From d8286fa2a5775ba7a4ab2c9fe5fb9376d500bfd8 Mon Sep 17 00:00:00 2001
From: Christoffer Lerno <christoffer@aegik.com>
Date: Fri, 19 Sep 2025 13:34:06 +0200
Subject: [PATCH] Add 'loop-vectorize', 'slp-vectorize', 'unroll-loops' and
 'merge-functions' optimization flags #2491.

---
 releasenotes.md                  |  1 +
 src/build/build_options.c        | 24 ++++++++++++++++++++++++
 src/build/project.c              | 18 ++++++++++++++++--
 src/build/project_manipulation.c |  4 ++++
 4 files changed, 45 insertions(+), 2 deletions(-)

diff --git a/releasenotes.md b/releasenotes.md
index 76d8d532b..09df6a9e0 100644
--- a/releasenotes.md
+++ b/releasenotes.md
@@ -10,6 +10,7 @@
 - Unify generic and regular module namespace.
 - `env::PROJECT_VERSION` now returns the version in project.json.
 - Comparing slices and arrays of user-defined types that implement == operator now works #2486.
+- Add 'loop-vectorize', 'slp-vectorize', 'unroll-loops' and 'merge-functions' optimization flags #2491.
 
 ### Fixes
 - Compiler assert with var x @noinit = 0 #2452
diff --git a/src/build/build_options.c b/src/build/build_options.c
index 190b01a1d..362c9c05e 100644
--- a/src/build/build_options.c
+++ b/src/build/build_options.c
@@ -128,6 +128,10 @@ static void usage(bool full)
 		print_opt("--panic-msg=<yes|no>", "Turn panic message output on or off.");
 		print_opt("--optlevel=<option>", "Code optimization level: none, less, more, max.");
 		print_opt("--optsize=<option>", "Code size optimization: none, small, tiny.");
+		print_opt("--unroll-loops=<yes|no>", "Enable loop unrolling.");
+		print_opt("--loop-vectorize=<yes|no>", "Enable loop auto-vectorization.");
+		print_opt("--slp-vectorize=<yes|no>", "Enable SLP (superword-level parallelism) auto-vectorization.");
+		print_opt("--merge-functions=<yes|no>", "Enable function merging.");
 		print_opt("--single-module=<yes|no>", "Compile all modules together, enables more inlining.");
 		print_opt("--show-backtrace=<yes|no>", "Show detailed backtrace on segfaults.");
 		print_opt("--lsp", "Emit data about errors suitable for a LSP.");
@@ -862,6 +866,26 @@ static void parse_option(BuildOptions *options)
 				options->optlevel = parse_opt_select(OptimizationLevel, argopt, optlevels);
 				return;
 			}
+			if ((argopt = match_argopt("merge-functions")))
+			{
+				options->merge_functions = parse_opt_select(MergeFunctions, argopt, on_off);
+				return;
+			}
+			if ((argopt = match_argopt("loop-vectorize")))
+			{
+				options->loop_vectorization = parse_opt_select(AutoVectorization, argopt, on_off);
+				return;
+			}
+			if ((argopt = match_argopt("unroll-loops")))
+			{
+				options->unroll_loops = parse_opt_select(UnrollLoops, argopt, on_off);
+				return;
+			}
+			if ((argopt = match_argopt("slp-vectorize")))
+			{
+				options->slp_vectorization = parse_opt_select(AutoVectorization, argopt, on_off);
+				return;
+			}
 			if ((argopt = match_argopt("safe")))
 			{
 				options->safety_level = parse_opt_select(SafetyLevel, argopt, on_off);
diff --git a/src/build/project.c b/src/build/project.c
index e0df5e42d..c844b7ed1 100644
--- a/src/build/project.c
+++ b/src/build/project.c
@@ -27,10 +27,13 @@ const char *project_default_keys[][2] = {
 		{"linker", "'builtin' for the builtin linker, 'cc' for the system linker or <path> to a custom compiler."},
 		{"linker-search-paths", "Linker search paths."},
 		{"linux-crt", "Set the directory to use for finding crt1.o and related files."},
+		{"linux-crtbegin", "Set the directory to use for finding crtbegin.o and related files."},
+		{"loop-vectorize", "Force enable/disable loop auto-vectorization."},
 		{"macos-min-version", "Set the minimum MacOS version to compile for."},
 		{"macos-sdk-version", "Set the MacOS SDK compiled for." },
 		{"macossdk", "Set the directory for the MacOS SDK for cross compilation."},
 		{"memory-env", "Set the memory environment: normal, small, tiny, none."},
+		{"merge-functions", "Force enable/disable function merging."},
 		{"no-entry", "Do not generate (or require) a main function."},
 		{"opt", "Optimization setting: O0, O1, O2, O3, O4, O5, Os, Oz."},
 		{"optlevel", "Code optimization level: none, less, more, max."},
@@ -46,6 +49,7 @@ const char *project_default_keys[][2] = {
 		{"script-dir", "The directory where 'exec' is run."},
 		{"show-backtrace", "Print backtrace on signals."},
 		{"single-module", "Compile all modules together, enables more inlining."},
+		{"slp-vectorize", "Force enable/disable SLP auto-vectorization."},
 		{"soft-float", "Output soft-float functions."},
 		{"sources", "Paths to project sources for all targets."},
 		{"strip-unused", "Strip unused code and globals from the output. (default: true)"},
@@ -55,6 +59,7 @@ const char *project_default_keys[][2] = {
 		{"test-sources", "Paths to project test sources for all targets."},
 		{"testfn", "Override the test function."},
 		{"trap-on-wrap", "Make signed and unsigned integer overflow generate a panic rather than wrapping."},
+		{"unroll-loops", "Force enable/disable loop unrolling optimization."},
 		{"use-stdlib", "Include the standard library (default: true)."},
 		{"vendor", "Vendor specific extensions, ignored by c3c."},
 		{"version", "Version using semantic versioning."},
@@ -65,7 +70,6 @@ const char *project_default_keys[][2] = {
 		{"x86-stack-struct-return", "Return structs on the stack for x86."},
 		{"x86cpu", "Set general level of x64 cpu: baseline, ssse3, sse4, avx1, avx2-v1, avx2-v2 (Skylake/Zen1+), avx512 (Icelake/Zen4+), native."},
 		{"x86vec", "Set max type of vector use: none, mmx, sse, avx, avx512, native."},
-		{"linux-crtbegin", "Set the directory to use for finding crtbegin.o and related files."},
 };
 
 const int project_default_keys_count = ELEMENTLEN(project_default_keys);
@@ -101,10 +105,13 @@ const char* project_target_keys[][2] = {
 		{"linker-search-paths", "Additional linker search paths for the target."},
 		{"linker-search-paths-override", "Linker search paths for this target, overriding global settings."},
 		{"linux-crt", "Set the directory to use for finding crt1.o and related files."},
+		{"linux-crtbegin", "Set the directory to use for finding crtbegin.o and related files."},
+		{"loop-vectorize", "Force enable/disable loop auto-vectorization."},
 		{"macos-min-version", "Set the minimum MacOS version to compile for."},
 		{"macos-sdk-version", "Set the MacOS SDK compiled for." },
 		{"macossdk", "Set the directory for the MacOS SDK for cross compilation."},
 		{"memory-env", "Set the memory environment: normal, small, tiny, none."},
+		{"merge-functions", "Force enable/disable function merging."},
 		{"name", "Set the name to be different from the target name."},
 		{"no-entry", "Do not generate (or require) a main function."},
 		{"opt", "Optimization setting: O0, O1, O2, O3, O4, O5, Os, Oz."},
@@ -121,6 +128,7 @@ const char* project_target_keys[][2] = {
 		{"script-dir", "The directory where 'exec' is run."},
 		{"show-backtrace", "Print backtrace on signals."},
 		{"single-module", "Compile all modules together, enables more inlining."},
+		{"slp-vectorize", "Force enable/disable SLP auto-vectorization."},
 		{"soft-float", "Output soft-float functions."},
 		{"sources", "Additional paths to project sources for the target."},
 		{"sources-override", "Paths to project sources for this target, overriding global settings."},
@@ -132,6 +140,7 @@ const char* project_target_keys[][2] = {
 		{"testfn", "Override the test function."},
 		{"trap-on-wrap", "Make signed and unsigned integer overflow generate a panic rather than wrapping."},
 		{"type", "Type of output, one of 'executable', 'static-lib', 'dynamic-lib', 'benchmark', 'test', 'object-files' and 'prepare'." },
+		{"unroll-loops", "Force enable/disable loop unrolling optimization."},
 		{"use-stdlib", "Include the standard library (default: true)."},
 		{"vendor", "Vendor specific extensions, ignored by c3c."},
 		{"version", "Version using semantic versioning."},
@@ -142,7 +151,7 @@ const char* project_target_keys[][2] = {
 		{"x86-stack-struct-return", "Return structs on the stack for x86."},
 		{"x86cpu", "Set general level of x64 cpu: baseline, ssse3, sse4, avx1, avx2-v1, avx2-v2 (Skylake/Zen1+), avx512 (Icelake/Zen4+), native."},
 		{"x86vec", "Set max type of vector use: none, mmx, sse, avx, avx512, native."},
-		{"linux-crtbegin", "Set the directory to use for finding crtbegin.o and related files."},
+
 };
 
 const int project_target_keys_count = ELEMENTLEN(project_target_keys);
@@ -267,6 +276,11 @@ static void load_into_build_target(BuildParseContext context, JSONObject *json,
 	// Size optimization
 	target->optsize = GET_SETTING(SizeOptimizationLevel, "optsize", optsizes, "`none`, `small`, `tiny`.");
 
+	target->loop_vectorization = (AutoVectorization)get_valid_bool(context, json, "loop-vectorize", target->loop_vectorization);
+	target->slp_vectorization = (AutoVectorization)get_valid_bool(context, json, "slp-vectorize", target->slp_vectorization);
+	target->unroll_loops = (UnrollLoops)get_valid_bool(context, json, "unroll-loops", target->unroll_loops);
+	target->merge_functions = (MergeFunctions)get_valid_bool(context, json, "merge-functions", target->merge_functions);
+
 	static const char *opt_settings[8] = {
 			[OPT_SETTING_O0] = "O0",
 			[OPT_SETTING_O1] = "O1",
diff --git a/src/build/project_manipulation.c b/src/build/project_manipulation.c
index 762553576..a75ee4a61 100644
--- a/src/build/project_manipulation.c
+++ b/src/build/project_manipulation.c
@@ -259,6 +259,10 @@ static void view_target(BuildParseContext context, JSONObject *target, bool verb
 	TARGET_VIEW_SETTING("x64 CPU level", "x86cpu", x86_cpu_set);
 	TARGET_VIEW_SETTING("Max vector use type", "x86vec", x86_vector_capability);
 	TARGET_VIEW_BOOL("Return structs on the stack", "x86-stack-struct-return");
+	TARGET_VIEW_BOOL("Unroll loops", "unroll-loops");
+	TARGET_VIEW_BOOL("SLP auto-vectorization", "slp-vectorize");
+	TARGET_VIEW_BOOL("Loop auto-vectorization", "loop-vectorize");
+	TARGET_VIEW_BOOL("Merge functions", "merge-functions");
 }