compiler: improve target features support

3 files changed, 181 insertions, 0 deletions

diff --git a/misc/featuregen/README.md b/misc/featuregen/README.md
new file mode 100644
index 000000000..22a798cca
--- /dev/null
+++ b/misc/featuregen/README.md
@@ -0,0 +1,28 @@
+# Featuregen
+
+This directory contains a python and CPP script that generates the needed information
+for features regarding microarchitecture and target features of the compiler.
+
+It is not pretty! But LLVM has no way to query this information with their C API.
+
+It generates these globals (intended for `src/build_settings.cpp`:
+
+- `target_microarch_list`: an array of strings indexed by the architecture, each string is a comma-seperated list of microarchitectures available on that architecture
+- `target_features_list`: an array of strings indexed by the architecture, each string is a comma-seperated list of target features available on that architecture
+- `target_microarch_counts`: an array of ints indexed by the architecture, each int represents the amount of microarchitectures available on that target, intended for easier iteration of the next global
+- `microarch_features_list`: an array of a tuple like struct where the first string is a microarchitecture and the second is a comma-seperated list of all features that are enabled by default for it
+
+In order to get the default features for a microarchitecture there is a small CPP program that takes
+a target triple and microarchitecture and spits out the default features, this is then parsed by the python script.
+
+This should be ran each time we update LLVM to stay in sync.
+
+If there are minor differences (like the Odin user using LLVM 14 and this table being generated on LLVM 17) it
+does not impact much at all, the only thing it will do is make LLVM print a message that the feature is ignored (if it was added between 14 and 17 in this case).
+
+## Usage
+
+1. Make sure the table of architectures at the top of the python script is up-to-date (the triple can be any valid triple for the architecture)
+1. `./build.sh`
+1. `python3 featuregen.py`
+1. Copy the output into `src/build_settings.cpp`
diff --git a/misc/featuregen/featuregen.cpp b/misc/featuregen/featuregen.cpp
new file mode 100644
index 000000000..a1d00ab31
--- /dev/null
+++ b/misc/featuregen/featuregen.cpp
@@ -0,0 +1,37 @@
+#include <llvm/MC/MCSubtargetInfo.h>
+#include <llvm/MC/TargetRegistry.h>
+#include <llvm/Support/raw_ostream.h>
+#include <llvm/ADT/ArrayRef.h>
+#include <llvm/Support/InitLLVM.h>
+#include <llvm/Support/TargetSelect.h>
+
+// Dumps the default set of supported features for the given microarch.
+int main(int argc, char **argv) {
+	if (argc < 3) {
+		llvm::errs() << "Error: first arg should be triple, second should be microarch\n";
+		return 1;
+	}
+
+	llvm::InitializeAllTargets();
+	llvm::InitializeAllTargetMCs();
+
+	std::string error;
+	const llvm::Target* target = llvm::TargetRegistry::lookupTarget(argv[1], error);
+
+	if (!target) {
+		llvm::errs() << "Error: " << error << "\n";
+		return 1;
+	}
+
+	auto STI = target->createMCSubtargetInfo(argv[1], argv[2], "");
+
+	std::string plus = "+";
+	llvm::ArrayRef<llvm::SubtargetFeatureKV> features = STI->getAllProcessorFeatures();
+	for (const auto& feature : features) {
+		if (STI->checkFeatures(plus + feature.Key)) {
+			llvm::outs() << feature.Key << "\n";
+		}
+	}
+
+	return 0;
+}
diff --git a/misc/featuregen/featuregen.py b/misc/featuregen/featuregen.py
new file mode 100644
index 000000000..da4cc68f5
--- /dev/null
+++ b/misc/featuregen/featuregen.py
@@ -0,0 +1,116 @@
+import subprocess
+import tempfile
+import os
+import sys
+
+archs = [
+	("amd64",     "linux_amd64", "x86_64-pc-linux-gnu", [], []),
+	("i386",      "linux_i386",  "i386-pc-linux-gnu",   [], []),
+	("arm32",     "linux_arm32", "arm-linux-gnu",       [], []),
+	("arm64",     "linux_arm64", "aarch64-linux-elf",   [], []),
+	("wasm32",    "js_wasm32",   "wasm32-js-js",        [], []),
+	("wasm64p32", "js_wasm64p32","wasm32-js-js",        [], []),
+];
+
+SEEKING_CPUS     = 0
+PARSING_CPUS     = 1
+PARSING_FEATURES = 2
+
+with tempfile.NamedTemporaryFile(suffix=".odin", delete=True) as temp_file:
+	temp_file.write(b"package main\n")
+
+	for arch, target, triple, cpus, features in archs:
+		cmd = ["odin", "build", temp_file.name, "-file", "-build-mode:llvm", "-out:temp", "-target-features:\"help\"", f"-target:\"{target}\""]
+		process = subprocess.Popen(cmd, stderr=subprocess.PIPE, text=True)
+
+		state = SEEKING_CPUS
+		for line in process.stderr:
+
+			if state == SEEKING_CPUS:
+				if line == "Available CPUs for this target:\n":
+					state = PARSING_CPUS
+			
+			elif state == PARSING_CPUS:
+				if line == "Available features for this target:\n":
+					state = PARSING_FEATURES
+					continue
+			
+				parts = line.split(" -", maxsplit=1)
+				if len(parts) < 2:
+					continue
+
+				cpu = parts[0].strip()
+				cpus.append(cpu)
+
+			elif state == PARSING_FEATURES:
+				if line == "\n" and len(features) > 0:
+					break
+
+				parts = line.split(" -", maxsplit=1)
+				if len(parts) < 2:
+					continue
+
+				feature = parts[0].strip()
+				features.append(feature)
+
+		process.wait()
+		if process.returncode != 0:
+			print(f"odin build returned with non-zero exit code {process.returncode}")
+			sys.exit(1)
+
+		os.remove("temp.ll")
+
+def print_default_features(triple, microarch):
+	cmd = ["./featuregen", triple, microarch]
+	process = subprocess.Popen(cmd, stdout=subprocess.PIPE, text=True)
+	first = True
+	for line in process.stdout:
+		print("" if first else ",", line.strip(), sep="", end="")
+		first = False
+	process.wait()
+	if process.returncode != 0:
+		print(f"featuregen returned with non-zero exit code {process.returncode}")
+		sys.exit(1)
+
+print("// Generated with the featuregen script in `misc/featuregen`")
+print("gb_global String target_microarch_list[TargetArch_COUNT] = {")
+print("\t// TargetArch_Invalid:")
+print('\tstr_lit(""),')
+for arch, target, triple, cpus, features in archs:
+	print(f"\t// TargetArch_{arch}:")
+	print(f'\tstr_lit("{','.join(cpus)}"),')
+print("};")
+
+print("")
+
+print("// Generated with the featuregen script in `misc/featuregen`")
+print("gb_global String target_features_list[TargetArch_COUNT] = {")
+print("\t// TargetArch_Invalid:")
+print('\tstr_lit(""),')
+for arch, target, triple, cpus, features in archs:
+	print(f"\t// TargetArch_{arch}:")
+	print(f'\tstr_lit("{','.join(features)}"),')
+print("};")
+
+print("")
+
+print("// Generated with the featuregen script in `misc/featuregen`")
+print("gb_global int target_microarch_counts[TargetArch_COUNT] = {")
+print("\t// TargetArch_Invalid:")
+print("\t0,")
+for arch, target, triple, cpus, feature in archs:
+	print(f"\t// TargetArch_{arch}:")
+	print(f"\t{len(cpus)},")
+print("};")
+
+print("")
+
+print("// Generated with the featuregen script in `misc/featuregen`")
+print("gb_global MicroarchFeatureList microarch_features_list[] = {")
+for arch, target, triple, cpus, features in archs:
+	print(f"\t// TargetArch_{arch}:")
+	for cpu in cpus:
+		print(f'\t{{ str_lit("{cpu}"), str_lit("', end="")
+		print_default_features(triple, cpu)
+		print('") },')
+print("};")