x64: Improve lowerings for `ctz`/`clz` instructions (#8673)

* x64: Add some more tests for `ctz`/`clz`

* x64: Improve lowerings for i8/i16/i128 `ctz` and `clz` intructions

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

@@ -2129,21 +2129,14 @@
 
 ;; Rules for `clz` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
-;; If available, we can use a plain lzcnt instruction here. Note no
-;; special handling is required for zero inputs, because the machine
-;; instruction does what the CLIF expects for zero, i.e. it returns
-;; zero.
-(rule 3 (lower (has_type (ty_32_or_64 ty) (clz src)))
-      (if-let $true (use_lzcnt))
-      (x64_lzcnt ty src))
-
 (rule 2 (lower (has_type (ty_32_or_64 ty) (clz src)))
       (do_clz ty ty src))
 
-(rule 1 (lower
+(rule 1 (lower (has_type (ty_8_or_16 ty) (clz src)))
-       (has_type (ty_8_or_16 ty)
+      (let ((extended Gpr (extend_to_gpr src $I64 (ExtendKind.Zero)))
-                 (clz src)))
+            (clz Gpr (do_clz $I64 $I64 extended)))
-      (do_clz $I32 ty (extend_to_gpr src $I32 (ExtendKind.Zero))))
+        (x64_sub $I64 clz (RegMemImm.Imm (u32_sub 64 (ty_bits ty))))))
+
 
 (rule 0 (lower
        (has_type $I128
@@ -2160,27 +2153,29 @@
 
 ;; Implementation helper for clz; operates on 32 or 64-bit units.
 (decl do_clz (Type Type Gpr) Gpr)
-(rule (do_clz ty orig_ty src)
+
+;; If available, we can use a plain lzcnt instruction here. Note no
+;; special handling is required for zero inputs, because the machine
+;; instruction does what the CLIF expects for zero, i.e. it returns
+;; zero.
+(rule 1 (do_clz ty orig_ty src)
+      (if-let $true (use_lzcnt))
+      (x64_lzcnt ty src))
+
+(rule 0 (do_clz ty orig_ty src)
       (let ((highest_bit_index Reg (bsr_or_else ty src (imm_i64 $I64 -1)))
             (bits_minus_1 Reg (imm ty (u64_sub (ty_bits_u64 orig_ty) 1))))
         (x64_sub ty bits_minus_1 highest_bit_index)))
 
 ;; Rules for `ctz` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
-;; Analogous to `clz` cases above, but using mirror instructions
-;; (tzcnt vs lzcnt, bsf vs bsr).
-
-(rule 3 (lower (has_type (ty_32_or_64 ty) (ctz src)))
-      (if-let $true (use_bmi1))
-      (x64_tzcnt ty src))
-
 (rule 2 (lower (has_type (ty_32_or_64 ty) (ctz src)))
       (do_ctz ty ty src))
 
-(rule 1 (lower
+(rule 1 (lower (has_type (ty_8_or_16 ty) (ctz src)))
-       (has_type (ty_8_or_16 ty)
+      (let ((extended Gpr (extend_to_gpr src $I32 (ExtendKind.Zero)))
-                 (ctz src)))
+            (stopbit Gpr (x64_or $I32 extended (RegMemImm.Imm (u32_shl 1 (ty_bits ty))))))
-      (do_ctz $I32 ty (extend_to_gpr src $I32 (ExtendKind.Zero))))
+        (do_ctz $I32 ty stopbit)))
 
 (rule 0 (lower
        (has_type $I128
@@ -2196,7 +2191,14 @@
         (value_regs result_lo (imm $I64 0))))
 
 (decl do_ctz (Type Type Gpr) Gpr)
-(rule (do_ctz ty orig_ty src)
+
+;; Analogous to `clz` cases above, but using mirror instructions
+;; (tzcnt vs lzcnt, bsf vs bsr).
+(rule 1 (do_ctz ty orig_ty src)
+      (if-let $true (use_bmi1))
+      (x64_tzcnt ty src))
+
+(rule 0 (do_ctz ty orig_ty src)
       (bsf_or_else ty src (imm $I64 (ty_bits_u64 orig_ty))))
 
 ;; Rules for `popcnt` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

cranelift/codegen/src/isle_prelude.rs

@@ -683,6 +683,11 @@ macro_rules! isle_common_prelude_methods {
             a & b
         }
 
+        #[inline]
+        fn u32_shl(&mut self, x: u32, y: u32) -> u32 {
+            x << y
+        }
+
         #[inline]
         fn s32_add_fallible(&mut self, a: i32, b: i32) -> Option<i32> {
             a.checked_add(b)

cranelift/codegen/src/prelude.isle

@@ -133,6 +133,9 @@
 (decl pure u32_and (u32 u32) u32)
 (extern constructor u32_and u32_and)
 
+(decl pure u32_shl (u32 u32) u32)
+(extern constructor u32_shl u32_shl)
+
 ;; Pure/fallible constructor that tries to add two `u32`s, interpreted
 ;; as signed values, and fails to match on overflow.
 (decl pure partial s32_add_fallible (i32 i32) i32)

cranelift/filetests/filetests/isa/x64/clz-lzcnt.clif

@@ -1,6 +1,43 @@
 test compile precise-output
+set enable_llvm_abi_extensions=true
 target x86_64 has_lzcnt
 
+
+function %clz(i128) -> i128 {
+block0(v0: i128):
+    v1 = clz v0
+    return v1
+}
+
+; VCode:
+;   pushq   %rbp
+;   movq    %rsp, %rbp
+; block0:
+;   lzcntq  %rsi, %rcx
+;   lzcntq  %rdi, %rax
+;   addq    %rax, $64, %rax
+;   cmpq    $64, %rcx
+;   cmovnzq %rcx, %rax, %rax
+;   xorq    %rdx, %rdx, %rdx
+;   movq    %rbp, %rsp
+;   popq    %rbp
+;   ret
+;
+; Disassembled:
+; block0: ; offset 0x0
+;   pushq %rbp
+;   movq %rsp, %rbp
+; block1: ; offset 0x4
+;   lzcntq %rsi, %rcx
+;   lzcntq %rdi, %rax
+;   addq $0x40, %rax
+;   cmpq $0x40, %rcx
+;   cmovneq %rcx, %rax
+;   xorq %rdx, %rdx
+;   movq %rbp, %rsp
+;   popq %rbp
+;   retq
+
 function %clz(i64) -> i64 {
 block0(v0: i64):
     v1 = clz v0
@@ -51,3 +88,61 @@ block0(v0: i32):
 ;   popq %rbp
 ;   retq
 
+function %clz(i16) -> i16 {
+block0(v0: i16):
+    v1 = clz v0
+    return v1
+}
+
+; VCode:
+;   pushq   %rbp
+;   movq    %rsp, %rbp
+; block0:
+;   movzwq  %di, %rax
+;   lzcntq  %rax, %rax
+;   subq    %rax, $48, %rax
+;   movq    %rbp, %rsp
+;   popq    %rbp
+;   ret
+;
+; Disassembled:
+; block0: ; offset 0x0
+;   pushq %rbp
+;   movq %rsp, %rbp
+; block1: ; offset 0x4
+;   movzwq %di, %rax
+;   lzcntq %rax, %rax
+;   subq $0x30, %rax
+;   movq %rbp, %rsp
+;   popq %rbp
+;   retq
+
+function %clz(i8) -> i8 {
+block0(v0: i8):
+    v1 = clz v0
+    return v1
+}
+
+; VCode:
+;   pushq   %rbp
+;   movq    %rsp, %rbp
+; block0:
+;   movzbq  %dil, %rax
+;   lzcntq  %rax, %rax
+;   subq    %rax, $56, %rax
+;   movq    %rbp, %rsp
+;   popq    %rbp
+;   ret
+;
+; Disassembled:
+; block0: ; offset 0x0
+;   pushq %rbp
+;   movq %rsp, %rbp
+; block1: ; offset 0x4
+;   movzbq %dil, %rax
+;   lzcntq %rax, %rax
+;   subq $0x38, %rax
+;   movq %rbp, %rsp
+;   popq %rbp
+;   retq
+

cranelift/filetests/filetests/isa/x64/clz.clif

@@ -0,0 +1,198 @@
+test compile precise-output
+set enable_llvm_abi_extensions=true
+target x86_64
+
+
+function %clz(i128) -> i128 {
+block0(v0: i128):
+    v1 = clz v0
+    return v1
+}
+
+; VCode:
+;   pushq   %rbp
+;   movq    %rsp, %rbp
+; block0:
+;   movq    %rdi, %r8
+;   movabsq $-1, %rcx
+;   bsrq    %rsi, %r9
+;   cmovzq  %rcx, %r9, %r9
+;   movl    $63, %edi
+;   subq    %rdi, %r9, %rdi
+;   movabsq $-1, %rdx
+;   bsrq    %r8, %r10
+;   cmovzq  %rdx, %r10, %r10
+;   movl    $63, %eax
+;   subq    %rax, %r10, %rax
+;   addq    %rax, $64, %rax
+;   cmpq    $64, %rdi
+;   cmovnzq %rdi, %rax, %rax
+;   xorq    %rdx, %rdx, %rdx
+;   movq    %rbp, %rsp
+;   popq    %rbp
+;   ret
+;
+; Disassembled:
+; block0: ; offset 0x0
+;   pushq %rbp
+;   movq %rsp, %rbp
+; block1: ; offset 0x4
+;   movq %rdi, %r8
+;   movq $18446744073709551615, %rcx
+;   bsrq %rsi, %r9
+;   cmoveq %rcx, %r9
+;   movl $0x3f, %edi
+;   subq %r9, %rdi
+;   movq $18446744073709551615, %rdx
+;   bsrq %r8, %r10
+;   cmoveq %rdx, %r10
+;   movl $0x3f, %eax
+;   subq %r10, %rax
+;   addq $0x40, %rax
+;   cmpq $0x40, %rdi
+;   cmovneq %rdi, %rax
+;   xorq %rdx, %rdx
+;   movq %rbp, %rsp
+;   popq %rbp
+;   retq
+
+function %clz(i64) -> i64 {
+block0(v0: i64):
+    v1 = clz v0
+    return v1
+}
+
+; VCode:
+;   pushq   %rbp
+;   movq    %rsp, %rbp
+; block0:
+;   movabsq $-1, %rax
+;   bsrq    %rdi, %r8
+;   cmovzq  %rax, %r8, %r8
+;   movl    $63, %eax
+;   subq    %rax, %r8, %rax
+;   movq    %rbp, %rsp
+;   popq    %rbp
+;   ret
+;
+; Disassembled:
+; block0: ; offset 0x0
+;   pushq %rbp
+;   movq %rsp, %rbp
+; block1: ; offset 0x4
+;   movq $18446744073709551615, %rax
+;   bsrq %rdi, %r8
+;   cmoveq %rax, %r8
+;   movl $0x3f, %eax
+;   subq %r8, %rax
+;   movq %rbp, %rsp
+;   popq %rbp
+;   retq
+
+function %clz(i32) -> i32 {
+block0(v0: i32):
+    v1 = clz v0
+    return v1
+}
+
+; VCode:
+;   pushq   %rbp
+;   movq    %rsp, %rbp
+; block0:
+;   movabsq $-1, %rax
+;   bsrl    %edi, %r8d
+;   cmovzl  %eax, %r8d, %r8d
+;   movl    $31, %eax
+;   subl    %eax, %r8d, %eax
+;   movq    %rbp, %rsp
+;   popq    %rbp
+;   ret
+;
+; Disassembled:
+; block0: ; offset 0x0
+;   pushq %rbp
+;   movq %rsp, %rbp
+; block1: ; offset 0x4
+;   movq $18446744073709551615, %rax
+;   bsrl %edi, %r8d
+;   cmovel %eax, %r8d
+;   movl $0x1f, %eax
+;   subl %r8d, %eax
+;   movq %rbp, %rsp
+;   popq %rbp
+;   retq
+
+function %clz(i16) -> i16 {
+block0(v0: i16):
+    v1 = clz v0
+    return v1
+}
+
+; VCode:
+;   pushq   %rbp
+;   movq    %rsp, %rbp
+; block0:
+;   movzwq  %di, %rax
+;   movabsq $-1, %rdx
+;   bsrq    %rax, %r10
+;   cmovzq  %rdx, %r10, %r10
+;   movl    $63, %eax
+;   subq    %rax, %r10, %rax
+;   subq    %rax, $48, %rax
+;   movq    %rbp, %rsp
+;   popq    %rbp
+;   ret
+;
+; Disassembled:
+; block0: ; offset 0x0
+;   pushq %rbp
+;   movq %rsp, %rbp
+; block1: ; offset 0x4
+;   movzwq %di, %rax
+;   movq $18446744073709551615, %rdx
+;   bsrq %rax, %r10
+;   cmoveq %rdx, %r10
+;   movl $0x3f, %eax
+;   subq %r10, %rax
+;   subq $0x30, %rax
+;   movq %rbp, %rsp
+;   popq %rbp
+;   retq
+
+function %clz(i8) -> i8 {
+block0(v0: i8):
+    v1 = clz v0
+    return v1
+}
+
+; VCode:
+;   pushq   %rbp
+;   movq    %rsp, %rbp
+; block0:
+;   movzbq  %dil, %rax
+;   movabsq $-1, %rdx
+;   bsrq    %rax, %r10
+;   cmovzq  %rdx, %r10, %r10
+;   movl    $63, %eax
+;   subq    %rax, %r10, %rax
+;   subq    %rax, $56, %rax
+;   movq    %rbp, %rsp
+;   popq    %rbp
+;   ret
+;
+; Disassembled:
+; block0: ; offset 0x0
+;   pushq %rbp
+;   movq %rsp, %rbp
+; block1: ; offset 0x4
+;   movzbq %dil, %rax
+;   movq $18446744073709551615, %rdx
+;   bsrq %rax, %r10
+;   cmoveq %rdx, %r10
+;   movl $0x3f, %eax
+;   subq %r10, %rax
+;   subq $0x38, %rax
+;   movq %rbp, %rsp
+;   popq %rbp
+;   retq
+

cranelift/filetests/filetests/isa/x64/ctz-bmi1.clif

@@ -1,6 +1,43 @@
 test compile precise-output
+set enable_llvm_abi_extensions=true
 target x86_64 has_bmi1
 
+
+function %ctz(i128) -> i128 {
+block0(v0: i128):
+    v1 = ctz v0
+    return v1
+}
+
+; VCode:
+;   pushq   %rbp
+;   movq    %rsp, %rbp
+; block0:
+;   tzcntq  %rdi, %rax
+;   tzcntq  %rsi, %r9
+;   addq    %r9, $64, %r9
+;   cmpq    $64, %rax
+;   cmovzq  %r9, %rax, %rax
+;   xorq    %rdx, %rdx, %rdx
+;   movq    %rbp, %rsp
+;   popq    %rbp
+;   ret
+;
+; Disassembled:
+; block0: ; offset 0x0
+;   pushq %rbp
+;   movq %rsp, %rbp
+; block1: ; offset 0x4
+;   tzcntq %rdi, %rax
+;   tzcntq %rsi, %r9
+;   addq $0x40, %r9
+;   cmpq $0x40, %rax
+;   cmoveq %r9, %rax
+;   xorq %rdx, %rdx
+;   movq %rbp, %rsp
+;   popq %rbp
+;   retq
+
 function %ctz(i64) -> i64 {
 block0(v0: i64):
     v1 = ctz v0
@@ -51,3 +88,61 @@ block0(v0: i32):
 ;   popq %rbp
 ;   retq
 
+function %ctz(i16) -> i16 {
+block0(v0: i16):
+    v1 = ctz v0
+    return v1
+}
+
+; VCode:
+;   pushq   %rbp
+;   movq    %rsp, %rbp
+; block0:
+;   movzwl  %di, %ecx
+;   orl     %ecx, $65536, %ecx
+;   tzcntl  %ecx, %eax
+;   movq    %rbp, %rsp
+;   popq    %rbp
+;   ret
+;
+; Disassembled:
+; block0: ; offset 0x0
+;   pushq %rbp
+;   movq %rsp, %rbp
+; block1: ; offset 0x4
+;   movzwl %di, %ecx
+;   orl $0x10000, %ecx
+;   tzcntl %ecx, %eax
+;   movq %rbp, %rsp
+;   popq %rbp
+;   retq
+
+function %ctz(i8) -> i8 {
+block0(v0: i8):
+    v1 = ctz v0
+    return v1
+}
+
+; VCode:
+;   pushq   %rbp
+;   movq    %rsp, %rbp
+; block0:
+;   movzbl  %dil, %ecx
+;   orl     %ecx, $256, %ecx
+;   tzcntl  %ecx, %eax
+;   movq    %rbp, %rsp
+;   popq    %rbp
+;   ret
+;
+; Disassembled:
+; block0: ; offset 0x0
+;   pushq %rbp
+;   movq %rsp, %rbp
+; block1: ; offset 0x4
+;   movzbl %dil, %ecx
+;   orl $0x100, %ecx
+;   tzcntl %ecx, %eax
+;   movq %rbp, %rsp
+;   popq %rbp
+;   retq
+

cranelift/filetests/filetests/isa/x64/ctz.clif

@@ -0,0 +1,172 @@
+test compile precise-output
+set enable_llvm_abi_extensions=true
+target x86_64
+
+
+function %ctz(i128) -> i128 {
+block0(v0: i128):
+    v1 = ctz v0
+    return v1
+}
+
+; VCode:
+;   pushq   %rbp
+;   movq    %rsp, %rbp
+; block0:
+;   movl    $64, %ecx
+;   bsfq    %rdi, %rax
+;   cmovzq  %rcx, %rax, %rax
+;   movl    $64, %edi
+;   bsfq    %rsi, %rdx
+;   cmovzq  %rdi, %rdx, %rdx
+;   addq    %rdx, $64, %rdx
+;   cmpq    $64, %rax
+;   cmovzq  %rdx, %rax, %rax
+;   xorq    %rdx, %rdx, %rdx
+;   movq    %rbp, %rsp
+;   popq    %rbp
+;   ret
+;
+; Disassembled:
+; block0: ; offset 0x0
+;   pushq %rbp
+;   movq %rsp, %rbp
+; block1: ; offset 0x4
+;   movl $0x40, %ecx
+;   bsfq %rdi, %rax
+;   cmoveq %rcx, %rax
+;   movl $0x40, %edi
+;   bsfq %rsi, %rdx
+;   cmoveq %rdi, %rdx
+;   addq $0x40, %rdx
+;   cmpq $0x40, %rax
+;   cmoveq %rdx, %rax
+;   xorq %rdx, %rdx
+;   movq %rbp, %rsp
+;   popq %rbp
+;   retq
+
+function %ctz(i64) -> i64 {
+block0(v0: i64):
+    v1 = ctz v0
+    return v1
+}
+
+; VCode:
+;   pushq   %rbp
+;   movq    %rsp, %rbp
+; block0:
+;   movl    $64, %ecx
+;   bsfq    %rdi, %rax
+;   cmovzq  %rcx, %rax, %rax
+;   movq    %rbp, %rsp
+;   popq    %rbp
+;   ret
+;
+; Disassembled:
+; block0: ; offset 0x0
+;   pushq %rbp
+;   movq %rsp, %rbp
+; block1: ; offset 0x4
+;   movl $0x40, %ecx
+;   bsfq %rdi, %rax
+;   cmoveq %rcx, %rax
+;   movq %rbp, %rsp
+;   popq %rbp
+;   retq
+
+function %ctz(i32) -> i32 {
+block0(v0: i32):
+    v1 = ctz v0
+    return v1
+}
+
+; VCode:
+;   pushq   %rbp
+;   movq    %rsp, %rbp
+; block0:
+;   movl    $32, %ecx
+;   bsfl    %edi, %eax
+;   cmovzl  %ecx, %eax, %eax
+;   movq    %rbp, %rsp
+;   popq    %rbp
+;   ret
+;
+; Disassembled:
+; block0: ; offset 0x0
+;   pushq %rbp
+;   movq %rsp, %rbp
+; block1: ; offset 0x4
+;   movl $0x20, %ecx
+;   bsfl %edi, %eax
+;   cmovel %ecx, %eax
+;   movq %rbp, %rsp
+;   popq %rbp
+;   retq
+
+function %ctz(i16) -> i16 {
+block0(v0: i16):
+    v1 = ctz v0
+    return v1
+}
+
+; VCode:
+;   pushq   %rbp
+;   movq    %rsp, %rbp
+; block0:
+;   movzwl  %di, %ecx
+;   orl     %ecx, $65536, %ecx
+;   movl    $16, %r9d
+;   bsfl    %ecx, %eax
+;   cmovzl  %r9d, %eax, %eax
+;   movq    %rbp, %rsp
+;   popq    %rbp
+;   ret
+;
+; Disassembled:
+; block0: ; offset 0x0
+;   pushq %rbp
+;   movq %rsp, %rbp
+; block1: ; offset 0x4
+;   movzwl %di, %ecx
+;   orl $0x10000, %ecx
+;   movl $0x10, %r9d
+;   bsfl %ecx, %eax
+;   cmovel %r9d, %eax
+;   movq %rbp, %rsp
+;   popq %rbp
+;   retq
+
+function %ctz(i8) -> i8 {
+block0(v0: i8):
+    v1 = ctz v0
+    return v1
+}
+
+; VCode:
+;   pushq   %rbp
+;   movq    %rsp, %rbp
+; block0:
+;   movzbl  %dil, %ecx
+;   orl     %ecx, $256, %ecx
+;   movl    $8, %r9d
+;   bsfl    %ecx, %eax
+;   cmovzl  %r9d, %eax, %eax
+;   movq    %rbp, %rsp
+;   popq    %rbp
+;   ret
+;
+; Disassembled:
+; block0: ; offset 0x0
+;   pushq %rbp
+;   movq %rsp, %rbp
+; block1: ; offset 0x4
+;   movzbl %dil, %ecx
+;   orl $0x100, %ecx
+;   movl $8, %r9d
+;   bsfl %ecx, %eax
+;   cmovel %r9d, %eax
+;   movq %rbp, %rsp
+;   popq %rbp
+;   retq
+

cranelift/filetests/filetests/runtests/i128-bitops-count.clif

@@ -3,6 +3,9 @@ set enable_llvm_abi_extensions=true
 target aarch64
 target s390x
 target x86_64
+target x86_64 has_lzcnt
+target x86_64 has_bmi1
+target x86_64 has_popcnt has_sse42
 target riscv64
 target riscv64 has_zbb
 target riscv64 has_zbb has_zbs

cranelift/filetests/filetests/runtests/popcnt.clif

@@ -3,7 +3,7 @@ test run
 target aarch64
 target s390x
 target x86_64
-target x86_64 has_popcnt
+target x86_64 has_popcnt has_sse42
 target riscv64
 target riscv64 has_c has_zcb