math/mpx-mul4-{arm-neon,arm64-simd}.S, etc.: Add ARM versions of `mul4'.

[catacomb] / math / mpx.c

diff --git a/math/mpx.c b/math/mpx.c
index 4294845..d3d0a04 100644 (file)
--- a/math/mpx.c
+++ b/math/mpx.c
@@ -818,7 +818,7 @@ void mpx_usub(mpw *dv, mpw *dvl, const mpw *av, const mpw *avl,
 
 void mpx_usubn(mpw *dv, mpw *dvl, mpw n) { MPX_USUBN(dv, dvl, n); }
 
-/* --- @mpx_uaddnlsl@ --- *
+/* --- @mpx_usubnlsl@ --- *
  *
  * Arguments:  @mpw *dv, *dvl@ = destination and first argument vector
  *             @mpw a@ = second argument
@@ -931,6 +931,14 @@ static void simple_umul(mpw *dv, mpw *dvl, const mpw *av, const mpw *avl,
   MAYBE_UMUL4(amd64_avx)
 #endif
 
+#if CPUFAM_ARMEL
+  MAYBE_UMUL4(arm_neon)
+#endif
+
+#if CPUFAM_ARM64
+  MAYBE_UMUL4(arm64_simd)
+#endif
+
 static mpx_umul__functype *pick_umul(void)
 {
 #if CPUFAM_X86
@@ -945,6 +953,13 @@ static mpx_umul__functype *pick_umul(void)
   DISPATCH_PICK_COND(mpx_umul, maybe_umul4_amd64_sse2,
                     cpu_feature_p(CPUFEAT_X86_SSE2));
 #endif
+#if CPUFAM_ARMEL
+  DISPATCH_PICK_COND(mpx_umul, maybe_umul4_arm_neon,
+                    cpu_feature_p(CPUFEAT_ARM_NEON));
+#endif
+#if CPUFAM_ARM64
+  DISPATCH_PICK_COND(mpx_umul, maybe_umul4_arm64_simd, 1);
+#endif
   DISPATCH_PICK_FALLBACK(mpx_umul, simple_umul);
 }
 
@@ -1274,9 +1289,14 @@ mpw mpx_udivn(mpw *qv, mpw *qvl, const mpw *rv, const mpw *rvl, mpw d)
 
 #include <mLib/alloc.h>
 #include <mLib/dstr.h>
+#include <mLib/macros.h>
 #include <mLib/quis.h>
 #include <mLib/testrig.h>
 
+#ifdef ENABLE_ASM_DEBUG
+#  include "regdump.h"
+#endif
+
 #include "mpscan.h"
 
 #define ALLOC(v, vl, sz) do {                                          \
@@ -1366,7 +1386,7 @@ static int loadstore(dstr *v)
       ok = 0;
     MPX_OCTETS(oct, m, ml);
     mpx_storel(m, ml, d.buf, d.sz);
-    if (memcmp(d.buf, v->buf, oct) != 0) {
+    if (MEMCMP(d.buf, !=, v->buf, oct)) {
       dumpbits("\n*** storel failed", d.buf, d.sz);
       ok = 0;
     }
@@ -1376,7 +1396,7 @@ static int loadstore(dstr *v)
       ok = 0;
     MPX_OCTETS(oct, m, ml);
     mpx_storeb(m, ml, d.buf, d.sz);
-    if (memcmp(d.buf + d.sz - oct, v->buf + v->len - oct, oct) != 0) {
+    if (MEMCMP(d.buf + d.sz - oct, !=, v->buf + v->len - oct, oct)) {
       dumpbits("\n*** storeb failed", d.buf, d.sz);
       ok = 0;
     }
@@ -1410,14 +1430,14 @@ static int twocl(dstr *v)
 
     mpx_loadl(m, ml0, v[0].buf, v[0].len);
     mpx_storel2cn(m, ml0, d.buf, v[1].len);
-    if (memcmp(d.buf, v[1].buf, v[1].len)) {
+    if (MEMCMP(d.buf, !=, v[1].buf, v[1].len)) {
       dumpbits("\n*** storel2cn failed", d.buf, v[1].len);
       ok = 0;
     }
 
     mpx_loadl2cn(m, ml1, v[1].buf, v[1].len);
     mpx_storel(m, ml1, d.buf, v[0].len);
-    if (memcmp(d.buf, v[0].buf, v[0].len)) {
+    if (MEMCMP(d.buf, !=, v[0].buf, v[0].len)) {
       dumpbits("\n*** loadl2cn failed", d.buf, v[0].len);
       ok = 0;
     }
@@ -1454,14 +1474,14 @@ static int twocb(dstr *v)
 
     mpx_loadb(m, ml0, v[0].buf, v[0].len);
     mpx_storeb2cn(m, ml0, d.buf, v[1].len);
-    if (memcmp(d.buf, v[1].buf, v[1].len)) {
+    if (MEMCMP(d.buf, !=, v[1].buf, v[1].len)) {
       dumpbits("\n*** storeb2cn failed", d.buf, v[1].len);
       ok = 0;
     }
 
     mpx_loadb2cn(m, ml1, v[1].buf, v[1].len);
     mpx_storeb(m, ml1, d.buf, v[0].len);
-    if (memcmp(d.buf, v[0].buf, v[0].len)) {
+    if (MEMCMP(d.buf, !=, v[0].buf, v[0].len)) {
       dumpbits("\n*** loadb2cn failed", d.buf, v[0].len);
       ok = 0;
     }
@@ -1708,6 +1728,9 @@ static test_chunk defs[] = {
 
 int main(int argc, char *argv[])
 {
+#ifdef ENABLE_ASM_DEBUG
+  regdump_init();
+#endif
   test_run(argc, argv, defs, SRCDIR"/t/mpx");
   return (0);
 }