[simd/jit]: Implement floating point rounding instructions (#106 from…

… haoyu-zc/jit-f32x4-rounding)
titzer · Aug 5, 2023 · 35c6c1c · 35c6c1c
2 parents d941a11 + e9cc14b
commit 35c6c1c
Show file tree

Hide file tree

Showing 2 changed files with 25 additions and 17 deletions.
diff --git a/src/engine/x86-64/X86_64Interpreter.v3 b/src/engine/x86-64/X86_64Interpreter.v3
@@ -2600,22 +2600,16 @@ class X86_64InterpreterGen(ic: X86_64InterpreterCode, w: DataWriter) {
 			asm.movdqu_m_s(vsph[-1].value, r_xmm0);
 			endHandler();
 		}
-		for (t in [
-			(Opcode.F32X4_CEIL, X86_64Rounding.TO_POS_INF, asm.roundps_s_m),
-			(Opcode.F32X4_FLOOR, X86_64Rounding.TO_NEG_INF, asm.roundps_s_m),
-			(Opcode.F32X4_TRUNC, X86_64Rounding.TO_ZERO, asm.roundps_s_m),
-			(Opcode.F32X4_NEAREST, X86_64Rounding.TO_NEAREST, asm.roundps_s_m),
-
-			(Opcode.F64X2_CEIL, X86_64Rounding.TO_POS_INF, asm.roundpd_s_m),
-			(Opcode.F64X2_FLOOR, X86_64Rounding.TO_NEG_INF, asm.roundpd_s_m),
-			(Opcode.F64X2_TRUNC, X86_64Rounding.TO_ZERO, asm.roundpd_s_m),
-			(Opcode.F64X2_NEAREST, X86_64Rounding.TO_NEAREST, asm.roundpd_s_m)
-		]) {
-			bindHandler(t.0);
-			t.2(r_xmm0, vsph[-1].value, t.1);
-			asm.movdqu_m_s(vsph[-1].value, r_xmm0);
-			endHandler();
-		}
+
+		genSimdUnop(Opcode.F32X4_CEIL, asm.roundps_s_s(_, _, X86_64Rounding.TO_POS_INF));
+		genSimdUnop(Opcode.F32X4_FLOOR, asm.roundps_s_s(_, _, X86_64Rounding.TO_NEG_INF));
+		genSimdUnop(Opcode.F32X4_TRUNC, asm.roundps_s_s(_, _, X86_64Rounding.TO_ZERO));
+		genSimdUnop(Opcode.F32X4_NEAREST, asm.roundps_s_s(_, _, X86_64Rounding.TO_NEAREST));
+		genSimdUnop(Opcode.F64X2_CEIL, asm.roundpd_s_s(_, _, X86_64Rounding.TO_POS_INF));
+		genSimdUnop(Opcode.F64X2_FLOOR, asm.roundpd_s_s(_, _, X86_64Rounding.TO_NEG_INF));
+		genSimdUnop(Opcode.F64X2_TRUNC, asm.roundpd_s_s(_, _, X86_64Rounding.TO_ZERO));
+		genSimdUnop(Opcode.F64X2_NEAREST, asm.roundpd_s_s(_, _, X86_64Rounding.TO_NEAREST));
+
 		// Unary operations that need masks
 		for (t in [
 			(Opcode.F32X4_NEG, masm.emit_v128_negps),
@@ -2792,6 +2786,13 @@ class X86_64InterpreterGen(ic: X86_64InterpreterCode, w: DataWriter) {
 		decrementVsp();
 		endHandler();
 	}
+	def genSimdUnop<T>(opcode: Opcode, f: (X86_64Xmmr, X86_64Xmmr) -> T) {
+		bindHandler(opcode);
+		asm.movdqu_s_m(r_xmm0, vsph[-1].value);
+		f(r_xmm0, r_xmm0);
+		asm.movdqu_m_s(vsph[-1].value, r_xmm0);
+		endHandler();
+	}
 	def bindHandler(opcode: Opcode) {
 		if (FastIntTuning.handlerAlignment > 1) w.align(FastIntTuning.handlerAlignment);
 		var pos = w.atEnd().pos;

diff --git a/src/engine/x86-64/X86_64SinglePassCompiler.v3 b/src/engine/x86-64/X86_64SinglePassCompiler.v3
@@ -578,6 +578,10 @@ class X86_64SinglePassCompiler extends SinglePassCompiler {
 	def visit_F32X4_MIN() { do_op2_x_x_xtmp(ValueKind.V128, mmasm.emit_f32x4_min); }
 	def visit_F32X4_MAX() { do_op2_x_x_xtmp(ValueKind.V128, mmasm.emit_f32x4_max); }
 	def visit_F32X4_ABS() { do_op1_x_gtmp_xtmp(ValueKind.V128, mmasm.emit_v128_absps); }
+	def visit_F32X4_CEIL() { do_op1_x_x(ValueKind.V128, asm.roundps_s_s(_, _, X86_64Rounding.TO_POS_INF)); }
+	def visit_F32X4_FLOOR() { do_op1_x_x(ValueKind.V128, asm.roundps_s_s(_, _, X86_64Rounding.TO_NEG_INF)); }
+	def visit_F32X4_TRUNC() { do_op1_x_x(ValueKind.V128, asm.roundps_s_s(_, _, X86_64Rounding.TO_ZERO)); }
+	def visit_F32X4_NEAREST() { do_op1_x_x(ValueKind.V128, asm.roundps_s_s(_, _, X86_64Rounding.TO_NEAREST)); }
 
 	def visit_F64X2_ADD() { do_op2_x_x(ValueKind.V128, asm.addpd_s_s); }
 	def visit_F64X2_SUB() { do_op2_x_x(ValueKind.V128, asm.subpd_s_s); }
@@ -596,7 +600,10 @@ class X86_64SinglePassCompiler extends SinglePassCompiler {
 	def visit_F64X2_MIN() { do_op2_x_x_xtmp(ValueKind.V128, mmasm.emit_f64x2_min); }
 	def visit_F64X2_MAX() { do_op2_x_x_xtmp(ValueKind.V128, mmasm.emit_f64x2_max); }
 	def visit_F64X2_ABS() { do_op1_x_gtmp_xtmp(ValueKind.V128, mmasm.emit_v128_abspd); }
-
+	def visit_F64X2_CEIL() { do_op1_x_x(ValueKind.V128, asm.roundpd_s_s(_, _, X86_64Rounding.TO_POS_INF)); }
+	def visit_F64X2_FLOOR() { do_op1_x_x(ValueKind.V128, asm.roundpd_s_s(_, _, X86_64Rounding.TO_NEG_INF)); }
+	def visit_F64X2_TRUNC() { do_op1_x_x(ValueKind.V128, asm.roundpd_s_s(_, _, X86_64Rounding.TO_ZERO)); }
+	def visit_F64X2_NEAREST() { do_op1_x_x(ValueKind.V128, asm.roundpd_s_s(_, _, X86_64Rounding.TO_NEAREST)); }
 
 	def visit_V128_BITSELECT() {
 		var c = popReg();