Feedback‑Driven Compiler Optimizations for Cloud C/C++ Applications

This article explores how system‑level compiler and OS optimizations can improve the performance of modern cloud applications, using C/C++ workloads as a concrete example.

1. Characteristics of modern cloud workloads – Cloud services exhibit highly diversified workloads, flat hotspot distribution across many components, and extremely large binary sizes (tens to hundreds of megabytes). These factors cause severe instruction‑cache (I‑Cache) and instruction‑TLB pressure, leading to front‑end stalls that can occupy 15‑30% of execution time, far higher than on desktop or mobile workloads.

2. Top‑Down Micro‑Architecture Analysis (TMAM) – The TMAM methodology partitions execution into Frontend‑Bound, Bad‑Speculation, Backend‑Bound, and Retiring stages. Cloud databases R and M were analyzed, showing that Frontend‑Bound dominates and that I‑Cache miss rates are especially high for workloads such as MySQL.

3. Feedback‑driven optimization techniques – The article describes several feedback‑guided methods:

Profile‑Guided Optimization (PGO) – both Sampling PGO (using tools like Linux perf ) and Instrumentation PGO (adding recording points to the binary).

Binary Optimization and Layout Tool (BOLT) – reorders code based on collected profiles.

AutoFDO – integrates PGO and BOLT into a CI/CD pipeline for continuous optimization.

Sampling PGO collects runtime call information without modifying the binary, while Instrumentation PGO inserts probes and incurs higher overhead. Both approaches require realistic input data to be effective.

4. Specific compiler optimizations

Basic Block Reorder – Reorders hot basic blocks to improve I‑Cache locality. Example before/after layout is illustrated with a simple loop:

for (i = 0; i < n; i++) {
    guard(i < len);
    ...
}

After profiling, the guard can be moved outside the loop:

for (i = 0; i < n; i++) {
    guard(n-1 < len);
    ...
}

Partial Inlining – Uses profiling data to inline only the hot portions of a function. Example:

void foo() {
    bar();
    // rest of foo
}

void bar() {
    if (X) return;
    // rest of bar
}

After partial inlining, the hot return path is inlined into foo :

void foo() {
    if (!X) bar.outlined();
    // rest of foo
}

void bar.outlined() {
    // rest of bar
}

Branch optimization – Eliminates redundant branches, reduces branch execution frequency, and improves branch‑prediction accuracy using profiling‑guided layout changes.

5. Function Reorder – Reorders functions based on a weighted call‑graph derived from profiling. Techniques such as Facebook’s HFSort are mentioned, and the workflow includes collecting profiling data, building the call graph, applying a placement algorithm, and relinking.

6. Practical results – The optimizations were applied to two internal databases (R and M). TMAM analysis showed a reduction of Frontend‑Bound stalls by up to 18% for MySQL read workloads. Figures (omitted) illustrate the performance gains.

7. Conclusion – System‑level analysis of cloud workloads and tight integration of business‑specific profiling with compiler/OS optimizations can yield substantial performance improvements, providing a competitive edge for cloud services.