cranelift: Port bitselect over to ISLE on x64

cranelift/codegen/src/isa/x64/inst.isle

@@ -462,7 +462,6 @@
 ;; Performs an xor operation of the two operands specified
 (decl sse_xor (Type Reg RegMem) Reg)
 (rule (sse_xor ty x y) (xmm_rm_r ty (sse_xor_op ty) x y))
-
 ;; Determine the appropriate operation to compare two vectors of the specified
 ;; type.
 (decl sse_cmp_op (Type) SseOpcode)

cranelift/codegen/src/isa/x64/lower.isle

@@ -360,18 +360,14 @@
 
 ;; SSE.
 
-(rule (lower (has_type $F32X4 (band x y)))
-      (value_reg (andps (put_in_reg x)
-                        (put_in_reg_mem y))))
-
-(rule (lower (has_type $F64X2 (band x y)))
-      (value_reg (andpd (put_in_reg x)
-                        (put_in_reg_mem y))))
+(decl sse_and (Type Reg RegMem) Reg)
+(rule (sse_and $F32X4 x y) (andps x y))
+(rule (sse_and $F64X2 x y) (andpd x y))
+(rule (sse_and (multi_lane _bits _lanes) x y) (pand x y))
 
-(rule (lower (has_type (multi_lane _bits _lanes)
+(rule (lower (has_type ty @ (multi_lane _bits _lanes)
                        (band x y)))
-      (value_reg (pand (put_in_reg x)
-                       (put_in_reg_mem y))))
+      (value_reg (sse_and ty (put_in_reg x) (put_in_reg_mem y))))
 
 ;; `{i,b}128`.
 
@@ -436,18 +432,14 @@
 
 ;; SSE.
 
-(rule (lower (has_type $F32X4 (bor x y)))
-      (value_reg (orps (put_in_reg x)
-                       (put_in_reg_mem y))))
-
-(rule (lower (has_type $F64X2 (bor x y)))
-      (value_reg (orpd (put_in_reg x)
-                       (put_in_reg_mem y))))
+(decl sse_or (Type Reg RegMem) Reg)
+(rule (sse_or $F32X4 x y) (orps x y))
+(rule (sse_or $F64X2 x y) (orpd x y))
+(rule (sse_or (multi_lane _bits _lanes) x y) (por x y))
 
-(rule (lower (has_type (multi_lane _bits _lanes)
+(rule (lower (has_type ty @ (multi_lane _bits _lanes)
                        (bor x y)))
-      (value_reg (por (put_in_reg x)
-                      (put_in_reg_mem y))))
+      (value_reg (sse_or ty (put_in_reg x) (put_in_reg_mem y))))
 
 ;; `{i,b}128`.
 
@@ -960,22 +952,22 @@
 
 ;;;; Rules for `band_not` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
+(decl sse_and_not (Type Reg RegMem) Reg)
+(rule (sse_and_not $F32X4 x y) (andnps x y))
+(rule (sse_and_not $F64X2 x y) (andnpd x y))
+(rule (sse_and_not (multi_lane _bits _lanes) x y) (pandn x y))
+
 ;; Note the flipping of operands below. CLIF specifies
 ;;
 ;;   band_not(x, y) = and(x, not(y))
 ;;
 ;; while x86 does
 ;;
 ;;   pandn(x, y) = and(not(x), y)
-
-(rule (lower (has_type $F32X4 (band_not x y)))
-      (value_reg (andnps (put_in_reg y) (put_in_reg_mem x))))
-
-(rule (lower (has_type $F64X2 (band_not x y)))
-      (value_reg (andnpd (put_in_reg y) (put_in_reg_mem x))))
-
-(rule (lower (has_type (multi_lane _bits _lanes) (band_not x y)))
-      (value_reg (pandn (put_in_reg y) (put_in_reg_mem x))))
+(rule (lower (has_type ty (band_not x y)))
+      (value_reg (sse_and_not ty
+                              (put_in_reg y)
+                              (put_in_reg_mem x))))
 
 ;;;; Rules for `iabs` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
@@ -1044,6 +1036,20 @@
 (rule (lower (has_type ty @ (multi_lane _bits _lanes) (bnot x)))
       (value_reg (sse_xor ty (put_in_reg x) (RegMem.Reg (vector_all_ones ty)))))
 
+;;;; Rules for `bitselect` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule (lower (has_type ty @ (multi_lane _bits _lanes)
+                       (bitselect condition
+                                  if_true
+                                  if_false)))
+      ;; a = and condition, if_true
+      ;; b = and_not condition, if_false
+      ;; return or a, b
+      (let ((cond_reg Reg (put_in_reg condition))
+            (a Reg (sse_and ty cond_reg (put_in_reg_mem if_true)))
+            (b Reg (sse_and_not ty cond_reg (put_in_reg_mem if_false))))
+        (value_reg (sse_or ty a (RegMem.Reg b)))))
+
 ;;;; Rules for `insertlane` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
 (rule (lower (insertlane vec @ (value_type ty) val (u8_from_uimm8 idx)))

cranelift/codegen/src/isa/x64/lower.rs

@@ -1534,30 +1534,8 @@ fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
         | Opcode::Umax
         | Opcode::Imin
         | Opcode::Umin
-        | Opcode::Bnot => implemented_in_isle(ctx),
-
-        Opcode::Bitselect => {
-            let ty = ty.unwrap();
-            let condition = put_input_in_reg(ctx, inputs[0]);
-            let if_true = put_input_in_reg(ctx, inputs[1]);
-            let if_false = input_to_reg_mem(ctx, inputs[2]);
-            let dst = get_output_reg(ctx, outputs[0]).only_reg().unwrap();
-
-            if ty.is_vector() {
-                let tmp1 = ctx.alloc_tmp(ty).only_reg().unwrap();
-                ctx.emit(Inst::gen_move(tmp1, if_true, ty));
-                ctx.emit(Inst::and(ty, RegMem::reg(condition.clone()), tmp1));
-
-                let tmp2 = ctx.alloc_tmp(ty).only_reg().unwrap();
-                ctx.emit(Inst::gen_move(tmp2, condition, ty));
-                ctx.emit(Inst::and_not(ty, if_false, tmp2));
-
-                ctx.emit(Inst::gen_move(dst, tmp2.to_reg(), ty));
-                ctx.emit(Inst::or(ty, RegMem::from(tmp1), dst));
-            } else {
-                unimplemented!("no lowering for scalar bitselect instruction")
-            }
-        }
+        | Opcode::Bnot
+        | Opcode::Bitselect => implemented_in_isle(ctx),
 
         Opcode::Vselect => {
             let ty = ty.unwrap();

cranelift/codegen/src/isa/x64/lower/isle/generated_code.manifest

@@ -1,4 +1,4 @@
 src/clif.isle f176ef3bba99365
 src/prelude.isle babc931e5dc5b4cf
-src/isa/x64/inst.isle fb5d3ac8e68c46d2
-src/isa/x64/lower.isle d39e01add89178d5
+src/isa/x64/inst.isle 9951164661578952
+src/isa/x64/lower.isle 57c693599ccffefc