@@ -32,4 +32,6 @@ int arch_sf_invsqrt_common(float32 *Rs, float32 *Rd, int *adjust,
extern const uint8_t recip_lookup_table[128];
+extern const uint8_t invsqrt_lookup_table[128];
+
#endif
@@ -216,6 +216,22 @@
tcg_temp_free_i64(tmp); \
} while (0)
+/*
+ * Approximation of the reciprocal square root
+ * r1,p0 = sfinvsqrta(r0)
+ *
+ * The helper packs the 2 32-bit results into a 64-bit value,
+ * so unpack them into the proper results.
+ */
+#define fGEN_TCG_F2_sfinvsqrta(SHORTCODE) \
+ do { \
+ TCGv_i64 tmp = tcg_temp_new_i64(); \
+ gen_helper_sfinvsqrta(tmp, cpu_env, RsV); \
+ tcg_gen_extrh_i64_i32(RdV, tmp); \
+ tcg_gen_extrl_i64_i32(PeV, tmp); \
+ tcg_temp_free_i64(tmp); \
+ } while (0)
+
/* Floating point */
#define fGEN_TCG_F2_conv_sf2df(SHORTCODE) \
gen_helper_conv_sf2df(RddV, cpu_env, RsV)
@@ -25,6 +25,7 @@ DEF_HELPER_FLAGS_3(debug_commit_end, TCG_CALL_NO_WG, void, env, int, int)
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)
/* Floating point */
DEF_HELPER_2(conv_sf2df, f64, env, f32)
@@ -247,7 +247,7 @@ int arch_sf_invsqrt_common(float32 *Rs, float32 *Rd, int *adjust,
int r_exp;
int ret = 0;
RsV = *Rs;
- if (float32_is_infinity(RsV)) {
+ if (float32_is_any_nan(RsV)) {
if (extract32(RsV, 22, 1) == 0) {
float_raise(float_flag_invalid, fp_status);
}
@@ -299,3 +299,22 @@ const uint8_t recip_lookup_table[128] = {
0x011, 0x00f, 0x00e, 0x00d, 0x00c, 0x00b, 0x00a, 0x009,
0x008, 0x007, 0x006, 0x005, 0x004, 0x003, 0x002, 0x000,
};
+
+const uint8_t invsqrt_lookup_table[128] = {
+ 0x069, 0x066, 0x063, 0x061, 0x05e, 0x05b, 0x059, 0x057,
+ 0x054, 0x052, 0x050, 0x04d, 0x04b, 0x049, 0x047, 0x045,
+ 0x043, 0x041, 0x03f, 0x03d, 0x03b, 0x039, 0x037, 0x036,
+ 0x034, 0x032, 0x030, 0x02f, 0x02d, 0x02c, 0x02a, 0x028,
+ 0x027, 0x025, 0x024, 0x022, 0x021, 0x01f, 0x01e, 0x01d,
+ 0x01b, 0x01a, 0x019, 0x017, 0x016, 0x015, 0x014, 0x012,
+ 0x011, 0x010, 0x00f, 0x00d, 0x00c, 0x00b, 0x00a, 0x009,
+ 0x008, 0x007, 0x006, 0x005, 0x004, 0x003, 0x002, 0x001,
+ 0x0fe, 0x0fa, 0x0f6, 0x0f3, 0x0ef, 0x0eb, 0x0e8, 0x0e4,
+ 0x0e1, 0x0de, 0x0db, 0x0d7, 0x0d4, 0x0d1, 0x0ce, 0x0cb,
+ 0x0c9, 0x0c6, 0x0c3, 0x0c0, 0x0be, 0x0bb, 0x0b8, 0x0b6,
+ 0x0b3, 0x0b1, 0x0af, 0x0ac, 0x0aa, 0x0a8, 0x0a5, 0x0a3,
+ 0x0a1, 0x09f, 0x09d, 0x09b, 0x099, 0x097, 0x095, 0x093,
+ 0x091, 0x08f, 0x08d, 0x08b, 0x089, 0x087, 0x086, 0x084,
+ 0x082, 0x080, 0x07f, 0x07d, 0x07b, 0x07a, 0x078, 0x077,
+ 0x075, 0x074, 0x072, 0x071, 0x06f, 0x06e, 0x06c, 0x06b,
+};
@@ -326,6 +326,27 @@ uint64_t HELPER(sfrecipa)(CPUHexagonState *env, float32 RsV, float32 RtV)
return ((uint64_t)RdV << 32) | PeV;
}
+uint64_t HELPER(sfinvsqrta)(CPUHexagonState *env, float32 RsV)
+{
+ int PeV = 0;
+ float32 RdV;
+ int idx;
+ int adjust;
+ int mant;
+ int exp;
+
+ arch_fpop_start(env);
+ if (arch_sf_invsqrt_common(&RsV, &RdV, &adjust, &env->fp_status)) {
+ PeV = adjust;
+ idx = (RsV >> 17) & 0x7f;
+ mant = (invsqrt_lookup_table[idx] << 15);
+ exp = SF_BIAS - ((float32_getexp(RsV) - SF_BIAS) >> 1) - 1;
+ RdV = build_float32(extract32(RsV, 31, 1), exp, mant);
+ }
+ arch_fpop_end(env);
+ return ((uint64_t)RdV << 32) | PeV;
+}
+
/*
* mem_noshuf
* Section 5.5 of the Hexagon V67 Programmer's Reference Manual
@@ -441,6 +441,20 @@ static void check_canonical_NaN(void)
check_fpstatus(usr, 0);
}
+static void check_invsqrta(void)
+{
+ int result;
+ int predval;
+
+ asm volatile("%0,p0 = sfinvsqrta(%2)\n\t"
+ "%1 = p0\n\t"
+ : "+r"(result), "=r"(predval)
+ : "r"(0x7f800000)
+ : "p0");
+ check32(result, 0xff800000);
+ check32(predval, 0x0);
+}
+
static void check_float2int_convs()
{
int res32;
@@ -590,6 +604,7 @@ int main()
check_dfminmax();
check_recip_exception();
check_canonical_NaN();
+ check_invsqrta();
check_float2int_convs();
puts(err ? "FAIL" : "PASS");
@@ -31,6 +31,20 @@ static int sfrecipa(int Rs, int Rt, int *pred_result)
return result;
}
+static int sfinvsqrta(int Rs, int *pred_result)
+{
+ int result;
+ int predval;
+
+ asm volatile("%0,p0 = sfinvsqrta(%2)\n\t"
+ "%1 = p0\n\t"
+ : "+r"(result), "=r"(predval)
+ : "r"(Rs)
+ : "p0");
+ *pred_result = predval;
+ return result;
+}
+
int err;
static void check(int val, int expect)
@@ -59,9 +73,24 @@ static void test_sfrecipa()
check_p(pred_result, 0x00);
}
+static void test_sfinvsqrta()
+{
+ int res;
+ int pred_result;
+
+ res = sfinvsqrta(0x04030201, &pred_result);
+ check(res, 0x4d330000);
+ check_p(pred_result, 0xe0);
+
+ res = sfinvsqrta(0x0, &pred_result);
+ check(res, 0x3f800000);
+ check_p(pred_result, 0x0);
+}
+
int main()
{
test_sfrecipa();
+ test_sfinvsqrta();
puts(err ? "FAIL" : "PASS");
return err;
@@ -1642,6 +1642,7 @@ SH2_RR_ENC(F2_conv_sf2w, "1011","100","-","000","ddddd")
SH2_RR_ENC(F2_conv_sf2uw_chop, "1011","011","-","001","ddddd")
SH2_RR_ENC(F2_conv_sf2w_chop, "1011","100","-","001","ddddd")
SH2_RR_ENC(F2_sffixupr, "1011","101","-","000","ddddd")
+SH2_RR_ENC(F2_sfinvsqrta, "1011","111","-","0ee","ddddd")
DEF_FIELDROW_DESC32(ICLASS_S2op" 1100 -------- PP------ --------","[#12] Rd=(Rs,#u6)")
@@ -178,6 +178,22 @@ Q6INSN(F2_sffixupd,"Rd32=sffixupd(Rs32,Rt32)",ATTRIBS(),
RdV = RtV;
})
+Q6INSN(F2_sfinvsqrta,"Rd32,Pe4=sfinvsqrta(Rs32)",ATTRIBS(),
+"Reciprocal Square Root Approximation",
+{
+ fHIDE(int idx;)
+ fHIDE(int adjust;)
+ fHIDE(int mant;)
+ fHIDE(int exp;)
+ if (fSF_INVSQRT_COMMON(RsV,RdV,adjust)) {
+ PeV = adjust;
+ idx = (RsV >> 17) & 0x7f;
+ mant = (fSF_INVSQRT_LOOKUP(idx) << 15);
+ exp = fSF_BIAS() - ((fSF_GETEXP(RsV) - fSF_BIAS()) >> 1) - 1;
+ RdV = fMAKESF(fGETBIT(31,RsV),exp,mant);
+ }
+})
+
Q6INSN(F2_sffixupr,"Rd32=sffixupr(Rs32)",ATTRIBS(),
"Fix Up Radicand",
{