S390 [liftoff]: Implement simd integer Q-format rounding mul

Change-Id: I0d6258cae0e75b6bd4916cb4c559161f7eac2170
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3465739Reviewed-by: Junliang Yan <junyan@redhat.com>
Commit-Queue: Milad Farazmand <mfarazma@redhat.com>
Cr-Commit-Position: refs/heads/main@{#79128}

src/codegen/s390/macro-assembler-s390.cc

@@ -6015,6 +6015,29 @@ void TurboAssembler::I32x4DotI16x8S(Simd128Register dst, Simd128Register src1,
   va(dst, scratch, dst, Condition(0), Condition(0), Condition(2));
 }
 
+#define Q15_MUL_ROAUND(accumulator, src1, src2, const_val, scratch, unpack) \
+  unpack(scratch, src1, Condition(0), Condition(0), Condition(1));          \
+  unpack(accumulator, src2, Condition(0), Condition(0), Condition(1));      \
+  vml(accumulator, scratch, accumulator, Condition(0), Condition(0),        \
+      Condition(2));                                                        \
+  va(accumulator, accumulator, const_val, Condition(0), Condition(0),       \
+     Condition(2));                                                         \
+  vrepi(scratch, Operand(15), Condition(2));                                \
+  vesrav(accumulator, accumulator, scratch, Condition(0), Condition(0),     \
+         Condition(2));
+void TurboAssembler::I16x8Q15MulRSatS(Simd128Register dst, Simd128Register src1,
+                                      Simd128Register src2,
+                                      Simd128Register scratch1,
+                                      Simd128Register scratch2,
+                                      Simd128Register scratch3) {
+  DCHECK(!AreAliased(src1, src2, scratch1, scratch2, scratch3));
+  vrepi(scratch1, Operand(0x4000), Condition(2));
+  Q15_MUL_ROAUND(scratch2, src1, src2, scratch1, scratch3, vupl)
+  Q15_MUL_ROAUND(dst, src1, src2, scratch1, scratch3, vuph)
+  vpks(dst, dst, scratch2, Condition(0), Condition(2));
+}
+#undef Q15_MUL_ROAUND
+
 // Vector LE Load and Transform instructions.
 #ifdef V8_TARGET_BIG_ENDIAN
 #define IS_BIG_ENDIAN true

src/codegen/s390/macro-assembler-s390.h

@@ -1155,6 +1155,9 @@ class V8_EXPORT_PRIVATE TurboAssembler : public TurboAssemblerBase {
                     Simd128Register scratch3);
   void I32x4DotI16x8S(Simd128Register dst, Simd128Register src1,
                       Simd128Register src2, Simd128Register scratch);
+  void I16x8Q15MulRSatS(Simd128Register dst, Simd128Register src1,
+                        Simd128Register src2, Simd128Register scratch1,
+                        Simd128Register scratch2, Simd128Register scratch3);
   void S128Select(Simd128Register dst, Simd128Register src1,
                   Simd128Register src2, Simd128Register mask);
 

src/compiler/backend/s390/code-generator-s390.cc

@@ -2975,32 +2975,13 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
                         i.InputSimd128Register(1), kScratchDoubleReg);
       break;
     }
-#define Q15_MUL_ROAUND(accumulator, unpack)                                   \
-  __ unpack(tempFPReg1, src0, Condition(0), Condition(0), Condition(1));      \
-  __ unpack(accumulator, src1, Condition(0), Condition(0), Condition(1));     \
-  __ vml(accumulator, tempFPReg1, accumulator, Condition(0), Condition(0),    \
-         Condition(2));                                                       \
-  __ va(accumulator, accumulator, tempFPReg2, Condition(0), Condition(0),     \
-        Condition(2));                                                        \
-  __ vrepi(tempFPReg1, Operand(15), Condition(2));                            \
-  __ vesrav(accumulator, accumulator, tempFPReg1, Condition(0), Condition(0), \
-            Condition(2));
     case kS390_I16x8Q15MulRSatS: {
-      Simd128Register dst = i.OutputSimd128Register();
-      Simd128Register src0 = i.InputSimd128Register(0);
-      Simd128Register src1 = i.InputSimd128Register(1);
-      Simd128Register tempFPReg1 = i.ToSimd128Register(instr->TempAt(0));
-      Simd128Register tempFPReg2 = i.ToSimd128Register(instr->TempAt(1));
-      DCHECK_NE(src1, tempFPReg1);
-      DCHECK_NE(src0, tempFPReg2);
-      DCHECK_NE(src1, tempFPReg2);
-      __ vrepi(tempFPReg2, Operand(0x4000), Condition(2));
-      Q15_MUL_ROAUND(kScratchDoubleReg, vupl)
-      Q15_MUL_ROAUND(dst, vuph)
-      __ vpks(dst, dst, kScratchDoubleReg, Condition(0), Condition(2));
-      break;
-    }
-#undef Q15_MUL_ROAUND
+      __ I16x8Q15MulRSatS(i.OutputSimd128Register(), i.InputSimd128Register(0),
+                          i.InputSimd128Register(1), kScratchDoubleReg,
+                          i.ToSimd128Register(instr->TempAt(0)),
+                          i.ToSimd128Register(instr->TempAt(1)));
+      break;
+    }
     case kS390_I8x16Popcnt: {
       __ I8x16Popcnt(i.OutputSimd128Register(), i.InputSimd128Register(0));
       break;

src/wasm/baseline/s390/liftoff-assembler-s390.h

@@ -2639,7 +2639,16 @@ void LiftoffAssembler::emit_i16x8_bitmask(LiftoffRegister dst,
 void LiftoffAssembler::emit_i16x8_q15mulr_sat_s(LiftoffRegister dst,
                                                 LiftoffRegister src1,
                                                 LiftoffRegister src2) {
-  bailout(kSimd, "i16x8_q15mulr_sat_s");
+  Simd128Register s1 = src1.fp();
+  Simd128Register s2 = src2.fp();
+  Simd128Register dest = dst.fp();
+  // Make sure temp registers are unique.
+  Simd128Register temp1 =
+      GetUnusedRegister(kFpReg, LiftoffRegList::ForRegs(dest, s1, s2)).fp();
+  Simd128Register temp2 =
+      GetUnusedRegister(kFpReg, LiftoffRegList::ForRegs(dest, s1, s2, temp1))
+          .fp();
+  I16x8Q15MulRSatS(dest, s1, s2, kScratchDoubleReg, temp1, temp2);
 }
 
 void LiftoffAssembler::emit_i8x16_shuffle(LiftoffRegister dst,