@@ -199,6 +199,11 @@
* Mathematical operations with more than one definition require
* special handling
*/
+#define fGEN_TCG_A5_ACS(SHORTCODE) \
+ do { \
+ gen_helper_vacsh_pred(PeV, cpu_env, RxxV, RssV, RttV); \
+ gen_helper_vacsh_val(RxxV, cpu_env, RxxV, RssV, RttV); \
+ } while (0)
/*
* Approximate reciprocal
@@ -26,6 +26,8 @@ DEF_HELPER_2(commit_store, void, env, int)
DEF_HELPER_FLAGS_4(fcircadd, TCG_CALL_NO_RWG_SE, s32, s32, s32, s32, s32)
DEF_HELPER_3(sfrecipa, i64, env, f32, f32)
DEF_HELPER_2(sfinvsqrta, i64, env, f32)
+DEF_HELPER_4(vacsh_val, s64, env, s64, s64, s64)
+DEF_HELPER_FLAGS_4(vacsh_pred, TCG_CALL_NO_RWG_SE, s32, env, s64, s64, s64)
/* Floating point */
DEF_HELPER_2(conv_sf2df, f64, env, f32)
@@ -347,6 +347,39 @@ uint64_t HELPER(sfinvsqrta)(CPUHexagonState *env, float32 RsV)
return ((uint64_t)RdV << 32) | PeV;
}
+int64_t HELPER(vacsh_val)(CPUHexagonState *env,
+ int64_t RxxV, int64_t RssV, int64_t RttV)
+{
+ for (int i = 0; i < 4; i++) {
+ int xv = sextract64(RxxV, i * 16, 16);
+ int sv = sextract64(RssV, i * 16, 16);
+ int tv = sextract64(RttV, i * 16, 16);
+ int max;
+ xv = xv + tv;
+ sv = sv - tv;
+ max = xv > sv ? xv : sv;
+ /* Note that fSATH can set the OVF bit in usr */
+ RxxV = deposit64(RxxV, i * 16, 16, fSATH(max));
+ }
+ return RxxV;
+}
+
+int32_t HELPER(vacsh_pred)(CPUHexagonState *env,
+ int64_t RxxV, int64_t RssV, int64_t RttV)
+{
+ int32_t PeV = 0;
+ for (int i = 0; i < 4; i++) {
+ int xv = sextract64(RxxV, i * 16, 16);
+ int sv = sextract64(RssV, i * 16, 16);
+ int tv = sextract64(RttV, i * 16, 16);
+ xv = xv + tv;
+ sv = sv - tv;
+ PeV = deposit32(PeV, i * 2, 1, (xv > sv));
+ PeV = deposit32(PeV, i * 2 + 1, 1, (xv > sv));
+ }
+ return PeV;
+}
+
/*
* mem_noshuf
* Section 5.5 of the Hexagon V67 Programmer's Reference Manual
@@ -45,8 +45,41 @@ static int sfinvsqrta(int Rs, int *pred_result)
return result;
}
+static long long vacsh(long long Rxx, long long Rss, long long Rtt,
+ int *pred_result, int *ovf_result)
+{
+ long long result = Rxx;
+ int predval;
+ int usr;
+
+ /*
+ * This instruction can set bit 0 (OVF/overflow) in usr
+ * Clear the bit first, then return that bit to the caller
+ */
+ asm volatile("r2 = usr\n\t"
+ "r2 = clrbit(r2, #0)\n\t" /* clear overflow bit */
+ "usr = r2\n\t"
+ "%0,p0 = vacsh(%3, %4)\n\t"
+ "%1 = p0\n\t"
+ "%2 = usr\n\t"
+ : "+r"(result), "=r"(predval), "=r"(usr)
+ : "r"(Rss), "r"(Rtt)
+ : "r2", "p0", "usr");
+ *pred_result = predval;
+ *ovf_result = (usr & 1);
+ return result;
+}
+
int err;
+static void check_ll(long long val, long long expect)
+{
+ if (val != expect) {
+ printf("ERROR: 0x%016llx != 0x%016llx\n", val, expect);
+ err++;
+ }
+}
+
static void check(int val, int expect)
{
if (val != expect) {
@@ -87,10 +120,46 @@ static void test_sfinvsqrta()
check_p(pred_result, 0x0);
}
+static void test_vacsh()
+{
+ long long res64;
+ int pred_result;
+ int ovf_result;
+
+ res64 = vacsh(0x0004000300020001LL,
+ 0x0001000200030004LL,
+ 0x0000000000000000LL, &pred_result, &ovf_result);
+ check_ll(res64, 0x0004000300030004LL);
+ check_p(pred_result, 0xf0);
+ check(ovf_result, 0);
+
+ res64 = vacsh(0x0004000300020001LL,
+ 0x0001000200030004LL,
+ 0x000affff000d0000LL, &pred_result, &ovf_result);
+ check_ll(res64, 0x000e0003000f0004LL);
+ check_p(pred_result, 0xcc);
+ check(ovf_result, 0);
+
+ res64 = vacsh(0x00047fff00020001LL,
+ 0x00017fff00030004LL,
+ 0x000a0fff000d0000LL, &pred_result, &ovf_result);
+ check_ll(res64, 0x000e7fff000f0004LL);
+ check_p(pred_result, 0xfc);
+ check(ovf_result, 1);
+
+ res64 = vacsh(0x0004000300020001LL,
+ 0x0001000200030009LL,
+ 0x000affff000d0001LL, &pred_result, &ovf_result);
+ check_ll(res64, 0x000e0003000f0008LL);
+ check_p(pred_result, 0xcc);
+ check(ovf_result, 0);
+}
+
int main()
{
test_sfrecipa();
test_sfinvsqrta();
+ test_vacsh();
puts(err ? "FAIL" : "PASS");
return err;
@@ -1240,6 +1240,25 @@ MINMAX(uw,WORD,UWORD,2)
#undef VMINORMAX3
+Q6INSN(A5_ACS,"Rxx32,Pe4=vacsh(Rss32,Rtt32)",ATTRIBS(),
+"Add Compare and Select elements of two vectors, record the maximums and the decisions ",
+{
+ fHIDE(int i;)
+ fHIDE(int xv;)
+ fHIDE(int sv;)
+ fHIDE(int tv;)
+ for (i = 0; i < 4; i++) {
+ xv = (int) fGETHALF(i,RxxV);
+ sv = (int) fGETHALF(i,RssV);
+ tv = (int) fGETHALF(i,RttV);
+ xv = xv + tv; //assumes 17bit datapath
+ sv = sv - tv; //assumes 17bit datapath
+ fSETBIT(i*2, PeV, (xv > sv));
+ fSETBIT(i*2+1,PeV, (xv > sv));
+ fSETHALF(i, RxxV, fSATH(fMAX(xv,sv)));
+ }
+})
+
/**********************************************/
/* Vector Min/Max */
/**********************************************/
@@ -1017,6 +1017,7 @@ MPY_ENC(M7_dcmpyiwc_acc, "1010","xxxxx","1","0","1","0","10")
+MPY_ENC(A5_ACS, "1010","xxxxx","0","1","0","1","ee")
/*
*/