<float> in(n);
+    for (int i = 0; i < n; ++i) {
+        in[i] = sinf(float(i) * freq * M_PI * 2.f / float(rate));
+    }
+    float *inp = in.data();
+
+    stretcher.setMaxProcessSize(n);
+    stretcher.setExpectedInputDuration(n);
+
+    stretcher.study(&inp, n, true);
+    stretcher.process(&inp, n, true);
+}
+
 BOOST_AUTO_TEST_SUITE_END()