8318158: RISC-V: implement roundD/roundF intrinsics

Co-authored-by: Vladimir Kempik <vkempik@openjdk.org>
Reviewed-by: luhenry, fyang, mli

Author: Olga Mikhaltsova

src/hotspot/cpu/riscv/macroAssembler_riscv.cpp

@@ -4481,6 +4481,57 @@ void MacroAssembler::zero_dcache_blocks(Register base, Register cnt, Register tm
   bge(cnt, tmp1, loop);
 }
 
+// java.lang.Math.round(float a)
+// Returns the closest int to the argument, with ties rounding to positive infinity.
+void MacroAssembler::java_round_float(Register dst, FloatRegister src, FloatRegister ftmp) {
+  // this instructions calling sequence provides performance improvement on all tested devices;
+  // don't change it without re-verification
+  Label done;
+  mv(t0, jint_cast(0.5f));
+  fmv_w_x(ftmp, t0);
+
+  // dst = 0 if NaN
+  feq_s(t0, src, src); // replacing fclass with feq as performance optimization
+  mv(dst, zr);
+  beqz(t0, done);
+
+  // dst = (src + 0.5f) rounded down towards negative infinity
+  //   Adding 0.5f to some floats exceeds the precision limits for a float and rounding takes place.
+  //   RDN is required for fadd_s, RNE gives incorrect results:
+  //     --------------------------------------------------------------------
+  //     fadd.s rne (src + 0.5f): src = 8388609.000000  ftmp = 8388610.000000
+  //     fcvt.w.s rdn: ftmp = 8388610.000000 dst = 8388610
+  //     --------------------------------------------------------------------
+  //     fadd.s rdn (src + 0.5f): src = 8388609.000000  ftmp = 8388609.000000
+  //     fcvt.w.s rdn: ftmp = 8388609.000000 dst = 8388609
+  //     --------------------------------------------------------------------
+  fadd_s(ftmp, src, ftmp, RoundingMode::rdn);
+  fcvt_w_s(dst, ftmp, RoundingMode::rdn);
+
+  bind(done);
+}
+
+// java.lang.Math.round(double a)
+// Returns the closest long to the argument, with ties rounding to positive infinity.
+void MacroAssembler::java_round_double(Register dst, FloatRegister src, FloatRegister ftmp) {
+  // this instructions calling sequence provides performance improvement on all tested devices;
+  // don't change it without re-verification
+  Label done;
+  mv(t0, julong_cast(0.5));
+  fmv_d_x(ftmp, t0);
+
+  // dst = 0 if NaN
+  feq_d(t0, src, src); // replacing fclass with feq as performance optimization
+  mv(dst, zr);
+  beqz(t0, done);
+
+  // dst = (src + 0.5) rounded down towards negative infinity
+  fadd_d(ftmp, src, ftmp, RoundingMode::rdn); // RDN is required here otherwise some inputs produce incorrect results
+  fcvt_l_d(dst, ftmp, RoundingMode::rdn);
+
+  bind(done);
+}
+
 #define FCVT_SAFE(FLOATCVT, FLOATSIG)                                                     \
 void MacroAssembler::FLOATCVT##_safe(Register dst, FloatRegister src, Register tmp) {     \
   Label done;                                                                             \

src/hotspot/cpu/riscv/macroAssembler_riscv.hpp

@@ -1265,6 +1265,9 @@ class MacroAssembler: public Assembler {
   void fcvt_w_d_safe(Register dst, FloatRegister src, Register tmp = t0);
   void fcvt_l_d_safe(Register dst, FloatRegister src, Register tmp = t0);
 
+  void java_round_float(Register dst, FloatRegister src, FloatRegister ftmp);
+  void java_round_double(Register dst, FloatRegister src, FloatRegister ftmp);
+
   // vector load/store unit-stride instructions
   void vlex_v(VectorRegister vd, Register base, Assembler::SEW sew, VectorMask vm = unmasked) {
     switch (sew) {

src/hotspot/cpu/riscv/riscv.ad

@@ -8417,6 +8417,34 @@ instruct convN2I(iRegINoSp dst, iRegN src)
   ins_pipe(ialu_reg);
 %}
 
+instruct round_double_reg(iRegLNoSp dst, fRegD src, fRegD ftmp) %{
+  match(Set dst (RoundD src));
+
+  ins_cost(XFER_COST + BRANCH_COST);
+  effect(TEMP ftmp);
+  format %{ "java_round_double $dst, $src\t#@round_double_reg" %}
+
+  ins_encode %{
+    __ java_round_double($dst$$Register, as_FloatRegister($src$$reg), as_FloatRegister($ftmp$$reg));
+  %}
+
+  ins_pipe(pipe_slow);
+%}
+
+instruct round_float_reg(iRegINoSp dst, fRegF src, fRegF ftmp) %{
+  match(Set dst (RoundF src));
+
+  ins_cost(XFER_COST + BRANCH_COST);
+  effect(TEMP ftmp);
+  format %{ "java_round_float $dst, $src\t#@round_float_reg" %}
+
+  ins_encode %{
+    __ java_round_float($dst$$Register, as_FloatRegister($src$$reg), as_FloatRegister($ftmp$$reg));
+  %}
+
+  ins_pipe(pipe_slow);
+%}
+
 // Convert oop pointer into compressed form
 instruct encodeHeapOop(iRegNNoSp dst, iRegP src) %{
   match(Set dst (EncodeP src));