blob: a59223515fb6867f1cb8fefb1d77256a07e5b669 [file] [log] [blame]
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc -mtriple=riscv32 -mattr=+m,+experimental-v -riscv-v-vector-bits-min=128 -riscv-v-fixed-length-vector-lmul-max=2 -verify-machineinstrs < %s | FileCheck %s --check-prefixes=LMULMAX2-RV32,LMULMAX2-RV32I
; RUN: llc -mtriple=riscv64 -mattr=+m,+experimental-v -riscv-v-vector-bits-min=128 -riscv-v-fixed-length-vector-lmul-max=2 -verify-machineinstrs < %s | FileCheck %s --check-prefixes=LMULMAX2-RV64,LMULMAX2-RV64I
; RUN: llc -mtriple=riscv32 -mattr=+m,+experimental-v -riscv-v-vector-bits-min=128 -riscv-v-fixed-length-vector-lmul-max=1 -verify-machineinstrs < %s | FileCheck %s --check-prefixes=LMULMAX1-RV32
; RUN: llc -mtriple=riscv64 -mattr=+m,+experimental-v -riscv-v-vector-bits-min=128 -riscv-v-fixed-length-vector-lmul-max=1 -verify-machineinstrs < %s | FileCheck %s --check-prefixes=LMULMAX1-RV64
; RUN: llc -mtriple=riscv32 -mattr=+m,+experimental-v,+d -riscv-v-vector-bits-min=128 -riscv-v-fixed-length-vector-lmul-max=2 -verify-machineinstrs < %s | FileCheck %s --check-prefixes=LMULMAX2-RV32,LMULMAX2-RV32D
; RUN: llc -mtriple=riscv64 -mattr=+m,+experimental-v,+d -riscv-v-vector-bits-min=128 -riscv-v-fixed-length-vector-lmul-max=2 -verify-machineinstrs < %s | FileCheck %s --check-prefixes=LMULMAX2-RV64,LMULMAX2-RV64D
; RUN: llc -mtriple=riscv32 -mattr=+m,+experimental-v,+d -riscv-v-vector-bits-min=128 -riscv-v-fixed-length-vector-lmul-max=1 -verify-machineinstrs < %s | FileCheck %s --check-prefixes=LMULMAX1-RV32
; RUN: llc -mtriple=riscv64 -mattr=+m,+experimental-v,+d -riscv-v-vector-bits-min=128 -riscv-v-fixed-length-vector-lmul-max=1 -verify-machineinstrs < %s | FileCheck %s --check-prefixes=LMULMAX1-RV64
; RUN: llc -mtriple=riscv32 -mattr=+m,+experimental-v,+d -riscv-v-vector-bits-min=128 -riscv-v-fixed-length-vector-lmul-max=8 -verify-machineinstrs < %s | FileCheck %s --check-prefixes=LMULMAX8-RV32
; RUN: llc -mtriple=riscv64 -mattr=+m,+experimental-v,+d -riscv-v-vector-bits-min=128 -riscv-v-fixed-length-vector-lmul-max=8 -verify-machineinstrs < %s | FileCheck %s --check-prefixes=LMULMAX8-RV64
define void @cttz_v16i8(<16 x i8>* %x, <16 x i8>* %y) nounwind {
; LMULMAX2-RV32-LABEL: cttz_v16i8:
; LMULMAX2-RV32: # %bb.0:
; LMULMAX2-RV32-NEXT: vsetivli zero, 16, e8, m1, ta, mu
; LMULMAX2-RV32-NEXT: vle8.v v8, (a0)
; LMULMAX2-RV32-NEXT: li a1, 1
; LMULMAX2-RV32-NEXT: vsub.vx v9, v8, a1
; LMULMAX2-RV32-NEXT: vxor.vi v8, v8, -1
; LMULMAX2-RV32-NEXT: vand.vv v8, v8, v9
; LMULMAX2-RV32-NEXT: vsrl.vi v9, v8, 1
; LMULMAX2-RV32-NEXT: li a1, 85
; LMULMAX2-RV32-NEXT: vand.vx v9, v9, a1
; LMULMAX2-RV32-NEXT: vsub.vv v8, v8, v9
; LMULMAX2-RV32-NEXT: li a1, 51
; LMULMAX2-RV32-NEXT: vand.vx v9, v8, a1
; LMULMAX2-RV32-NEXT: vsrl.vi v8, v8, 2
; LMULMAX2-RV32-NEXT: vand.vx v8, v8, a1
; LMULMAX2-RV32-NEXT: vadd.vv v8, v9, v8
; LMULMAX2-RV32-NEXT: vsrl.vi v9, v8, 4
; LMULMAX2-RV32-NEXT: vadd.vv v8, v8, v9
; LMULMAX2-RV32-NEXT: vand.vi v8, v8, 15
; LMULMAX2-RV32-NEXT: vse8.v v8, (a0)
; LMULMAX2-RV32-NEXT: ret
;
; LMULMAX2-RV64-LABEL: cttz_v16i8:
; LMULMAX2-RV64: # %bb.0:
; LMULMAX2-RV64-NEXT: vsetivli zero, 16, e8, m1, ta, mu
; LMULMAX2-RV64-NEXT: vle8.v v8, (a0)
; LMULMAX2-RV64-NEXT: li a1, 1
; LMULMAX2-RV64-NEXT: vsub.vx v9, v8, a1
; LMULMAX2-RV64-NEXT: vxor.vi v8, v8, -1
; LMULMAX2-RV64-NEXT: vand.vv v8, v8, v9
; LMULMAX2-RV64-NEXT: vsrl.vi v9, v8, 1
; LMULMAX2-RV64-NEXT: li a1, 85
; LMULMAX2-RV64-NEXT: vand.vx v9, v9, a1
; LMULMAX2-RV64-NEXT: vsub.vv v8, v8, v9
; LMULMAX2-RV64-NEXT: li a1, 51
; LMULMAX2-RV64-NEXT: vand.vx v9, v8, a1
; LMULMAX2-RV64-NEXT: vsrl.vi v8, v8, 2
; LMULMAX2-RV64-NEXT: vand.vx v8, v8, a1
; LMULMAX2-RV64-NEXT: vadd.vv v8, v9, v8
; LMULMAX2-RV64-NEXT: vsrl.vi v9, v8, 4
; LMULMAX2-RV64-NEXT: vadd.vv v8, v8, v9
; LMULMAX2-RV64-NEXT: vand.vi v8, v8, 15
; LMULMAX2-RV64-NEXT: vse8.v v8, (a0)
; LMULMAX2-RV64-NEXT: ret
;
; LMULMAX1-RV32-LABEL: cttz_v16i8:
; LMULMAX1-RV32: # %bb.0:
; LMULMAX1-RV32-NEXT: vsetivli zero, 16, e8, m1, ta, mu
; LMULMAX1-RV32-NEXT: vle8.v v8, (a0)
; LMULMAX1-RV32-NEXT: li a1, 1
; LMULMAX1-RV32-NEXT: vsub.vx v9, v8, a1
; LMULMAX1-RV32-NEXT: vxor.vi v8, v8, -1
; LMULMAX1-RV32-NEXT: vand.vv v8, v8, v9
; LMULMAX1-RV32-NEXT: vsrl.vi v9, v8, 1
; LMULMAX1-RV32-NEXT: li a1, 85
; LMULMAX1-RV32-NEXT: vand.vx v9, v9, a1
; LMULMAX1-RV32-NEXT: vsub.vv v8, v8, v9
; LMULMAX1-RV32-NEXT: li a1, 51
; LMULMAX1-RV32-NEXT: vand.vx v9, v8, a1
; LMULMAX1-RV32-NEXT: vsrl.vi v8, v8, 2
; LMULMAX1-RV32-NEXT: vand.vx v8, v8, a1
; LMULMAX1-RV32-NEXT: vadd.vv v8, v9, v8
; LMULMAX1-RV32-NEXT: vsrl.vi v9, v8, 4
; LMULMAX1-RV32-NEXT: vadd.vv v8, v8, v9
; LMULMAX1-RV32-NEXT: vand.vi v8, v8, 15
; LMULMAX1-RV32-NEXT: vse8.v v8, (a0)
; LMULMAX1-RV32-NEXT: ret
;
; LMULMAX1-RV64-LABEL: cttz_v16i8:
; LMULMAX1-RV64: # %bb.0:
; LMULMAX1-RV64-NEXT: vsetivli zero, 16, e8, m1, ta, mu
; LMULMAX1-RV64-NEXT: vle8.v v8, (a0)
; LMULMAX1-RV64-NEXT: li a1, 1
; LMULMAX1-RV64-NEXT: vsub.vx v9, v8, a1
; LMULMAX1-RV64-NEXT: vxor.vi v8, v8, -1
; LMULMAX1-RV64-NEXT: vand.vv v8, v8, v9
; LMULMAX1-RV64-NEXT: vsrl.vi v9, v8, 1
; LMULMAX1-RV64-NEXT: li a1, 85
; LMULMAX1-RV64-NEXT: vand.vx v9, v9, a1
; LMULMAX1-RV64-NEXT: vsub.vv v8, v8, v9
; LMULMAX1-RV64-NEXT: li a1, 51
; LMULMAX1-RV64-NEXT: vand.vx v9, v8, a1
; LMULMAX1-RV64-NEXT: vsrl.vi v8, v8, 2
; LMULMAX1-RV64-NEXT: vand.vx v8, v8, a1
; LMULMAX1-RV64-NEXT: vadd.vv v8, v9, v8
; LMULMAX1-RV64-NEXT: vsrl.vi v9, v8, 4
; LMULMAX1-RV64-NEXT: vadd.vv v8, v8, v9
; LMULMAX1-RV64-NEXT: vand.vi v8, v8, 15
; LMULMAX1-RV64-NEXT: vse8.v v8, (a0)
; LMULMAX1-RV64-NEXT: ret
;
; LMULMAX8-RV32-LABEL: cttz_v16i8:
; LMULMAX8-RV32: # %bb.0:
; LMULMAX8-RV32-NEXT: vsetivli zero, 16, e8, m1, ta, mu
; LMULMAX8-RV32-NEXT: vle8.v v8, (a0)
; LMULMAX8-RV32-NEXT: vrsub.vi v9, v8, 0
; LMULMAX8-RV32-NEXT: vand.vv v9, v8, v9
; LMULMAX8-RV32-NEXT: vsetvli zero, zero, e32, m4, ta, mu
; LMULMAX8-RV32-NEXT: vzext.vf4 v12, v9
; LMULMAX8-RV32-NEXT: vfcvt.f.xu.v v12, v12
; LMULMAX8-RV32-NEXT: vsetvli zero, zero, e16, m2, ta, mu
; LMULMAX8-RV32-NEXT: vnsrl.wi v10, v12, 23
; LMULMAX8-RV32-NEXT: vsetvli zero, zero, e8, m1, ta, mu
; LMULMAX8-RV32-NEXT: vnsrl.wi v9, v10, 0
; LMULMAX8-RV32-NEXT: li a1, 127
; LMULMAX8-RV32-NEXT: vmseq.vi v0, v8, 0
; LMULMAX8-RV32-NEXT: vsub.vx v8, v9, a1
; LMULMAX8-RV32-NEXT: vmerge.vim v8, v8, 8, v0
; LMULMAX8-RV32-NEXT: vse8.v v8, (a0)
; LMULMAX8-RV32-NEXT: ret
;
; LMULMAX8-RV64-LABEL: cttz_v16i8:
; LMULMAX8-RV64: # %bb.0:
; LMULMAX8-RV64-NEXT: vsetivli zero, 16, e8, m1, ta, mu
; LMULMAX8-RV64-NEXT: vle8.v v8, (a0)
; LMULMAX8-RV64-NEXT: vrsub.vi v9, v8, 0
; LMULMAX8-RV64-NEXT: vand.vv v9, v8, v9
; LMULMAX8-RV64-NEXT: vsetvli zero, zero, e32, m4, ta, mu
; LMULMAX8-RV64-NEXT: vzext.vf4 v12, v9
; LMULMAX8-RV64-NEXT: vfcvt.f.xu.v v12, v12
; LMULMAX8-RV64-NEXT: vsetvli zero, zero, e16, m2, ta, mu
; LMULMAX8-RV64-NEXT: vnsrl.wi v10, v12, 23
; LMULMAX8-RV64-NEXT: vsetvli zero, zero, e8, m1, ta, mu
; LMULMAX8-RV64-NEXT: vnsrl.wi v9, v10, 0
; LMULMAX8-RV64-NEXT: li a1, 127
; LMULMAX8-RV64-NEXT: vmseq.vi v0, v8, 0
; LMULMAX8-RV64-NEXT: vsub.vx v8, v9, a1
; LMULMAX8-RV64-NEXT: vmerge.vim v8, v8, 8, v0
; LMULMAX8-RV64-NEXT: vse8.v v8, (a0)
; LMULMAX8-RV64-NEXT: ret
%a = load <16 x i8>, <16 x i8>* %x
%b = load <16 x i8>, <16 x i8>* %y
%c = call <16 x i8> @llvm.cttz.v16i8(<16 x i8> %a, i1 false)
store <16 x i8> %c, <16 x i8>* %x
ret void
}
declare <16 x i8> @llvm.cttz.v16i8(<16 x i8>, i1)
define void @cttz_v8i16(<8 x i16>* %x, <8 x i16>* %y) nounwind {
; LMULMAX2-RV32I-LABEL: cttz_v8i16:
; LMULMAX2-RV32I: # %bb.0:
; LMULMAX2-RV32I-NEXT: vsetivli zero, 8, e16, m1, ta, mu
; LMULMAX2-RV32I-NEXT: vle16.v v8, (a0)
; LMULMAX2-RV32I-NEXT: li a1, 1
; LMULMAX2-RV32I-NEXT: vsub.vx v9, v8, a1
; LMULMAX2-RV32I-NEXT: vxor.vi v8, v8, -1
; LMULMAX2-RV32I-NEXT: vand.vv v8, v8, v9
; LMULMAX2-RV32I-NEXT: vsrl.vi v9, v8, 1
; LMULMAX2-RV32I-NEXT: lui a1, 5
; LMULMAX2-RV32I-NEXT: addi a1, a1, 1365
; LMULMAX2-RV32I-NEXT: vand.vx v9, v9, a1
; LMULMAX2-RV32I-NEXT: vsub.vv v8, v8, v9
; LMULMAX2-RV32I-NEXT: lui a1, 3
; LMULMAX2-RV32I-NEXT: addi a1, a1, 819
; LMULMAX2-RV32I-NEXT: vand.vx v9, v8, a1
; LMULMAX2-RV32I-NEXT: vsrl.vi v8, v8, 2
; LMULMAX2-RV32I-NEXT: vand.vx v8, v8, a1
; LMULMAX2-RV32I-NEXT: vadd.vv v8, v9, v8
; LMULMAX2-RV32I-NEXT: vsrl.vi v9, v8, 4
; LMULMAX2-RV32I-NEXT: vadd.vv v8, v8, v9
; LMULMAX2-RV32I-NEXT: lui a1, 1
; LMULMAX2-RV32I-NEXT: addi a1, a1, -241
; LMULMAX2-RV32I-NEXT: vand.vx v8, v8, a1
; LMULMAX2-RV32I-NEXT: li a1, 257
; LMULMAX2-RV32I-NEXT: vmul.vx v8, v8, a1
; LMULMAX2-RV32I-NEXT: vsrl.vi v8, v8, 8
; LMULMAX2-RV32I-NEXT: vse16.v v8, (a0)
; LMULMAX2-RV32I-NEXT: ret
;
; LMULMAX2-RV64I-LABEL: cttz_v8i16:
; LMULMAX2-RV64I: # %bb.0:
; LMULMAX2-RV64I-NEXT: vsetivli zero, 8, e16, m1, ta, mu
; LMULMAX2-RV64I-NEXT: vle16.v v8, (a0)
; LMULMAX2-RV64I-NEXT: li a1, 1
; LMULMAX2-RV64I-NEXT: vsub.vx v9, v8, a1
; LMULMAX2-RV64I-NEXT: vxor.vi v8, v8, -1
; LMULMAX2-RV64I-NEXT: vand.vv v8, v8, v9
; LMULMAX2-RV64I-NEXT: vsrl.vi v9, v8, 1
; LMULMAX2-RV64I-NEXT: lui a1, 5
; LMULMAX2-RV64I-NEXT: addiw a1, a1, 1365
; LMULMAX2-RV64I-NEXT: vand.vx v9, v9, a1
; LMULMAX2-RV64I-NEXT: vsub.vv v8, v8, v9
; LMULMAX2-RV64I-NEXT: lui a1, 3
; LMULMAX2-RV64I-NEXT: addiw a1, a1, 819
; LMULMAX2-RV64I-NEXT: vand.vx v9, v8, a1
; LMULMAX2-RV64I-NEXT: vsrl.vi v8, v8, 2
; LMULMAX2-RV64I-NEXT: vand.vx v8, v8, a1
; LMULMAX2-RV64I-NEXT: vadd.vv v8, v9, v8
; LMULMAX2-RV64I-NEXT: vsrl.vi v9, v8, 4
; LMULMAX2-RV64I-NEXT: vadd.vv v8, v8, v9
; LMULMAX2-RV64I-NEXT: lui a1, 1
; LMULMAX2-RV64I-NEXT: addiw a1, a1, -241
; LMULMAX2-RV64I-NEXT: vand.vx v8, v8, a1
; LMULMAX2-RV64I-NEXT: li a1, 257
; LMULMAX2-RV64I-NEXT: vmul.vx v8, v8, a1
; LMULMAX2-RV64I-NEXT: vsrl.vi v8, v8, 8
; LMULMAX2-RV64I-NEXT: vse16.v v8, (a0)
; LMULMAX2-RV64I-NEXT: ret
;
; LMULMAX1-RV32-LABEL: cttz_v8i16:
; LMULMAX1-RV32: # %bb.0:
; LMULMAX1-RV32-NEXT: vsetivli zero, 8, e16, m1, ta, mu
; LMULMAX1-RV32-NEXT: vle16.v v8, (a0)
; LMULMAX1-RV32-NEXT: li a1, 1
; LMULMAX1-RV32-NEXT: vsub.vx v9, v8, a1
; LMULMAX1-RV32-NEXT: vxor.vi v8, v8, -1
; LMULMAX1-RV32-NEXT: vand.vv v8, v8, v9
; LMULMAX1-RV32-NEXT: vsrl.vi v9, v8, 1
; LMULMAX1-RV32-NEXT: lui a1, 5
; LMULMAX1-RV32-NEXT: addi a1, a1, 1365
; LMULMAX1-RV32-NEXT: vand.vx v9, v9, a1
; LMULMAX1-RV32-NEXT: vsub.vv v8, v8, v9
; LMULMAX1-RV32-NEXT: lui a1, 3
; LMULMAX1-RV32-NEXT: addi a1, a1, 819
; LMULMAX1-RV32-NEXT: vand.vx v9, v8, a1
; LMULMAX1-RV32-NEXT: vsrl.vi v8, v8, 2
; LMULMAX1-RV32-NEXT: vand.vx v8, v8, a1
; LMULMAX1-RV32-NEXT: vadd.vv v8, v9, v8
; LMULMAX1-RV32-NEXT: vsrl.vi v9, v8, 4
; LMULMAX1-RV32-NEXT: vadd.vv v8, v8, v9
; LMULMAX1-RV32-NEXT: lui a1, 1
; LMULMAX1-RV32-NEXT: addi a1, a1, -241
; LMULMAX1-RV32-NEXT: vand.vx v8, v8, a1
; LMULMAX1-RV32-NEXT: li a1, 257
; LMULMAX1-RV32-NEXT: vmul.vx v8, v8, a1
; LMULMAX1-RV32-NEXT: vsrl.vi v8, v8, 8
; LMULMAX1-RV32-NEXT: vse16.v v8, (a0)
; LMULMAX1-RV32-NEXT: ret
;
; LMULMAX1-RV64-LABEL: cttz_v8i16:
; LMULMAX1-RV64: # %bb.0:
; LMULMAX1-RV64-NEXT: vsetivli zero, 8, e16, m1, ta, mu
; LMULMAX1-RV64-NEXT: vle16.v v8, (a0)
; LMULMAX1-RV64-NEXT: li a1, 1
; LMULMAX1-RV64-NEXT: vsub.vx v9, v8, a1
; LMULMAX1-RV64-NEXT: vxor.vi v8, v8, -1
; LMULMAX1-RV64-NEXT: vand.vv v8, v8, v9
; LMULMAX1-RV64-NEXT: vsrl.vi v9, v8, 1
; LMULMAX1-RV64-NEXT: lui a1, 5
; LMULMAX1-RV64-NEXT: addiw a1, a1, 1365
; LMULMAX1-RV64-NEXT: vand.vx v9, v9, a1
; LMULMAX1-RV64-NEXT: vsub.vv v8, v8, v9
; LMULMAX1-RV64-NEXT: lui a1, 3
; LMULMAX1-RV64-NEXT: addiw a1, a1, 819
; LMULMAX1-RV64-NEXT: vand.vx v9, v8, a1
; LMULMAX1-RV64-NEXT: vsrl.vi v8, v8, 2
; LMULMAX1-RV64-NEXT: vand.vx v8, v8, a1
; LMULMAX1-RV64-NEXT: vadd.vv v8, v9, v8
; LMULMAX1-RV64-NEXT: vsrl.vi v9, v8, 4
; LMULMAX1-RV64-NEXT: vadd.vv v8, v8, v9
; LMULMAX1-RV64-NEXT: lui a1, 1
; LMULMAX1-RV64-NEXT: addiw a1, a1, -241
; LMULMAX1-RV64-NEXT: vand.vx v8, v8, a1
; LMULMAX1-RV64-NEXT: li a1, 257
; LMULMAX1-RV64-NEXT: vmul.vx v8, v8, a1
; LMULMAX1-RV64-NEXT: vsrl.vi v8, v8, 8
; LMULMAX1-RV64-NEXT: vse16.v v8, (a0)
; LMULMAX1-RV64-NEXT: ret
;
; LMULMAX2-RV32D-LABEL: cttz_v8i16:
; LMULMAX2-RV32D: # %bb.0:
; LMULMAX2-RV32D-NEXT: vsetivli zero, 8, e16, m1, ta, mu
; LMULMAX2-RV32D-NEXT: vle16.v v8, (a0)
; LMULMAX2-RV32D-NEXT: vrsub.vi v9, v8, 0
; LMULMAX2-RV32D-NEXT: vand.vv v9, v8, v9
; LMULMAX2-RV32D-NEXT: vfwcvt.f.xu.v v10, v9
; LMULMAX2-RV32D-NEXT: vnsrl.wi v9, v10, 23
; LMULMAX2-RV32D-NEXT: li a1, 127
; LMULMAX2-RV32D-NEXT: vsub.vx v9, v9, a1
; LMULMAX2-RV32D-NEXT: vmseq.vi v0, v8, 0
; LMULMAX2-RV32D-NEXT: li a1, 16
; LMULMAX2-RV32D-NEXT: vmerge.vxm v8, v9, a1, v0
; LMULMAX2-RV32D-NEXT: vse16.v v8, (a0)
; LMULMAX2-RV32D-NEXT: ret
;
; LMULMAX2-RV64D-LABEL: cttz_v8i16:
; LMULMAX2-RV64D: # %bb.0:
; LMULMAX2-RV64D-NEXT: vsetivli zero, 8, e16, m1, ta, mu
; LMULMAX2-RV64D-NEXT: vle16.v v8, (a0)
; LMULMAX2-RV64D-NEXT: vrsub.vi v9, v8, 0
; LMULMAX2-RV64D-NEXT: vand.vv v9, v8, v9
; LMULMAX2-RV64D-NEXT: vfwcvt.f.xu.v v10, v9
; LMULMAX2-RV64D-NEXT: vnsrl.wi v9, v10, 23
; LMULMAX2-RV64D-NEXT: li a1, 127
; LMULMAX2-RV64D-NEXT: vsub.vx v9, v9, a1
; LMULMAX2-RV64D-NEXT: vmseq.vi v0, v8, 0
; LMULMAX2-RV64D-NEXT: li a1, 16
; LMULMAX2-RV64D-NEXT: vmerge.vxm v8, v9, a1, v0
; LMULMAX2-RV64D-NEXT: vse16.v v8, (a0)
; LMULMAX2-RV64D-NEXT: ret
;
; LMULMAX8-RV32-LABEL: cttz_v8i16:
; LMULMAX8-RV32: # %bb.0:
; LMULMAX8-RV32-NEXT: vsetivli zero, 8, e16, m1, ta, mu
; LMULMAX8-RV32-NEXT: vle16.v v8, (a0)
; LMULMAX8-RV32-NEXT: vrsub.vi v9, v8, 0
; LMULMAX8-RV32-NEXT: vand.vv v9, v8, v9
; LMULMAX8-RV32-NEXT: vfwcvt.f.xu.v v10, v9
; LMULMAX8-RV32-NEXT: vnsrl.wi v9, v10, 23
; LMULMAX8-RV32-NEXT: li a1, 127
; LMULMAX8-RV32-NEXT: vsub.vx v9, v9, a1
; LMULMAX8-RV32-NEXT: vmseq.vi v0, v8, 0
; LMULMAX8-RV32-NEXT: li a1, 16
; LMULMAX8-RV32-NEXT: vmerge.vxm v8, v9, a1, v0
; LMULMAX8-RV32-NEXT: vse16.v v8, (a0)
; LMULMAX8-RV32-NEXT: ret
;
; LMULMAX8-RV64-LABEL: cttz_v8i16:
; LMULMAX8-RV64: # %bb.0:
; LMULMAX8-RV64-NEXT: vsetivli zero, 8, e16, m1, ta, mu
; LMULMAX8-RV64-NEXT: vle16.v v8, (a0)
; LMULMAX8-RV64-NEXT: vrsub.vi v9, v8, 0
; LMULMAX8-RV64-NEXT: vand.vv v9, v8, v9
; LMULMAX8-RV64-NEXT: vfwcvt.f.xu.v v10, v9
; LMULMAX8-RV64-NEXT: vnsrl.wi v9, v10, 23
; LMULMAX8-RV64-NEXT: li a1, 127
; LMULMAX8-RV64-NEXT: vsub.vx v9, v9, a1
; LMULMAX8-RV64-NEXT: vmseq.vi v0, v8, 0
; LMULMAX8-RV64-NEXT: li a1, 16
; LMULMAX8-RV64-NEXT: vmerge.vxm v8, v9, a1, v0
; LMULMAX8-RV64-NEXT: vse16.v v8, (a0)
; LMULMAX8-RV64-NEXT: ret
%a = load <8 x i16>, <8 x i16>* %x
%b = load <8 x i16>, <8 x i16>* %y
%c = call <8 x i16> @llvm.cttz.v8i16(<8 x i16> %a, i1 false)
store <8 x i16> %c, <8 x i16>* %x
ret void
}
declare <8 x i16> @llvm.cttz.v8i16(<8 x i16>, i1)
define void @cttz_v4i32(<4 x i32>* %x, <4 x i32>* %y) nounwind {
; LMULMAX2-RV32I-LABEL: cttz_v4i32:
; LMULMAX2-RV32I: # %bb.0:
; LMULMAX2-RV32I-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX2-RV32I-NEXT: vle32.v v8, (a0)
; LMULMAX2-RV32I-NEXT: li a1, 1
; LMULMAX2-RV32I-NEXT: vsub.vx v9, v8, a1
; LMULMAX2-RV32I-NEXT: vxor.vi v8, v8, -1
; LMULMAX2-RV32I-NEXT: vand.vv v8, v8, v9
; LMULMAX2-RV32I-NEXT: vsrl.vi v9, v8, 1
; LMULMAX2-RV32I-NEXT: lui a1, 349525
; LMULMAX2-RV32I-NEXT: addi a1, a1, 1365
; LMULMAX2-RV32I-NEXT: vand.vx v9, v9, a1
; LMULMAX2-RV32I-NEXT: vsub.vv v8, v8, v9
; LMULMAX2-RV32I-NEXT: lui a1, 209715
; LMULMAX2-RV32I-NEXT: addi a1, a1, 819
; LMULMAX2-RV32I-NEXT: vand.vx v9, v8, a1
; LMULMAX2-RV32I-NEXT: vsrl.vi v8, v8, 2
; LMULMAX2-RV32I-NEXT: vand.vx v8, v8, a1
; LMULMAX2-RV32I-NEXT: vadd.vv v8, v9, v8
; LMULMAX2-RV32I-NEXT: vsrl.vi v9, v8, 4
; LMULMAX2-RV32I-NEXT: vadd.vv v8, v8, v9
; LMULMAX2-RV32I-NEXT: lui a1, 61681
; LMULMAX2-RV32I-NEXT: addi a1, a1, -241
; LMULMAX2-RV32I-NEXT: vand.vx v8, v8, a1
; LMULMAX2-RV32I-NEXT: lui a1, 4112
; LMULMAX2-RV32I-NEXT: addi a1, a1, 257
; LMULMAX2-RV32I-NEXT: vmul.vx v8, v8, a1
; LMULMAX2-RV32I-NEXT: vsrl.vi v8, v8, 24
; LMULMAX2-RV32I-NEXT: vse32.v v8, (a0)
; LMULMAX2-RV32I-NEXT: ret
;
; LMULMAX2-RV64I-LABEL: cttz_v4i32:
; LMULMAX2-RV64I: # %bb.0:
; LMULMAX2-RV64I-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX2-RV64I-NEXT: vle32.v v8, (a0)
; LMULMAX2-RV64I-NEXT: li a1, 1
; LMULMAX2-RV64I-NEXT: vsub.vx v9, v8, a1
; LMULMAX2-RV64I-NEXT: vxor.vi v8, v8, -1
; LMULMAX2-RV64I-NEXT: vand.vv v8, v8, v9
; LMULMAX2-RV64I-NEXT: vsrl.vi v9, v8, 1
; LMULMAX2-RV64I-NEXT: lui a1, 349525
; LMULMAX2-RV64I-NEXT: addiw a1, a1, 1365
; LMULMAX2-RV64I-NEXT: vand.vx v9, v9, a1
; LMULMAX2-RV64I-NEXT: vsub.vv v8, v8, v9
; LMULMAX2-RV64I-NEXT: lui a1, 209715
; LMULMAX2-RV64I-NEXT: addiw a1, a1, 819
; LMULMAX2-RV64I-NEXT: vand.vx v9, v8, a1
; LMULMAX2-RV64I-NEXT: vsrl.vi v8, v8, 2
; LMULMAX2-RV64I-NEXT: vand.vx v8, v8, a1
; LMULMAX2-RV64I-NEXT: vadd.vv v8, v9, v8
; LMULMAX2-RV64I-NEXT: vsrl.vi v9, v8, 4
; LMULMAX2-RV64I-NEXT: vadd.vv v8, v8, v9
; LMULMAX2-RV64I-NEXT: lui a1, 61681
; LMULMAX2-RV64I-NEXT: addiw a1, a1, -241
; LMULMAX2-RV64I-NEXT: vand.vx v8, v8, a1
; LMULMAX2-RV64I-NEXT: lui a1, 4112
; LMULMAX2-RV64I-NEXT: addiw a1, a1, 257
; LMULMAX2-RV64I-NEXT: vmul.vx v8, v8, a1
; LMULMAX2-RV64I-NEXT: vsrl.vi v8, v8, 24
; LMULMAX2-RV64I-NEXT: vse32.v v8, (a0)
; LMULMAX2-RV64I-NEXT: ret
;
; LMULMAX1-RV32-LABEL: cttz_v4i32:
; LMULMAX1-RV32: # %bb.0:
; LMULMAX1-RV32-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX1-RV32-NEXT: vle32.v v8, (a0)
; LMULMAX1-RV32-NEXT: li a1, 1
; LMULMAX1-RV32-NEXT: vsub.vx v9, v8, a1
; LMULMAX1-RV32-NEXT: vxor.vi v8, v8, -1
; LMULMAX1-RV32-NEXT: vand.vv v8, v8, v9
; LMULMAX1-RV32-NEXT: vsrl.vi v9, v8, 1
; LMULMAX1-RV32-NEXT: lui a1, 349525
; LMULMAX1-RV32-NEXT: addi a1, a1, 1365
; LMULMAX1-RV32-NEXT: vand.vx v9, v9, a1
; LMULMAX1-RV32-NEXT: vsub.vv v8, v8, v9
; LMULMAX1-RV32-NEXT: lui a1, 209715
; LMULMAX1-RV32-NEXT: addi a1, a1, 819
; LMULMAX1-RV32-NEXT: vand.vx v9, v8, a1
; LMULMAX1-RV32-NEXT: vsrl.vi v8, v8, 2
; LMULMAX1-RV32-NEXT: vand.vx v8, v8, a1
; LMULMAX1-RV32-NEXT: vadd.vv v8, v9, v8
; LMULMAX1-RV32-NEXT: vsrl.vi v9, v8, 4
; LMULMAX1-RV32-NEXT: vadd.vv v8, v8, v9
; LMULMAX1-RV32-NEXT: lui a1, 61681
; LMULMAX1-RV32-NEXT: addi a1, a1, -241
; LMULMAX1-RV32-NEXT: vand.vx v8, v8, a1
; LMULMAX1-RV32-NEXT: lui a1, 4112
; LMULMAX1-RV32-NEXT: addi a1, a1, 257
; LMULMAX1-RV32-NEXT: vmul.vx v8, v8, a1
; LMULMAX1-RV32-NEXT: vsrl.vi v8, v8, 24
; LMULMAX1-RV32-NEXT: vse32.v v8, (a0)
; LMULMAX1-RV32-NEXT: ret
;
; LMULMAX1-RV64-LABEL: cttz_v4i32:
; LMULMAX1-RV64: # %bb.0:
; LMULMAX1-RV64-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX1-RV64-NEXT: vle32.v v8, (a0)
; LMULMAX1-RV64-NEXT: li a1, 1
; LMULMAX1-RV64-NEXT: vsub.vx v9, v8, a1
; LMULMAX1-RV64-NEXT: vxor.vi v8, v8, -1
; LMULMAX1-RV64-NEXT: vand.vv v8, v8, v9
; LMULMAX1-RV64-NEXT: vsrl.vi v9, v8, 1
; LMULMAX1-RV64-NEXT: lui a1, 349525
; LMULMAX1-RV64-NEXT: addiw a1, a1, 1365
; LMULMAX1-RV64-NEXT: vand.vx v9, v9, a1
; LMULMAX1-RV64-NEXT: vsub.vv v8, v8, v9
; LMULMAX1-RV64-NEXT: lui a1, 209715
; LMULMAX1-RV64-NEXT: addiw a1, a1, 819
; LMULMAX1-RV64-NEXT: vand.vx v9, v8, a1
; LMULMAX1-RV64-NEXT: vsrl.vi v8, v8, 2
; LMULMAX1-RV64-NEXT: vand.vx v8, v8, a1
; LMULMAX1-RV64-NEXT: vadd.vv v8, v9, v8
; LMULMAX1-RV64-NEXT: vsrl.vi v9, v8, 4
; LMULMAX1-RV64-NEXT: vadd.vv v8, v8, v9
; LMULMAX1-RV64-NEXT: lui a1, 61681
; LMULMAX1-RV64-NEXT: addiw a1, a1, -241
; LMULMAX1-RV64-NEXT: vand.vx v8, v8, a1
; LMULMAX1-RV64-NEXT: lui a1, 4112
; LMULMAX1-RV64-NEXT: addiw a1, a1, 257
; LMULMAX1-RV64-NEXT: vmul.vx v8, v8, a1
; LMULMAX1-RV64-NEXT: vsrl.vi v8, v8, 24
; LMULMAX1-RV64-NEXT: vse32.v v8, (a0)
; LMULMAX1-RV64-NEXT: ret
;
; LMULMAX2-RV32D-LABEL: cttz_v4i32:
; LMULMAX2-RV32D: # %bb.0:
; LMULMAX2-RV32D-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX2-RV32D-NEXT: vle32.v v8, (a0)
; LMULMAX2-RV32D-NEXT: vrsub.vi v9, v8, 0
; LMULMAX2-RV32D-NEXT: vand.vv v9, v8, v9
; LMULMAX2-RV32D-NEXT: vfwcvt.f.xu.v v10, v9
; LMULMAX2-RV32D-NEXT: li a1, 52
; LMULMAX2-RV32D-NEXT: vnsrl.wx v9, v10, a1
; LMULMAX2-RV32D-NEXT: li a1, 1023
; LMULMAX2-RV32D-NEXT: vsub.vx v9, v9, a1
; LMULMAX2-RV32D-NEXT: vmseq.vi v0, v8, 0
; LMULMAX2-RV32D-NEXT: li a1, 32
; LMULMAX2-RV32D-NEXT: vmerge.vxm v8, v9, a1, v0
; LMULMAX2-RV32D-NEXT: vse32.v v8, (a0)
; LMULMAX2-RV32D-NEXT: ret
;
; LMULMAX2-RV64D-LABEL: cttz_v4i32:
; LMULMAX2-RV64D: # %bb.0:
; LMULMAX2-RV64D-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX2-RV64D-NEXT: vle32.v v8, (a0)
; LMULMAX2-RV64D-NEXT: vrsub.vi v9, v8, 0
; LMULMAX2-RV64D-NEXT: vand.vv v9, v8, v9
; LMULMAX2-RV64D-NEXT: vfwcvt.f.xu.v v10, v9
; LMULMAX2-RV64D-NEXT: li a1, 52
; LMULMAX2-RV64D-NEXT: vnsrl.wx v9, v10, a1
; LMULMAX2-RV64D-NEXT: li a1, 1023
; LMULMAX2-RV64D-NEXT: vsub.vx v9, v9, a1
; LMULMAX2-RV64D-NEXT: vmseq.vi v0, v8, 0
; LMULMAX2-RV64D-NEXT: li a1, 32
; LMULMAX2-RV64D-NEXT: vmerge.vxm v8, v9, a1, v0
; LMULMAX2-RV64D-NEXT: vse32.v v8, (a0)
; LMULMAX2-RV64D-NEXT: ret
;
; LMULMAX8-RV32-LABEL: cttz_v4i32:
; LMULMAX8-RV32: # %bb.0:
; LMULMAX8-RV32-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX8-RV32-NEXT: vle32.v v8, (a0)
; LMULMAX8-RV32-NEXT: vrsub.vi v9, v8, 0
; LMULMAX8-RV32-NEXT: vand.vv v9, v8, v9
; LMULMAX8-RV32-NEXT: vfwcvt.f.xu.v v10, v9
; LMULMAX8-RV32-NEXT: li a1, 52
; LMULMAX8-RV32-NEXT: vnsrl.wx v9, v10, a1
; LMULMAX8-RV32-NEXT: li a1, 1023
; LMULMAX8-RV32-NEXT: vsub.vx v9, v9, a1
; LMULMAX8-RV32-NEXT: vmseq.vi v0, v8, 0
; LMULMAX8-RV32-NEXT: li a1, 32
; LMULMAX8-RV32-NEXT: vmerge.vxm v8, v9, a1, v0
; LMULMAX8-RV32-NEXT: vse32.v v8, (a0)
; LMULMAX8-RV32-NEXT: ret
;
; LMULMAX8-RV64-LABEL: cttz_v4i32:
; LMULMAX8-RV64: # %bb.0:
; LMULMAX8-RV64-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX8-RV64-NEXT: vle32.v v8, (a0)
; LMULMAX8-RV64-NEXT: vrsub.vi v9, v8, 0
; LMULMAX8-RV64-NEXT: vand.vv v9, v8, v9
; LMULMAX8-RV64-NEXT: vfwcvt.f.xu.v v10, v9
; LMULMAX8-RV64-NEXT: li a1, 52
; LMULMAX8-RV64-NEXT: vnsrl.wx v9, v10, a1
; LMULMAX8-RV64-NEXT: li a1, 1023
; LMULMAX8-RV64-NEXT: vsub.vx v9, v9, a1
; LMULMAX8-RV64-NEXT: vmseq.vi v0, v8, 0
; LMULMAX8-RV64-NEXT: li a1, 32
; LMULMAX8-RV64-NEXT: vmerge.vxm v8, v9, a1, v0
; LMULMAX8-RV64-NEXT: vse32.v v8, (a0)
; LMULMAX8-RV64-NEXT: ret
%a = load <4 x i32>, <4 x i32>* %x
%b = load <4 x i32>, <4 x i32>* %y
%c = call <4 x i32> @llvm.cttz.v4i32(<4 x i32> %a, i1 false)
store <4 x i32> %c, <4 x i32>* %x
ret void
}
declare <4 x i32> @llvm.cttz.v4i32(<4 x i32>, i1)
define void @cttz_v2i64(<2 x i64>* %x, <2 x i64>* %y) nounwind {
; LMULMAX2-RV32-LABEL: cttz_v2i64:
; LMULMAX2-RV32: # %bb.0:
; LMULMAX2-RV32-NEXT: vsetivli zero, 2, e64, m1, ta, mu
; LMULMAX2-RV32-NEXT: vle64.v v8, (a0)
; LMULMAX2-RV32-NEXT: li a1, 1
; LMULMAX2-RV32-NEXT: vsub.vx v9, v8, a1
; LMULMAX2-RV32-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX2-RV32-NEXT: vmv.v.i v10, -1
; LMULMAX2-RV32-NEXT: vsetivli zero, 2, e64, m1, ta, mu
; LMULMAX2-RV32-NEXT: vxor.vv v8, v8, v10
; LMULMAX2-RV32-NEXT: vand.vv v8, v8, v9
; LMULMAX2-RV32-NEXT: vsrl.vi v9, v8, 1
; LMULMAX2-RV32-NEXT: lui a1, 349525
; LMULMAX2-RV32-NEXT: addi a1, a1, 1365
; LMULMAX2-RV32-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX2-RV32-NEXT: vmv.v.x v10, a1
; LMULMAX2-RV32-NEXT: vsetivli zero, 2, e64, m1, ta, mu
; LMULMAX2-RV32-NEXT: vand.vv v9, v9, v10
; LMULMAX2-RV32-NEXT: vsub.vv v8, v8, v9
; LMULMAX2-RV32-NEXT: lui a1, 209715
; LMULMAX2-RV32-NEXT: addi a1, a1, 819
; LMULMAX2-RV32-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX2-RV32-NEXT: vmv.v.x v9, a1
; LMULMAX2-RV32-NEXT: vsetivli zero, 2, e64, m1, ta, mu
; LMULMAX2-RV32-NEXT: vand.vv v10, v8, v9
; LMULMAX2-RV32-NEXT: vsrl.vi v8, v8, 2
; LMULMAX2-RV32-NEXT: vand.vv v8, v8, v9
; LMULMAX2-RV32-NEXT: vadd.vv v8, v10, v8
; LMULMAX2-RV32-NEXT: vsrl.vi v9, v8, 4
; LMULMAX2-RV32-NEXT: vadd.vv v8, v8, v9
; LMULMAX2-RV32-NEXT: lui a1, 61681
; LMULMAX2-RV32-NEXT: addi a1, a1, -241
; LMULMAX2-RV32-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX2-RV32-NEXT: vmv.v.x v9, a1
; LMULMAX2-RV32-NEXT: vsetivli zero, 2, e64, m1, ta, mu
; LMULMAX2-RV32-NEXT: vand.vv v8, v8, v9
; LMULMAX2-RV32-NEXT: lui a1, 4112
; LMULMAX2-RV32-NEXT: addi a1, a1, 257
; LMULMAX2-RV32-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX2-RV32-NEXT: vmv.v.x v9, a1
; LMULMAX2-RV32-NEXT: vsetivli zero, 2, e64, m1, ta, mu
; LMULMAX2-RV32-NEXT: vmul.vv v8, v8, v9
; LMULMAX2-RV32-NEXT: li a1, 56
; LMULMAX2-RV32-NEXT: vsrl.vx v8, v8, a1
; LMULMAX2-RV32-NEXT: vse64.v v8, (a0)
; LMULMAX2-RV32-NEXT: ret
;
; LMULMAX2-RV64-LABEL: cttz_v2i64:
; LMULMAX2-RV64: # %bb.0:
; LMULMAX2-RV64-NEXT: vsetivli zero, 2, e64, m1, ta, mu
; LMULMAX2-RV64-NEXT: vle64.v v8, (a0)
; LMULMAX2-RV64-NEXT: li a1, 1
; LMULMAX2-RV64-NEXT: vsub.vx v9, v8, a1
; LMULMAX2-RV64-NEXT: vxor.vi v8, v8, -1
; LMULMAX2-RV64-NEXT: vand.vv v8, v8, v9
; LMULMAX2-RV64-NEXT: vsrl.vi v9, v8, 1
; LMULMAX2-RV64-NEXT: lui a1, 21845
; LMULMAX2-RV64-NEXT: addiw a1, a1, 1365
; LMULMAX2-RV64-NEXT: slli a1, a1, 12
; LMULMAX2-RV64-NEXT: addi a1, a1, 1365
; LMULMAX2-RV64-NEXT: slli a1, a1, 12
; LMULMAX2-RV64-NEXT: addi a1, a1, 1365
; LMULMAX2-RV64-NEXT: slli a1, a1, 12
; LMULMAX2-RV64-NEXT: addi a1, a1, 1365
; LMULMAX2-RV64-NEXT: vand.vx v9, v9, a1
; LMULMAX2-RV64-NEXT: vsub.vv v8, v8, v9
; LMULMAX2-RV64-NEXT: lui a1, 13107
; LMULMAX2-RV64-NEXT: addiw a1, a1, 819
; LMULMAX2-RV64-NEXT: slli a1, a1, 12
; LMULMAX2-RV64-NEXT: addi a1, a1, 819
; LMULMAX2-RV64-NEXT: slli a1, a1, 12
; LMULMAX2-RV64-NEXT: addi a1, a1, 819
; LMULMAX2-RV64-NEXT: slli a1, a1, 12
; LMULMAX2-RV64-NEXT: addi a1, a1, 819
; LMULMAX2-RV64-NEXT: vand.vx v9, v8, a1
; LMULMAX2-RV64-NEXT: vsrl.vi v8, v8, 2
; LMULMAX2-RV64-NEXT: vand.vx v8, v8, a1
; LMULMAX2-RV64-NEXT: vadd.vv v8, v9, v8
; LMULMAX2-RV64-NEXT: vsrl.vi v9, v8, 4
; LMULMAX2-RV64-NEXT: vadd.vv v8, v8, v9
; LMULMAX2-RV64-NEXT: lui a1, 3855
; LMULMAX2-RV64-NEXT: addiw a1, a1, 241
; LMULMAX2-RV64-NEXT: slli a1, a1, 12
; LMULMAX2-RV64-NEXT: addi a1, a1, -241
; LMULMAX2-RV64-NEXT: slli a1, a1, 12
; LMULMAX2-RV64-NEXT: addi a1, a1, 241
; LMULMAX2-RV64-NEXT: slli a1, a1, 12
; LMULMAX2-RV64-NEXT: addi a1, a1, -241
; LMULMAX2-RV64-NEXT: vand.vx v8, v8, a1
; LMULMAX2-RV64-NEXT: lui a1, 4112
; LMULMAX2-RV64-NEXT: addiw a1, a1, 257
; LMULMAX2-RV64-NEXT: slli a1, a1, 16
; LMULMAX2-RV64-NEXT: addi a1, a1, 257
; LMULMAX2-RV64-NEXT: slli a1, a1, 16
; LMULMAX2-RV64-NEXT: addi a1, a1, 257
; LMULMAX2-RV64-NEXT: vmul.vx v8, v8, a1
; LMULMAX2-RV64-NEXT: li a1, 56
; LMULMAX2-RV64-NEXT: vsrl.vx v8, v8, a1
; LMULMAX2-RV64-NEXT: vse64.v v8, (a0)
; LMULMAX2-RV64-NEXT: ret
;
; LMULMAX1-RV32-LABEL: cttz_v2i64:
; LMULMAX1-RV32: # %bb.0:
; LMULMAX1-RV32-NEXT: vsetivli zero, 2, e64, m1, ta, mu
; LMULMAX1-RV32-NEXT: vle64.v v8, (a0)
; LMULMAX1-RV32-NEXT: li a1, 1
; LMULMAX1-RV32-NEXT: vsub.vx v9, v8, a1
; LMULMAX1-RV32-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX1-RV32-NEXT: vmv.v.i v10, -1
; LMULMAX1-RV32-NEXT: vsetivli zero, 2, e64, m1, ta, mu
; LMULMAX1-RV32-NEXT: vxor.vv v8, v8, v10
; LMULMAX1-RV32-NEXT: vand.vv v8, v8, v9
; LMULMAX1-RV32-NEXT: vsrl.vi v9, v8, 1
; LMULMAX1-RV32-NEXT: lui a1, 349525
; LMULMAX1-RV32-NEXT: addi a1, a1, 1365
; LMULMAX1-RV32-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX1-RV32-NEXT: vmv.v.x v10, a1
; LMULMAX1-RV32-NEXT: vsetivli zero, 2, e64, m1, ta, mu
; LMULMAX1-RV32-NEXT: vand.vv v9, v9, v10
; LMULMAX1-RV32-NEXT: vsub.vv v8, v8, v9
; LMULMAX1-RV32-NEXT: lui a1, 209715
; LMULMAX1-RV32-NEXT: addi a1, a1, 819
; LMULMAX1-RV32-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX1-RV32-NEXT: vmv.v.x v9, a1
; LMULMAX1-RV32-NEXT: vsetivli zero, 2, e64, m1, ta, mu
; LMULMAX1-RV32-NEXT: vand.vv v10, v8, v9
; LMULMAX1-RV32-NEXT: vsrl.vi v8, v8, 2
; LMULMAX1-RV32-NEXT: vand.vv v8, v8, v9
; LMULMAX1-RV32-NEXT: vadd.vv v8, v10, v8
; LMULMAX1-RV32-NEXT: vsrl.vi v9, v8, 4
; LMULMAX1-RV32-NEXT: vadd.vv v8, v8, v9
; LMULMAX1-RV32-NEXT: lui a1, 61681
; LMULMAX1-RV32-NEXT: addi a1, a1, -241
; LMULMAX1-RV32-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX1-RV32-NEXT: vmv.v.x v9, a1
; LMULMAX1-RV32-NEXT: vsetivli zero, 2, e64, m1, ta, mu
; LMULMAX1-RV32-NEXT: vand.vv v8, v8, v9
; LMULMAX1-RV32-NEXT: lui a1, 4112
; LMULMAX1-RV32-NEXT: addi a1, a1, 257
; LMULMAX1-RV32-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX1-RV32-NEXT: vmv.v.x v9, a1
; LMULMAX1-RV32-NEXT: vsetivli zero, 2, e64, m1, ta, mu
; LMULMAX1-RV32-NEXT: vmul.vv v8, v8, v9
; LMULMAX1-RV32-NEXT: li a1, 56
; LMULMAX1-RV32-NEXT: vsrl.vx v8, v8, a1
; LMULMAX1-RV32-NEXT: vse64.v v8, (a0)
; LMULMAX1-RV32-NEXT: ret
;
; LMULMAX1-RV64-LABEL: cttz_v2i64:
; LMULMAX1-RV64: # %bb.0:
; LMULMAX1-RV64-NEXT: vsetivli zero, 2, e64, m1, ta, mu
; LMULMAX1-RV64-NEXT: vle64.v v8, (a0)
; LMULMAX1-RV64-NEXT: li a1, 1
; LMULMAX1-RV64-NEXT: vsub.vx v9, v8, a1
; LMULMAX1-RV64-NEXT: vxor.vi v8, v8, -1
; LMULMAX1-RV64-NEXT: vand.vv v8, v8, v9
; LMULMAX1-RV64-NEXT: vsrl.vi v9, v8, 1
; LMULMAX1-RV64-NEXT: lui a1, 21845
; LMULMAX1-RV64-NEXT: addiw a1, a1, 1365
; LMULMAX1-RV64-NEXT: slli a1, a1, 12
; LMULMAX1-RV64-NEXT: addi a1, a1, 1365
; LMULMAX1-RV64-NEXT: slli a1, a1, 12
; LMULMAX1-RV64-NEXT: addi a1, a1, 1365
; LMULMAX1-RV64-NEXT: slli a1, a1, 12
; LMULMAX1-RV64-NEXT: addi a1, a1, 1365
; LMULMAX1-RV64-NEXT: vand.vx v9, v9, a1
; LMULMAX1-RV64-NEXT: vsub.vv v8, v8, v9
; LMULMAX1-RV64-NEXT: lui a1, 13107
; LMULMAX1-RV64-NEXT: addiw a1, a1, 819
; LMULMAX1-RV64-NEXT: slli a1, a1, 12
; LMULMAX1-RV64-NEXT: addi a1, a1, 819
; LMULMAX1-RV64-NEXT: slli a1, a1, 12
; LMULMAX1-RV64-NEXT: addi a1, a1, 819
; LMULMAX1-RV64-NEXT: slli a1, a1, 12
; LMULMAX1-RV64-NEXT: addi a1, a1, 819
; LMULMAX1-RV64-NEXT: vand.vx v9, v8, a1
; LMULMAX1-RV64-NEXT: vsrl.vi v8, v8, 2
; LMULMAX1-RV64-NEXT: vand.vx v8, v8, a1
; LMULMAX1-RV64-NEXT: vadd.vv v8, v9, v8
; LMULMAX1-RV64-NEXT: vsrl.vi v9, v8, 4
; LMULMAX1-RV64-NEXT: vadd.vv v8, v8, v9
; LMULMAX1-RV64-NEXT: lui a1, 3855
; LMULMAX1-RV64-NEXT: addiw a1, a1, 241
; LMULMAX1-RV64-NEXT: slli a1, a1, 12
; LMULMAX1-RV64-NEXT: addi a1, a1, -241
; LMULMAX1-RV64-NEXT: slli a1, a1, 12
; LMULMAX1-RV64-NEXT: addi a1, a1, 241
; LMULMAX1-RV64-NEXT: slli a1, a1, 12
; LMULMAX1-RV64-NEXT: addi a1, a1, -241
; LMULMAX1-RV64-NEXT: vand.vx v8, v8, a1
; LMULMAX1-RV64-NEXT: lui a1, 4112
; LMULMAX1-RV64-NEXT: addiw a1, a1, 257
; LMULMAX1-RV64-NEXT: slli a1, a1, 16
; LMULMAX1-RV64-NEXT: addi a1, a1, 257
; LMULMAX1-RV64-NEXT: slli a1, a1, 16
; LMULMAX1-RV64-NEXT: addi a1, a1, 257
; LMULMAX1-RV64-NEXT: vmul.vx v8, v8, a1
; LMULMAX1-RV64-NEXT: li a1, 56
; LMULMAX1-RV64-NEXT: vsrl.vx v8, v8, a1
; LMULMAX1-RV64-NEXT: vse64.v v8, (a0)
; LMULMAX1-RV64-NEXT: ret
;
; LMULMAX8-RV32-LABEL: cttz_v2i64:
; LMULMAX8-RV32: # %bb.0:
; LMULMAX8-RV32-NEXT: vsetivli zero, 2, e64, m1, ta, mu
; LMULMAX8-RV32-NEXT: vle64.v v8, (a0)
; LMULMAX8-RV32-NEXT: li a1, 1
; LMULMAX8-RV32-NEXT: vsub.vx v9, v8, a1
; LMULMAX8-RV32-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX8-RV32-NEXT: vmv.v.i v10, -1
; LMULMAX8-RV32-NEXT: vsetivli zero, 2, e64, m1, ta, mu
; LMULMAX8-RV32-NEXT: vxor.vv v8, v8, v10
; LMULMAX8-RV32-NEXT: vand.vv v8, v8, v9
; LMULMAX8-RV32-NEXT: vsrl.vi v9, v8, 1
; LMULMAX8-RV32-NEXT: lui a1, 349525
; LMULMAX8-RV32-NEXT: addi a1, a1, 1365
; LMULMAX8-RV32-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX8-RV32-NEXT: vmv.v.x v10, a1
; LMULMAX8-RV32-NEXT: vsetivli zero, 2, e64, m1, ta, mu
; LMULMAX8-RV32-NEXT: vand.vv v9, v9, v10
; LMULMAX8-RV32-NEXT: vsub.vv v8, v8, v9
; LMULMAX8-RV32-NEXT: lui a1, 209715
; LMULMAX8-RV32-NEXT: addi a1, a1, 819
; LMULMAX8-RV32-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX8-RV32-NEXT: vmv.v.x v9, a1
; LMULMAX8-RV32-NEXT: vsetivli zero, 2, e64, m1, ta, mu
; LMULMAX8-RV32-NEXT: vand.vv v10, v8, v9
; LMULMAX8-RV32-NEXT: vsrl.vi v8, v8, 2
; LMULMAX8-RV32-NEXT: vand.vv v8, v8, v9
; LMULMAX8-RV32-NEXT: vadd.vv v8, v10, v8
; LMULMAX8-RV32-NEXT: vsrl.vi v9, v8, 4
; LMULMAX8-RV32-NEXT: vadd.vv v8, v8, v9
; LMULMAX8-RV32-NEXT: lui a1, 61681
; LMULMAX8-RV32-NEXT: addi a1, a1, -241
; LMULMAX8-RV32-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX8-RV32-NEXT: vmv.v.x v9, a1
; LMULMAX8-RV32-NEXT: vsetivli zero, 2, e64, m1, ta, mu
; LMULMAX8-RV32-NEXT: vand.vv v8, v8, v9
; LMULMAX8-RV32-NEXT: lui a1, 4112
; LMULMAX8-RV32-NEXT: addi a1, a1, 257
; LMULMAX8-RV32-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX8-RV32-NEXT: vmv.v.x v9, a1
; LMULMAX8-RV32-NEXT: vsetivli zero, 2, e64, m1, ta, mu
; LMULMAX8-RV32-NEXT: vmul.vv v8, v8, v9
; LMULMAX8-RV32-NEXT: li a1, 56
; LMULMAX8-RV32-NEXT: vsrl.vx v8, v8, a1
; LMULMAX8-RV32-NEXT: vse64.v v8, (a0)
; LMULMAX8-RV32-NEXT: ret
;
; LMULMAX8-RV64-LABEL: cttz_v2i64:
; LMULMAX8-RV64: # %bb.0:
; LMULMAX8-RV64-NEXT: vsetivli zero, 2, e64, m1, ta, mu
; LMULMAX8-RV64-NEXT: vle64.v v8, (a0)
; LMULMAX8-RV64-NEXT: li a1, 1
; LMULMAX8-RV64-NEXT: vsub.vx v9, v8, a1
; LMULMAX8-RV64-NEXT: vxor.vi v8, v8, -1
; LMULMAX8-RV64-NEXT: vand.vv v8, v8, v9
; LMULMAX8-RV64-NEXT: vsrl.vi v9, v8, 1
; LMULMAX8-RV64-NEXT: lui a1, 21845
; LMULMAX8-RV64-NEXT: addiw a1, a1, 1365
; LMULMAX8-RV64-NEXT: slli a1, a1, 12
; LMULMAX8-RV64-NEXT: addi a1, a1, 1365
; LMULMAX8-RV64-NEXT: slli a1, a1, 12
; LMULMAX8-RV64-NEXT: addi a1, a1, 1365
; LMULMAX8-RV64-NEXT: slli a1, a1, 12
; LMULMAX8-RV64-NEXT: addi a1, a1, 1365
; LMULMAX8-RV64-NEXT: vand.vx v9, v9, a1
; LMULMAX8-RV64-NEXT: vsub.vv v8, v8, v9
; LMULMAX8-RV64-NEXT: lui a1, 13107
; LMULMAX8-RV64-NEXT: addiw a1, a1, 819
; LMULMAX8-RV64-NEXT: slli a1, a1, 12
; LMULMAX8-RV64-NEXT: addi a1, a1, 819
; LMULMAX8-RV64-NEXT: slli a1, a1, 12
; LMULMAX8-RV64-NEXT: addi a1, a1, 819
; LMULMAX8-RV64-NEXT: slli a1, a1, 12
; LMULMAX8-RV64-NEXT: addi a1, a1, 819
; LMULMAX8-RV64-NEXT: vand.vx v9, v8, a1
; LMULMAX8-RV64-NEXT: vsrl.vi v8, v8, 2
; LMULMAX8-RV64-NEXT: vand.vx v8, v8, a1
; LMULMAX8-RV64-NEXT: vadd.vv v8, v9, v8
; LMULMAX8-RV64-NEXT: vsrl.vi v9, v8, 4
; LMULMAX8-RV64-NEXT: vadd.vv v8, v8, v9
; LMULMAX8-RV64-NEXT: lui a1, 3855
; LMULMAX8-RV64-NEXT: addiw a1, a1, 241
; LMULMAX8-RV64-NEXT: slli a1, a1, 12
; LMULMAX8-RV64-NEXT: addi a1, a1, -241
; LMULMAX8-RV64-NEXT: slli a1, a1, 12
; LMULMAX8-RV64-NEXT: addi a1, a1, 241
; LMULMAX8-RV64-NEXT: slli a1, a1, 12
; LMULMAX8-RV64-NEXT: addi a1, a1, -241
; LMULMAX8-RV64-NEXT: vand.vx v8, v8, a1
; LMULMAX8-RV64-NEXT: lui a1, 4112
; LMULMAX8-RV64-NEXT: addiw a1, a1, 257
; LMULMAX8-RV64-NEXT: slli a1, a1, 16
; LMULMAX8-RV64-NEXT: addi a1, a1, 257
; LMULMAX8-RV64-NEXT: slli a1, a1, 16
; LMULMAX8-RV64-NEXT: addi a1, a1, 257
; LMULMAX8-RV64-NEXT: vmul.vx v8, v8, a1
; LMULMAX8-RV64-NEXT: li a1, 56
; LMULMAX8-RV64-NEXT: vsrl.vx v8, v8, a1
; LMULMAX8-RV64-NEXT: vse64.v v8, (a0)
; LMULMAX8-RV64-NEXT: ret
%a = load <2 x i64>, <2 x i64>* %x
%b = load <2 x i64>, <2 x i64>* %y
%c = call <2 x i64> @llvm.cttz.v2i64(<2 x i64> %a, i1 false)
store <2 x i64> %c, <2 x i64>* %x
ret void
}
declare <2 x i64> @llvm.cttz.v2i64(<2 x i64>, i1)
define void @cttz_v32i8(<32 x i8>* %x, <32 x i8>* %y) nounwind {
; LMULMAX2-RV32-LABEL: cttz_v32i8:
; LMULMAX2-RV32: # %bb.0:
; LMULMAX2-RV32-NEXT: li a1, 32
; LMULMAX2-RV32-NEXT: vsetvli zero, a1, e8, m2, ta, mu
; LMULMAX2-RV32-NEXT: vle8.v v8, (a0)
; LMULMAX2-RV32-NEXT: li a1, 1
; LMULMAX2-RV32-NEXT: vsub.vx v10, v8, a1
; LMULMAX2-RV32-NEXT: vxor.vi v8, v8, -1
; LMULMAX2-RV32-NEXT: vand.vv v8, v8, v10
; LMULMAX2-RV32-NEXT: vsrl.vi v10, v8, 1
; LMULMAX2-RV32-NEXT: li a1, 85
; LMULMAX2-RV32-NEXT: vand.vx v10, v10, a1
; LMULMAX2-RV32-NEXT: vsub.vv v8, v8, v10
; LMULMAX2-RV32-NEXT: li a1, 51
; LMULMAX2-RV32-NEXT: vand.vx v10, v8, a1
; LMULMAX2-RV32-NEXT: vsrl.vi v8, v8, 2
; LMULMAX2-RV32-NEXT: vand.vx v8, v8, a1
; LMULMAX2-RV32-NEXT: vadd.vv v8, v10, v8
; LMULMAX2-RV32-NEXT: vsrl.vi v10, v8, 4
; LMULMAX2-RV32-NEXT: vadd.vv v8, v8, v10
; LMULMAX2-RV32-NEXT: vand.vi v8, v8, 15
; LMULMAX2-RV32-NEXT: vse8.v v8, (a0)
; LMULMAX2-RV32-NEXT: ret
;
; LMULMAX2-RV64-LABEL: cttz_v32i8:
; LMULMAX2-RV64: # %bb.0:
; LMULMAX2-RV64-NEXT: li a1, 32
; LMULMAX2-RV64-NEXT: vsetvli zero, a1, e8, m2, ta, mu
; LMULMAX2-RV64-NEXT: vle8.v v8, (a0)
; LMULMAX2-RV64-NEXT: li a1, 1
; LMULMAX2-RV64-NEXT: vsub.vx v10, v8, a1
; LMULMAX2-RV64-NEXT: vxor.vi v8, v8, -1
; LMULMAX2-RV64-NEXT: vand.vv v8, v8, v10
; LMULMAX2-RV64-NEXT: vsrl.vi v10, v8, 1
; LMULMAX2-RV64-NEXT: li a1, 85
; LMULMAX2-RV64-NEXT: vand.vx v10, v10, a1
; LMULMAX2-RV64-NEXT: vsub.vv v8, v8, v10
; LMULMAX2-RV64-NEXT: li a1, 51
; LMULMAX2-RV64-NEXT: vand.vx v10, v8, a1
; LMULMAX2-RV64-NEXT: vsrl.vi v8, v8, 2
; LMULMAX2-RV64-NEXT: vand.vx v8, v8, a1
; LMULMAX2-RV64-NEXT: vadd.vv v8, v10, v8
; LMULMAX2-RV64-NEXT: vsrl.vi v10, v8, 4
; LMULMAX2-RV64-NEXT: vadd.vv v8, v8, v10
; LMULMAX2-RV64-NEXT: vand.vi v8, v8, 15
; LMULMAX2-RV64-NEXT: vse8.v v8, (a0)
; LMULMAX2-RV64-NEXT: ret
;
; LMULMAX1-RV32-LABEL: cttz_v32i8:
; LMULMAX1-RV32: # %bb.0:
; LMULMAX1-RV32-NEXT: vsetivli zero, 16, e8, m1, ta, mu
; LMULMAX1-RV32-NEXT: addi a1, a0, 16
; LMULMAX1-RV32-NEXT: vle8.v v8, (a1)
; LMULMAX1-RV32-NEXT: vle8.v v9, (a0)
; LMULMAX1-RV32-NEXT: li a2, 1
; LMULMAX1-RV32-NEXT: vsub.vx v10, v8, a2
; LMULMAX1-RV32-NEXT: vxor.vi v8, v8, -1
; LMULMAX1-RV32-NEXT: vand.vv v8, v8, v10
; LMULMAX1-RV32-NEXT: vsrl.vi v10, v8, 1
; LMULMAX1-RV32-NEXT: li a3, 85
; LMULMAX1-RV32-NEXT: vand.vx v10, v10, a3
; LMULMAX1-RV32-NEXT: vsub.vv v8, v8, v10
; LMULMAX1-RV32-NEXT: li a4, 51
; LMULMAX1-RV32-NEXT: vand.vx v10, v8, a4
; LMULMAX1-RV32-NEXT: vsrl.vi v8, v8, 2
; LMULMAX1-RV32-NEXT: vand.vx v8, v8, a4
; LMULMAX1-RV32-NEXT: vadd.vv v8, v10, v8
; LMULMAX1-RV32-NEXT: vsrl.vi v10, v8, 4
; LMULMAX1-RV32-NEXT: vadd.vv v8, v8, v10
; LMULMAX1-RV32-NEXT: vand.vi v8, v8, 15
; LMULMAX1-RV32-NEXT: vsub.vx v10, v9, a2
; LMULMAX1-RV32-NEXT: vxor.vi v9, v9, -1
; LMULMAX1-RV32-NEXT: vand.vv v9, v9, v10
; LMULMAX1-RV32-NEXT: vsrl.vi v10, v9, 1
; LMULMAX1-RV32-NEXT: vand.vx v10, v10, a3
; LMULMAX1-RV32-NEXT: vsub.vv v9, v9, v10
; LMULMAX1-RV32-NEXT: vand.vx v10, v9, a4
; LMULMAX1-RV32-NEXT: vsrl.vi v9, v9, 2
; LMULMAX1-RV32-NEXT: vand.vx v9, v9, a4
; LMULMAX1-RV32-NEXT: vadd.vv v9, v10, v9
; LMULMAX1-RV32-NEXT: vsrl.vi v10, v9, 4
; LMULMAX1-RV32-NEXT: vadd.vv v9, v9, v10
; LMULMAX1-RV32-NEXT: vand.vi v9, v9, 15
; LMULMAX1-RV32-NEXT: vse8.v v9, (a0)
; LMULMAX1-RV32-NEXT: vse8.v v8, (a1)
; LMULMAX1-RV32-NEXT: ret
;
; LMULMAX1-RV64-LABEL: cttz_v32i8:
; LMULMAX1-RV64: # %bb.0:
; LMULMAX1-RV64-NEXT: vsetivli zero, 16, e8, m1, ta, mu
; LMULMAX1-RV64-NEXT: addi a1, a0, 16
; LMULMAX1-RV64-NEXT: vle8.v v8, (a1)
; LMULMAX1-RV64-NEXT: vle8.v v9, (a0)
; LMULMAX1-RV64-NEXT: li a2, 1
; LMULMAX1-RV64-NEXT: vsub.vx v10, v8, a2
; LMULMAX1-RV64-NEXT: vxor.vi v8, v8, -1
; LMULMAX1-RV64-NEXT: vand.vv v8, v8, v10
; LMULMAX1-RV64-NEXT: vsrl.vi v10, v8, 1
; LMULMAX1-RV64-NEXT: li a3, 85
; LMULMAX1-RV64-NEXT: vand.vx v10, v10, a3
; LMULMAX1-RV64-NEXT: vsub.vv v8, v8, v10
; LMULMAX1-RV64-NEXT: li a4, 51
; LMULMAX1-RV64-NEXT: vand.vx v10, v8, a4
; LMULMAX1-RV64-NEXT: vsrl.vi v8, v8, 2
; LMULMAX1-RV64-NEXT: vand.vx v8, v8, a4
; LMULMAX1-RV64-NEXT: vadd.vv v8, v10, v8
; LMULMAX1-RV64-NEXT: vsrl.vi v10, v8, 4
; LMULMAX1-RV64-NEXT: vadd.vv v8, v8, v10
; LMULMAX1-RV64-NEXT: vand.vi v8, v8, 15
; LMULMAX1-RV64-NEXT: vsub.vx v10, v9, a2
; LMULMAX1-RV64-NEXT: vxor.vi v9, v9, -1
; LMULMAX1-RV64-NEXT: vand.vv v9, v9, v10
; LMULMAX1-RV64-NEXT: vsrl.vi v10, v9, 1
; LMULMAX1-RV64-NEXT: vand.vx v10, v10, a3
; LMULMAX1-RV64-NEXT: vsub.vv v9, v9, v10
; LMULMAX1-RV64-NEXT: vand.vx v10, v9, a4
; LMULMAX1-RV64-NEXT: vsrl.vi v9, v9, 2
; LMULMAX1-RV64-NEXT: vand.vx v9, v9, a4
; LMULMAX1-RV64-NEXT: vadd.vv v9, v10, v9
; LMULMAX1-RV64-NEXT: vsrl.vi v10, v9, 4
; LMULMAX1-RV64-NEXT: vadd.vv v9, v9, v10
; LMULMAX1-RV64-NEXT: vand.vi v9, v9, 15
; LMULMAX1-RV64-NEXT: vse8.v v9, (a0)
; LMULMAX1-RV64-NEXT: vse8.v v8, (a1)
; LMULMAX1-RV64-NEXT: ret
;
; LMULMAX8-RV32-LABEL: cttz_v32i8:
; LMULMAX8-RV32: # %bb.0:
; LMULMAX8-RV32-NEXT: li a1, 32
; LMULMAX8-RV32-NEXT: vsetvli zero, a1, e8, m2, ta, mu
; LMULMAX8-RV32-NEXT: vle8.v v8, (a0)
; LMULMAX8-RV32-NEXT: vrsub.vi v10, v8, 0
; LMULMAX8-RV32-NEXT: vand.vv v10, v8, v10
; LMULMAX8-RV32-NEXT: vsetvli zero, zero, e32, m8, ta, mu
; LMULMAX8-RV32-NEXT: vzext.vf4 v16, v10
; LMULMAX8-RV32-NEXT: vfcvt.f.xu.v v16, v16
; LMULMAX8-RV32-NEXT: vsetvli zero, zero, e16, m4, ta, mu
; LMULMAX8-RV32-NEXT: vnsrl.wi v12, v16, 23
; LMULMAX8-RV32-NEXT: vsetvli zero, zero, e8, m2, ta, mu
; LMULMAX8-RV32-NEXT: vnsrl.wi v10, v12, 0
; LMULMAX8-RV32-NEXT: li a1, 127
; LMULMAX8-RV32-NEXT: vmseq.vi v0, v8, 0
; LMULMAX8-RV32-NEXT: vsub.vx v8, v10, a1
; LMULMAX8-RV32-NEXT: vmerge.vim v8, v8, 8, v0
; LMULMAX8-RV32-NEXT: vse8.v v8, (a0)
; LMULMAX8-RV32-NEXT: ret
;
; LMULMAX8-RV64-LABEL: cttz_v32i8:
; LMULMAX8-RV64: # %bb.0:
; LMULMAX8-RV64-NEXT: li a1, 32
; LMULMAX8-RV64-NEXT: vsetvli zero, a1, e8, m2, ta, mu
; LMULMAX8-RV64-NEXT: vle8.v v8, (a0)
; LMULMAX8-RV64-NEXT: vrsub.vi v10, v8, 0
; LMULMAX8-RV64-NEXT: vand.vv v10, v8, v10
; LMULMAX8-RV64-NEXT: vsetvli zero, zero, e32, m8, ta, mu
; LMULMAX8-RV64-NEXT: vzext.vf4 v16, v10
; LMULMAX8-RV64-NEXT: vfcvt.f.xu.v v16, v16
; LMULMAX8-RV64-NEXT: vsetvli zero, zero, e16, m4, ta, mu
; LMULMAX8-RV64-NEXT: vnsrl.wi v12, v16, 23
; LMULMAX8-RV64-NEXT: vsetvli zero, zero, e8, m2, ta, mu
; LMULMAX8-RV64-NEXT: vnsrl.wi v10, v12, 0
; LMULMAX8-RV64-NEXT: li a1, 127
; LMULMAX8-RV64-NEXT: vmseq.vi v0, v8, 0
; LMULMAX8-RV64-NEXT: vsub.vx v8, v10, a1
; LMULMAX8-RV64-NEXT: vmerge.vim v8, v8, 8, v0
; LMULMAX8-RV64-NEXT: vse8.v v8, (a0)
; LMULMAX8-RV64-NEXT: ret
%a = load <32 x i8>, <32 x i8>* %x
%b = load <32 x i8>, <32 x i8>* %y
%c = call <32 x i8> @llvm.cttz.v32i8(<32 x i8> %a, i1 false)
store <32 x i8> %c, <32 x i8>* %x
ret void
}
declare <32 x i8> @llvm.cttz.v32i8(<32 x i8>, i1)
define void @cttz_v16i16(<16 x i16>* %x, <16 x i16>* %y) nounwind {
; LMULMAX2-RV32-LABEL: cttz_v16i16:
; LMULMAX2-RV32: # %bb.0:
; LMULMAX2-RV32-NEXT: vsetivli zero, 16, e16, m2, ta, mu
; LMULMAX2-RV32-NEXT: vle16.v v8, (a0)
; LMULMAX2-RV32-NEXT: li a1, 1
; LMULMAX2-RV32-NEXT: vsub.vx v10, v8, a1
; LMULMAX2-RV32-NEXT: vxor.vi v8, v8, -1
; LMULMAX2-RV32-NEXT: vand.vv v8, v8, v10
; LMULMAX2-RV32-NEXT: vsrl.vi v10, v8, 1
; LMULMAX2-RV32-NEXT: lui a1, 5
; LMULMAX2-RV32-NEXT: addi a1, a1, 1365
; LMULMAX2-RV32-NEXT: vand.vx v10, v10, a1
; LMULMAX2-RV32-NEXT: vsub.vv v8, v8, v10
; LMULMAX2-RV32-NEXT: lui a1, 3
; LMULMAX2-RV32-NEXT: addi a1, a1, 819
; LMULMAX2-RV32-NEXT: vand.vx v10, v8, a1
; LMULMAX2-RV32-NEXT: vsrl.vi v8, v8, 2
; LMULMAX2-RV32-NEXT: vand.vx v8, v8, a1
; LMULMAX2-RV32-NEXT: vadd.vv v8, v10, v8
; LMULMAX2-RV32-NEXT: vsrl.vi v10, v8, 4
; LMULMAX2-RV32-NEXT: vadd.vv v8, v8, v10
; LMULMAX2-RV32-NEXT: lui a1, 1
; LMULMAX2-RV32-NEXT: addi a1, a1, -241
; LMULMAX2-RV32-NEXT: vand.vx v8, v8, a1
; LMULMAX2-RV32-NEXT: li a1, 257
; LMULMAX2-RV32-NEXT: vmul.vx v8, v8, a1
; LMULMAX2-RV32-NEXT: vsrl.vi v8, v8, 8
; LMULMAX2-RV32-NEXT: vse16.v v8, (a0)
; LMULMAX2-RV32-NEXT: ret
;
; LMULMAX2-RV64-LABEL: cttz_v16i16:
; LMULMAX2-RV64: # %bb.0:
; LMULMAX2-RV64-NEXT: vsetivli zero, 16, e16, m2, ta, mu
; LMULMAX2-RV64-NEXT: vle16.v v8, (a0)
; LMULMAX2-RV64-NEXT: li a1, 1
; LMULMAX2-RV64-NEXT: vsub.vx v10, v8, a1
; LMULMAX2-RV64-NEXT: vxor.vi v8, v8, -1
; LMULMAX2-RV64-NEXT: vand.vv v8, v8, v10
; LMULMAX2-RV64-NEXT: vsrl.vi v10, v8, 1
; LMULMAX2-RV64-NEXT: lui a1, 5
; LMULMAX2-RV64-NEXT: addiw a1, a1, 1365
; LMULMAX2-RV64-NEXT: vand.vx v10, v10, a1
; LMULMAX2-RV64-NEXT: vsub.vv v8, v8, v10
; LMULMAX2-RV64-NEXT: lui a1, 3
; LMULMAX2-RV64-NEXT: addiw a1, a1, 819
; LMULMAX2-RV64-NEXT: vand.vx v10, v8, a1
; LMULMAX2-RV64-NEXT: vsrl.vi v8, v8, 2
; LMULMAX2-RV64-NEXT: vand.vx v8, v8, a1
; LMULMAX2-RV64-NEXT: vadd.vv v8, v10, v8
; LMULMAX2-RV64-NEXT: vsrl.vi v10, v8, 4
; LMULMAX2-RV64-NEXT: vadd.vv v8, v8, v10
; LMULMAX2-RV64-NEXT: lui a1, 1
; LMULMAX2-RV64-NEXT: addiw a1, a1, -241
; LMULMAX2-RV64-NEXT: vand.vx v8, v8, a1
; LMULMAX2-RV64-NEXT: li a1, 257
; LMULMAX2-RV64-NEXT: vmul.vx v8, v8, a1
; LMULMAX2-RV64-NEXT: vsrl.vi v8, v8, 8
; LMULMAX2-RV64-NEXT: vse16.v v8, (a0)
; LMULMAX2-RV64-NEXT: ret
;
; LMULMAX1-RV32-LABEL: cttz_v16i16:
; LMULMAX1-RV32: # %bb.0:
; LMULMAX1-RV32-NEXT: vsetivli zero, 8, e16, m1, ta, mu
; LMULMAX1-RV32-NEXT: addi a1, a0, 16
; LMULMAX1-RV32-NEXT: vle16.v v8, (a1)
; LMULMAX1-RV32-NEXT: vle16.v v9, (a0)
; LMULMAX1-RV32-NEXT: li a6, 1
; LMULMAX1-RV32-NEXT: vsub.vx v10, v8, a6
; LMULMAX1-RV32-NEXT: vxor.vi v8, v8, -1
; LMULMAX1-RV32-NEXT: vand.vv v8, v8, v10
; LMULMAX1-RV32-NEXT: vsrl.vi v10, v8, 1
; LMULMAX1-RV32-NEXT: lui a3, 5
; LMULMAX1-RV32-NEXT: addi a3, a3, 1365
; LMULMAX1-RV32-NEXT: vand.vx v10, v10, a3
; LMULMAX1-RV32-NEXT: vsub.vv v8, v8, v10
; LMULMAX1-RV32-NEXT: lui a4, 3
; LMULMAX1-RV32-NEXT: addi a4, a4, 819
; LMULMAX1-RV32-NEXT: vand.vx v10, v8, a4
; LMULMAX1-RV32-NEXT: vsrl.vi v8, v8, 2
; LMULMAX1-RV32-NEXT: vand.vx v8, v8, a4
; LMULMAX1-RV32-NEXT: vadd.vv v8, v10, v8
; LMULMAX1-RV32-NEXT: vsrl.vi v10, v8, 4
; LMULMAX1-RV32-NEXT: vadd.vv v8, v8, v10
; LMULMAX1-RV32-NEXT: lui a5, 1
; LMULMAX1-RV32-NEXT: addi a5, a5, -241
; LMULMAX1-RV32-NEXT: vand.vx v8, v8, a5
; LMULMAX1-RV32-NEXT: li a2, 257
; LMULMAX1-RV32-NEXT: vmul.vx v8, v8, a2
; LMULMAX1-RV32-NEXT: vsrl.vi v8, v8, 8
; LMULMAX1-RV32-NEXT: vsub.vx v10, v9, a6
; LMULMAX1-RV32-NEXT: vxor.vi v9, v9, -1
; LMULMAX1-RV32-NEXT: vand.vv v9, v9, v10
; LMULMAX1-RV32-NEXT: vsrl.vi v10, v9, 1
; LMULMAX1-RV32-NEXT: vand.vx v10, v10, a3
; LMULMAX1-RV32-NEXT: vsub.vv v9, v9, v10
; LMULMAX1-RV32-NEXT: vand.vx v10, v9, a4
; LMULMAX1-RV32-NEXT: vsrl.vi v9, v9, 2
; LMULMAX1-RV32-NEXT: vand.vx v9, v9, a4
; LMULMAX1-RV32-NEXT: vadd.vv v9, v10, v9
; LMULMAX1-RV32-NEXT: vsrl.vi v10, v9, 4
; LMULMAX1-RV32-NEXT: vadd.vv v9, v9, v10
; LMULMAX1-RV32-NEXT: vand.vx v9, v9, a5
; LMULMAX1-RV32-NEXT: vmul.vx v9, v9, a2
; LMULMAX1-RV32-NEXT: vsrl.vi v9, v9, 8
; LMULMAX1-RV32-NEXT: vse16.v v9, (a0)
; LMULMAX1-RV32-NEXT: vse16.v v8, (a1)
; LMULMAX1-RV32-NEXT: ret
;
; LMULMAX1-RV64-LABEL: cttz_v16i16:
; LMULMAX1-RV64: # %bb.0:
; LMULMAX1-RV64-NEXT: vsetivli zero, 8, e16, m1, ta, mu
; LMULMAX1-RV64-NEXT: addi a1, a0, 16
; LMULMAX1-RV64-NEXT: vle16.v v8, (a1)
; LMULMAX1-RV64-NEXT: vle16.v v9, (a0)
; LMULMAX1-RV64-NEXT: li a6, 1
; LMULMAX1-RV64-NEXT: vsub.vx v10, v8, a6
; LMULMAX1-RV64-NEXT: vxor.vi v8, v8, -1
; LMULMAX1-RV64-NEXT: vand.vv v8, v8, v10
; LMULMAX1-RV64-NEXT: vsrl.vi v10, v8, 1
; LMULMAX1-RV64-NEXT: lui a3, 5
; LMULMAX1-RV64-NEXT: addiw a3, a3, 1365
; LMULMAX1-RV64-NEXT: vand.vx v10, v10, a3
; LMULMAX1-RV64-NEXT: vsub.vv v8, v8, v10
; LMULMAX1-RV64-NEXT: lui a4, 3
; LMULMAX1-RV64-NEXT: addiw a4, a4, 819
; LMULMAX1-RV64-NEXT: vand.vx v10, v8, a4
; LMULMAX1-RV64-NEXT: vsrl.vi v8, v8, 2
; LMULMAX1-RV64-NEXT: vand.vx v8, v8, a4
; LMULMAX1-RV64-NEXT: vadd.vv v8, v10, v8
; LMULMAX1-RV64-NEXT: vsrl.vi v10, v8, 4
; LMULMAX1-RV64-NEXT: vadd.vv v8, v8, v10
; LMULMAX1-RV64-NEXT: lui a5, 1
; LMULMAX1-RV64-NEXT: addiw a5, a5, -241
; LMULMAX1-RV64-NEXT: vand.vx v8, v8, a5
; LMULMAX1-RV64-NEXT: li a2, 257
; LMULMAX1-RV64-NEXT: vmul.vx v8, v8, a2
; LMULMAX1-RV64-NEXT: vsrl.vi v8, v8, 8
; LMULMAX1-RV64-NEXT: vsub.vx v10, v9, a6
; LMULMAX1-RV64-NEXT: vxor.vi v9, v9, -1
; LMULMAX1-RV64-NEXT: vand.vv v9, v9, v10
; LMULMAX1-RV64-NEXT: vsrl.vi v10, v9, 1
; LMULMAX1-RV64-NEXT: vand.vx v10, v10, a3
; LMULMAX1-RV64-NEXT: vsub.vv v9, v9, v10
; LMULMAX1-RV64-NEXT: vand.vx v10, v9, a4
; LMULMAX1-RV64-NEXT: vsrl.vi v9, v9, 2
; LMULMAX1-RV64-NEXT: vand.vx v9, v9, a4
; LMULMAX1-RV64-NEXT: vadd.vv v9, v10, v9
; LMULMAX1-RV64-NEXT: vsrl.vi v10, v9, 4
; LMULMAX1-RV64-NEXT: vadd.vv v9, v9, v10
; LMULMAX1-RV64-NEXT: vand.vx v9, v9, a5
; LMULMAX1-RV64-NEXT: vmul.vx v9, v9, a2
; LMULMAX1-RV64-NEXT: vsrl.vi v9, v9, 8
; LMULMAX1-RV64-NEXT: vse16.v v9, (a0)
; LMULMAX1-RV64-NEXT: vse16.v v8, (a1)
; LMULMAX1-RV64-NEXT: ret
;
; LMULMAX8-RV32-LABEL: cttz_v16i16:
; LMULMAX8-RV32: # %bb.0:
; LMULMAX8-RV32-NEXT: vsetivli zero, 16, e16, m2, ta, mu
; LMULMAX8-RV32-NEXT: vle16.v v8, (a0)
; LMULMAX8-RV32-NEXT: vrsub.vi v10, v8, 0
; LMULMAX8-RV32-NEXT: vand.vv v10, v8, v10
; LMULMAX8-RV32-NEXT: vfwcvt.f.xu.v v12, v10
; LMULMAX8-RV32-NEXT: vnsrl.wi v10, v12, 23
; LMULMAX8-RV32-NEXT: li a1, 127
; LMULMAX8-RV32-NEXT: vsub.vx v10, v10, a1
; LMULMAX8-RV32-NEXT: vmseq.vi v0, v8, 0
; LMULMAX8-RV32-NEXT: li a1, 16
; LMULMAX8-RV32-NEXT: vmerge.vxm v8, v10, a1, v0
; LMULMAX8-RV32-NEXT: vse16.v v8, (a0)
; LMULMAX8-RV32-NEXT: ret
;
; LMULMAX8-RV64-LABEL: cttz_v16i16:
; LMULMAX8-RV64: # %bb.0:
; LMULMAX8-RV64-NEXT: vsetivli zero, 16, e16, m2, ta, mu
; LMULMAX8-RV64-NEXT: vle16.v v8, (a0)
; LMULMAX8-RV64-NEXT: vrsub.vi v10, v8, 0
; LMULMAX8-RV64-NEXT: vand.vv v10, v8, v10
; LMULMAX8-RV64-NEXT: vfwcvt.f.xu.v v12, v10
; LMULMAX8-RV64-NEXT: vnsrl.wi v10, v12, 23
; LMULMAX8-RV64-NEXT: li a1, 127
; LMULMAX8-RV64-NEXT: vsub.vx v10, v10, a1
; LMULMAX8-RV64-NEXT: vmseq.vi v0, v8, 0
; LMULMAX8-RV64-NEXT: li a1, 16
; LMULMAX8-RV64-NEXT: vmerge.vxm v8, v10, a1, v0
; LMULMAX8-RV64-NEXT: vse16.v v8, (a0)
; LMULMAX8-RV64-NEXT: ret
%a = load <16 x i16>, <16 x i16>* %x
%b = load <16 x i16>, <16 x i16>* %y
%c = call <16 x i16> @llvm.cttz.v16i16(<16 x i16> %a, i1 false)
store <16 x i16> %c, <16 x i16>* %x
ret void
}
declare <16 x i16> @llvm.cttz.v16i16(<16 x i16>, i1)
define void @cttz_v8i32(<8 x i32>* %x, <8 x i32>* %y) nounwind {
; LMULMAX2-RV32-LABEL: cttz_v8i32:
; LMULMAX2-RV32: # %bb.0:
; LMULMAX2-RV32-NEXT: vsetivli zero, 8, e32, m2, ta, mu
; LMULMAX2-RV32-NEXT: vle32.v v8, (a0)
; LMULMAX2-RV32-NEXT: li a1, 1
; LMULMAX2-RV32-NEXT: vsub.vx v10, v8, a1
; LMULMAX2-RV32-NEXT: vxor.vi v8, v8, -1
; LMULMAX2-RV32-NEXT: vand.vv v8, v8, v10
; LMULMAX2-RV32-NEXT: vsrl.vi v10, v8, 1
; LMULMAX2-RV32-NEXT: lui a1, 349525
; LMULMAX2-RV32-NEXT: addi a1, a1, 1365
; LMULMAX2-RV32-NEXT: vand.vx v10, v10, a1
; LMULMAX2-RV32-NEXT: vsub.vv v8, v8, v10
; LMULMAX2-RV32-NEXT: lui a1, 209715
; LMULMAX2-RV32-NEXT: addi a1, a1, 819
; LMULMAX2-RV32-NEXT: vand.vx v10, v8, a1
; LMULMAX2-RV32-NEXT: vsrl.vi v8, v8, 2
; LMULMAX2-RV32-NEXT: vand.vx v8, v8, a1
; LMULMAX2-RV32-NEXT: vadd.vv v8, v10, v8
; LMULMAX2-RV32-NEXT: vsrl.vi v10, v8, 4
; LMULMAX2-RV32-NEXT: vadd.vv v8, v8, v10
; LMULMAX2-RV32-NEXT: lui a1, 61681
; LMULMAX2-RV32-NEXT: addi a1, a1, -241
; LMULMAX2-RV32-NEXT: vand.vx v8, v8, a1
; LMULMAX2-RV32-NEXT: lui a1, 4112
; LMULMAX2-RV32-NEXT: addi a1, a1, 257
; LMULMAX2-RV32-NEXT: vmul.vx v8, v8, a1
; LMULMAX2-RV32-NEXT: vsrl.vi v8, v8, 24
; LMULMAX2-RV32-NEXT: vse32.v v8, (a0)
; LMULMAX2-RV32-NEXT: ret
;
; LMULMAX2-RV64-LABEL: cttz_v8i32:
; LMULMAX2-RV64: # %bb.0:
; LMULMAX2-RV64-NEXT: vsetivli zero, 8, e32, m2, ta, mu
; LMULMAX2-RV64-NEXT: vle32.v v8, (a0)
; LMULMAX2-RV64-NEXT: li a1, 1
; LMULMAX2-RV64-NEXT: vsub.vx v10, v8, a1
; LMULMAX2-RV64-NEXT: vxor.vi v8, v8, -1
; LMULMAX2-RV64-NEXT: vand.vv v8, v8, v10
; LMULMAX2-RV64-NEXT: vsrl.vi v10, v8, 1
; LMULMAX2-RV64-NEXT: lui a1, 349525
; LMULMAX2-RV64-NEXT: addiw a1, a1, 1365
; LMULMAX2-RV64-NEXT: vand.vx v10, v10, a1
; LMULMAX2-RV64-NEXT: vsub.vv v8, v8, v10
; LMULMAX2-RV64-NEXT: lui a1, 209715
; LMULMAX2-RV64-NEXT: addiw a1, a1, 819
; LMULMAX2-RV64-NEXT: vand.vx v10, v8, a1
; LMULMAX2-RV64-NEXT: vsrl.vi v8, v8, 2
; LMULMAX2-RV64-NEXT: vand.vx v8, v8, a1
; LMULMAX2-RV64-NEXT: vadd.vv v8, v10, v8
; LMULMAX2-RV64-NEXT: vsrl.vi v10, v8, 4
; LMULMAX2-RV64-NEXT: vadd.vv v8, v8, v10
; LMULMAX2-RV64-NEXT: lui a1, 61681
; LMULMAX2-RV64-NEXT: addiw a1, a1, -241
; LMULMAX2-RV64-NEXT: vand.vx v8, v8, a1
; LMULMAX2-RV64-NEXT: lui a1, 4112
; LMULMAX2-RV64-NEXT: addiw a1, a1, 257
; LMULMAX2-RV64-NEXT: vmul.vx v8, v8, a1
; LMULMAX2-RV64-NEXT: vsrl.vi v8, v8, 24
; LMULMAX2-RV64-NEXT: vse32.v v8, (a0)
; LMULMAX2-RV64-NEXT: ret
;
; LMULMAX1-RV32-LABEL: cttz_v8i32:
; LMULMAX1-RV32: # %bb.0:
; LMULMAX1-RV32-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX1-RV32-NEXT: addi a1, a0, 16
; LMULMAX1-RV32-NEXT: vle32.v v8, (a1)
; LMULMAX1-RV32-NEXT: vle32.v v9, (a0)
; LMULMAX1-RV32-NEXT: li a6, 1
; LMULMAX1-RV32-NEXT: vsub.vx v10, v8, a6
; LMULMAX1-RV32-NEXT: vxor.vi v8, v8, -1
; LMULMAX1-RV32-NEXT: vand.vv v8, v8, v10
; LMULMAX1-RV32-NEXT: vsrl.vi v10, v8, 1
; LMULMAX1-RV32-NEXT: lui a3, 349525
; LMULMAX1-RV32-NEXT: addi a3, a3, 1365
; LMULMAX1-RV32-NEXT: vand.vx v10, v10, a3
; LMULMAX1-RV32-NEXT: vsub.vv v8, v8, v10
; LMULMAX1-RV32-NEXT: lui a4, 209715
; LMULMAX1-RV32-NEXT: addi a4, a4, 819
; LMULMAX1-RV32-NEXT: vand.vx v10, v8, a4
; LMULMAX1-RV32-NEXT: vsrl.vi v8, v8, 2
; LMULMAX1-RV32-NEXT: vand.vx v8, v8, a4
; LMULMAX1-RV32-NEXT: vadd.vv v8, v10, v8
; LMULMAX1-RV32-NEXT: vsrl.vi v10, v8, 4
; LMULMAX1-RV32-NEXT: vadd.vv v8, v8, v10
; LMULMAX1-RV32-NEXT: lui a5, 61681
; LMULMAX1-RV32-NEXT: addi a5, a5, -241
; LMULMAX1-RV32-NEXT: vand.vx v8, v8, a5
; LMULMAX1-RV32-NEXT: lui a2, 4112
; LMULMAX1-RV32-NEXT: addi a2, a2, 257
; LMULMAX1-RV32-NEXT: vmul.vx v8, v8, a2
; LMULMAX1-RV32-NEXT: vsrl.vi v8, v8, 24
; LMULMAX1-RV32-NEXT: vsub.vx v10, v9, a6
; LMULMAX1-RV32-NEXT: vxor.vi v9, v9, -1
; LMULMAX1-RV32-NEXT: vand.vv v9, v9, v10
; LMULMAX1-RV32-NEXT: vsrl.vi v10, v9, 1
; LMULMAX1-RV32-NEXT: vand.vx v10, v10, a3
; LMULMAX1-RV32-NEXT: vsub.vv v9, v9, v10
; LMULMAX1-RV32-NEXT: vand.vx v10, v9, a4
; LMULMAX1-RV32-NEXT: vsrl.vi v9, v9, 2
; LMULMAX1-RV32-NEXT: vand.vx v9, v9, a4
; LMULMAX1-RV32-NEXT: vadd.vv v9, v10, v9
; LMULMAX1-RV32-NEXT: vsrl.vi v10, v9, 4
; LMULMAX1-RV32-NEXT: vadd.vv v9, v9, v10
; LMULMAX1-RV32-NEXT: vand.vx v9, v9, a5
; LMULMAX1-RV32-NEXT: vmul.vx v9, v9, a2
; LMULMAX1-RV32-NEXT: vsrl.vi v9, v9, 24
; LMULMAX1-RV32-NEXT: vse32.v v9, (a0)
; LMULMAX1-RV32-NEXT: vse32.v v8, (a1)
; LMULMAX1-RV32-NEXT: ret
;
; LMULMAX1-RV64-LABEL: cttz_v8i32:
; LMULMAX1-RV64: # %bb.0:
; LMULMAX1-RV64-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX1-RV64-NEXT: addi a1, a0, 16
; LMULMAX1-RV64-NEXT: vle32.v v8, (a1)
; LMULMAX1-RV64-NEXT: vle32.v v9, (a0)
; LMULMAX1-RV64-NEXT: li a6, 1
; LMULMAX1-RV64-NEXT: vsub.vx v10, v8, a6
; LMULMAX1-RV64-NEXT: vxor.vi v8, v8, -1
; LMULMAX1-RV64-NEXT: vand.vv v8, v8, v10
; LMULMAX1-RV64-NEXT: vsrl.vi v10, v8, 1
; LMULMAX1-RV64-NEXT: lui a3, 349525
; LMULMAX1-RV64-NEXT: addiw a3, a3, 1365
; LMULMAX1-RV64-NEXT: vand.vx v10, v10, a3
; LMULMAX1-RV64-NEXT: vsub.vv v8, v8, v10
; LMULMAX1-RV64-NEXT: lui a4, 209715
; LMULMAX1-RV64-NEXT: addiw a4, a4, 819
; LMULMAX1-RV64-NEXT: vand.vx v10, v8, a4
; LMULMAX1-RV64-NEXT: vsrl.vi v8, v8, 2
; LMULMAX1-RV64-NEXT: vand.vx v8, v8, a4
; LMULMAX1-RV64-NEXT: vadd.vv v8, v10, v8
; LMULMAX1-RV64-NEXT: vsrl.vi v10, v8, 4
; LMULMAX1-RV64-NEXT: vadd.vv v8, v8, v10
; LMULMAX1-RV64-NEXT: lui a5, 61681
; LMULMAX1-RV64-NEXT: addiw a5, a5, -241
; LMULMAX1-RV64-NEXT: vand.vx v8, v8, a5
; LMULMAX1-RV64-NEXT: lui a2, 4112
; LMULMAX1-RV64-NEXT: addiw a2, a2, 257
; LMULMAX1-RV64-NEXT: vmul.vx v8, v8, a2
; LMULMAX1-RV64-NEXT: vsrl.vi v8, v8, 24
; LMULMAX1-RV64-NEXT: vsub.vx v10, v9, a6
; LMULMAX1-RV64-NEXT: vxor.vi v9, v9, -1
; LMULMAX1-RV64-NEXT: vand.vv v9, v9, v10
; LMULMAX1-RV64-NEXT: vsrl.vi v10, v9, 1
; LMULMAX1-RV64-NEXT: vand.vx v10, v10, a3
; LMULMAX1-RV64-NEXT: vsub.vv v9, v9, v10
; LMULMAX1-RV64-NEXT: vand.vx v10, v9, a4
; LMULMAX1-RV64-NEXT: vsrl.vi v9, v9, 2
; LMULMAX1-RV64-NEXT: vand.vx v9, v9, a4
; LMULMAX1-RV64-NEXT: vadd.vv v9, v10, v9
; LMULMAX1-RV64-NEXT: vsrl.vi v10, v9, 4
; LMULMAX1-RV64-NEXT: vadd.vv v9, v9, v10
; LMULMAX1-RV64-NEXT: vand.vx v9, v9, a5
; LMULMAX1-RV64-NEXT: vmul.vx v9, v9, a2
; LMULMAX1-RV64-NEXT: vsrl.vi v9, v9, 24
; LMULMAX1-RV64-NEXT: vse32.v v9, (a0)
; LMULMAX1-RV64-NEXT: vse32.v v8, (a1)
; LMULMAX1-RV64-NEXT: ret
;
; LMULMAX8-RV32-LABEL: cttz_v8i32:
; LMULMAX8-RV32: # %bb.0:
; LMULMAX8-RV32-NEXT: vsetivli zero, 8, e32, m2, ta, mu
; LMULMAX8-RV32-NEXT: vle32.v v8, (a0)
; LMULMAX8-RV32-NEXT: vrsub.vi v10, v8, 0
; LMULMAX8-RV32-NEXT: vand.vv v10, v8, v10
; LMULMAX8-RV32-NEXT: vfwcvt.f.xu.v v12, v10
; LMULMAX8-RV32-NEXT: li a1, 52
; LMULMAX8-RV32-NEXT: vnsrl.wx v10, v12, a1
; LMULMAX8-RV32-NEXT: li a1, 1023
; LMULMAX8-RV32-NEXT: vsub.vx v10, v10, a1
; LMULMAX8-RV32-NEXT: vmseq.vi v0, v8, 0
; LMULMAX8-RV32-NEXT: li a1, 32
; LMULMAX8-RV32-NEXT: vmerge.vxm v8, v10, a1, v0
; LMULMAX8-RV32-NEXT: vse32.v v8, (a0)
; LMULMAX8-RV32-NEXT: ret
;
; LMULMAX8-RV64-LABEL: cttz_v8i32:
; LMULMAX8-RV64: # %bb.0:
; LMULMAX8-RV64-NEXT: vsetivli zero, 8, e32, m2, ta, mu
; LMULMAX8-RV64-NEXT: vle32.v v8, (a0)
; LMULMAX8-RV64-NEXT: vrsub.vi v10, v8, 0
; LMULMAX8-RV64-NEXT: vand.vv v10, v8, v10
; LMULMAX8-RV64-NEXT: vfwcvt.f.xu.v v12, v10
; LMULMAX8-RV64-NEXT: li a1, 52
; LMULMAX8-RV64-NEXT: vnsrl.wx v10, v12, a1
; LMULMAX8-RV64-NEXT: li a1, 1023
; LMULMAX8-RV64-NEXT: vsub.vx v10, v10, a1
; LMULMAX8-RV64-NEXT: vmseq.vi v0, v8, 0
; LMULMAX8-RV64-NEXT: li a1, 32
; LMULMAX8-RV64-NEXT: vmerge.vxm v8, v10, a1, v0
; LMULMAX8-RV64-NEXT: vse32.v v8, (a0)
; LMULMAX8-RV64-NEXT: ret
%a = load <8 x i32>, <8 x i32>* %x
%b = load <8 x i32>, <8 x i32>* %y
%c = call <8 x i32> @llvm.cttz.v8i32(<8 x i32> %a, i1 false)
store <8 x i32> %c, <8 x i32>* %x
ret void
}
declare <8 x i32> @llvm.cttz.v8i32(<8 x i32>, i1)
define void @cttz_v4i64(<4 x i64>* %x, <4 x i64>* %y) nounwind {
; LMULMAX2-RV32-LABEL: cttz_v4i64:
; LMULMAX2-RV32: # %bb.0:
; LMULMAX2-RV32-NEXT: vsetivli zero, 4, e64, m2, ta, mu
; LMULMAX2-RV32-NEXT: vle64.v v8, (a0)
; LMULMAX2-RV32-NEXT: li a1, 1
; LMULMAX2-RV32-NEXT: vsub.vx v10, v8, a1
; LMULMAX2-RV32-NEXT: vsetivli zero, 8, e32, m2, ta, mu
; LMULMAX2-RV32-NEXT: vmv.v.i v12, -1
; LMULMAX2-RV32-NEXT: vsetivli zero, 4, e64, m2, ta, mu
; LMULMAX2-RV32-NEXT: vxor.vv v8, v8, v12
; LMULMAX2-RV32-NEXT: vand.vv v8, v8, v10
; LMULMAX2-RV32-NEXT: vsrl.vi v10, v8, 1
; LMULMAX2-RV32-NEXT: lui a1, 349525
; LMULMAX2-RV32-NEXT: addi a1, a1, 1365
; LMULMAX2-RV32-NEXT: vsetivli zero, 8, e32, m2, ta, mu
; LMULMAX2-RV32-NEXT: vmv.v.x v12, a1
; LMULMAX2-RV32-NEXT: vsetivli zero, 4, e64, m2, ta, mu
; LMULMAX2-RV32-NEXT: vand.vv v10, v10, v12
; LMULMAX2-RV32-NEXT: vsub.vv v8, v8, v10
; LMULMAX2-RV32-NEXT: lui a1, 209715
; LMULMAX2-RV32-NEXT: addi a1, a1, 819
; LMULMAX2-RV32-NEXT: vsetivli zero, 8, e32, m2, ta, mu
; LMULMAX2-RV32-NEXT: vmv.v.x v10, a1
; LMULMAX2-RV32-NEXT: vsetivli zero, 4, e64, m2, ta, mu
; LMULMAX2-RV32-NEXT: vand.vv v12, v8, v10
; LMULMAX2-RV32-NEXT: vsrl.vi v8, v8, 2
; LMULMAX2-RV32-NEXT: vand.vv v8, v8, v10
; LMULMAX2-RV32-NEXT: vadd.vv v8, v12, v8
; LMULMAX2-RV32-NEXT: vsrl.vi v10, v8, 4
; LMULMAX2-RV32-NEXT: vadd.vv v8, v8, v10
; LMULMAX2-RV32-NEXT: lui a1, 61681
; LMULMAX2-RV32-NEXT: addi a1, a1, -241
; LMULMAX2-RV32-NEXT: vsetivli zero, 8, e32, m2, ta, mu
; LMULMAX2-RV32-NEXT: vmv.v.x v10, a1
; LMULMAX2-RV32-NEXT: vsetivli zero, 4, e64, m2, ta, mu
; LMULMAX2-RV32-NEXT: vand.vv v8, v8, v10
; LMULMAX2-RV32-NEXT: lui a1, 4112
; LMULMAX2-RV32-NEXT: addi a1, a1, 257
; LMULMAX2-RV32-NEXT: vsetivli zero, 8, e32, m2, ta, mu
; LMULMAX2-RV32-NEXT: vmv.v.x v10, a1
; LMULMAX2-RV32-NEXT: vsetivli zero, 4, e64, m2, ta, mu
; LMULMAX2-RV32-NEXT: vmul.vv v8, v8, v10
; LMULMAX2-RV32-NEXT: li a1, 56
; LMULMAX2-RV32-NEXT: vsrl.vx v8, v8, a1
; LMULMAX2-RV32-NEXT: vse64.v v8, (a0)
; LMULMAX2-RV32-NEXT: ret
;
; LMULMAX2-RV64-LABEL: cttz_v4i64:
; LMULMAX2-RV64: # %bb.0:
; LMULMAX2-RV64-NEXT: vsetivli zero, 4, e64, m2, ta, mu
; LMULMAX2-RV64-NEXT: vle64.v v8, (a0)
; LMULMAX2-RV64-NEXT: li a1, 1
; LMULMAX2-RV64-NEXT: vsub.vx v10, v8, a1
; LMULMAX2-RV64-NEXT: vxor.vi v8, v8, -1
; LMULMAX2-RV64-NEXT: vand.vv v8, v8, v10
; LMULMAX2-RV64-NEXT: vsrl.vi v10, v8, 1
; LMULMAX2-RV64-NEXT: lui a1, 21845
; LMULMAX2-RV64-NEXT: addiw a1, a1, 1365
; LMULMAX2-RV64-NEXT: slli a1, a1, 12
; LMULMAX2-RV64-NEXT: addi a1, a1, 1365
; LMULMAX2-RV64-NEXT: slli a1, a1, 12
; LMULMAX2-RV64-NEXT: addi a1, a1, 1365
; LMULMAX2-RV64-NEXT: slli a1, a1, 12
; LMULMAX2-RV64-NEXT: addi a1, a1, 1365
; LMULMAX2-RV64-NEXT: vand.vx v10, v10, a1
; LMULMAX2-RV64-NEXT: vsub.vv v8, v8, v10
; LMULMAX2-RV64-NEXT: lui a1, 13107
; LMULMAX2-RV64-NEXT: addiw a1, a1, 819
; LMULMAX2-RV64-NEXT: slli a1, a1, 12
; LMULMAX2-RV64-NEXT: addi a1, a1, 819
; LMULMAX2-RV64-NEXT: slli a1, a1, 12
; LMULMAX2-RV64-NEXT: addi a1, a1, 819
; LMULMAX2-RV64-NEXT: slli a1, a1, 12
; LMULMAX2-RV64-NEXT: addi a1, a1, 819
; LMULMAX2-RV64-NEXT: vand.vx v10, v8, a1
; LMULMAX2-RV64-NEXT: vsrl.vi v8, v8, 2
; LMULMAX2-RV64-NEXT: vand.vx v8, v8, a1
; LMULMAX2-RV64-NEXT: vadd.vv v8, v10, v8
; LMULMAX2-RV64-NEXT: vsrl.vi v10, v8, 4
; LMULMAX2-RV64-NEXT: vadd.vv v8, v8, v10
; LMULMAX2-RV64-NEXT: lui a1, 3855
; LMULMAX2-RV64-NEXT: addiw a1, a1, 241
; LMULMAX2-RV64-NEXT: slli a1, a1, 12
; LMULMAX2-RV64-NEXT: addi a1, a1, -241
; LMULMAX2-RV64-NEXT: slli a1, a1, 12
; LMULMAX2-RV64-NEXT: addi a1, a1, 241
; LMULMAX2-RV64-NEXT: slli a1, a1, 12
; LMULMAX2-RV64-NEXT: addi a1, a1, -241
; LMULMAX2-RV64-NEXT: vand.vx v8, v8, a1
; LMULMAX2-RV64-NEXT: lui a1, 4112
; LMULMAX2-RV64-NEXT: addiw a1, a1, 257
; LMULMAX2-RV64-NEXT: slli a1, a1, 16
; LMULMAX2-RV64-NEXT: addi a1, a1, 257
; LMULMAX2-RV64-NEXT: slli a1, a1, 16
; LMULMAX2-RV64-NEXT: addi a1, a1, 257
; LMULMAX2-RV64-NEXT: vmul.vx v8, v8, a1
; LMULMAX2-RV64-NEXT: li a1, 56
; LMULMAX2-RV64-NEXT: vsrl.vx v8, v8, a1
; LMULMAX2-RV64-NEXT: vse64.v v8, (a0)
; LMULMAX2-RV64-NEXT: ret
;
; LMULMAX1-RV32-LABEL: cttz_v4i64:
; LMULMAX1-RV32: # %bb.0:
; LMULMAX1-RV32-NEXT: vsetivli zero, 2, e64, m1, ta, mu
; LMULMAX1-RV32-NEXT: addi a1, a0, 16
; LMULMAX1-RV32-NEXT: vle64.v v8, (a1)
; LMULMAX1-RV32-NEXT: vle64.v v9, (a0)
; LMULMAX1-RV32-NEXT: li a2, 1
; LMULMAX1-RV32-NEXT: vsub.vx v10, v8, a2
; LMULMAX1-RV32-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX1-RV32-NEXT: vmv.v.i v11, -1
; LMULMAX1-RV32-NEXT: vsetivli zero, 2, e64, m1, ta, mu
; LMULMAX1-RV32-NEXT: vxor.vv v8, v8, v11
; LMULMAX1-RV32-NEXT: vand.vv v8, v8, v10
; LMULMAX1-RV32-NEXT: vsrl.vi v10, v8, 1
; LMULMAX1-RV32-NEXT: lui a3, 349525
; LMULMAX1-RV32-NEXT: addi a3, a3, 1365
; LMULMAX1-RV32-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX1-RV32-NEXT: vmv.v.x v12, a3
; LMULMAX1-RV32-NEXT: vsetivli zero, 2, e64, m1, ta, mu
; LMULMAX1-RV32-NEXT: vand.vv v10, v10, v12
; LMULMAX1-RV32-NEXT: vsub.vv v8, v8, v10
; LMULMAX1-RV32-NEXT: lui a3, 209715
; LMULMAX1-RV32-NEXT: addi a3, a3, 819
; LMULMAX1-RV32-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX1-RV32-NEXT: vmv.v.x v10, a3
; LMULMAX1-RV32-NEXT: vsetivli zero, 2, e64, m1, ta, mu
; LMULMAX1-RV32-NEXT: vand.vv v13, v8, v10
; LMULMAX1-RV32-NEXT: vsrl.vi v8, v8, 2
; LMULMAX1-RV32-NEXT: vand.vv v8, v8, v10
; LMULMAX1-RV32-NEXT: vadd.vv v8, v13, v8
; LMULMAX1-RV32-NEXT: vsrl.vi v13, v8, 4
; LMULMAX1-RV32-NEXT: vadd.vv v8, v8, v13
; LMULMAX1-RV32-NEXT: lui a3, 61681
; LMULMAX1-RV32-NEXT: addi a3, a3, -241
; LMULMAX1-RV32-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX1-RV32-NEXT: vmv.v.x v13, a3
; LMULMAX1-RV32-NEXT: vsetivli zero, 2, e64, m1, ta, mu
; LMULMAX1-RV32-NEXT: vand.vv v8, v8, v13
; LMULMAX1-RV32-NEXT: lui a3, 4112
; LMULMAX1-RV32-NEXT: addi a3, a3, 257
; LMULMAX1-RV32-NEXT: vsetivli zero, 4, e32, m1, ta, mu
; LMULMAX1-RV32-NEXT: vmv.v.x v14, a3
; LMULMAX1-RV32-NEXT: vsetivli zero, 2, e64, m1, ta, mu
; LMULMAX1-RV32-NEXT: vmul.vv v8, v8, v14
; LMULMAX1-RV32-NEXT: li a3, 56
; LMULMAX1-RV32-NEXT: vsrl.vx v8, v8, a3
; LMULMAX1-RV32-NEXT: vsub.vx v15, v9, a2
; LMULMAX1-RV32-NEXT: vxor.vv v9, v9, v11
; LMULMAX1-RV32-NEXT: vand.vv v9, v9, v15
; LMULMAX1-RV32-NEXT: vsrl.vi v11, v9, 1
; LMULMAX1-RV32-NEXT: vand.vv v11, v11, v12
; LMULMAX1-RV32-NEXT: vsub.vv v9, v9, v11
; LMULMAX1-RV32-NEXT: vand.vv v11, v9, v10
; LMULMAX1-RV32-NEXT: vsrl.vi v9, v9, 2
; LMULMAX1-RV32-NEXT: vand.vv v9, v9, v10
; LMULMAX1-RV32-NEXT: vadd.vv v9, v11, v9
; LMULMAX1-RV32-NEXT: vsrl.vi v10, v9, 4
; LMULMAX1-RV32-NEXT: vadd.vv v9, v9, v10
; LMULMAX1-RV32-NEXT: vand.vv v9, v9, v13
; LMULMAX1-RV32-NEXT: vmul.vv v9, v9, v14
; LMULMAX1-RV32-NEXT: vsrl.vx v9, v9, a3
; LMULMAX1-RV32-NEXT: vse64.v v9, (a0)
; LMULMAX1-RV32-NEXT: vse64.v v8, (a1)
; LMULMAX1-RV32-NEXT: ret
;
; LMULMAX1-RV64-LABEL: cttz_v4i64:
; LMULMAX1-RV64: # %bb.0:
; LMULMAX1-RV64-NEXT: vsetivli zero, 2, e64, m1, ta, mu
; LMULMAX1-RV64-NEXT: addi a7, a0, 16
; LMULMAX1-RV64-NEXT: vle64.v v8, (a7)
; LMULMAX1-RV64-NEXT: vle64.v v9, (a0)
; LMULMAX1-RV64-NEXT: li a6, 1
; LMULMAX1-RV64-NEXT: vsub.vx v10, v8, a6
; LMULMAX1-RV64-NEXT: vxor.vi v8, v8, -1
; LMULMAX1-RV64-NEXT: vand.vv v8, v8, v10
; LMULMAX1-RV64-NEXT: vsrl.vi v10, v8, 1
; LMULMAX1-RV64-NEXT: lui a3, 21845
; LMULMAX1-RV64-NEXT: addiw a3, a3, 1365
; LMULMAX1-RV64-NEXT: slli a3, a3, 12
; LMULMAX1-RV64-NEXT: addi a3, a3, 1365
; LMULMAX1-RV64-NEXT: slli a3, a3, 12
; LMULMAX1-RV64-NEXT: addi a3, a3, 1365
; LMULMAX1-RV64-NEXT: slli a3, a3, 12
; LMULMAX1-RV64-NEXT: addi a3, a3, 1365
; LMULMAX1-RV64-NEXT: vand.vx v10, v10, a3
; LMULMAX1-RV64-NEXT: vsub.vv v8, v8, v10
; LMULMAX1-RV64-NEXT: lui a4, 13107
; LMULMAX1-RV64-NEXT: addiw a4, a4, 819
; LMULMAX1-RV64-NEXT: slli a4, a4, 12
; LMULMAX1-RV64-NEXT: addi a4, a4, 819
; LMULMAX1-RV64-NEXT: slli a4, a4, 12
; LMULMAX1-RV64-NEXT: addi a4, a4, 819
; LMULMAX1-RV64-NEXT: slli a4, a4, 12
; LMULMAX1-RV64-NEXT: addi a4, a4, 819
; LMULMAX1-RV64-NEXT: vand.vx v10, v8, a4
; LMULMAX1-RV64-NEXT: vsrl.vi v8, v8, 2
; LMULMAX1-RV64-NEXT: vand.vx v8, v8, a4
; LMULMAX1-RV64-NEXT: vadd.vv v8, v10, v8
; LMULMAX1-RV64-NEXT: vsrl.vi v10, v8, 4
; LMULMAX1-RV64-NEXT: vadd.vv v8, v8, v10
; LMULMAX1-RV64-NEXT: lui a5, 3855
; LMULMAX1-RV64-NEXT: addiw a5, a5, 241
; LMULMAX1-RV64-NEXT: slli a5, a5, 12
; LMULMAX1-RV64-NEXT: addi a5, a5, -241
; LMULMAX1-RV64-NEXT: slli a5, a5, 12
; LMULMAX1-RV64-NEXT: addi a5, a5, 241
; LMULMAX1-RV64-NEXT: slli a5, a5, 12
; LMULMAX1-RV64-NEXT: addi a5, a5, -241
; LMULMAX1-RV64-NEXT: vand.vx v8, v8, a5
; LMULMAX1-RV64-NEXT: lui a2, 4112
; LMULMAX1-RV64-NEXT: addiw a2, a2, 257
; LMULMAX1-RV64-NEXT: slli a2, a2, 16
; LMULMAX1-RV64-NEXT: addi a2, a2, 257
; LMULMAX1-RV64-NEXT: slli a2, a2, 16
; LMULMAX1-RV64-NEXT: addi a2, a2, 257
; LMULMAX1-RV64-NEXT: vmul.vx v8, v8, a2
; LMULMAX1-RV64-NEXT: li a1, 56
; LMULMAX1-RV64-NEXT: vsrl.vx v8, v8, a1
; LMULMAX1-RV64-NEXT: vsub.vx v10, v9, a6
; LMULMAX1-RV64-NEXT: vxor.vi v9, v9, -1
; LMULMAX1-RV64-NEXT: vand.vv v9, v9, v10
; LMULMAX1-RV64-NEXT: vsrl.vi v10, v9, 1
; LMULMAX1-RV64-NEXT: vand.vx v10, v10, a3
; LMULMAX1-RV64-NEXT: vsub.vv v9, v9, v10
; LMULMAX1-RV64-NEXT: vand.vx v10, v9, a4
; LMULMAX1-RV64-NEXT: vsrl.vi v9, v9, 2
; LMULMAX1-RV64-NEXT: vand.vx v9, v9, a4
; LMULMAX1-RV64-NEXT: vadd.vv v9, v10, v9
; LMULMAX1-RV64-NEXT: vsrl.vi v10, v9, 4
; LMULMAX1-RV64-NEXT: vadd.vv v9, v9, v10
; LMULMAX1-RV64-NEXT: vand.vx v9, v9, a5
; LMULMAX1-RV64-NEXT: vmul.vx v9, v9, a2
; LMULMAX1-RV64-NEXT: vsrl.vx v9, v9, a1
; LMULMAX1-RV64-NEXT: vse64.v v9, (a0)
; LMULMAX1-RV64-NEXT: vse64.v v8, (a7)
; LMULMAX1-RV64-NEXT: ret
;
; LMULMAX8-RV32-LABEL: cttz_v4i64:
; LMULMAX8-RV32: # %bb.0:
; LMULMAX8-RV32-NEXT: vsetivli zero, 4, e64, m2, ta, mu
; LMULMAX8-RV32-NEXT: vle64.v v8, (a0)
; LMULMAX8-RV32-NEXT: li a1, 1
; LMULMAX8-RV32-NEXT: vsub.vx v10, v8, a1
; LMULMAX8-RV32-NEXT: vsetivli zero, 8, e32, m2, ta, mu
; LMULMAX8-RV32-NEXT: vmv.v.i v12, -1
; LMULMAX8-RV32-NEXT: vsetivli zero, 4, e64, m2, ta, mu
; LMULMAX8-RV32-NEXT: vxor.vv v8, v8, v12
; LMULMAX8-RV32-NEXT: vand.vv v8, v8, v10
; LMULMAX8-RV32-NEXT: vsrl.vi v10, v8, 1
; LMULMAX8-RV32-NEXT: lui a1, 349525
; LMULMAX8-RV32-NEXT: addi a1, a1, 1365
; LMULMAX8-RV32-NEXT: vsetivli zero, 8, e32, m2, ta, mu
; LMULMAX8-RV32-NEXT: vmv.v.x v12, a1
; LMULMAX8-RV32-NEXT: vsetivli zero, 4, e64, m2, ta, mu
; LMULMAX8-RV32-NEXT: vand.vv v10, v10, v12
; LMULMAX8-RV32-NEXT: vsub.vv v8, v8, v10
; LMULMAX8-RV32-NEXT: lui a1, 209715
; LMULMAX8-RV32-NEXT: addi a1, a1, 819
; LMULMAX8-RV32-NEXT: vsetivli zero, 8, e32, m2, ta, mu
; LMULMAX8-RV32-NEXT: vmv.v.x v10, a1
; LMULMAX8-RV32-NEXT: vsetivli zero, 4, e64, m2, ta, mu
; LMULMAX8-RV32-NEXT: vand.vv v12, v8, v10
; LMULMAX8-RV32-NEXT: vsrl.vi v8, v8, 2
; LMULMAX8-RV32-NEXT: vand.vv v8, v8, v10
; LMULMAX8-RV32-NEXT: vadd.vv v8, v12, v8
; LMULMAX8-RV32-NEXT: vsrl.vi v10, v8, 4
; LMULMAX8-RV32-NEXT: vadd.vv v8, v8, v10
; LMULMAX8-RV32-NEXT: lui a1, 61681
; LMULMAX8-RV32-NEXT: addi a1, a1, -241
; LMULMAX8-RV32-NEXT: vsetivli zero, 8, e32, m2, ta, mu
; LMULMAX8-RV32-NEXT: vmv.v.x v10, a1
; LMULMAX8-RV32-NEXT: vsetivli zero, 4, e64, m2, ta, mu
; LMULMAX8-RV32-NEXT: vand.vv v8, v8, v10
; LMULMAX8-RV32-NEXT: lui a1, 4112
; LMULMAX8-RV32-NEXT: addi a1, a1, 257
; LMULMAX8-RV32-NEXT: vsetivli zero, 8, e32, m2, ta, mu
; LMULMAX8-RV32-NEXT: vmv.v.x v10, a1
; LMULMAX8-RV32-NEXT: vsetivli zero, 4, e64, m2, ta, mu
; LMULMAX8-RV32-NEXT: vmul.vv v8, v8, v10
; LMULMAX8-RV32-NEXT: li a1, 56
; LMULMAX8-RV32-NEXT: vsrl.vx v8, v8, a1
; LMULMAX8-RV32-NEXT: vse64.v v8, (a0)
; LMULMAX8-RV32-NEXT: ret
;
; LMULMAX8-RV64-LABEL: cttz_v4i64:
; LMULMAX8-RV64: # %bb.0:
; LMULMAX8-RV64-NEXT: vsetivli zero, 4, e64, m2, ta, mu
; LMULMAX8-RV64-NEXT: vle64.v v8, (a0)
; LMULMAX8-RV64-NEXT: li a1, 1
; LMULMAX8-RV64-NEXT: vsub.vx v10, v8, a1
; LMULMAX8-RV64-NEXT: vxor.vi v8, v8, -1
; LMULMAX8-RV64-NEXT: vand.vv v8, v8, v10
; LMULMAX8-RV64-NEXT: vsrl.vi v10, v8, 1
; LMULMAX8-RV64-NEXT: lui a1, 21845
; LMULMAX8-RV64-NEXT: addiw a1, a1, 1365
; LMULMAX8-RV64-NEXT: slli a1, a1, 12
; LMULMAX8-RV64-NEXT: addi a1, a1, 1365
; LMULMAX8-RV64-NEXT: slli a1, a1, 12
; LMULMAX8-RV64-NEXT: addi a1, a1, 1365
; LMULMAX8-RV64-NEXT: slli a1, a1, 12
; LMULMAX8-RV64-NEXT: addi a1, a1, 1365
; LMULMAX8-RV64-NEXT: vand.vx v10, v10, a1
; LMULMAX8-RV64-NEXT: vsub.vv v8, v8, v10
; LMULMAX8-RV64-NEXT: lui a1, 13107
; LMULMAX8-RV64-NEXT: addiw a1, a1, 819
; LMULMAX8-RV64-NEXT: slli a1, a1, 12
; LMULMAX8-RV64-NEXT: addi a1, a1, 819
; LMULMAX8-RV64-NEXT: slli a1, a1, 12
; LMULMAX8-RV64-NEXT: addi a1, a1, 819
; LMULMAX8-RV64-NEXT: slli a1, a1, 12
; LMULMAX8-RV64-NEXT: addi a1, a1, 819
; LMULMAX8-RV64-NEXT: vand.vx v10, v8, a1
; LMULMAX8-RV64-NEXT: vsrl.vi v8, v8, 2
; LMULMAX8-RV64-NEXT: vand.vx v8, v8, a1
; LMULMAX8-RV64-NEXT: vadd.vv v8, v10, v8
; LMULMAX8-RV64-NEXT: vsrl.vi v10, v8, 4
; LMULMAX8-RV64-NEXT: vadd.vv v8, v8, v10
; LMULMAX8-RV64-NEXT: lui a1, 3855
; LMULMAX8-RV64-NEXT: addiw a1, a1, 241
; LMULMAX8-RV64-NEXT: slli a1, a1, 12
; LMULMAX8-RV64-NEXT: addi a1, a1, -241
; LMULMAX8-RV64-NEXT: slli a1, a1, 12
; LMULMAX8-RV64-NEXT: addi a1, a1, 241
; LMULMAX8-RV64-NEXT: slli a1, a1, 12
; LMULMAX8-RV64-NEXT: addi a1, a1, -241
; LMULMAX8-RV64-NEXT: vand.vx v8, v8, a1
; LMULMAX8-RV64-NEXT: lui a1, 4112
; LMULMAX8-RV64-NEXT: addiw a1, a1, 257
; LMULMAX8-RV64-NEXT: slli a1, a1, 16
; LMULMAX8-RV64-NEXT: addi a1, a1, 257
; LMULMAX8-RV64-NEXT: slli a1, a1, 16
; LMULMAX8-RV64-NEXT: addi a1, a1, 257
; LMULMAX8-RV64-NEXT: vmul.vx v8, v8, a1
; LMULMAX8-RV64-NEXT: li a1, 56
; LMULMAX8-RV64-NEXT: vsrl.vx v8, v8, a1
; LMULMAX8-RV64-NEXT: vse64.v v8, (a0)
; LMULMAX8-RV64-NEXT: ret
%a = load <4 x i64>, <4 x i64>* %x
%b = load <4 x i64>, <4 x i64>* %y
%c = call <4 x i64> @llvm.cttz.v4i64(<4 x i64> %a, i1 false)
store <4 x i64> %c, <4 x i64>* %x
ret void
}
declare <4 x i64> @llvm.cttz.v4i64(<4 x i64>, i1)