arm32/itx: Trim dav1d_inv_wht4_1d_c, saves 68 bytes

Author: Nathan E. Egge

src/itx_1d.c

@@ -1016,7 +1016,10 @@ void dav1d_inv_identity32_1d_c(int32_t *const c, const ptrdiff_t stride,
         c[stride * i] *= 4;
 }
 
-#if !(HAVE_ASM && TRIM_DSP_FUNCTIONS && ARCH_AARCH64)
+#if !(HAVE_ASM && TRIM_DSP_FUNCTIONS && ( \
+  ARCH_AARCH64 || \
+  (ARCH_ARM && (defined(__ARM_NEON) || defined(__APPLE__) || defined(_WIN32))) \
+))
 void dav1d_inv_wht4_1d_c(int32_t *const c, const ptrdiff_t stride) {
     assert(stride > 0);
     const int in0 = c[0 * stride], in1 = c[1 * stride];

src/itx_tmpl.c

@@ -159,7 +159,10 @@ inv_txfm_fn64(64, 16, 2)
 inv_txfm_fn64(64, 32, 1)
 inv_txfm_fn64(64, 64, 2)
 
-#if !(HAVE_ASM && TRIM_DSP_FUNCTIONS && ARCH_AARCH64)
+#if !(HAVE_ASM && TRIM_DSP_FUNCTIONS && ( \
+  ARCH_AARCH64 || \
+  (ARCH_ARM && (defined(__ARM_NEON) || defined(__APPLE__) || defined(_WIN32))) \
+))
 static void inv_txfm_add_wht_wht_4x4_c(pixel *dst, const ptrdiff_t stride,
                                        coef *const coeff, const int eob
                                        HIGHBD_DECL_SUFFIX)
@@ -234,7 +237,10 @@ COLD void bitfn(dav1d_itx_dsp_init)(Dav1dInvTxfmDSPContext *const c, int bpc) {
     c->itxfm_add[pfx##TX_##w##X##h][V_ADST] = \
         inv_txfm_add_identity_adst_##w##x##h##_c; \
 
-#if !(HAVE_ASM && TRIM_DSP_FUNCTIONS && ARCH_AARCH64)
+#if !(HAVE_ASM && TRIM_DSP_FUNCTIONS && ( \
+  ARCH_AARCH64 || \
+  (ARCH_ARM && (defined(__ARM_NEON) || defined(__APPLE__) || defined(_WIN32))) \
+))
     c->itxfm_add[TX_4X4][WHT_WHT] = inv_txfm_add_wht_wht_4x4_c;
 #endif
     assign_itx_all_fn84( 4,  4, );