Directly generate correctly formatted intrinsics file and regenerate intrinsics

src/intrinsic/archs.rs

@@ -388,6 +388,7 @@ fn map_arch_intrinsic(full_name: &str) -> &'static str {
                     "s.barrier.signal.isfirst" => "__builtin_amdgcn_s_barrier_signal_isfirst",
                     "s.barrier.signal.var" => "__builtin_amdgcn_s_barrier_signal_var",
                     "s.barrier.wait" => "__builtin_amdgcn_s_barrier_wait",
+                    "s.bitreplicate" => "__builtin_amdgcn_s_bitreplicate",
                     "s.buffer.prefetch.data" => "__builtin_amdgcn_s_buffer_prefetch_data",
                     "s.cluster.barrier" => "__builtin_amdgcn_s_cluster_barrier",
                     "s.dcache.inv" => "__builtin_amdgcn_s_dcache_inv",
@@ -8712,10 +8713,6 @@ fn map_arch_intrinsic(full_name: &str) -> &'static str {
                     "bmi.bextr.64" => "__builtin_ia32_bextr_u64",
                     "bmi.bzhi.32" => "__builtin_ia32_bzhi_si",
                     "bmi.bzhi.64" => "__builtin_ia32_bzhi_di",
-                    "bmi.pdep.32" => "__builtin_ia32_pdep_si",
-                    "bmi.pdep.64" => "__builtin_ia32_pdep_di",
-                    "bmi.pext.32" => "__builtin_ia32_pext_si",
-                    "bmi.pext.64" => "__builtin_ia32_pext_di",
                     "cldemote" => "__builtin_ia32_cldemote",
                     "clflushopt" => "__builtin_ia32_clflushopt",
                     "clrssbsy" => "__builtin_ia32_clrssbsy",

src/intrinsic/old_archs.rs

@@ -1240,6 +1240,10 @@ pub(crate) fn old_archs(arch: &str, name: &str) -> ArchCheckResult {
             "avx512.vbroadcast.sd.pd.512" => "__builtin_ia32_vbroadcastsd_pd512",
             "avx512.vbroadcast.ss.512" => "__builtin_ia32_vbroadcastss512",
             "avx512.vbroadcast.ss.ps.512" => "__builtin_ia32_vbroadcastss_ps512",
+            "bmi.pdep.32" => "__builtin_ia32_pdep_si",
+            "bmi.pdep.64" => "__builtin_ia32_pdep_di",
+            "bmi.pext.32" => "__builtin_ia32_pext_si",
+            "bmi.pext.64" => "__builtin_ia32_pext_di",
             "fma.mask.vfmadd.pd.512" => "__builtin_ia32_vfmaddpd512_mask",
             "fma.mask.vfmadd.ps.512" => "__builtin_ia32_vfmaddps512_mask",
             "fma.mask.vfmaddsub.pd.512" => "__builtin_ia32_vfmaddsubpd512_mask",

tools/generate_intrinsics.py

@@ -84,6 +84,10 @@ def update_intrinsics(llvm_path):
     # This speeds up the comparison, and makes our code considerably smaller.
     # Since all intrinsic names start with "llvm.", we skip that prefix.
     print("Updating content of `{}`...".format(output_file))
+    indent4 = "    "
+    indent8 = indent4 + indent4
+    indent12 = indent8 + indent4
+    indent16 = indent12 + indent4
     with open(output_file, "w", encoding="utf8") as out:
         out.write("""// File generated by `rustc_codegen_gcc/tools/generate_intrinsics.py`
 // DO NOT EDIT IT!
@@ -95,33 +99,35 @@ def update_intrinsics(llvm_path):
     if let ArchCheckResult::Ok(res) = old_arch_res {
         return res;
     }
-match arch {""")
+    match arch {
+""")
         for arch in archs:
             if len(intrinsics[arch]) == 0:
                 continue
             attribute = "#[expect(non_snake_case)]" if arch[0].isupper() else ""
-            out.write("\"{}\" => {{ {} fn {}(name: &str,full_name:&str) -> &'static str {{ match name {{".format(arch, attribute, arch))
+            out.write(f"""{indent4}"{arch}" => {{
+{indent8}{attribute} fn {arch}(name: &str,full_name:&str) -> &'static str {{
+{indent12}match name {{""")
             intrinsics[arch].sort(key=lambda x: (x[0], x[1]))
-            out.write('    // {}\n'.format(arch))
+            out.write(f'{indent16}// {arch}\n')
             for entry in intrinsics[arch]:
                 llvm_name = entry[0].removeprefix("llvm.");
                 llvm_name = llvm_name.removeprefix(arch);
                 llvm_name = llvm_name.removeprefix(".");
                 if "_round_mask" in entry[1]:
-                    out.write('    // [INVALID CONVERSION]: "{}" => "{}",\n'.format(llvm_name, entry[1]))
+                    out.write(f'{indent16}// [INVALID CONVERSION]: "{llvm_name}" => "{entry[1]}",\n')
                 else:
-                    out.write('    "{}" => "{}",\n'.format(llvm_name, entry[1]))
-            out.write('    _ => unimplemented!("***** unsupported LLVM intrinsic {full_name}"),\n')
-            out.write("}} }} {}(name,full_name) }}\n,".format(arch))
-        out.write("""    _ => {
-        match old_arch_res {
-            ArchCheckResult::UnknownIntrinsic => unimplemented!("***** unsupported LLVM intrinsic {full_name}"),
-            ArchCheckResult::UnknownArch => unimplemented!("***** unsupported LLVM architecture {arch}, intrinsic: {full_name}"),
-            ArchCheckResult::Ok(_) => unreachable!(),
-        }
-    }""")
+                    out.write(f'{indent16}"{llvm_name}" => "{entry[1]}",\n')
+            out.write(f'{indent16}_ => unimplemented!("***** unsupported LLVM intrinsic {{full_name}}"),\n')
+            out.write(f"{indent16}}}\n{indent12}}}\n{indent8}{arch}(name,full_name)\n{indent8}}}\n,")
+        out.write(f"""{indent4}_ => {{
+{indent8}match old_arch_res {{
+{indent8}ArchCheckResult::UnknownIntrinsic => unimplemented!("***** unsupported LLVM intrinsic {{full_name}}"),
+{indent8}ArchCheckResult::UnknownArch => unimplemented!("***** unsupported LLVM architecture {{arch}}, intrinsic: {{full_name}}"),
+{indent8}ArchCheckResult::Ok(_) => unreachable!(),
+{indent4}}}
+}}""")
         out.write("}\n}")
-    subprocess.call(["rustfmt", output_file])
     print("Done!")