compilersutra · compilersutra · Sep 1, 2026 · Aug 29, 2026
diff --git a/docs/articles.md b/docs/articles.md
@@ -278,13 +278,20 @@ No article has been published from the Jenkins queue yet.
 
 <tr>
 <td>When Clang -O2 Gets Slower: From Perf Counters to MachineBlockPlacement</td>
-<td>Perf counters → MBP layout → branch misses (parse loop)</td>
 <td><a href="./when-o2-layout-hurts-machineblockplacement">Article</a></td>
 <td>Coming Soon</td>
 <td>Coming Soon</td>
 <td>Coming Soon</td>
 </tr>
 
+<tr>
+<td>Why MachineBlockPlacement Made the Wrong Bet: Static Probabilities, Fall-Through, and PGO</td>
+<td><a href="./machineblockplacement-wrong-bet-static-probabilities-pgo">Article</a></td>
+<td>Coming Soon</td>
+<td>Coming Soon</td>
+<td>Coming Soon</td>
+</tr>
+
 <tr>
 <td>std::byte vs unsigned char: Do Modern Compilers Optimize Them Differently?</td>
 <td><a href="./std-byte-vs-unsigned-char">Article</a></td>

diff --git a/docs/articles/machineblockplacement-wrong-bet-static-probabilities-pgo.md b/docs/articles/machineblockplacement-wrong-bet-static-probabilities-pgo.md
@@ -0,0 +1,276 @@
+---
+title: "Why MachineBlockPlacement Made the Wrong Bet: Static Probabilities, Fall-Through, and PGO"
+description: "Part 2 of the parse-loop case study: BranchProbabilityInfo without PGO, the backward shared-increment layout, and how profile branch_weights restore O1-like misses."
+keywords:
+  - MachineBlockPlacement branch probability
+  - BranchProbabilityInfo llvm
+  - clang profile guided optimization parse loop
+  - branch_weights llvm ir
+  - static branch probability compiler
+  - disable-block-placement clang
+  - clang O2 slower than O1
+  - fall-through branch layout
+  - compilersutra llvm case study
+  - pgo branch misprediction
+---
+
+import AdBanner from '@site/src/components/AdBanner';
+import Head from '@docusaurus/Head';
+
+<Head>
+  <meta name="description" content="Part 2: why MachineBlockPlacement laid out a CSV parse loop badly without profile data, and how branch_weights from PGO fix branch misses." />
+</Head>
+
+# Why MachineBlockPlacement Made the Wrong Bet: Static Probabilities, Fall-Through, and PGO
+
+Part 1 named the pass. [**When Clang `-O2` Gets Slower**](/docs/articles/when-o2-layout-hurts-machineblockplacement) showed that **MachineBlockPlacement** is where the parse-loop regression lives — `-mllvm -disable-block-placement` and `opt-bisect` limit **309** are the proof.
+
+What part 1 did **not** show is **what branch probabilities MachineBlockPlacement consumed** before it laid out the separator check as backward taken jumps.
+
+That is what this article traces: workload counts → IR metadata → layout → misses, with and without PGO.
+
+:::info Mechanism in 30 seconds
+
+<pre>
+No PGO
+  ↓
+switch has no branch_weights
+  ↓
+static BranchProbabilityInfo guess
+  ↓
+MBP → backward je to shared inc
+  ↓
+~2.6× branch misses (Clang 18 -O2)
+
+PGO
+  ↓
+branch_weights 0 / 2097152 / 1048576
+  ↓
+forward je + fall-through inc
+  ↓
+misses ≈ O1 (MBP on or off)
+</pre>
+
+:::
+
+### Evidence 1 — Part 1 baseline (Clang 24 stage1, n=100)
+
+**`-O1`:** 0.657M branch misses, 64.9 ms · **`-O2`:** 1.730M, 71.9 ms · **`-O2 -mllvm -disable-block-placement`:** 0.660M, 66.3 ms. Pass identified: **MachineBlockPlacement**. Full hunt: [part 1](/docs/articles/when-o2-layout-hurts-machineblockplacement).
+
+### Evidence 2 — Workload branch mix
+
+Lines are `a,b,c\n`. At the separator check, **`,` is ~67%** and **`\n` is ~33%** — both hot, not one cold branch. Counts: **2 097 152 / 1 048 576 / ~0** (comma / newline / neither) for `N = 1048576`.
+
+### Evidence 3 — IR without profile
+
+`-O2` IR has a `switch i8` for `,` and `\n` into a **shared increment block**, with **no `branch_weights`**. MachineBlockPlacement runs on static **BranchProbabilityInfo** after that.
+
+### Evidence 4 — IR with PGO
+
+`-fprofile-use` attaches `!59 = !{!"branch_weights", i32 0, i32 2097152, i32 1048576}` — matching Evidence 2 exactly.
+
+### Evidence 5 — Layout and matrix (revalidated 2026-08-28)
+
+**Clang 18 `-O2`, no profile:** backward `je` → **1.724M** misses, **69.3 ms** (n=100). **`-O2 -fprofile-use`:** forward fall-through inc → **0.653M** misses, **63.3 ms**; MBP on vs off unchanged once weights exist.
+
+:::tip Takeaway
+**MachineBlockPlacement is not missing from `-O2` — profile metadata is.** Without `branch_weights`, static guesses on a two-hot shared-target switch are enough to pick a layout this CPU mispredicts heavily.
+:::
+
+The point is not “turn off block placement in production.” The point is **what input the pass needs** — and that PGO (or better static weights) supplies it.
+
+<div>
+  <AdBanner />
+</div>
+
+## 1. Branch mix at the separator
+
+Same hot kernel as [part 1](/docs/articles/when-o2-layout-hurts-machineblockplacement):
+
+```c
+if (p < end && (*p == ',' || *p == '\n'))
+  p++;
+```
+
+Input: `N` lines of `a,b,c\n` (seed `29`). Three fields ⇒ **three separator checks per line**.
+
+| Outcome | When | Count (`N = 1048576`) | Share |
+|---------|------|----------------------:|------:|
+| `,` | after field 1 or 2 | 2 097 152 | **66.7%** |
+| `\n` | after field 3 | 1 048 576 | **33.3%** |
+| neither | edge / bad line | ~0 on this input | ~0% |
+
+**3 145 728** separator checks total. Both `,` and `\n` are hot; default is cold.
+
+LLVM lowers the test to a **`switch i8`** with **one shared target** for comma and newline — the increment block. That CFG shape is where **BranchProbabilityInfo** feeds **MachineBlockPlacement**: two frequent edges into the same block, plus a cold default.
+
+**Artifacts:** <a href="/files/articles/machineblockplacement-wrong-bet-static-probabilities-pgo/" target="_blank" rel="noopener">bundle</a> · <a href="/files/articles/machineblockplacement-wrong-bet-static-probabilities-pgo/highlights.json" target="_blank" rel="noopener">highlights.json</a> · <a href="/files/articles/machineblockplacement-wrong-bet-static-probabilities-pgo/source/machine_block_placement_csv_parse.c" target="_blank" rel="noopener">source</a>
+
+## 2. Without PGO: no `branch_weights`
+
+Emit IR at `-O2` (Clang 24 stage1, no profile):
+
+```llvm
+; ir/parse.O2.ll — separator (abbrev.)
+%61 = load i8, ptr %54, align 1
+switch i8 %61, label %64 [
+  i8 44, label %62    ; ','
+  i8 10, label %62    ; '\n'
+]
+```
+
+No `!prof`, no `branch_weights`. The middle-end already had a similar `switch` at `-O1`; part 1 ruled out SimplifyCFG / jump-threading as the miss driver. The regression lands **after** IR is stable — when **MachineBlockPlacement** reads **static** branch probabilities on the machine CFG.
+
+Without profile metadata, those probabilities come from LLVM heuristics (switch defaults, loop structure, symmetry guesses). On this switch, the static guess is wrong enough that MBP **outlines the shared increment** and emits **backward taken `je`** on both hot compares — the layout from part 1 §8.
+
+## 3. With PGO: measured weights on the switch
+
+Profile workflow (Clang **18.1.3** here — our stage1 tree lacks `libclang_rt.profile`; Ubuntu’s package links cleanly):
+
+```bash
+clang -O2 -fprofile-generate -o parse.gen machine_block_placement_csv_parse.c
+LLVM_PROFILE_FILE=default.profraw ./parse.gen 1048576
+llvm-profdata merge -output=default.profdata default.profraw
+clang -O2 -fprofile-use=default.profdata -o parse.O2.pgo machine_block_placement_csv_parse.c
+```
+
+Same `switch`, now annotated (<a href="/files/articles/machineblockplacement-wrong-bet-static-probabilities-pgo/ir/parse.O2_pgo.ll" target="_blank" rel="noopener">`parse.O2_pgo.ll`</a>):
+
+```llvm
+switch i8 %60, label %63 [
+  i8 44, label %61
+  i8 10, label %61
+], !prof !59
+
+!59 = !{!"branch_weights", i32 0, i32 2097152, i32 1048576}
+;                      default   comma      newline
+```
+
+**0 / 2 097 152 / 1 048 576** — matches §1 by construction (one training run on the same input). PGO recorded the branch histogram MachineBlockPlacement needed; it did not invent a new heuristic.
+
+## 4. Assembly: layout tracks metadata
+
+**Clang 18 `-O2`, no profile** — backward shared increment (<a href="/files/articles/machineblockplacement-wrong-bet-static-probabilities-pgo/asm/parse.O2_c18.asm" target="_blank" rel="noopener">asm</a>):
+
+```text
+1320:  cmp    $0x2c,%edx
+1323:  je     12f0                ; backward → shared inc
+1325:  cmp    $0xa,%edx
+1328:  je     12f0
+132a:  jmp    12f3                ; neither
+```
+
+Goes with **1.724M** branch misses, **69.3 ms** (Clang 18, n=100 — revalidated 2026-08-28).
+
+**`-O2 -fprofile-use`, MBP on** — forward fall-through increment (<a href="/files/articles/machineblockplacement-wrong-bet-static-probabilities-pgo/asm/parse.O2_pgo.asm" target="_blank" rel="noopener">asm</a>):
+
+```text
+1225:  cmp    $0x2c,%edx
+1228:  je     +5
+122a:  cmp    $0xa,%edx
+122d:  jne    +3
+122f:  inc    %rcx                ; fall-through
+```
+
+Same shape as `-O1` / `-O2 -disable-block-placement` in part 1. Profile data recovered the layout **without** disabling the pass.
+
+## 5. MBP × PGO matrix
+
+**Machine:** AMD Ryzen 7 9700X · **Input:** `N = 1048576` · **Runs:** 100 (`perf stat -r 100`) · **Compiler:** Clang 18.1.3 · **Revalidated:** 2026-08-28
+
+| Build | MBP | PGO | Time | Branch misses | vs `-O1` misses |
+|-------|:---:|:---:|-----:|--------------:|----------------:|
+| `-O1` | — | OFF | **64.0 ms** | **0.654M** | 1.00× |
+| `-O2` | ON | OFF | 69.3 ms | **1.724M** | **2.64×** |
+| `-O2 -mllvm -disable-block-placement` | OFF | OFF | **63.7 ms** | **0.654M** | 1.00× |
+| `-O2 -fprofile-use` | ON | ON | **63.3 ms** | **0.653M** | 1.00× |
+| `-O2 -fprofile-use -mllvm -disable-block-placement` | OFF | ON | **63.7 ms** | **0.654M** | 1.00× |
+
+<a href="/files/articles/machineblockplacement-wrong-bet-static-probabilities-pgo/highlights.json" target="_blank" rel="noopener">highlights.json</a> · <a href="/files/articles/machineblockplacement-wrong-bet-static-probabilities-pgo/scripts/run_pgo_matrix.sh" target="_blank" rel="noopener">`run_pgo_matrix.sh`</a> · <a href="/files/articles/machineblockplacement-wrong-bet-static-probabilities-pgo/scripts/revalidate.sh" target="_blank" rel="noopener">`revalidate.sh`</a>
+
+**Without PGO:** only disabling MBP restores misses — static probabilities are the gap.
+
+**With PGO:** misses and time both return to `-O1` levels whether MBP is on or off. The **`branch_weights`** attachment drives layout; toggling placement is a second-order effect once counts are correct.
+
+### Part 1 stage1 (historical) vs current tip
+
+Part 1 primary table: Clang 24 stage1 **`196786fa…`**, n=100 — `-O2` **1.730M** misses, **71.9 ms**; `-O2np` **0.660M**, **66.3 ms** (<a href="/files/articles/when-o2-layout-hurts-machineblockplacement/highlights.json" target="_blank" rel="noopener">bundled</a>).
+
+**Revalidated today** on the same testcase, same machine:
+
+| Compiler | `-O2` misses | `-O2` time | Bad backward layout? |
+|----------|-------------:|-----------:|:--------------------:|
+| Clang 18.1.3 | **1.724M** (2.64×) | 69.3 ms | yes |
+| Stage1 **`b7dc8e35…`** (current tip) | **0.654M** (1.00×) | 63.6 ms | **no** |
+
+Mechanism in part 1/part 2 still holds on **Clang 18**. Current stage1 tip already emits forward-increment `-O2` on this box — use Clang 18 to reproduce the regression for teaching and bisect.
+
+## 6. Pipeline
+
+```text
+if (*p==',' || *p=='\n') p++
+  ↓
+switch i8 { 44→inc, 10→inc, default→skip }
+  ↓
+BranchProbabilityInfo  ← static guess (no PGO) or branch_weights (PGO)
+  ↓
+MachineBlockPlacement
+  ↓
+fall-through vs taken encoding at separator
+  ↓
+branch-miss counters
+```
+
+The “wrong bet” is not “delete MachineBlockPlacement.” It is **optimizing layout against branch odds that do not match a two-hot, shared-target char switch when profile metadata is absent.** PGO supplies the odds; `-disable-block-placement` is part 1’s diagnostic override.
+
+## 7. If you hit this in the field
+
+**No PGO, `-O2` slower, `branch-misses` spike on a parse-style loop:** compare against `-mllvm -disable-block-placement` on **that translation unit** to test a layout hypothesis. Diagnostic only — global placement-off can regress code that relied on fall-through layout elsewhere.
+
+**Shipping PGO:** train on representative inputs. Switches like §2 get `branch_weights`; MBP can stay enabled.
+
+**Upstream:** [llvm/llvm-project#218248](https://github.com/llvm/llvm-project/issues/218248). Durable fix is likely **better static weights for multi-arm char tests** (or honoring switch symmetry) when metadata is missing — not removing the pass.
+
+## 8. Reproduce
+
+Uses the **part 1 testcase** (`machine_block_placement_csv_parse.c`).
+
+```bash
+cd static/files/articles/machineblockplacement-wrong-bet-static-probabilities-pgo
+chmod +x scripts/run_pgo_matrix.sh scripts/revalidate.sh
+
+# Full matrix (build + perf n=100)
+./scripts/run_pgo_matrix.sh clang-18 1048576 100
+
+# Or revalidate from part-1 source path
+./scripts/revalidate.sh /tmp/mbb_revalidate
+
+# Correctness
+./results/perf/parse.O2_c18 1048576    # expect PASS sum=15723844160
+
+# IR / asm checks
+grep -A3 'switch i8' ir/parse.O2.ll
+grep 'branch_weights' ir/parse.O2_pgo.ll | head
+grep -A5 'cmp.*0x2c' asm/parse.O2_c18.asm
+grep -A5 'cmp.*0x2c' asm/parse.O2_pgo.asm
+```
+
+`kernel.perf_event_paranoid=0` if counters are blocked.
+
+## 9. Limitations
+
+- One kernel, Zen 5. Predictor mechanism inferred from counters + layout, not uarch simulation.
+- PGO trained on the **benchmark input** — best case for weight accuracy.
+- Correlation chain (metadata → layout → misses) is strong; full proof through predictor tables is not claimed.
+- Check your **compiler revision** — newer stage1 may no longer emit the bad default `-O2` layout.
+
+## References
+
+- [Part 1: When Clang `-O2` Gets Slower](/docs/articles/when-o2-layout-hurts-machineblockplacement)  
+- LLVM [MachineBlockPlacement](https://llvm.org/doxygen/MachineBlockPlacement_8cpp.html) · [BranchProbabilityInfo](https://llvm.org/doxygen/classllvm_1_1BranchProbabilityInfo.html)  
+- [llvm/llvm-project#218248](https://github.com/llvm/llvm-project/issues/218248)  
+- <a href="/files/articles/machineblockplacement-wrong-bet-static-probabilities-pgo/" target="_blank" rel="noopener">Part 2 artifacts</a> · <a href="/files/articles/when-o2-layout-hurts-machineblockplacement/" target="_blank" rel="noopener">Part 1 artifacts</a>
+
+:::note Lab note
+Matrix revalidated **2026-08-28**: Clang 18.1.3, `perf` n=100, testcase from <a href="/files/articles/when-o2-layout-hurts-machineblockplacement/source/machine_block_placement_csv_parse.c" target="_blank" rel="noopener">part 1 source</a>. All builds PASS `sum=15723844160` at `N=1048576`. Current stage1 tip (`b7dc8e35…`) does **not** reproduce the bad `-O2` layout on this machine.
+:::
diff --git a/docs/articles/when-o2-layout-hurts-machineblockplacement.md b/docs/articles/when-o2-layout-hurts-machineblockplacement.md
@@ -172,7 +172,7 @@ MachineBlockPlacement
 block order → fall-through / taken branches → machine code
 ```
 
-The follow-up article asks what probabilities LLVM assigns to `,`, `\n`, and “neither,” and whether that explains the shared-increment layout at `-O2`.
+The follow-up article — [**Why MachineBlockPlacement Made the Wrong Bet**](/docs/articles/machineblockplacement-wrong-bet-static-probabilities-pgo) — documents what probabilities LLVM assigns with and without PGO, and whether that explains the shared-increment layout at `-O2`.
 
 Full file with expected-sum checks: <a href="/files/articles/when-o2-layout-hurts-machineblockplacement/source/machine_block_placement_csv_parse.c" target="_blank" rel="noopener">machine_block_placement_csv_parse.c</a>. All measured data is in the <a href="/files/articles/when-o2-layout-hurts-machineblockplacement/" target="_blank" rel="noopener">artifact index</a>. A minimal pasteable `main` is also in the appendix below.
 
@@ -476,8 +476,8 @@ Follow-up: why MachineBlockPlacement chose the wrong layout **without** profile
 
 **What this suggests:** without PGO, static probabilities → bad layout → high misses. With PGO, layout is good whether MBP is on or off. Still open: exact `,` / `\n` / neither probabilities in BranchProbabilityInfo, and stage1 n=100 confirmation.
 
-:::tip Coming next on CompilerSutra
-**Why MachineBlockPlacement Made the Wrong Bet: Static Probabilities, Fall-Through, and PGO**
+:::tip Part 2 — now published
+**[Why MachineBlockPlacement Made the Wrong Bet: Static Probabilities, Fall-Through, and PGO](/docs/articles/machineblockplacement-wrong-bet-static-probabilities-pgo)** — workload branch mix, missing `branch_weights`, PGO matrix, and the layout fix.
 :::
 
 ## Appendix: pasteable testcase
@@ -539,7 +539,7 @@ Expected sum for `1048576` / seed `29`: `15723844160` (asserted in the full kit
 - Upstream report: [llvm/llvm-project#218248](https://github.com/llvm/llvm-project/issues/218248)
 - [`opt-bisect-limit`](https://llvm.org/docs/CommandGuide/opt.html) (same knob via `clang -mllvm`)
 - [CompilerSutra Perf](https://pypi.org/project/compilersutra-perf/) (`pip install compilersutra-perf`)
-- Related: [GCC vs Clang real benchmarks](/docs/articles/gcc_vs_clang_real_benchmarks_2026_reporter), [stencil pass trace](/docs/articles/where_gcc_and_clang_diverge_stencil_pass_trace)
+- Related: [Part 2 — static probabilities, PGO, and layout](/docs/articles/machineblockplacement-wrong-bet-static-probabilities-pgo), [GCC vs Clang real benchmarks](/docs/articles/gcc_vs_clang_real_benchmarks_2026_reporter), [stencil pass trace](/docs/articles/where_gcc_and_clang_diverge_stencil_pass_trace)
 
 :::note Lab note
 Primary: Clang 24 stage1, Experiment A/B on <a href="/files/articles/when-o2-layout-hurts-machineblockplacement/source/machine_block_placement_csv_parse.c" target="_blank" rel="noopener">machine_block_placement_csv_parse.c</a>. Bundled public artifacts: <a href="/files/articles/when-o2-layout-hurts-machineblockplacement/" target="_blank" rel="noopener">artifact index</a>. Secondary: Clang 18.1.3.

diff --git a/sidebars/site.js b/sidebars/site.js
@@ -210,6 +210,7 @@ const site = {
         'articles/gcc_vs_clang_stencil_ir_passes_part2b',
         'articles/where_gcc_and_clang_diverge_stencil_pass_trace',
         'articles/when-o2-layout-hurts-machineblockplacement',
+        'articles/machineblockplacement-wrong-bet-static-probabilities-pgo',
         'articles/hft_stdlib_restrictions',
         'articles/language_energy_efficiency_validation',
         {