cranelift/x64: implement cls for all integer types (#12644)

* cranelift/x64: implement cls for all integer types

* refactored to use x64_sar; moved tests to runtests

* move i128-cls tests into cls tests

Author: shreyasravi320

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

@@ -2332,6 +2332,41 @@
 (rule 0 (do_ctz ty orig_ty src)
       (bsf_or_else ty src (imm $I64 (ty_bits_u64 orig_ty))))
 
+;; Rules for `cls` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule 2 (lower (has_type (ty_32_or_64 ty) (cls src)))
+      (do_cls ty src))
+
+(rule 1 (lower (has_type (ty_8_or_16 ty) (cls src)))
+      (let ((extended Gpr (extend_to_gpr src $I32 (ExtendKind.Sign)))
+            (cls Gpr (do_cls $I32 extended)))
+        (x64_sub $I32 cls (RegMemImm.Imm (u32_wrapping_sub 32 (ty_bits ty))))))
+
+(rule 0 (lower
+       (has_type $I128
+                 (cls src)))
+      (let ((upper Gpr (do_cls $I64 (value_regs_get_gpr src 1)))
+            (sign_fill Gpr (x64_sarq_mi (value_regs_get_gpr src 1) 63))
+            (xored Gpr (x64_xor $I64 (value_regs_get_gpr src 0) sign_fill))
+            (lower Gpr (x64_add $I64
+                            (do_clz $I64 $I64 xored)
+                            (RegMemImm.Imm 63)))
+            (result_lo Gpr
+              (with_flags_reg
+               (x64_cmpq_mi_sxb upper 63)
+               (cmove $I64 (CC.NZ) upper lower))))
+        (value_regs result_lo (imm $I64 0))))
+
+;; Implementation helper for cls; operates on 32 or 64-bit units.
+(decl do_cls (Type Gpr) Gpr)
+
+;; cls is implemented via clz using the identity: cls(x) = clz(x ^ (x >> 1)) - 1
+(rule (do_cls ty src)
+      (let ((shifted Gpr (x64_sar ty src (Imm8Gpr.Imm8 1)))
+            (xored Gpr (x64_xor ty src (RegMemImm.Reg shifted)))
+            (clz Gpr (do_clz ty ty xored)))
+        (x64_sub ty clz (RegMemImm.Imm 1))))
+
 ;; Rules for `popcnt` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
 (rule 4 (lower (has_type (ty_32_or_64 ty) (popcnt src)))

cranelift/filetests/filetests/runtests/cls.clif

@@ -1,11 +1,13 @@
 test interpret
 test run
+set enable_llvm_abi_extensions=true
 target aarch64
+target x86_64
 target riscv64
 target riscv64 has_zbb
 target riscv64 has_c has_zcb
+set enable_multi_ret_implicit_sret
 target s390x
-; not implemented on `x86_64`
 
 function %cls_i8(i8) -> i8 {
 block0(v0: i8):
@@ -46,3 +48,22 @@ block0(v0: i64):
 ; run: %cls_i64(0x4000000000000000) == 0
 ; run: %cls_i64(-1) == 63
 ; run: %cls_i64(0) == 63
+
+function %cls_i128(i128) -> i128 {
+block0(v0: i128):
+    v1 = cls v0
+    return v1
+}
+; run: %cls_i128(1) == 126
+; run: %cls_i128(0x40000000_00000000_00000000_00000000) == 0
+; run: %cls_i128(0) == 127
+; run: %cls_i128(0x00000000_00000000_00000000_00000000) == 127
+; run: %cls_i128(0x00000000_00000000_FFFFFFFF_FFFFFFFF) == 63
+; run: %cls_i128(0xFFFFFFFF_FFFFFFFF_00000000_00000000) == 63
+; run: %cls_i128(0xFFFFFFFF_FFFFFFFF_FFFFFFFF_FFFFFFFF) == 127
+; run: %cls_i128(0x7FFFFFFF_FFFFFFFF_FFFFFFFF_FFFFFFFF) == 0
+; run: %cls_i128(0x3FFFFFFF_FFFFFFFF_FFFFFFFF_FFFFFFFF) == 1
+; run: %cls_i128(0xFFFFFFFF_FFFFFFFF_7FFFFFFF_FFFFFFFF) == 63
+; run: %cls_i128(0xC0000000_00000000_80000000_00000000) == 1
+; run: %cls_i128(0xC0000000_00000000_00000000_00000000) == 1
+; run: %cls_i128(0x80000000_00000000_80000000_00000000) == 0