[do not merge] inspecting llc's bejaviour with llvm.sext, llvm.zext and llvm.trunc

extension-tests/.gitignore

@@ -0,0 +1,6 @@
+trunc/
+sext/
+zext/
+trunc_lowered/
+sext_lowered/
+zext_lowered/

extension-tests/README.md

@@ -0,0 +1,2 @@
+We inspect closely the lowering of `llvm.sext`, `llvm.zext`, and `llvm.trunc` instructions with 
+`llc -march=riscv64 -mcpu=generic-rv64 -mattr=+m,+b -filetype=asm -O0`, for every combination of extension from `i1` to `i64`. 

extension-tests/cluster.py

@@ -0,0 +1,141 @@
+import os
+import subprocess
+from pathlib import Path
+from collections import defaultdict
+
+# zext 
+
+input_dir  = Path("zext")
+output_dir = Path("zext_lowered")
+input_dir.mkdir(exist_ok=True)
+output_dir.mkdir(exist_ok=True)
+
+def extract_instructions(asm_path):
+    instructions = []
+    for line in asm_path.read_text().splitlines():
+        stripped = line.strip()
+        # skip empty lines, comments, directives, labels
+        if not stripped:
+            continue
+        if stripped.startswith("#"):
+            continue
+        if stripped.startswith("."):
+            continue
+        if stripped.endswith(":"):
+            continue
+        # kill: comments mid-line (e.g. "# kill: def $x10...")
+        stripped = stripped.split("#")[0].strip()
+        if not stripped:
+            continue
+        instructions.append(stripped)
+    return tuple(instructions)
+
+errors = []
+
+clusters = defaultdict(list)
+
+for asm_file in sorted(output_dir.glob("*.s")):
+    instrs = extract_instructions(asm_file)
+    clusters[instrs].append(asm_file.stem)
+
+output_log_file = open("zext_clusters.txt", "w")
+
+output_log_file.write(f"Found {len(clusters)} distinct instruction pattern(s)\n")
+output_log_file.write("=" * 60)
+
+for instrs, files in sorted(clusters.items(), key=lambda x: len(x[1]), reverse=True):
+    output_log_file.write(f"\n[{len(files)} file(s)] Pattern:\n")
+    if instrs:
+        for instr in instrs:
+            output_log_file.write(f"    {instr}\n")
+    else:
+        output_log_file.write("    (no instructions / empty body)")
+    output_log_file.write(f"  Files: {', '.join(files)}")
+
+# sext 
+
+input_dir  = Path("sext")
+output_dir = Path("sext_lowered")
+input_dir.mkdir(exist_ok=True)
+output_dir.mkdir(exist_ok=True)
+
+def extract_instructions(asm_path):
+    instructions = []
+    for line in asm_path.read_text().splitlines():
+        stripped = line.strip()
+        # skip empty lines, comments, directives, labels
+        if not stripped:
+            continue
+        if stripped.startswith("#"):
+            continue
+        if stripped.startswith("."):
+            continue
+        if stripped.endswith(":"):
+            continue
+        # kill: comments mid-line (e.g. "# kill: def $x10...")
+        stripped = stripped.split("#")[0].strip()
+        if not stripped:
+            continue
+        instructions.append(stripped)
+    return tuple(instructions)
+
+errors = []
+
+clusters = defaultdict(list)
+
+for asm_file in sorted(output_dir.glob("*.s")):
+    instrs = extract_instructions(asm_file)
+    clusters[instrs].append(asm_file.stem)
+
+output_log_file = open("sext_clusters.txt", "w")
+
+output_log_file.write(f"Found {len(clusters)} distinct instruction pattern(s)\n")
+output_log_file.write("=" * 60)
+
+for instrs, files in sorted(clusters.items(), key=lambda x: len(x[1]), reverse=True):
+    output_log_file.write(f"\n[{len(files)} file(s)] Pattern:\n")
+    if instrs:
+        for instr in instrs:
+            output_log_file.write(f"    {instr}\n")
+    else:
+        output_log_file.write("    (no instructions / empty body)")
+    output_log_file.write(f"  Files: {',\n '.join(files)}")
+
+
+# trunc
+
+input_dir  = Path("trunc")
+output_dir = Path("trunc_lowered")
+input_dir.mkdir(exist_ok=True)
+output_dir.mkdir(exist_ok=True)
+
+def extract_instructions(asm_path):
+    instructions = []
+    for line in asm_path.read_text().splitlines():
+        stripped = line.strip()
+        if not stripped or stripped.startswith("#") or stripped.startswith(".") or stripped.endswith(":"):
+            continue
+        stripped = stripped.split("#")[0].strip()
+        if stripped:
+            instructions.append(stripped)
+    return tuple(instructions)
+
+clusters = defaultdict(list)
+for asm_file in sorted(output_dir.glob("*.s")):
+    clusters[extract_instructions(asm_file)].append(asm_file.stem)
+
+output_log_file = open("trunc_clusters.txt", "w")
+
+output_log_file.write(f"\nFound {len(clusters)} distinct instruction pattern(s)\n")
+output_log_file.write("=" * 60)
+for instrs, files in sorted(clusters.items(), key=lambda x: len(x[1]), reverse=True):
+    output_log_file.write(f"\n[{len(files)} file(s)] Pattern:\n")
+    for instr in instrs or ["(empty)"]:
+        output_log_file.write(f"    {instr}\n")
+    # group files by flag variant for readability
+    by_flag = defaultdict(list)
+    for f in files:
+        flag = f.split("_flags_")[-1]
+        by_flag[flag].append(f)
+    for flag, flist in sorted(by_flag.items()):
+        output_log_file.write(f"  [{flag}] {len(flist)} case(s): {', '.join(flist)}{'...' if len(flist) > 5 else ''}")