[wasm-simd][liftoff][arm][arm64] Implement pmin pmax

Implement f32x4.pmin, f32x4.pmax, f64x2.pmin, and f64x2.pmax for arm and
arm64.

Bug: v8:10904
Change-Id: Ife8b832dfc21850c2c292b0d6446b7dd3f6d9849
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2415109Reviewed-by: Bill Budge <bbudge@chromium.org>
Commit-Queue: Zhi An Ng <zhin@chromium.org>
Cr-Commit-Position: refs/heads/master@{#69975}

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

@@ -2407,12 +2407,34 @@ void LiftoffAssembler::emit_f64x2_max(LiftoffRegister dst, LiftoffRegister lhs,
 
 void LiftoffAssembler::emit_f64x2_pmin(LiftoffRegister dst, LiftoffRegister lhs,
                                        LiftoffRegister rhs) {
-  bailout(kSimd, "pmin unimplemented");
+  QwNeonRegister dest = liftoff::GetSimd128Register(dst);
+  QwNeonRegister left = liftoff::GetSimd128Register(lhs);
+  QwNeonRegister right = liftoff::GetSimd128Register(rhs);
+
+  if (dst != rhs) {
+    vmov(dest, left);
+  }
+
+  VFPCompareAndSetFlags(right.low(), left.low());
+  vmov(dest.low(), right.low(), mi);
+  VFPCompareAndSetFlags(right.high(), left.high());
+  vmov(dest.high(), right.high(), mi);
 }
 
 void LiftoffAssembler::emit_f64x2_pmax(LiftoffRegister dst, LiftoffRegister lhs,
                                        LiftoffRegister rhs) {
-  bailout(kSimd, "pmax unimplemented");
+  QwNeonRegister dest = liftoff::GetSimd128Register(dst);
+  QwNeonRegister left = liftoff::GetSimd128Register(lhs);
+  QwNeonRegister right = liftoff::GetSimd128Register(rhs);
+
+  if (dst != rhs) {
+    vmov(dest, left);
+  }
+
+  VFPCompareAndSetFlags(right.low(), left.low());
+  vmov(dest.low(), right.low(), gt);
+  VFPCompareAndSetFlags(right.high(), left.high());
+  vmov(dest.high(), right.high(), gt);
 }
 
 void LiftoffAssembler::emit_f32x4_splat(LiftoffRegister dst,
@@ -2536,12 +2558,40 @@ void LiftoffAssembler::emit_f32x4_max(LiftoffRegister dst, LiftoffRegister lhs,
 
 void LiftoffAssembler::emit_f32x4_pmin(LiftoffRegister dst, LiftoffRegister lhs,
                                        LiftoffRegister rhs) {
-  bailout(kSimd, "pmin unimplemented");
+  UseScratchRegisterScope temps(this);
+
+  QwNeonRegister tmp = liftoff::GetSimd128Register(dst);
+  if (dst == lhs || dst == rhs) {
+    tmp = temps.AcquireQ();
+  }
+
+  QwNeonRegister left = liftoff::GetSimd128Register(lhs);
+  QwNeonRegister right = liftoff::GetSimd128Register(rhs);
+  vcgt(tmp, left, right);
+  vbsl(tmp, right, left);
+
+  if (dst == lhs || dst == rhs) {
+    vmov(liftoff::GetSimd128Register(dst), tmp);
+  }
 }
 
 void LiftoffAssembler::emit_f32x4_pmax(LiftoffRegister dst, LiftoffRegister lhs,
                                        LiftoffRegister rhs) {
-  bailout(kSimd, "pmax unimplemented");
+  UseScratchRegisterScope temps(this);
+
+  QwNeonRegister tmp = liftoff::GetSimd128Register(dst);
+  if (dst == lhs || dst == rhs) {
+    tmp = temps.AcquireQ();
+  }
+
+  QwNeonRegister left = liftoff::GetSimd128Register(lhs);
+  QwNeonRegister right = liftoff::GetSimd128Register(rhs);
+  vcgt(tmp, right, left);
+  vbsl(tmp, right, left);
+
+  if (dst == lhs || dst == rhs) {
+    vmov(liftoff::GetSimd128Register(dst), tmp);
+  }
 }
 
 void LiftoffAssembler::emit_i64x2_splat(LiftoffRegister dst,

src/wasm/baseline/arm64/liftoff-assembler-arm64.h

@@ -1620,12 +1620,36 @@ void LiftoffAssembler::emit_f64x2_max(LiftoffRegister dst, LiftoffRegister lhs,
 
 void LiftoffAssembler::emit_f64x2_pmin(LiftoffRegister dst, LiftoffRegister lhs,
                                        LiftoffRegister rhs) {
-  bailout(kSimd, "pmin unimplemented");
+  UseScratchRegisterScope temps(this);
+
+  VRegister tmp = dst.fp();
+  if (dst == lhs || dst == rhs) {
+    tmp = temps.AcquireV(kFormat2D);
+  }
+
+  Fcmgt(tmp.V2D(), lhs.fp().V2D(), rhs.fp().V2D());
+  Bsl(tmp.V16B(), rhs.fp().V16B(), lhs.fp().V16B());
+
+  if (dst == lhs || dst == rhs) {
+    Mov(dst.fp(), tmp);
+  }
 }
 
 void LiftoffAssembler::emit_f64x2_pmax(LiftoffRegister dst, LiftoffRegister lhs,
                                        LiftoffRegister rhs) {
-  bailout(kSimd, "pmax unimplemented");
+  UseScratchRegisterScope temps(this);
+
+  VRegister tmp = dst.fp();
+  if (dst == lhs || dst == rhs) {
+    tmp = temps.AcquireV(kFormat2D);
+  }
+
+  Fcmgt(tmp.V2D(), rhs.fp().V2D(), lhs.fp().V2D());
+  Bsl(tmp.V16B(), rhs.fp().V16B(), lhs.fp().V16B());
+
+  if (dst == lhs || dst == rhs) {
+    Mov(dst.fp(), tmp);
+  }
 }
 
 void LiftoffAssembler::emit_f32x4_splat(LiftoffRegister dst,
@@ -1720,12 +1744,36 @@ void LiftoffAssembler::emit_f32x4_max(LiftoffRegister dst, LiftoffRegister lhs,
 
 void LiftoffAssembler::emit_f32x4_pmin(LiftoffRegister dst, LiftoffRegister lhs,
                                        LiftoffRegister rhs) {
-  bailout(kSimd, "pmin unimplemented");
+  UseScratchRegisterScope temps(this);
+
+  VRegister tmp = dst.fp();
+  if (dst == lhs || dst == rhs) {
+    tmp = temps.AcquireV(kFormat4S);
+  }
+
+  Fcmgt(tmp.V4S(), lhs.fp().V4S(), rhs.fp().V4S());
+  Bsl(tmp.V16B(), rhs.fp().V16B(), lhs.fp().V16B());
+
+  if (dst == lhs || dst == rhs) {
+    Mov(dst.fp(), tmp);
+  }
 }
 
 void LiftoffAssembler::emit_f32x4_pmax(LiftoffRegister dst, LiftoffRegister lhs,
                                        LiftoffRegister rhs) {
-  bailout(kSimd, "pmax unimplemented");
+  UseScratchRegisterScope temps(this);
+
+  VRegister tmp = dst.fp();
+  if (dst == lhs || dst == rhs) {
+    tmp = temps.AcquireV(kFormat4S);
+  }
+
+  Fcmgt(tmp.V4S(), rhs.fp().V4S(), lhs.fp().V4S());
+  Bsl(tmp.V16B(), rhs.fp().V16B(), lhs.fp().V16B());
+
+  if (dst == lhs || dst == rhs) {
+    Mov(dst.fp(), tmp);
+  }
 }
 
 void LiftoffAssembler::emit_i64x2_splat(LiftoffRegister dst,