term: move the term core to subprojects/term

The core now builds, tests, and benchmarks alone, as a Meson subproject
with no dependency on the main tree.

- include/term/ holds term.h and the header-only unicode/. src/ holds
  the core and its private headers. The main tree gets only
  include/term through term_dep, so it cannot reach a private header.
- The term tests, the term benchmarks, the Kitty parser suite, and
  tools/amalgamate.py and tools/generate_unicode.py move with it. The
  tests run as term:<name>.
- The main build passes its -march (arch_flag) and b_ndebug to the
  subproject. A directory configured before this commit needs
  meson configure -Dterm:arch_flag=-march=... once.
- The term tests compile with -UNDEBUG. They do work inside assert.
- The subproject exports term_test_dep with term_api.h, for the main
  tree's render, thumbnail, color, and selection tests.
- The Kitty bridge used struct vector2i and has not compiled since the
  public API redesign. It builds again. 2 of its 14 tests fail: REP
  after a tab, and a cursor report after a resize.
- dirgraph.py drops the term and unicode layers. Its self-test uses pty.
- The symbols test ignores the __odr_asan symbols that ASan adds.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

1 files changed, 117 insertions, 0 deletions

diff --git a/src/simd.h b/src/simd.h
new file mode 100644
index 0000000..0238500
--- /dev/null
+++ b/src/simd.h
@@ -0,0 +1,117 @@
+#ifndef TERM_SIMD_H
+#define TERM_SIMD_H
+
+#include "compat.h"
+
+#include <stdalign.h>
+#include <stdbool.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#if defined(__x86_64__) || defined(_M_X64)
+#include <immintrin.h>
+#endif
+#if defined(__aarch64__)
+#include <arm_neon.h>
+#endif
+
+/*
+ * x86-64 kernels come in three levels: base (SSE2), AVX2 (x86-64-v3), and
+ * AVX-512 (x86-64-v4 with VBMI and VBMI2). GCC and Clang compile every
+ * level with target attributes, and term_isa_init() picks one at run time.
+ * Other compilers compile the levels that their flags enable.
+ *
+ * A kernel at a level is a function marked TERM_TARGET_AVX2 or
+ * TERM_TARGET_AVX512. It exists when TERM_KERNEL_AVX2 or TERM_KERNEL_AVX512
+ * is defined. Call it only when term_use_avx2() or term_use_avx512() is
+ * true. When the compiler flags already enable a level, these checks are
+ * constants and the kernels inline into their callers.
+ */
+#if defined(__AVX512F__) && defined(__AVX512BW__) && defined(__AVX512CD__) && \
+    defined(__AVX512DQ__) && defined(__AVX512VL__) && defined(__AVX512VBMI__) && \
+    defined(__AVX512VBMI2__) && defined(__AVX2__) && defined(__BMI2__)
+#define TERM_ISA_FLOOR 2
+#elif defined(__AVX2__) && defined(__BMI__) && defined(__BMI2__) && defined(__FMA__) && \
+    defined(__LZCNT__)
+#define TERM_ISA_FLOOR 1
+#else
+#define TERM_ISA_FLOOR 0
+#endif
+
+/* Clang 18 to 21 also need evex512 for 512-bit vectors. GCC and Clang 22
+ * reject the name. Apple Clang 17 is LLVM 19. */
+#if defined(__clang__) && ((defined(__apple_build_version__) && __clang_major__ >= 17) || \
+    (!defined(__apple_build_version__) && __clang_major__ >= 18 && __clang_major__ <= 21))
+#define TERM_EVEX512 ",evex512"
+#else
+#define TERM_EVEX512 ""
+#endif
+
+#if (defined(__x86_64__) || defined(_M_X64)) && defined(TERM_COMPILER_GNU)
+#define TERM_TARGET_AVX2 \
+    __attribute__((target("avx2,bmi,bmi2,fma,lzcnt,popcnt,movbe,f16c")))
+#define TERM_TARGET_AVX512 \
+    __attribute__((target("avx2,bmi,bmi2,fma,lzcnt,popcnt,movbe,f16c,avx512f,avx512bw," \
+                          "avx512cd,avx512dq,avx512vl,avx512vbmi,avx512vbmi2" TERM_EVEX512)))
+#define TERM_KERNEL_AVX2 1
+#define TERM_KERNEL_AVX512 1
+#define TERM_ISA_RUNTIME 1
+#else
+#define TERM_TARGET_AVX2
+#define TERM_TARGET_AVX512
+#if TERM_ISA_FLOOR >= 1
+#define TERM_KERNEL_AVX2 1
+#endif
+#if TERM_ISA_FLOOR >= 2
+#define TERM_KERNEL_AVX512 1
+#endif
+#endif
+
+enum { TERM_ISA_BASE, TERM_ISA_AVX2, TERM_ISA_AVX512 };
+
+/* The level in use. Zero, the base level, until term_isa_init() runs. It
+ * is a plain int, so the compiler can hoist the checks out of loops: it is
+ * written once, before the release that publishes it, and read only after
+ * that. */
+extern int term_isa_level;
+
+/* Pick the level once. Any thread can call it. term_new() calls it before
+ * the first terminal exists. */
+void term_isa_init(void);
+
+static inline bool term_use_avx2(void)
+{
+#if TERM_ISA_FLOOR >= 1
+    return true;
+#elif defined(TERM_ISA_RUNTIME)
+    return term_isa_level >= TERM_ISA_AVX2;
+#else
+    return false;
+#endif
+}
+
+static inline bool term_use_avx512(void)
+{
+#if TERM_ISA_FLOOR >= 2
+    return true;
+#elif defined(TERM_ISA_RUNTIME)
+    return term_isa_level >= TERM_ISA_AVX512;
+#else
+    return false;
+#endif
+}
+
+#if defined(__aarch64__)
+static inline uint16_t term_lane_bits(uint8x16_t m)
+{
+    const uint8x16_t idx = {1, 2, 4, 8, 16, 32, 64, 128,
+                            1, 2, 4, 8, 16, 32, 64, 128};
+    uint8x16_t s = vandq_u8(m, idx);
+    s = vpaddq_u8(s, s);
+    s = vpaddq_u8(s, s);
+    s = vpaddq_u8(s, s);
+    return (uint16_t)(vgetq_lane_u8(s, 0) | (vgetq_lane_u8(s, 1) << 8));
+}
+#endif
+
+#endif