[llvm] 44383ac - [RISCV] Add fixed vector tests for ct[l, t]z_zero_undef
Luke Lau via llvm-commits
llvm-commits at lists.llvm.org
Tue Aug 8 01:47:06 PDT 2023
Author: Luke Lau
Date: 2023-08-08T09:46:55+01:00
New Revision: 44383ac7fdbb16fea048f60d8009599b9bf7d793
URL: https://github.com/llvm/llvm-project/commit/44383ac7fdbb16fea048f60d8009599b9bf7d793
DIFF: https://github.com/llvm/llvm-project/commit/44383ac7fdbb16fea048f60d8009599b9bf7d793.diff
LOG: [RISCV] Add fixed vector tests for ct[l,t]z_zero_undef
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D157293
Added:
Modified:
llvm/test/CodeGen/RISCV/rvv/fixed-vectors-ctlz.ll
llvm/test/CodeGen/RISCV/rvv/fixed-vectors-cttz.ll
Removed:
################################################################################
diff --git a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-ctlz.ll b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-ctlz.ll
index 7de1fb27c9a73e..69a71c2e458185 100644
--- a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-ctlz.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-ctlz.ll
@@ -1291,3 +1291,1215 @@ define void @ctlz_v4i64(ptr %x, ptr %y) nounwind {
ret void
}
declare <4 x i64> @llvm.ctlz.v4i64(<4 x i64>, i1)
+
+define void @ctlz_zero_undef_v16i8(ptr %x, ptr %y) nounwind {
+; CHECK-LABEL: ctlz_zero_undef_v16i8:
+; CHECK: # %bb.0:
+; CHECK-NEXT: vsetivli zero, 16, e8, m1, ta, ma
+; CHECK-NEXT: vle8.v v8, (a0)
+; CHECK-NEXT: vsrl.vi v9, v8, 1
+; CHECK-NEXT: vor.vv v8, v8, v9
+; CHECK-NEXT: vsrl.vi v9, v8, 2
+; CHECK-NEXT: vor.vv v8, v8, v9
+; CHECK-NEXT: vsrl.vi v9, v8, 4
+; CHECK-NEXT: vor.vv v8, v8, v9
+; CHECK-NEXT: vnot.v v8, v8
+; CHECK-NEXT: vsrl.vi v9, v8, 1
+; CHECK-NEXT: li a1, 85
+; CHECK-NEXT: vand.vx v9, v9, a1
+; CHECK-NEXT: vsub.vv v8, v8, v9
+; CHECK-NEXT: li a1, 51
+; CHECK-NEXT: vand.vx v9, v8, a1
+; CHECK-NEXT: vsrl.vi v8, v8, 2
+; CHECK-NEXT: vand.vx v8, v8, a1
+; CHECK-NEXT: vadd.vv v8, v9, v8
+; CHECK-NEXT: vsrl.vi v9, v8, 4
+; CHECK-NEXT: vadd.vv v8, v8, v9
+; CHECK-NEXT: vand.vi v8, v8, 15
+; CHECK-NEXT: vse8.v v8, (a0)
+; CHECK-NEXT: ret
+;
+; LMULMAX8-LABEL: ctlz_zero_undef_v16i8:
+; LMULMAX8: # %bb.0:
+; LMULMAX8-NEXT: vsetivli zero, 16, e16, m2, ta, ma
+; LMULMAX8-NEXT: vle8.v v8, (a0)
+; LMULMAX8-NEXT: vzext.vf2 v10, v8
+; LMULMAX8-NEXT: vfwcvt.f.xu.v v12, v10
+; LMULMAX8-NEXT: vnsrl.wi v8, v12, 23
+; LMULMAX8-NEXT: vsetvli zero, zero, e8, m1, ta, ma
+; LMULMAX8-NEXT: vnsrl.wi v10, v8, 0
+; LMULMAX8-NEXT: li a1, 134
+; LMULMAX8-NEXT: vrsub.vx v8, v10, a1
+; LMULMAX8-NEXT: vse8.v v8, (a0)
+; LMULMAX8-NEXT: ret
+ %a = load <16 x i8>, ptr %x
+ %b = load <16 x i8>, ptr %y
+ %c = call <16 x i8> @llvm.ctlz.v16i8(<16 x i8> %a, i1 true)
+ store <16 x i8> %c, ptr %x
+ ret void
+}
+
+define void @ctlz_zero_undef_v8i16(ptr %x, ptr %y) nounwind {
+; LMULMAX2-RV32I-LABEL: ctlz_zero_undef_v8i16:
+; LMULMAX2-RV32I: # %bb.0:
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 8, e16, m1, ta, ma
+; LMULMAX2-RV32I-NEXT: vle16.v v8, (a0)
+; LMULMAX2-RV32I-NEXT: vsrl.vi v9, v8, 1
+; LMULMAX2-RV32I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV32I-NEXT: vsrl.vi v9, v8, 2
+; LMULMAX2-RV32I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV32I-NEXT: vsrl.vi v9, v8, 4
+; LMULMAX2-RV32I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV32I-NEXT: vsrl.vi v9, v8, 8
+; LMULMAX2-RV32I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV32I-NEXT: vnot.v v8, v8
+; 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: ctlz_zero_undef_v8i16:
+; LMULMAX2-RV64I: # %bb.0:
+; LMULMAX2-RV64I-NEXT: vsetivli zero, 8, e16, m1, ta, ma
+; LMULMAX2-RV64I-NEXT: vle16.v v8, (a0)
+; LMULMAX2-RV64I-NEXT: vsrl.vi v9, v8, 1
+; LMULMAX2-RV64I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV64I-NEXT: vsrl.vi v9, v8, 2
+; LMULMAX2-RV64I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV64I-NEXT: vsrl.vi v9, v8, 4
+; LMULMAX2-RV64I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV64I-NEXT: vsrl.vi v9, v8, 8
+; LMULMAX2-RV64I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV64I-NEXT: vnot.v v8, v8
+; 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: ctlz_zero_undef_v8i16:
+; LMULMAX1-RV32: # %bb.0:
+; LMULMAX1-RV32-NEXT: vsetivli zero, 8, e16, m1, ta, ma
+; LMULMAX1-RV32-NEXT: vle16.v v8, (a0)
+; LMULMAX1-RV32-NEXT: vsrl.vi v9, v8, 1
+; LMULMAX1-RV32-NEXT: vor.vv v8, v8, v9
+; LMULMAX1-RV32-NEXT: vsrl.vi v9, v8, 2
+; LMULMAX1-RV32-NEXT: vor.vv v8, v8, v9
+; LMULMAX1-RV32-NEXT: vsrl.vi v9, v8, 4
+; LMULMAX1-RV32-NEXT: vor.vv v8, v8, v9
+; LMULMAX1-RV32-NEXT: vsrl.vi v9, v8, 8
+; LMULMAX1-RV32-NEXT: vor.vv v8, v8, v9
+; LMULMAX1-RV32-NEXT: vnot.v v8, v8
+; 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: ctlz_zero_undef_v8i16:
+; LMULMAX1-RV64: # %bb.0:
+; LMULMAX1-RV64-NEXT: vsetivli zero, 8, e16, m1, ta, ma
+; LMULMAX1-RV64-NEXT: vle16.v v8, (a0)
+; LMULMAX1-RV64-NEXT: vsrl.vi v9, v8, 1
+; LMULMAX1-RV64-NEXT: vor.vv v8, v8, v9
+; LMULMAX1-RV64-NEXT: vsrl.vi v9, v8, 2
+; LMULMAX1-RV64-NEXT: vor.vv v8, v8, v9
+; LMULMAX1-RV64-NEXT: vsrl.vi v9, v8, 4
+; LMULMAX1-RV64-NEXT: vor.vv v8, v8, v9
+; LMULMAX1-RV64-NEXT: vsrl.vi v9, v8, 8
+; LMULMAX1-RV64-NEXT: vor.vv v8, v8, v9
+; LMULMAX1-RV64-NEXT: vnot.v v8, v8
+; 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-RV32F-LABEL: ctlz_zero_undef_v8i16:
+; LMULMAX2-RV32F: # %bb.0:
+; LMULMAX2-RV32F-NEXT: vsetivli zero, 8, e16, m1, ta, ma
+; LMULMAX2-RV32F-NEXT: vle16.v v8, (a0)
+; LMULMAX2-RV32F-NEXT: vfwcvt.f.xu.v v10, v8
+; LMULMAX2-RV32F-NEXT: vnsrl.wi v8, v10, 23
+; LMULMAX2-RV32F-NEXT: li a1, 142
+; LMULMAX2-RV32F-NEXT: vrsub.vx v8, v8, a1
+; LMULMAX2-RV32F-NEXT: vse16.v v8, (a0)
+; LMULMAX2-RV32F-NEXT: ret
+;
+; LMULMAX2-RV64F-LABEL: ctlz_zero_undef_v8i16:
+; LMULMAX2-RV64F: # %bb.0:
+; LMULMAX2-RV64F-NEXT: vsetivli zero, 8, e16, m1, ta, ma
+; LMULMAX2-RV64F-NEXT: vle16.v v8, (a0)
+; LMULMAX2-RV64F-NEXT: vfwcvt.f.xu.v v10, v8
+; LMULMAX2-RV64F-NEXT: vnsrl.wi v8, v10, 23
+; LMULMAX2-RV64F-NEXT: li a1, 142
+; LMULMAX2-RV64F-NEXT: vrsub.vx v8, v8, a1
+; LMULMAX2-RV64F-NEXT: vse16.v v8, (a0)
+; LMULMAX2-RV64F-NEXT: ret
+;
+; LMULMAX2-RV32D-LABEL: ctlz_zero_undef_v8i16:
+; LMULMAX2-RV32D: # %bb.0:
+; LMULMAX2-RV32D-NEXT: vsetivli zero, 8, e16, m1, ta, ma
+; LMULMAX2-RV32D-NEXT: vle16.v v8, (a0)
+; LMULMAX2-RV32D-NEXT: vfwcvt.f.xu.v v10, v8
+; LMULMAX2-RV32D-NEXT: vnsrl.wi v8, v10, 23
+; LMULMAX2-RV32D-NEXT: li a1, 142
+; LMULMAX2-RV32D-NEXT: vrsub.vx v8, v8, a1
+; LMULMAX2-RV32D-NEXT: vse16.v v8, (a0)
+; LMULMAX2-RV32D-NEXT: ret
+;
+; LMULMAX2-RV64D-LABEL: ctlz_zero_undef_v8i16:
+; LMULMAX2-RV64D: # %bb.0:
+; LMULMAX2-RV64D-NEXT: vsetivli zero, 8, e16, m1, ta, ma
+; LMULMAX2-RV64D-NEXT: vle16.v v8, (a0)
+; LMULMAX2-RV64D-NEXT: vfwcvt.f.xu.v v10, v8
+; LMULMAX2-RV64D-NEXT: vnsrl.wi v8, v10, 23
+; LMULMAX2-RV64D-NEXT: li a1, 142
+; LMULMAX2-RV64D-NEXT: vrsub.vx v8, v8, a1
+; LMULMAX2-RV64D-NEXT: vse16.v v8, (a0)
+; LMULMAX2-RV64D-NEXT: ret
+;
+; LMULMAX8-LABEL: ctlz_zero_undef_v8i16:
+; LMULMAX8: # %bb.0:
+; LMULMAX8-NEXT: vsetivli zero, 8, e16, m1, ta, ma
+; LMULMAX8-NEXT: vle16.v v8, (a0)
+; LMULMAX8-NEXT: vfwcvt.f.xu.v v10, v8
+; LMULMAX8-NEXT: vnsrl.wi v8, v10, 23
+; LMULMAX8-NEXT: li a1, 142
+; LMULMAX8-NEXT: vrsub.vx v8, v8, a1
+; LMULMAX8-NEXT: vse16.v v8, (a0)
+; LMULMAX8-NEXT: ret
+ %a = load <8 x i16>, ptr %x
+ %b = load <8 x i16>, ptr %y
+ %c = call <8 x i16> @llvm.ctlz.v8i16(<8 x i16> %a, i1 true)
+ store <8 x i16> %c, ptr %x
+ ret void
+}
+
+define void @ctlz_zero_undef_v4i32(ptr %x, ptr %y) nounwind {
+; LMULMAX2-RV32I-LABEL: ctlz_zero_undef_v4i32:
+; LMULMAX2-RV32I: # %bb.0:
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 4, e32, m1, ta, ma
+; LMULMAX2-RV32I-NEXT: vle32.v v8, (a0)
+; LMULMAX2-RV32I-NEXT: vsrl.vi v9, v8, 1
+; LMULMAX2-RV32I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV32I-NEXT: vsrl.vi v9, v8, 2
+; LMULMAX2-RV32I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV32I-NEXT: vsrl.vi v9, v8, 4
+; LMULMAX2-RV32I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV32I-NEXT: vsrl.vi v9, v8, 8
+; LMULMAX2-RV32I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV32I-NEXT: vsrl.vi v9, v8, 16
+; LMULMAX2-RV32I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV32I-NEXT: vnot.v v8, v8
+; 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: ctlz_zero_undef_v4i32:
+; LMULMAX2-RV64I: # %bb.0:
+; LMULMAX2-RV64I-NEXT: vsetivli zero, 4, e32, m1, ta, ma
+; LMULMAX2-RV64I-NEXT: vle32.v v8, (a0)
+; LMULMAX2-RV64I-NEXT: vsrl.vi v9, v8, 1
+; LMULMAX2-RV64I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV64I-NEXT: vsrl.vi v9, v8, 2
+; LMULMAX2-RV64I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV64I-NEXT: vsrl.vi v9, v8, 4
+; LMULMAX2-RV64I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV64I-NEXT: vsrl.vi v9, v8, 8
+; LMULMAX2-RV64I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV64I-NEXT: vsrl.vi v9, v8, 16
+; LMULMAX2-RV64I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV64I-NEXT: vnot.v v8, v8
+; 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
+;
+; LMULMAX2-RV32F-LABEL: ctlz_zero_undef_v4i32:
+; LMULMAX2-RV32F: # %bb.0:
+; LMULMAX2-RV32F-NEXT: vsetivli zero, 4, e32, m1, ta, ma
+; LMULMAX2-RV32F-NEXT: vle32.v v8, (a0)
+; LMULMAX2-RV32F-NEXT: fsrmi a1, 1
+; LMULMAX2-RV32F-NEXT: vfcvt.f.xu.v v8, v8
+; LMULMAX2-RV32F-NEXT: fsrm a1
+; LMULMAX2-RV32F-NEXT: vsrl.vi v8, v8, 23
+; LMULMAX2-RV32F-NEXT: li a1, 158
+; LMULMAX2-RV32F-NEXT: vrsub.vx v8, v8, a1
+; LMULMAX2-RV32F-NEXT: vse32.v v8, (a0)
+; LMULMAX2-RV32F-NEXT: ret
+;
+; LMULMAX2-RV64F-LABEL: ctlz_zero_undef_v4i32:
+; LMULMAX2-RV64F: # %bb.0:
+; LMULMAX2-RV64F-NEXT: vsetivli zero, 4, e32, m1, ta, ma
+; LMULMAX2-RV64F-NEXT: vle32.v v8, (a0)
+; LMULMAX2-RV64F-NEXT: fsrmi a1, 1
+; LMULMAX2-RV64F-NEXT: vfcvt.f.xu.v v8, v8
+; LMULMAX2-RV64F-NEXT: fsrm a1
+; LMULMAX2-RV64F-NEXT: vsrl.vi v8, v8, 23
+; LMULMAX2-RV64F-NEXT: li a1, 158
+; LMULMAX2-RV64F-NEXT: vrsub.vx v8, v8, a1
+; LMULMAX2-RV64F-NEXT: vse32.v v8, (a0)
+; LMULMAX2-RV64F-NEXT: ret
+;
+; LMULMAX2-RV32D-LABEL: ctlz_zero_undef_v4i32:
+; LMULMAX2-RV32D: # %bb.0:
+; LMULMAX2-RV32D-NEXT: vsetivli zero, 4, e32, m1, ta, ma
+; LMULMAX2-RV32D-NEXT: vle32.v v8, (a0)
+; LMULMAX2-RV32D-NEXT: vfwcvt.f.xu.v v10, v8
+; LMULMAX2-RV32D-NEXT: li a1, 52
+; LMULMAX2-RV32D-NEXT: vnsrl.wx v8, v10, a1
+; LMULMAX2-RV32D-NEXT: li a1, 1054
+; LMULMAX2-RV32D-NEXT: vrsub.vx v8, v8, a1
+; LMULMAX2-RV32D-NEXT: vse32.v v8, (a0)
+; LMULMAX2-RV32D-NEXT: ret
+;
+; LMULMAX2-RV64D-LABEL: ctlz_zero_undef_v4i32:
+; LMULMAX2-RV64D: # %bb.0:
+; LMULMAX2-RV64D-NEXT: vsetivli zero, 4, e32, m1, ta, ma
+; LMULMAX2-RV64D-NEXT: vle32.v v8, (a0)
+; LMULMAX2-RV64D-NEXT: vfwcvt.f.xu.v v10, v8
+; LMULMAX2-RV64D-NEXT: li a1, 52
+; LMULMAX2-RV64D-NEXT: vnsrl.wx v8, v10, a1
+; LMULMAX2-RV64D-NEXT: li a1, 1054
+; LMULMAX2-RV64D-NEXT: vrsub.vx v8, v8, a1
+; LMULMAX2-RV64D-NEXT: vse32.v v8, (a0)
+; LMULMAX2-RV64D-NEXT: ret
+;
+; LMULMAX8-LABEL: ctlz_zero_undef_v4i32:
+; LMULMAX8: # %bb.0:
+; LMULMAX8-NEXT: vsetivli zero, 4, e32, m1, ta, ma
+; LMULMAX8-NEXT: vle32.v v8, (a0)
+; LMULMAX8-NEXT: vfwcvt.f.xu.v v10, v8
+; LMULMAX8-NEXT: li a1, 52
+; LMULMAX8-NEXT: vnsrl.wx v8, v10, a1
+; LMULMAX8-NEXT: li a1, 1054
+; LMULMAX8-NEXT: vrsub.vx v8, v8, a1
+; LMULMAX8-NEXT: vse32.v v8, (a0)
+; LMULMAX8-NEXT: ret
+ %a = load <4 x i32>, ptr %x
+ %b = load <4 x i32>, ptr %y
+ %c = call <4 x i32> @llvm.ctlz.v4i32(<4 x i32> %a, i1 true)
+ store <4 x i32> %c, ptr %x
+ ret void
+}
+
+define void @ctlz_zero_undef_v2i64(ptr %x, ptr %y) nounwind {
+; LMULMAX2-RV32I-LABEL: ctlz_zero_undef_v2i64:
+; LMULMAX2-RV32I: # %bb.0:
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; LMULMAX2-RV32I-NEXT: vle64.v v8, (a0)
+; LMULMAX2-RV32I-NEXT: vsrl.vi v9, v8, 1
+; LMULMAX2-RV32I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV32I-NEXT: vsrl.vi v9, v8, 2
+; LMULMAX2-RV32I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV32I-NEXT: vsrl.vi v9, v8, 4
+; LMULMAX2-RV32I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV32I-NEXT: vsrl.vi v9, v8, 8
+; LMULMAX2-RV32I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV32I-NEXT: vsrl.vi v9, v8, 16
+; LMULMAX2-RV32I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV32I-NEXT: li a1, 32
+; LMULMAX2-RV32I-NEXT: vsrl.vx v9, v8, a1
+; LMULMAX2-RV32I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 4, e32, m1, ta, ma
+; LMULMAX2-RV32I-NEXT: vmv.v.i v9, -1
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; LMULMAX2-RV32I-NEXT: vxor.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: vsetivli zero, 4, e32, m1, ta, ma
+; LMULMAX2-RV32I-NEXT: vmv.v.x v10, a1
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; LMULMAX2-RV32I-NEXT: vand.vv v9, v9, v10
+; LMULMAX2-RV32I-NEXT: vsub.vv v8, v8, v9
+; LMULMAX2-RV32I-NEXT: lui a1, 209715
+; LMULMAX2-RV32I-NEXT: addi a1, a1, 819
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 4, e32, m1, ta, ma
+; LMULMAX2-RV32I-NEXT: vmv.v.x v9, a1
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; LMULMAX2-RV32I-NEXT: vand.vv v10, v8, v9
+; LMULMAX2-RV32I-NEXT: vsrl.vi v8, v8, 2
+; LMULMAX2-RV32I-NEXT: vand.vv v8, v8, v9
+; LMULMAX2-RV32I-NEXT: vadd.vv v8, v10, 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: vsetivli zero, 4, e32, m1, ta, ma
+; LMULMAX2-RV32I-NEXT: vmv.v.x v9, a1
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; LMULMAX2-RV32I-NEXT: vand.vv v8, v8, v9
+; LMULMAX2-RV32I-NEXT: lui a1, 4112
+; LMULMAX2-RV32I-NEXT: addi a1, a1, 257
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 4, e32, m1, ta, ma
+; LMULMAX2-RV32I-NEXT: vmv.v.x v9, a1
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; LMULMAX2-RV32I-NEXT: vmul.vv v8, v8, v9
+; LMULMAX2-RV32I-NEXT: li a1, 56
+; LMULMAX2-RV32I-NEXT: vsrl.vx v8, v8, a1
+; LMULMAX2-RV32I-NEXT: vse64.v v8, (a0)
+; LMULMAX2-RV32I-NEXT: ret
+;
+; LMULMAX2-RV64I-LABEL: ctlz_zero_undef_v2i64:
+; LMULMAX2-RV64I: # %bb.0:
+; LMULMAX2-RV64I-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; LMULMAX2-RV64I-NEXT: vle64.v v8, (a0)
+; LMULMAX2-RV64I-NEXT: vsrl.vi v9, v8, 1
+; LMULMAX2-RV64I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV64I-NEXT: vsrl.vi v9, v8, 2
+; LMULMAX2-RV64I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV64I-NEXT: vsrl.vi v9, v8, 4
+; LMULMAX2-RV64I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV64I-NEXT: vsrl.vi v9, v8, 8
+; LMULMAX2-RV64I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV64I-NEXT: vsrl.vi v9, v8, 16
+; LMULMAX2-RV64I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV64I-NEXT: li a1, 32
+; LMULMAX2-RV64I-NEXT: vsrl.vx v9, v8, a1
+; LMULMAX2-RV64I-NEXT: vor.vv v8, v8, v9
+; LMULMAX2-RV64I-NEXT: vnot.v v8, v8
+; LMULMAX2-RV64I-NEXT: vsrl.vi v9, v8, 1
+; LMULMAX2-RV64I-NEXT: lui a1, 349525
+; LMULMAX2-RV64I-NEXT: addiw a1, a1, 1365
+; LMULMAX2-RV64I-NEXT: slli a2, a1, 32
+; LMULMAX2-RV64I-NEXT: add a1, a1, a2
+; 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: slli a2, a1, 32
+; LMULMAX2-RV64I-NEXT: add a1, a1, a2
+; 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: slli a2, a1, 32
+; LMULMAX2-RV64I-NEXT: add a1, a1, a2
+; LMULMAX2-RV64I-NEXT: vand.vx v8, v8, a1
+; LMULMAX2-RV64I-NEXT: lui a1, 4112
+; LMULMAX2-RV64I-NEXT: addiw a1, a1, 257
+; LMULMAX2-RV64I-NEXT: slli a2, a1, 32
+; LMULMAX2-RV64I-NEXT: add a1, a1, a2
+; LMULMAX2-RV64I-NEXT: vmul.vx v8, v8, a1
+; LMULMAX2-RV64I-NEXT: li a1, 56
+; LMULMAX2-RV64I-NEXT: vsrl.vx v8, v8, a1
+; LMULMAX2-RV64I-NEXT: vse64.v v8, (a0)
+; LMULMAX2-RV64I-NEXT: ret
+;
+; LMULMAX2-RV32F-LABEL: ctlz_zero_undef_v2i64:
+; LMULMAX2-RV32F: # %bb.0:
+; LMULMAX2-RV32F-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; LMULMAX2-RV32F-NEXT: vle64.v v8, (a0)
+; LMULMAX2-RV32F-NEXT: li a1, 190
+; LMULMAX2-RV32F-NEXT: vmv.v.x v9, a1
+; LMULMAX2-RV32F-NEXT: fsrmi a1, 1
+; LMULMAX2-RV32F-NEXT: vsetvli zero, zero, e32, mf2, ta, ma
+; LMULMAX2-RV32F-NEXT: vfncvt.f.xu.w v10, v8
+; LMULMAX2-RV32F-NEXT: fsrm a1
+; LMULMAX2-RV32F-NEXT: vsrl.vi v8, v10, 23
+; LMULMAX2-RV32F-NEXT: vwsubu.wv v9, v9, v8
+; LMULMAX2-RV32F-NEXT: vse64.v v9, (a0)
+; LMULMAX2-RV32F-NEXT: ret
+;
+; LMULMAX2-RV64F-LABEL: ctlz_zero_undef_v2i64:
+; LMULMAX2-RV64F: # %bb.0:
+; LMULMAX2-RV64F-NEXT: vsetivli zero, 2, e32, mf2, ta, ma
+; LMULMAX2-RV64F-NEXT: vle64.v v8, (a0)
+; LMULMAX2-RV64F-NEXT: li a1, 190
+; LMULMAX2-RV64F-NEXT: vmv.v.x v9, a1
+; LMULMAX2-RV64F-NEXT: fsrmi a1, 1
+; LMULMAX2-RV64F-NEXT: vfncvt.f.xu.w v10, v8
+; LMULMAX2-RV64F-NEXT: fsrm a1
+; LMULMAX2-RV64F-NEXT: vsrl.vi v8, v10, 23
+; LMULMAX2-RV64F-NEXT: vwsubu.vv v10, v9, v8
+; LMULMAX2-RV64F-NEXT: vse64.v v10, (a0)
+; LMULMAX2-RV64F-NEXT: ret
+;
+; LMULMAX2-RV32D-LABEL: ctlz_zero_undef_v2i64:
+; LMULMAX2-RV32D: # %bb.0:
+; LMULMAX2-RV32D-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; LMULMAX2-RV32D-NEXT: vle64.v v8, (a0)
+; LMULMAX2-RV32D-NEXT: fsrmi a1, 1
+; LMULMAX2-RV32D-NEXT: vfcvt.f.xu.v v8, v8
+; LMULMAX2-RV32D-NEXT: fsrm a1
+; LMULMAX2-RV32D-NEXT: li a1, 52
+; LMULMAX2-RV32D-NEXT: vsrl.vx v8, v8, a1
+; LMULMAX2-RV32D-NEXT: li a1, 1086
+; LMULMAX2-RV32D-NEXT: vrsub.vx v8, v8, a1
+; LMULMAX2-RV32D-NEXT: vse64.v v8, (a0)
+; LMULMAX2-RV32D-NEXT: ret
+;
+; LMULMAX2-RV64D-LABEL: ctlz_zero_undef_v2i64:
+; LMULMAX2-RV64D: # %bb.0:
+; LMULMAX2-RV64D-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; LMULMAX2-RV64D-NEXT: vle64.v v8, (a0)
+; LMULMAX2-RV64D-NEXT: fsrmi a1, 1
+; LMULMAX2-RV64D-NEXT: vfcvt.f.xu.v v8, v8
+; LMULMAX2-RV64D-NEXT: fsrm a1
+; LMULMAX2-RV64D-NEXT: li a1, 52
+; LMULMAX2-RV64D-NEXT: vsrl.vx v8, v8, a1
+; LMULMAX2-RV64D-NEXT: li a1, 1086
+; LMULMAX2-RV64D-NEXT: vrsub.vx v8, v8, a1
+; LMULMAX2-RV64D-NEXT: vse64.v v8, (a0)
+; LMULMAX2-RV64D-NEXT: ret
+;
+; LMULMAX8-LABEL: ctlz_zero_undef_v2i64:
+; LMULMAX8: # %bb.0:
+; LMULMAX8-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; LMULMAX8-NEXT: vle64.v v8, (a0)
+; LMULMAX8-NEXT: fsrmi a1, 1
+; LMULMAX8-NEXT: vfcvt.f.xu.v v8, v8
+; LMULMAX8-NEXT: fsrm a1
+; LMULMAX8-NEXT: li a1, 52
+; LMULMAX8-NEXT: vsrl.vx v8, v8, a1
+; LMULMAX8-NEXT: li a1, 1086
+; LMULMAX8-NEXT: vrsub.vx v8, v8, a1
+; LMULMAX8-NEXT: vse64.v v8, (a0)
+; LMULMAX8-NEXT: ret
+ %a = load <2 x i64>, ptr %x
+ %b = load <2 x i64>, ptr %y
+ %c = call <2 x i64> @llvm.ctlz.v2i64(<2 x i64> %a, i1 true)
+ store <2 x i64> %c, ptr %x
+ ret void
+}
+
+define void @ctlz_zero_undef_v32i8(ptr %x, ptr %y) nounwind {
+; LMULMAX2-LABEL: ctlz_zero_undef_v32i8:
+; LMULMAX2: # %bb.0:
+; LMULMAX2-NEXT: li a1, 32
+; LMULMAX2-NEXT: vsetvli zero, a1, e8, m2, ta, ma
+; LMULMAX2-NEXT: vle8.v v8, (a0)
+; LMULMAX2-NEXT: vsrl.vi v10, v8, 1
+; LMULMAX2-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-NEXT: vsrl.vi v10, v8, 2
+; LMULMAX2-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-NEXT: vsrl.vi v10, v8, 4
+; LMULMAX2-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-NEXT: vnot.v v8, v8
+; LMULMAX2-NEXT: vsrl.vi v10, v8, 1
+; LMULMAX2-NEXT: li a1, 85
+; LMULMAX2-NEXT: vand.vx v10, v10, a1
+; LMULMAX2-NEXT: vsub.vv v8, v8, v10
+; LMULMAX2-NEXT: li a1, 51
+; LMULMAX2-NEXT: vand.vx v10, v8, a1
+; LMULMAX2-NEXT: vsrl.vi v8, v8, 2
+; LMULMAX2-NEXT: vand.vx v8, v8, a1
+; LMULMAX2-NEXT: vadd.vv v8, v10, v8
+; LMULMAX2-NEXT: vsrl.vi v10, v8, 4
+; LMULMAX2-NEXT: vadd.vv v8, v8, v10
+; LMULMAX2-NEXT: vand.vi v8, v8, 15
+; LMULMAX2-NEXT: vse8.v v8, (a0)
+; LMULMAX2-NEXT: ret
+;
+; LMULMAX1-LABEL: ctlz_zero_undef_v32i8:
+; LMULMAX1: # %bb.0:
+; LMULMAX1-NEXT: vsetivli zero, 16, e8, m1, ta, ma
+; LMULMAX1-NEXT: addi a1, a0, 16
+; LMULMAX1-NEXT: vle8.v v8, (a1)
+; LMULMAX1-NEXT: vle8.v v9, (a0)
+; LMULMAX1-NEXT: vsrl.vi v10, v8, 1
+; LMULMAX1-NEXT: vor.vv v8, v8, v10
+; LMULMAX1-NEXT: vsrl.vi v10, v8, 2
+; LMULMAX1-NEXT: vor.vv v8, v8, v10
+; LMULMAX1-NEXT: vsrl.vi v10, v8, 4
+; LMULMAX1-NEXT: vor.vv v8, v8, v10
+; LMULMAX1-NEXT: vnot.v v8, v8
+; LMULMAX1-NEXT: vsrl.vi v10, v8, 1
+; LMULMAX1-NEXT: li a2, 85
+; LMULMAX1-NEXT: vand.vx v10, v10, a2
+; LMULMAX1-NEXT: vsub.vv v8, v8, v10
+; LMULMAX1-NEXT: li a3, 51
+; LMULMAX1-NEXT: vand.vx v10, v8, a3
+; LMULMAX1-NEXT: vsrl.vi v8, v8, 2
+; LMULMAX1-NEXT: vand.vx v8, v8, a3
+; LMULMAX1-NEXT: vadd.vv v8, v10, v8
+; LMULMAX1-NEXT: vsrl.vi v10, v8, 4
+; LMULMAX1-NEXT: vadd.vv v8, v8, v10
+; LMULMAX1-NEXT: vand.vi v8, v8, 15
+; LMULMAX1-NEXT: vsrl.vi v10, v9, 1
+; LMULMAX1-NEXT: vor.vv v9, v9, v10
+; LMULMAX1-NEXT: vsrl.vi v10, v9, 2
+; LMULMAX1-NEXT: vor.vv v9, v9, v10
+; LMULMAX1-NEXT: vsrl.vi v10, v9, 4
+; LMULMAX1-NEXT: vor.vv v9, v9, v10
+; LMULMAX1-NEXT: vnot.v v9, v9
+; LMULMAX1-NEXT: vsrl.vi v10, v9, 1
+; LMULMAX1-NEXT: vand.vx v10, v10, a2
+; LMULMAX1-NEXT: vsub.vv v9, v9, v10
+; LMULMAX1-NEXT: vand.vx v10, v9, a3
+; LMULMAX1-NEXT: vsrl.vi v9, v9, 2
+; LMULMAX1-NEXT: vand.vx v9, v9, a3
+; LMULMAX1-NEXT: vadd.vv v9, v10, v9
+; LMULMAX1-NEXT: vsrl.vi v10, v9, 4
+; LMULMAX1-NEXT: vadd.vv v9, v9, v10
+; LMULMAX1-NEXT: vand.vi v9, v9, 15
+; LMULMAX1-NEXT: vse8.v v9, (a0)
+; LMULMAX1-NEXT: vse8.v v8, (a1)
+; LMULMAX1-NEXT: ret
+;
+; LMULMAX8-LABEL: ctlz_zero_undef_v32i8:
+; LMULMAX8: # %bb.0:
+; LMULMAX8-NEXT: li a1, 32
+; LMULMAX8-NEXT: vsetvli zero, a1, e16, m4, ta, ma
+; LMULMAX8-NEXT: vle8.v v8, (a0)
+; LMULMAX8-NEXT: vzext.vf2 v12, v8
+; LMULMAX8-NEXT: vfwcvt.f.xu.v v16, v12
+; LMULMAX8-NEXT: vnsrl.wi v8, v16, 23
+; LMULMAX8-NEXT: vsetvli zero, zero, e8, m2, ta, ma
+; LMULMAX8-NEXT: vnsrl.wi v12, v8, 0
+; LMULMAX8-NEXT: li a1, 134
+; LMULMAX8-NEXT: vrsub.vx v8, v12, a1
+; LMULMAX8-NEXT: vse8.v v8, (a0)
+; LMULMAX8-NEXT: ret
+ %a = load <32 x i8>, ptr %x
+ %b = load <32 x i8>, ptr %y
+ %c = call <32 x i8> @llvm.ctlz.v32i8(<32 x i8> %a, i1 true)
+ store <32 x i8> %c, ptr %x
+ ret void
+}
+
+define void @ctlz_zero_undef_v16i16(ptr %x, ptr %y) nounwind {
+; LMULMAX2-RV32-LABEL: ctlz_zero_undef_v16i16:
+; LMULMAX2-RV32: # %bb.0:
+; LMULMAX2-RV32-NEXT: vsetivli zero, 16, e16, m2, ta, ma
+; LMULMAX2-RV32-NEXT: vle16.v v8, (a0)
+; LMULMAX2-RV32-NEXT: vsrl.vi v10, v8, 1
+; LMULMAX2-RV32-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV32-NEXT: vsrl.vi v10, v8, 2
+; LMULMAX2-RV32-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV32-NEXT: vsrl.vi v10, v8, 4
+; LMULMAX2-RV32-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV32-NEXT: vsrl.vi v10, v8, 8
+; LMULMAX2-RV32-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV32-NEXT: vnot.v v8, v8
+; 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: ctlz_zero_undef_v16i16:
+; LMULMAX2-RV64: # %bb.0:
+; LMULMAX2-RV64-NEXT: vsetivli zero, 16, e16, m2, ta, ma
+; LMULMAX2-RV64-NEXT: vle16.v v8, (a0)
+; LMULMAX2-RV64-NEXT: vsrl.vi v10, v8, 1
+; LMULMAX2-RV64-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV64-NEXT: vsrl.vi v10, v8, 2
+; LMULMAX2-RV64-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV64-NEXT: vsrl.vi v10, v8, 4
+; LMULMAX2-RV64-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV64-NEXT: vsrl.vi v10, v8, 8
+; LMULMAX2-RV64-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV64-NEXT: vnot.v v8, v8
+; 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: ctlz_zero_undef_v16i16:
+; LMULMAX1-RV32: # %bb.0:
+; LMULMAX1-RV32-NEXT: vsetivli zero, 8, e16, m1, ta, ma
+; LMULMAX1-RV32-NEXT: addi a1, a0, 16
+; LMULMAX1-RV32-NEXT: vle16.v v8, (a1)
+; LMULMAX1-RV32-NEXT: vle16.v v9, (a0)
+; LMULMAX1-RV32-NEXT: vsrl.vi v10, v8, 1
+; LMULMAX1-RV32-NEXT: vor.vv v8, v8, v10
+; LMULMAX1-RV32-NEXT: vsrl.vi v10, v8, 2
+; LMULMAX1-RV32-NEXT: vor.vv v8, v8, v10
+; LMULMAX1-RV32-NEXT: vsrl.vi v10, v8, 4
+; LMULMAX1-RV32-NEXT: vor.vv v8, v8, v10
+; LMULMAX1-RV32-NEXT: vsrl.vi v10, v8, 8
+; LMULMAX1-RV32-NEXT: vor.vv v8, v8, v10
+; LMULMAX1-RV32-NEXT: vnot.v v8, v8
+; LMULMAX1-RV32-NEXT: vsrl.vi v10, v8, 1
+; LMULMAX1-RV32-NEXT: lui a2, 5
+; LMULMAX1-RV32-NEXT: addi a2, a2, 1365
+; LMULMAX1-RV32-NEXT: vand.vx v10, v10, a2
+; LMULMAX1-RV32-NEXT: vsub.vv v8, v8, v10
+; LMULMAX1-RV32-NEXT: lui a3, 3
+; LMULMAX1-RV32-NEXT: addi a3, a3, 819
+; LMULMAX1-RV32-NEXT: vand.vx v10, v8, a3
+; LMULMAX1-RV32-NEXT: vsrl.vi v8, v8, 2
+; LMULMAX1-RV32-NEXT: vand.vx v8, v8, a3
+; 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 a4, 1
+; LMULMAX1-RV32-NEXT: addi a4, a4, -241
+; LMULMAX1-RV32-NEXT: vand.vx v8, v8, a4
+; LMULMAX1-RV32-NEXT: li a5, 257
+; LMULMAX1-RV32-NEXT: vmul.vx v8, v8, a5
+; LMULMAX1-RV32-NEXT: vsrl.vi v8, v8, 8
+; LMULMAX1-RV32-NEXT: vsrl.vi v10, v9, 1
+; LMULMAX1-RV32-NEXT: vor.vv v9, v9, v10
+; LMULMAX1-RV32-NEXT: vsrl.vi v10, v9, 2
+; LMULMAX1-RV32-NEXT: vor.vv v9, v9, v10
+; LMULMAX1-RV32-NEXT: vsrl.vi v10, v9, 4
+; LMULMAX1-RV32-NEXT: vor.vv v9, v9, v10
+; LMULMAX1-RV32-NEXT: vsrl.vi v10, v9, 8
+; LMULMAX1-RV32-NEXT: vor.vv v9, v9, v10
+; LMULMAX1-RV32-NEXT: vnot.v v9, v9
+; LMULMAX1-RV32-NEXT: vsrl.vi v10, v9, 1
+; LMULMAX1-RV32-NEXT: vand.vx v10, v10, a2
+; LMULMAX1-RV32-NEXT: vsub.vv v9, v9, v10
+; LMULMAX1-RV32-NEXT: vand.vx v10, v9, a3
+; LMULMAX1-RV32-NEXT: vsrl.vi v9, v9, 2
+; LMULMAX1-RV32-NEXT: vand.vx v9, v9, a3
+; 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, a4
+; LMULMAX1-RV32-NEXT: vmul.vx v9, v9, a5
+; 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: ctlz_zero_undef_v16i16:
+; LMULMAX1-RV64: # %bb.0:
+; LMULMAX1-RV64-NEXT: vsetivli zero, 8, e16, m1, ta, ma
+; LMULMAX1-RV64-NEXT: addi a1, a0, 16
+; LMULMAX1-RV64-NEXT: vle16.v v8, (a1)
+; LMULMAX1-RV64-NEXT: vle16.v v9, (a0)
+; LMULMAX1-RV64-NEXT: vsrl.vi v10, v8, 1
+; LMULMAX1-RV64-NEXT: vor.vv v8, v8, v10
+; LMULMAX1-RV64-NEXT: vsrl.vi v10, v8, 2
+; LMULMAX1-RV64-NEXT: vor.vv v8, v8, v10
+; LMULMAX1-RV64-NEXT: vsrl.vi v10, v8, 4
+; LMULMAX1-RV64-NEXT: vor.vv v8, v8, v10
+; LMULMAX1-RV64-NEXT: vsrl.vi v10, v8, 8
+; LMULMAX1-RV64-NEXT: vor.vv v8, v8, v10
+; LMULMAX1-RV64-NEXT: vnot.v v8, v8
+; LMULMAX1-RV64-NEXT: vsrl.vi v10, v8, 1
+; LMULMAX1-RV64-NEXT: lui a2, 5
+; LMULMAX1-RV64-NEXT: addiw a2, a2, 1365
+; LMULMAX1-RV64-NEXT: vand.vx v10, v10, a2
+; LMULMAX1-RV64-NEXT: vsub.vv v8, v8, v10
+; LMULMAX1-RV64-NEXT: lui a3, 3
+; LMULMAX1-RV64-NEXT: addiw a3, a3, 819
+; LMULMAX1-RV64-NEXT: vand.vx v10, v8, a3
+; LMULMAX1-RV64-NEXT: vsrl.vi v8, v8, 2
+; LMULMAX1-RV64-NEXT: vand.vx v8, v8, a3
+; 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 a4, 1
+; LMULMAX1-RV64-NEXT: addiw a4, a4, -241
+; LMULMAX1-RV64-NEXT: vand.vx v8, v8, a4
+; LMULMAX1-RV64-NEXT: li a5, 257
+; LMULMAX1-RV64-NEXT: vmul.vx v8, v8, a5
+; LMULMAX1-RV64-NEXT: vsrl.vi v8, v8, 8
+; LMULMAX1-RV64-NEXT: vsrl.vi v10, v9, 1
+; LMULMAX1-RV64-NEXT: vor.vv v9, v9, v10
+; LMULMAX1-RV64-NEXT: vsrl.vi v10, v9, 2
+; LMULMAX1-RV64-NEXT: vor.vv v9, v9, v10
+; LMULMAX1-RV64-NEXT: vsrl.vi v10, v9, 4
+; LMULMAX1-RV64-NEXT: vor.vv v9, v9, v10
+; LMULMAX1-RV64-NEXT: vsrl.vi v10, v9, 8
+; LMULMAX1-RV64-NEXT: vor.vv v9, v9, v10
+; LMULMAX1-RV64-NEXT: vnot.v v9, v9
+; LMULMAX1-RV64-NEXT: vsrl.vi v10, v9, 1
+; LMULMAX1-RV64-NEXT: vand.vx v10, v10, a2
+; LMULMAX1-RV64-NEXT: vsub.vv v9, v9, v10
+; LMULMAX1-RV64-NEXT: vand.vx v10, v9, a3
+; LMULMAX1-RV64-NEXT: vsrl.vi v9, v9, 2
+; LMULMAX1-RV64-NEXT: vand.vx v9, v9, a3
+; 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, a4
+; LMULMAX1-RV64-NEXT: vmul.vx v9, v9, a5
+; 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-LABEL: ctlz_zero_undef_v16i16:
+; LMULMAX8: # %bb.0:
+; LMULMAX8-NEXT: vsetivli zero, 16, e16, m2, ta, ma
+; LMULMAX8-NEXT: vle16.v v8, (a0)
+; LMULMAX8-NEXT: vfwcvt.f.xu.v v12, v8
+; LMULMAX8-NEXT: vnsrl.wi v8, v12, 23
+; LMULMAX8-NEXT: li a1, 142
+; LMULMAX8-NEXT: vrsub.vx v8, v8, a1
+; LMULMAX8-NEXT: vse16.v v8, (a0)
+; LMULMAX8-NEXT: ret
+ %a = load <16 x i16>, ptr %x
+ %b = load <16 x i16>, ptr %y
+ %c = call <16 x i16> @llvm.ctlz.v16i16(<16 x i16> %a, i1 true)
+ store <16 x i16> %c, ptr %x
+ ret void
+}
+
+define void @ctlz_zero_undef_v8i32(ptr %x, ptr %y) nounwind {
+; LMULMAX2-RV32I-LABEL: ctlz_zero_undef_v8i32:
+; LMULMAX2-RV32I: # %bb.0:
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; LMULMAX2-RV32I-NEXT: vle32.v v8, (a0)
+; LMULMAX2-RV32I-NEXT: vsrl.vi v10, v8, 1
+; LMULMAX2-RV32I-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV32I-NEXT: vsrl.vi v10, v8, 2
+; LMULMAX2-RV32I-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV32I-NEXT: vsrl.vi v10, v8, 4
+; LMULMAX2-RV32I-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV32I-NEXT: vsrl.vi v10, v8, 8
+; LMULMAX2-RV32I-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV32I-NEXT: vsrl.vi v10, v8, 16
+; LMULMAX2-RV32I-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV32I-NEXT: vnot.v v8, v8
+; LMULMAX2-RV32I-NEXT: vsrl.vi v10, v8, 1
+; LMULMAX2-RV32I-NEXT: lui a1, 349525
+; LMULMAX2-RV32I-NEXT: addi a1, a1, 1365
+; LMULMAX2-RV32I-NEXT: vand.vx v10, v10, a1
+; LMULMAX2-RV32I-NEXT: vsub.vv v8, v8, v10
+; LMULMAX2-RV32I-NEXT: lui a1, 209715
+; LMULMAX2-RV32I-NEXT: addi a1, a1, 819
+; LMULMAX2-RV32I-NEXT: vand.vx v10, v8, a1
+; LMULMAX2-RV32I-NEXT: vsrl.vi v8, v8, 2
+; LMULMAX2-RV32I-NEXT: vand.vx v8, v8, a1
+; LMULMAX2-RV32I-NEXT: vadd.vv v8, v10, v8
+; LMULMAX2-RV32I-NEXT: vsrl.vi v10, v8, 4
+; LMULMAX2-RV32I-NEXT: vadd.vv v8, v8, v10
+; 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: ctlz_zero_undef_v8i32:
+; LMULMAX2-RV64I: # %bb.0:
+; LMULMAX2-RV64I-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; LMULMAX2-RV64I-NEXT: vle32.v v8, (a0)
+; LMULMAX2-RV64I-NEXT: vsrl.vi v10, v8, 1
+; LMULMAX2-RV64I-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV64I-NEXT: vsrl.vi v10, v8, 2
+; LMULMAX2-RV64I-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV64I-NEXT: vsrl.vi v10, v8, 4
+; LMULMAX2-RV64I-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV64I-NEXT: vsrl.vi v10, v8, 8
+; LMULMAX2-RV64I-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV64I-NEXT: vsrl.vi v10, v8, 16
+; LMULMAX2-RV64I-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV64I-NEXT: vnot.v v8, v8
+; LMULMAX2-RV64I-NEXT: vsrl.vi v10, v8, 1
+; LMULMAX2-RV64I-NEXT: lui a1, 349525
+; LMULMAX2-RV64I-NEXT: addiw a1, a1, 1365
+; LMULMAX2-RV64I-NEXT: vand.vx v10, v10, a1
+; LMULMAX2-RV64I-NEXT: vsub.vv v8, v8, v10
+; LMULMAX2-RV64I-NEXT: lui a1, 209715
+; LMULMAX2-RV64I-NEXT: addiw a1, a1, 819
+; LMULMAX2-RV64I-NEXT: vand.vx v10, v8, a1
+; LMULMAX2-RV64I-NEXT: vsrl.vi v8, v8, 2
+; LMULMAX2-RV64I-NEXT: vand.vx v8, v8, a1
+; LMULMAX2-RV64I-NEXT: vadd.vv v8, v10, v8
+; LMULMAX2-RV64I-NEXT: vsrl.vi v10, v8, 4
+; LMULMAX2-RV64I-NEXT: vadd.vv v8, v8, v10
+; 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
+;
+; LMULMAX2-RV32F-LABEL: ctlz_zero_undef_v8i32:
+; LMULMAX2-RV32F: # %bb.0:
+; LMULMAX2-RV32F-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; LMULMAX2-RV32F-NEXT: vle32.v v8, (a0)
+; LMULMAX2-RV32F-NEXT: fsrmi a1, 1
+; LMULMAX2-RV32F-NEXT: vfcvt.f.xu.v v8, v8
+; LMULMAX2-RV32F-NEXT: fsrm a1
+; LMULMAX2-RV32F-NEXT: vsrl.vi v8, v8, 23
+; LMULMAX2-RV32F-NEXT: li a1, 158
+; LMULMAX2-RV32F-NEXT: vrsub.vx v8, v8, a1
+; LMULMAX2-RV32F-NEXT: vse32.v v8, (a0)
+; LMULMAX2-RV32F-NEXT: ret
+;
+; LMULMAX2-RV64F-LABEL: ctlz_zero_undef_v8i32:
+; LMULMAX2-RV64F: # %bb.0:
+; LMULMAX2-RV64F-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; LMULMAX2-RV64F-NEXT: vle32.v v8, (a0)
+; LMULMAX2-RV64F-NEXT: fsrmi a1, 1
+; LMULMAX2-RV64F-NEXT: vfcvt.f.xu.v v8, v8
+; LMULMAX2-RV64F-NEXT: fsrm a1
+; LMULMAX2-RV64F-NEXT: vsrl.vi v8, v8, 23
+; LMULMAX2-RV64F-NEXT: li a1, 158
+; LMULMAX2-RV64F-NEXT: vrsub.vx v8, v8, a1
+; LMULMAX2-RV64F-NEXT: vse32.v v8, (a0)
+; LMULMAX2-RV64F-NEXT: ret
+;
+; LMULMAX2-RV32D-LABEL: ctlz_zero_undef_v8i32:
+; LMULMAX2-RV32D: # %bb.0:
+; LMULMAX2-RV32D-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; LMULMAX2-RV32D-NEXT: vle32.v v8, (a0)
+; LMULMAX2-RV32D-NEXT: fsrmi a1, 1
+; LMULMAX2-RV32D-NEXT: vfcvt.f.xu.v v8, v8
+; LMULMAX2-RV32D-NEXT: fsrm a1
+; LMULMAX2-RV32D-NEXT: vsrl.vi v8, v8, 23
+; LMULMAX2-RV32D-NEXT: li a1, 158
+; LMULMAX2-RV32D-NEXT: vrsub.vx v8, v8, a1
+; LMULMAX2-RV32D-NEXT: vse32.v v8, (a0)
+; LMULMAX2-RV32D-NEXT: ret
+;
+; LMULMAX2-RV64D-LABEL: ctlz_zero_undef_v8i32:
+; LMULMAX2-RV64D: # %bb.0:
+; LMULMAX2-RV64D-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; LMULMAX2-RV64D-NEXT: vle32.v v8, (a0)
+; LMULMAX2-RV64D-NEXT: fsrmi a1, 1
+; LMULMAX2-RV64D-NEXT: vfcvt.f.xu.v v8, v8
+; LMULMAX2-RV64D-NEXT: fsrm a1
+; LMULMAX2-RV64D-NEXT: vsrl.vi v8, v8, 23
+; LMULMAX2-RV64D-NEXT: li a1, 158
+; LMULMAX2-RV64D-NEXT: vrsub.vx v8, v8, a1
+; LMULMAX2-RV64D-NEXT: vse32.v v8, (a0)
+; LMULMAX2-RV64D-NEXT: ret
+;
+; LMULMAX8-LABEL: ctlz_zero_undef_v8i32:
+; LMULMAX8: # %bb.0:
+; LMULMAX8-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; LMULMAX8-NEXT: vle32.v v8, (a0)
+; LMULMAX8-NEXT: vfwcvt.f.xu.v v12, v8
+; LMULMAX8-NEXT: li a1, 52
+; LMULMAX8-NEXT: vnsrl.wx v8, v12, a1
+; LMULMAX8-NEXT: li a1, 1054
+; LMULMAX8-NEXT: vrsub.vx v8, v8, a1
+; LMULMAX8-NEXT: vse32.v v8, (a0)
+; LMULMAX8-NEXT: ret
+ %a = load <8 x i32>, ptr %x
+ %b = load <8 x i32>, ptr %y
+ %c = call <8 x i32> @llvm.ctlz.v8i32(<8 x i32> %a, i1 true)
+ store <8 x i32> %c, ptr %x
+ ret void
+}
+
+define void @ctlz_zero_undef_v4i64(ptr %x, ptr %y) nounwind {
+; LMULMAX2-RV32I-LABEL: ctlz_zero_undef_v4i64:
+; LMULMAX2-RV32I: # %bb.0:
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 4, e64, m2, ta, ma
+; LMULMAX2-RV32I-NEXT: vle64.v v8, (a0)
+; LMULMAX2-RV32I-NEXT: vsrl.vi v10, v8, 1
+; LMULMAX2-RV32I-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV32I-NEXT: vsrl.vi v10, v8, 2
+; LMULMAX2-RV32I-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV32I-NEXT: vsrl.vi v10, v8, 4
+; LMULMAX2-RV32I-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV32I-NEXT: vsrl.vi v10, v8, 8
+; LMULMAX2-RV32I-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV32I-NEXT: vsrl.vi v10, v8, 16
+; LMULMAX2-RV32I-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV32I-NEXT: li a1, 32
+; LMULMAX2-RV32I-NEXT: vsrl.vx v10, v8, a1
+; LMULMAX2-RV32I-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; LMULMAX2-RV32I-NEXT: vmv.v.i v10, -1
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 4, e64, m2, ta, ma
+; LMULMAX2-RV32I-NEXT: vxor.vv v8, v8, v10
+; LMULMAX2-RV32I-NEXT: vsrl.vi v10, v8, 1
+; LMULMAX2-RV32I-NEXT: lui a1, 349525
+; LMULMAX2-RV32I-NEXT: addi a1, a1, 1365
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; LMULMAX2-RV32I-NEXT: vmv.v.x v12, a1
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 4, e64, m2, ta, ma
+; LMULMAX2-RV32I-NEXT: vand.vv v10, v10, v12
+; LMULMAX2-RV32I-NEXT: vsub.vv v8, v8, v10
+; LMULMAX2-RV32I-NEXT: lui a1, 209715
+; LMULMAX2-RV32I-NEXT: addi a1, a1, 819
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; LMULMAX2-RV32I-NEXT: vmv.v.x v10, a1
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 4, e64, m2, ta, ma
+; LMULMAX2-RV32I-NEXT: vand.vv v12, v8, v10
+; LMULMAX2-RV32I-NEXT: vsrl.vi v8, v8, 2
+; LMULMAX2-RV32I-NEXT: vand.vv v8, v8, v10
+; LMULMAX2-RV32I-NEXT: vadd.vv v8, v12, v8
+; LMULMAX2-RV32I-NEXT: vsrl.vi v10, v8, 4
+; LMULMAX2-RV32I-NEXT: vadd.vv v8, v8, v10
+; LMULMAX2-RV32I-NEXT: lui a1, 61681
+; LMULMAX2-RV32I-NEXT: addi a1, a1, -241
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; LMULMAX2-RV32I-NEXT: vmv.v.x v10, a1
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 4, e64, m2, ta, ma
+; LMULMAX2-RV32I-NEXT: vand.vv v8, v8, v10
+; LMULMAX2-RV32I-NEXT: lui a1, 4112
+; LMULMAX2-RV32I-NEXT: addi a1, a1, 257
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; LMULMAX2-RV32I-NEXT: vmv.v.x v10, a1
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 4, e64, m2, ta, ma
+; LMULMAX2-RV32I-NEXT: vmul.vv v8, v8, v10
+; LMULMAX2-RV32I-NEXT: li a1, 56
+; LMULMAX2-RV32I-NEXT: vsrl.vx v8, v8, a1
+; LMULMAX2-RV32I-NEXT: vse64.v v8, (a0)
+; LMULMAX2-RV32I-NEXT: ret
+;
+; LMULMAX2-RV64I-LABEL: ctlz_zero_undef_v4i64:
+; LMULMAX2-RV64I: # %bb.0:
+; LMULMAX2-RV64I-NEXT: vsetivli zero, 4, e64, m2, ta, ma
+; LMULMAX2-RV64I-NEXT: vle64.v v8, (a0)
+; LMULMAX2-RV64I-NEXT: vsrl.vi v10, v8, 1
+; LMULMAX2-RV64I-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV64I-NEXT: vsrl.vi v10, v8, 2
+; LMULMAX2-RV64I-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV64I-NEXT: vsrl.vi v10, v8, 4
+; LMULMAX2-RV64I-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV64I-NEXT: vsrl.vi v10, v8, 8
+; LMULMAX2-RV64I-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV64I-NEXT: vsrl.vi v10, v8, 16
+; LMULMAX2-RV64I-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV64I-NEXT: li a1, 32
+; LMULMAX2-RV64I-NEXT: vsrl.vx v10, v8, a1
+; LMULMAX2-RV64I-NEXT: vor.vv v8, v8, v10
+; LMULMAX2-RV64I-NEXT: vnot.v v8, v8
+; LMULMAX2-RV64I-NEXT: vsrl.vi v10, v8, 1
+; LMULMAX2-RV64I-NEXT: lui a1, 349525
+; LMULMAX2-RV64I-NEXT: addiw a1, a1, 1365
+; LMULMAX2-RV64I-NEXT: slli a2, a1, 32
+; LMULMAX2-RV64I-NEXT: add a1, a1, a2
+; LMULMAX2-RV64I-NEXT: vand.vx v10, v10, a1
+; LMULMAX2-RV64I-NEXT: vsub.vv v8, v8, v10
+; LMULMAX2-RV64I-NEXT: lui a1, 209715
+; LMULMAX2-RV64I-NEXT: addiw a1, a1, 819
+; LMULMAX2-RV64I-NEXT: slli a2, a1, 32
+; LMULMAX2-RV64I-NEXT: add a1, a1, a2
+; LMULMAX2-RV64I-NEXT: vand.vx v10, v8, a1
+; LMULMAX2-RV64I-NEXT: vsrl.vi v8, v8, 2
+; LMULMAX2-RV64I-NEXT: vand.vx v8, v8, a1
+; LMULMAX2-RV64I-NEXT: vadd.vv v8, v10, v8
+; LMULMAX2-RV64I-NEXT: vsrl.vi v10, v8, 4
+; LMULMAX2-RV64I-NEXT: vadd.vv v8, v8, v10
+; LMULMAX2-RV64I-NEXT: lui a1, 61681
+; LMULMAX2-RV64I-NEXT: addiw a1, a1, -241
+; LMULMAX2-RV64I-NEXT: slli a2, a1, 32
+; LMULMAX2-RV64I-NEXT: add a1, a1, a2
+; LMULMAX2-RV64I-NEXT: vand.vx v8, v8, a1
+; LMULMAX2-RV64I-NEXT: lui a1, 4112
+; LMULMAX2-RV64I-NEXT: addiw a1, a1, 257
+; LMULMAX2-RV64I-NEXT: slli a2, a1, 32
+; LMULMAX2-RV64I-NEXT: add a1, a1, a2
+; LMULMAX2-RV64I-NEXT: vmul.vx v8, v8, a1
+; LMULMAX2-RV64I-NEXT: li a1, 56
+; LMULMAX2-RV64I-NEXT: vsrl.vx v8, v8, a1
+; LMULMAX2-RV64I-NEXT: vse64.v v8, (a0)
+; LMULMAX2-RV64I-NEXT: ret
+;
+; LMULMAX2-RV32F-LABEL: ctlz_zero_undef_v4i64:
+; LMULMAX2-RV32F: # %bb.0:
+; LMULMAX2-RV32F-NEXT: vsetivli zero, 4, e64, m2, ta, ma
+; LMULMAX2-RV32F-NEXT: vle64.v v8, (a0)
+; LMULMAX2-RV32F-NEXT: li a1, 190
+; LMULMAX2-RV32F-NEXT: vmv.v.x v10, a1
+; LMULMAX2-RV32F-NEXT: fsrmi a1, 1
+; LMULMAX2-RV32F-NEXT: vsetvli zero, zero, e32, m1, ta, ma
+; LMULMAX2-RV32F-NEXT: vfncvt.f.xu.w v12, v8
+; LMULMAX2-RV32F-NEXT: fsrm a1
+; LMULMAX2-RV32F-NEXT: vsrl.vi v8, v12, 23
+; LMULMAX2-RV32F-NEXT: vwsubu.wv v10, v10, v8
+; LMULMAX2-RV32F-NEXT: vse64.v v10, (a0)
+; LMULMAX2-RV32F-NEXT: ret
+;
+; LMULMAX2-RV64F-LABEL: ctlz_zero_undef_v4i64:
+; LMULMAX2-RV64F: # %bb.0:
+; LMULMAX2-RV64F-NEXT: vsetivli zero, 4, e32, m1, ta, ma
+; LMULMAX2-RV64F-NEXT: vle64.v v8, (a0)
+; LMULMAX2-RV64F-NEXT: li a1, 190
+; LMULMAX2-RV64F-NEXT: vmv.v.x v10, a1
+; LMULMAX2-RV64F-NEXT: fsrmi a1, 1
+; LMULMAX2-RV64F-NEXT: vfncvt.f.xu.w v11, v8
+; LMULMAX2-RV64F-NEXT: fsrm a1
+; LMULMAX2-RV64F-NEXT: vsrl.vi v8, v11, 23
+; LMULMAX2-RV64F-NEXT: vwsubu.vv v12, v10, v8
+; LMULMAX2-RV64F-NEXT: vse64.v v12, (a0)
+; LMULMAX2-RV64F-NEXT: ret
+;
+; LMULMAX2-RV32D-LABEL: ctlz_zero_undef_v4i64:
+; LMULMAX2-RV32D: # %bb.0:
+; LMULMAX2-RV32D-NEXT: vsetivli zero, 4, e64, m2, ta, ma
+; LMULMAX2-RV32D-NEXT: vle64.v v8, (a0)
+; LMULMAX2-RV32D-NEXT: fsrmi a1, 1
+; LMULMAX2-RV32D-NEXT: vfcvt.f.xu.v v8, v8
+; LMULMAX2-RV32D-NEXT: fsrm a1
+; LMULMAX2-RV32D-NEXT: li a1, 52
+; LMULMAX2-RV32D-NEXT: vsrl.vx v8, v8, a1
+; LMULMAX2-RV32D-NEXT: li a1, 1086
+; LMULMAX2-RV32D-NEXT: vrsub.vx v8, v8, a1
+; LMULMAX2-RV32D-NEXT: vse64.v v8, (a0)
+; LMULMAX2-RV32D-NEXT: ret
+;
+; LMULMAX2-RV64D-LABEL: ctlz_zero_undef_v4i64:
+; LMULMAX2-RV64D: # %bb.0:
+; LMULMAX2-RV64D-NEXT: vsetivli zero, 4, e64, m2, ta, ma
+; LMULMAX2-RV64D-NEXT: vle64.v v8, (a0)
+; LMULMAX2-RV64D-NEXT: fsrmi a1, 1
+; LMULMAX2-RV64D-NEXT: vfcvt.f.xu.v v8, v8
+; LMULMAX2-RV64D-NEXT: fsrm a1
+; LMULMAX2-RV64D-NEXT: li a1, 52
+; LMULMAX2-RV64D-NEXT: vsrl.vx v8, v8, a1
+; LMULMAX2-RV64D-NEXT: li a1, 1086
+; LMULMAX2-RV64D-NEXT: vrsub.vx v8, v8, a1
+; LMULMAX2-RV64D-NEXT: vse64.v v8, (a0)
+; LMULMAX2-RV64D-NEXT: ret
+;
+; LMULMAX8-LABEL: ctlz_zero_undef_v4i64:
+; LMULMAX8: # %bb.0:
+; LMULMAX8-NEXT: vsetivli zero, 4, e64, m2, ta, ma
+; LMULMAX8-NEXT: vle64.v v8, (a0)
+; LMULMAX8-NEXT: fsrmi a1, 1
+; LMULMAX8-NEXT: vfcvt.f.xu.v v8, v8
+; LMULMAX8-NEXT: fsrm a1
+; LMULMAX8-NEXT: li a1, 52
+; LMULMAX8-NEXT: vsrl.vx v8, v8, a1
+; LMULMAX8-NEXT: li a1, 1086
+; LMULMAX8-NEXT: vrsub.vx v8, v8, a1
+; LMULMAX8-NEXT: vse64.v v8, (a0)
+; LMULMAX8-NEXT: ret
+ %a = load <4 x i64>, ptr %x
+ %b = load <4 x i64>, ptr %y
+ %c = call <4 x i64> @llvm.ctlz.v4i64(<4 x i64> %a, i1 true)
+ store <4 x i64> %c, ptr %x
+ ret void
+}
diff --git a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-cttz.ll b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-cttz.ll
index 7a9220ec39f74b..5f48a274df3481 100644
--- a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-cttz.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-cttz.ll
@@ -1298,3 +1298,1192 @@ define void @cttz_v4i64(ptr %x, ptr %y) nounwind {
ret void
}
declare <4 x i64> @llvm.cttz.v4i64(<4 x i64>, i1)
+
+define void @cttz_zero_undef_v16i8(ptr %x, ptr %y) nounwind {
+; CHECK-LABEL: cttz_zero_undef_v16i8:
+; CHECK: # %bb.0:
+; CHECK-NEXT: vsetivli zero, 16, e8, m1, ta, ma
+; CHECK-NEXT: vle8.v v8, (a0)
+; CHECK-NEXT: li a1, 1
+; CHECK-NEXT: vsub.vx v9, v8, a1
+; CHECK-NEXT: vnot.v v8, v8
+; CHECK-NEXT: vand.vv v8, v8, v9
+; CHECK-NEXT: vsrl.vi v9, v8, 1
+; CHECK-NEXT: li a1, 85
+; CHECK-NEXT: vand.vx v9, v9, a1
+; CHECK-NEXT: vsub.vv v8, v8, v9
+; CHECK-NEXT: li a1, 51
+; CHECK-NEXT: vand.vx v9, v8, a1
+; CHECK-NEXT: vsrl.vi v8, v8, 2
+; CHECK-NEXT: vand.vx v8, v8, a1
+; CHECK-NEXT: vadd.vv v8, v9, v8
+; CHECK-NEXT: vsrl.vi v9, v8, 4
+; CHECK-NEXT: vadd.vv v8, v8, v9
+; CHECK-NEXT: vand.vi v8, v8, 15
+; CHECK-NEXT: vse8.v v8, (a0)
+; CHECK-NEXT: ret
+;
+; LMULMAX8-LABEL: cttz_zero_undef_v16i8:
+; LMULMAX8: # %bb.0:
+; LMULMAX8-NEXT: vsetivli zero, 16, e8, m1, ta, ma
+; LMULMAX8-NEXT: vle8.v v8, (a0)
+; LMULMAX8-NEXT: vrsub.vi v9, v8, 0
+; LMULMAX8-NEXT: vand.vv v8, v8, v9
+; LMULMAX8-NEXT: vsetvli zero, zero, e16, m2, ta, ma
+; LMULMAX8-NEXT: vzext.vf2 v10, v8
+; LMULMAX8-NEXT: vfwcvt.f.xu.v v12, v10
+; LMULMAX8-NEXT: vnsrl.wi v8, v12, 23
+; LMULMAX8-NEXT: vsetvli zero, zero, e8, m1, ta, ma
+; LMULMAX8-NEXT: vnsrl.wi v10, v8, 0
+; LMULMAX8-NEXT: li a1, 127
+; LMULMAX8-NEXT: vsub.vx v8, v10, a1
+; LMULMAX8-NEXT: vse8.v v8, (a0)
+; LMULMAX8-NEXT: ret
+ %a = load <16 x i8>, ptr %x
+ %b = load <16 x i8>, ptr %y
+ %c = call <16 x i8> @llvm.cttz.v16i8(<16 x i8> %a, i1 true)
+ store <16 x i8> %c, ptr %x
+ ret void
+}
+
+define void @cttz_zero_undef_v8i16(ptr %x, ptr %y) nounwind {
+; LMULMAX2-RV32I-LABEL: cttz_zero_undef_v8i16:
+; LMULMAX2-RV32I: # %bb.0:
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 8, e16, m1, ta, ma
+; LMULMAX2-RV32I-NEXT: vle16.v v8, (a0)
+; LMULMAX2-RV32I-NEXT: li a1, 1
+; LMULMAX2-RV32I-NEXT: vsub.vx v9, v8, a1
+; LMULMAX2-RV32I-NEXT: vnot.v v8, v8
+; 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_zero_undef_v8i16:
+; LMULMAX2-RV64I: # %bb.0:
+; LMULMAX2-RV64I-NEXT: vsetivli zero, 8, e16, m1, ta, ma
+; LMULMAX2-RV64I-NEXT: vle16.v v8, (a0)
+; LMULMAX2-RV64I-NEXT: li a1, 1
+; LMULMAX2-RV64I-NEXT: vsub.vx v9, v8, a1
+; LMULMAX2-RV64I-NEXT: vnot.v v8, v8
+; 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_zero_undef_v8i16:
+; LMULMAX1-RV32: # %bb.0:
+; LMULMAX1-RV32-NEXT: vsetivli zero, 8, e16, m1, ta, ma
+; LMULMAX1-RV32-NEXT: vle16.v v8, (a0)
+; LMULMAX1-RV32-NEXT: li a1, 1
+; LMULMAX1-RV32-NEXT: vsub.vx v9, v8, a1
+; LMULMAX1-RV32-NEXT: vnot.v v8, v8
+; 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_zero_undef_v8i16:
+; LMULMAX1-RV64: # %bb.0:
+; LMULMAX1-RV64-NEXT: vsetivli zero, 8, e16, m1, ta, ma
+; LMULMAX1-RV64-NEXT: vle16.v v8, (a0)
+; LMULMAX1-RV64-NEXT: li a1, 1
+; LMULMAX1-RV64-NEXT: vsub.vx v9, v8, a1
+; LMULMAX1-RV64-NEXT: vnot.v v8, v8
+; 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-RV32F-LABEL: cttz_zero_undef_v8i16:
+; LMULMAX2-RV32F: # %bb.0:
+; LMULMAX2-RV32F-NEXT: vsetivli zero, 8, e16, m1, ta, ma
+; LMULMAX2-RV32F-NEXT: vle16.v v8, (a0)
+; LMULMAX2-RV32F-NEXT: vrsub.vi v9, v8, 0
+; LMULMAX2-RV32F-NEXT: vand.vv v8, v8, v9
+; LMULMAX2-RV32F-NEXT: vfwcvt.f.xu.v v10, v8
+; LMULMAX2-RV32F-NEXT: vnsrl.wi v8, v10, 23
+; LMULMAX2-RV32F-NEXT: li a1, 127
+; LMULMAX2-RV32F-NEXT: vsub.vx v8, v8, a1
+; LMULMAX2-RV32F-NEXT: vse16.v v8, (a0)
+; LMULMAX2-RV32F-NEXT: ret
+;
+; LMULMAX2-RV64F-LABEL: cttz_zero_undef_v8i16:
+; LMULMAX2-RV64F: # %bb.0:
+; LMULMAX2-RV64F-NEXT: vsetivli zero, 8, e16, m1, ta, ma
+; LMULMAX2-RV64F-NEXT: vle16.v v8, (a0)
+; LMULMAX2-RV64F-NEXT: vrsub.vi v9, v8, 0
+; LMULMAX2-RV64F-NEXT: vand.vv v8, v8, v9
+; LMULMAX2-RV64F-NEXT: vfwcvt.f.xu.v v10, v8
+; LMULMAX2-RV64F-NEXT: vnsrl.wi v8, v10, 23
+; LMULMAX2-RV64F-NEXT: li a1, 127
+; LMULMAX2-RV64F-NEXT: vsub.vx v8, v8, a1
+; LMULMAX2-RV64F-NEXT: vse16.v v8, (a0)
+; LMULMAX2-RV64F-NEXT: ret
+;
+; LMULMAX2-RV32D-LABEL: cttz_zero_undef_v8i16:
+; LMULMAX2-RV32D: # %bb.0:
+; LMULMAX2-RV32D-NEXT: vsetivli zero, 8, e16, m1, ta, ma
+; LMULMAX2-RV32D-NEXT: vle16.v v8, (a0)
+; LMULMAX2-RV32D-NEXT: vrsub.vi v9, v8, 0
+; LMULMAX2-RV32D-NEXT: vand.vv v8, v8, v9
+; LMULMAX2-RV32D-NEXT: vfwcvt.f.xu.v v10, v8
+; LMULMAX2-RV32D-NEXT: vnsrl.wi v8, v10, 23
+; LMULMAX2-RV32D-NEXT: li a1, 127
+; LMULMAX2-RV32D-NEXT: vsub.vx v8, v8, a1
+; LMULMAX2-RV32D-NEXT: vse16.v v8, (a0)
+; LMULMAX2-RV32D-NEXT: ret
+;
+; LMULMAX2-RV64D-LABEL: cttz_zero_undef_v8i16:
+; LMULMAX2-RV64D: # %bb.0:
+; LMULMAX2-RV64D-NEXT: vsetivli zero, 8, e16, m1, ta, ma
+; LMULMAX2-RV64D-NEXT: vle16.v v8, (a0)
+; LMULMAX2-RV64D-NEXT: vrsub.vi v9, v8, 0
+; LMULMAX2-RV64D-NEXT: vand.vv v8, v8, v9
+; LMULMAX2-RV64D-NEXT: vfwcvt.f.xu.v v10, v8
+; LMULMAX2-RV64D-NEXT: vnsrl.wi v8, v10, 23
+; LMULMAX2-RV64D-NEXT: li a1, 127
+; LMULMAX2-RV64D-NEXT: vsub.vx v8, v8, a1
+; LMULMAX2-RV64D-NEXT: vse16.v v8, (a0)
+; LMULMAX2-RV64D-NEXT: ret
+;
+; LMULMAX8-LABEL: cttz_zero_undef_v8i16:
+; LMULMAX8: # %bb.0:
+; LMULMAX8-NEXT: vsetivli zero, 8, e16, m1, ta, ma
+; LMULMAX8-NEXT: vle16.v v8, (a0)
+; LMULMAX8-NEXT: vrsub.vi v9, v8, 0
+; LMULMAX8-NEXT: vand.vv v8, v8, v9
+; LMULMAX8-NEXT: vfwcvt.f.xu.v v10, v8
+; LMULMAX8-NEXT: vnsrl.wi v8, v10, 23
+; LMULMAX8-NEXT: li a1, 127
+; LMULMAX8-NEXT: vsub.vx v8, v8, a1
+; LMULMAX8-NEXT: vse16.v v8, (a0)
+; LMULMAX8-NEXT: ret
+ %a = load <8 x i16>, ptr %x
+ %b = load <8 x i16>, ptr %y
+ %c = call <8 x i16> @llvm.cttz.v8i16(<8 x i16> %a, i1 true)
+ store <8 x i16> %c, ptr %x
+ ret void
+}
+
+define void @cttz_zero_undef_v4i32(ptr %x, ptr %y) nounwind {
+; LMULMAX2-RV32I-LABEL: cttz_zero_undef_v4i32:
+; LMULMAX2-RV32I: # %bb.0:
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 4, e32, m1, ta, ma
+; LMULMAX2-RV32I-NEXT: vle32.v v8, (a0)
+; LMULMAX2-RV32I-NEXT: li a1, 1
+; LMULMAX2-RV32I-NEXT: vsub.vx v9, v8, a1
+; LMULMAX2-RV32I-NEXT: vnot.v v8, v8
+; 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_zero_undef_v4i32:
+; LMULMAX2-RV64I: # %bb.0:
+; LMULMAX2-RV64I-NEXT: vsetivli zero, 4, e32, m1, ta, ma
+; LMULMAX2-RV64I-NEXT: vle32.v v8, (a0)
+; LMULMAX2-RV64I-NEXT: li a1, 1
+; LMULMAX2-RV64I-NEXT: vsub.vx v9, v8, a1
+; LMULMAX2-RV64I-NEXT: vnot.v v8, v8
+; 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
+;
+; LMULMAX2-RV32F-LABEL: cttz_zero_undef_v4i32:
+; LMULMAX2-RV32F: # %bb.0:
+; LMULMAX2-RV32F-NEXT: vsetivli zero, 4, e32, m1, ta, ma
+; LMULMAX2-RV32F-NEXT: vle32.v v8, (a0)
+; LMULMAX2-RV32F-NEXT: vrsub.vi v9, v8, 0
+; LMULMAX2-RV32F-NEXT: vand.vv v8, v8, v9
+; LMULMAX2-RV32F-NEXT: fsrmi a1, 1
+; LMULMAX2-RV32F-NEXT: vfcvt.f.xu.v v8, v8
+; LMULMAX2-RV32F-NEXT: fsrm a1
+; LMULMAX2-RV32F-NEXT: vsrl.vi v8, v8, 23
+; LMULMAX2-RV32F-NEXT: li a1, 127
+; LMULMAX2-RV32F-NEXT: vsub.vx v8, v8, a1
+; LMULMAX2-RV32F-NEXT: vse32.v v8, (a0)
+; LMULMAX2-RV32F-NEXT: ret
+;
+; LMULMAX2-RV64F-LABEL: cttz_zero_undef_v4i32:
+; LMULMAX2-RV64F: # %bb.0:
+; LMULMAX2-RV64F-NEXT: vsetivli zero, 4, e32, m1, ta, ma
+; LMULMAX2-RV64F-NEXT: vle32.v v8, (a0)
+; LMULMAX2-RV64F-NEXT: vrsub.vi v9, v8, 0
+; LMULMAX2-RV64F-NEXT: vand.vv v8, v8, v9
+; LMULMAX2-RV64F-NEXT: fsrmi a1, 1
+; LMULMAX2-RV64F-NEXT: vfcvt.f.xu.v v8, v8
+; LMULMAX2-RV64F-NEXT: fsrm a1
+; LMULMAX2-RV64F-NEXT: vsrl.vi v8, v8, 23
+; LMULMAX2-RV64F-NEXT: li a1, 127
+; LMULMAX2-RV64F-NEXT: vsub.vx v8, v8, a1
+; LMULMAX2-RV64F-NEXT: vse32.v v8, (a0)
+; LMULMAX2-RV64F-NEXT: ret
+;
+; LMULMAX2-RV32D-LABEL: cttz_zero_undef_v4i32:
+; LMULMAX2-RV32D: # %bb.0:
+; LMULMAX2-RV32D-NEXT: vsetivli zero, 4, e32, m1, ta, ma
+; LMULMAX2-RV32D-NEXT: vle32.v v8, (a0)
+; LMULMAX2-RV32D-NEXT: vrsub.vi v9, v8, 0
+; LMULMAX2-RV32D-NEXT: vand.vv v8, v8, v9
+; LMULMAX2-RV32D-NEXT: vfwcvt.f.xu.v v10, v8
+; LMULMAX2-RV32D-NEXT: li a1, 52
+; LMULMAX2-RV32D-NEXT: vnsrl.wx v8, v10, a1
+; LMULMAX2-RV32D-NEXT: li a1, 1023
+; LMULMAX2-RV32D-NEXT: vsub.vx v8, v8, a1
+; LMULMAX2-RV32D-NEXT: vse32.v v8, (a0)
+; LMULMAX2-RV32D-NEXT: ret
+;
+; LMULMAX2-RV64D-LABEL: cttz_zero_undef_v4i32:
+; LMULMAX2-RV64D: # %bb.0:
+; LMULMAX2-RV64D-NEXT: vsetivli zero, 4, e32, m1, ta, ma
+; LMULMAX2-RV64D-NEXT: vle32.v v8, (a0)
+; LMULMAX2-RV64D-NEXT: vrsub.vi v9, v8, 0
+; LMULMAX2-RV64D-NEXT: vand.vv v8, v8, v9
+; LMULMAX2-RV64D-NEXT: vfwcvt.f.xu.v v10, v8
+; LMULMAX2-RV64D-NEXT: li a1, 52
+; LMULMAX2-RV64D-NEXT: vnsrl.wx v8, v10, a1
+; LMULMAX2-RV64D-NEXT: li a1, 1023
+; LMULMAX2-RV64D-NEXT: vsub.vx v8, v8, a1
+; LMULMAX2-RV64D-NEXT: vse32.v v8, (a0)
+; LMULMAX2-RV64D-NEXT: ret
+;
+; LMULMAX8-LABEL: cttz_zero_undef_v4i32:
+; LMULMAX8: # %bb.0:
+; LMULMAX8-NEXT: vsetivli zero, 4, e32, m1, ta, ma
+; LMULMAX8-NEXT: vle32.v v8, (a0)
+; LMULMAX8-NEXT: vrsub.vi v9, v8, 0
+; LMULMAX8-NEXT: vand.vv v8, v8, v9
+; LMULMAX8-NEXT: vfwcvt.f.xu.v v10, v8
+; LMULMAX8-NEXT: li a1, 52
+; LMULMAX8-NEXT: vnsrl.wx v8, v10, a1
+; LMULMAX8-NEXT: li a1, 1023
+; LMULMAX8-NEXT: vsub.vx v8, v8, a1
+; LMULMAX8-NEXT: vse32.v v8, (a0)
+; LMULMAX8-NEXT: ret
+ %a = load <4 x i32>, ptr %x
+ %b = load <4 x i32>, ptr %y
+ %c = call <4 x i32> @llvm.cttz.v4i32(<4 x i32> %a, i1 true)
+ store <4 x i32> %c, ptr %x
+ ret void
+}
+
+define void @cttz_zero_undef_v2i64(ptr %x, ptr %y) nounwind {
+; LMULMAX2-RV32I-LABEL: cttz_zero_undef_v2i64:
+; LMULMAX2-RV32I: # %bb.0:
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; LMULMAX2-RV32I-NEXT: vle64.v v8, (a0)
+; LMULMAX2-RV32I-NEXT: li a1, 1
+; LMULMAX2-RV32I-NEXT: vsub.vx v9, v8, a1
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 4, e32, m1, ta, ma
+; LMULMAX2-RV32I-NEXT: vmv.v.i v10, -1
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; LMULMAX2-RV32I-NEXT: vxor.vv v8, v8, v10
+; 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: vsetivli zero, 4, e32, m1, ta, ma
+; LMULMAX2-RV32I-NEXT: vmv.v.x v10, a1
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; LMULMAX2-RV32I-NEXT: vand.vv v9, v9, v10
+; LMULMAX2-RV32I-NEXT: vsub.vv v8, v8, v9
+; LMULMAX2-RV32I-NEXT: lui a1, 209715
+; LMULMAX2-RV32I-NEXT: addi a1, a1, 819
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 4, e32, m1, ta, ma
+; LMULMAX2-RV32I-NEXT: vmv.v.x v9, a1
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; LMULMAX2-RV32I-NEXT: vand.vv v10, v8, v9
+; LMULMAX2-RV32I-NEXT: vsrl.vi v8, v8, 2
+; LMULMAX2-RV32I-NEXT: vand.vv v8, v8, v9
+; LMULMAX2-RV32I-NEXT: vadd.vv v8, v10, 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: vsetivli zero, 4, e32, m1, ta, ma
+; LMULMAX2-RV32I-NEXT: vmv.v.x v9, a1
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; LMULMAX2-RV32I-NEXT: vand.vv v8, v8, v9
+; LMULMAX2-RV32I-NEXT: lui a1, 4112
+; LMULMAX2-RV32I-NEXT: addi a1, a1, 257
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 4, e32, m1, ta, ma
+; LMULMAX2-RV32I-NEXT: vmv.v.x v9, a1
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; LMULMAX2-RV32I-NEXT: vmul.vv v8, v8, v9
+; LMULMAX2-RV32I-NEXT: li a1, 56
+; LMULMAX2-RV32I-NEXT: vsrl.vx v8, v8, a1
+; LMULMAX2-RV32I-NEXT: vse64.v v8, (a0)
+; LMULMAX2-RV32I-NEXT: ret
+;
+; LMULMAX2-RV64I-LABEL: cttz_zero_undef_v2i64:
+; LMULMAX2-RV64I: # %bb.0:
+; LMULMAX2-RV64I-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; LMULMAX2-RV64I-NEXT: vle64.v v8, (a0)
+; LMULMAX2-RV64I-NEXT: li a1, 1
+; LMULMAX2-RV64I-NEXT: vsub.vx v9, v8, a1
+; LMULMAX2-RV64I-NEXT: vnot.v v8, v8
+; 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: slli a2, a1, 32
+; LMULMAX2-RV64I-NEXT: add a1, a1, a2
+; 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: slli a2, a1, 32
+; LMULMAX2-RV64I-NEXT: add a1, a1, a2
+; 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: slli a2, a1, 32
+; LMULMAX2-RV64I-NEXT: add a1, a1, a2
+; LMULMAX2-RV64I-NEXT: vand.vx v8, v8, a1
+; LMULMAX2-RV64I-NEXT: lui a1, 4112
+; LMULMAX2-RV64I-NEXT: addiw a1, a1, 257
+; LMULMAX2-RV64I-NEXT: slli a2, a1, 32
+; LMULMAX2-RV64I-NEXT: add a1, a1, a2
+; LMULMAX2-RV64I-NEXT: vmul.vx v8, v8, a1
+; LMULMAX2-RV64I-NEXT: li a1, 56
+; LMULMAX2-RV64I-NEXT: vsrl.vx v8, v8, a1
+; LMULMAX2-RV64I-NEXT: vse64.v v8, (a0)
+; LMULMAX2-RV64I-NEXT: ret
+;
+; LMULMAX2-RV32F-LABEL: cttz_zero_undef_v2i64:
+; LMULMAX2-RV32F: # %bb.0:
+; LMULMAX2-RV32F-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; LMULMAX2-RV32F-NEXT: vle64.v v8, (a0)
+; LMULMAX2-RV32F-NEXT: vsetivli zero, 4, e32, m1, ta, ma
+; LMULMAX2-RV32F-NEXT: vmv.v.i v9, 0
+; LMULMAX2-RV32F-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; LMULMAX2-RV32F-NEXT: vsub.vv v9, v9, v8
+; LMULMAX2-RV32F-NEXT: vand.vv v8, v8, v9
+; LMULMAX2-RV32F-NEXT: fsrmi a1, 1
+; LMULMAX2-RV32F-NEXT: vsetvli zero, zero, e32, mf2, ta, ma
+; LMULMAX2-RV32F-NEXT: vfncvt.f.xu.w v9, v8
+; LMULMAX2-RV32F-NEXT: fsrm a1
+; LMULMAX2-RV32F-NEXT: vsrl.vi v8, v9, 23
+; LMULMAX2-RV32F-NEXT: vsetvli zero, zero, e64, m1, ta, ma
+; LMULMAX2-RV32F-NEXT: vzext.vf2 v9, v8
+; LMULMAX2-RV32F-NEXT: li a1, 127
+; LMULMAX2-RV32F-NEXT: vsub.vx v8, v9, a1
+; LMULMAX2-RV32F-NEXT: vse64.v v8, (a0)
+; LMULMAX2-RV32F-NEXT: ret
+;
+; LMULMAX2-RV64F-LABEL: cttz_zero_undef_v2i64:
+; LMULMAX2-RV64F: # %bb.0:
+; LMULMAX2-RV64F-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; LMULMAX2-RV64F-NEXT: vle64.v v8, (a0)
+; LMULMAX2-RV64F-NEXT: vrsub.vi v9, v8, 0
+; LMULMAX2-RV64F-NEXT: vand.vv v8, v8, v9
+; LMULMAX2-RV64F-NEXT: fsrmi a1, 1
+; LMULMAX2-RV64F-NEXT: vsetvli zero, zero, e32, mf2, ta, ma
+; LMULMAX2-RV64F-NEXT: vfncvt.f.xu.w v9, v8
+; LMULMAX2-RV64F-NEXT: fsrm a1
+; LMULMAX2-RV64F-NEXT: vsrl.vi v8, v9, 23
+; LMULMAX2-RV64F-NEXT: li a1, 127
+; LMULMAX2-RV64F-NEXT: vwsubu.vx v9, v8, a1
+; LMULMAX2-RV64F-NEXT: vse64.v v9, (a0)
+; LMULMAX2-RV64F-NEXT: ret
+;
+; LMULMAX2-RV32D-LABEL: cttz_zero_undef_v2i64:
+; LMULMAX2-RV32D: # %bb.0:
+; LMULMAX2-RV32D-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; LMULMAX2-RV32D-NEXT: vle64.v v8, (a0)
+; LMULMAX2-RV32D-NEXT: vsetivli zero, 4, e32, m1, ta, ma
+; LMULMAX2-RV32D-NEXT: vmv.v.i v9, 0
+; LMULMAX2-RV32D-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; LMULMAX2-RV32D-NEXT: vsub.vv v9, v9, v8
+; LMULMAX2-RV32D-NEXT: vand.vv v8, v8, v9
+; LMULMAX2-RV32D-NEXT: fsrmi a1, 1
+; LMULMAX2-RV32D-NEXT: vfcvt.f.xu.v v8, v8
+; LMULMAX2-RV32D-NEXT: fsrm a1
+; LMULMAX2-RV32D-NEXT: li a1, 52
+; LMULMAX2-RV32D-NEXT: vsrl.vx v8, v8, a1
+; LMULMAX2-RV32D-NEXT: li a1, 1023
+; LMULMAX2-RV32D-NEXT: vsub.vx v8, v8, a1
+; LMULMAX2-RV32D-NEXT: vse64.v v8, (a0)
+; LMULMAX2-RV32D-NEXT: ret
+;
+; LMULMAX2-RV64D-LABEL: cttz_zero_undef_v2i64:
+; LMULMAX2-RV64D: # %bb.0:
+; LMULMAX2-RV64D-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; LMULMAX2-RV64D-NEXT: vle64.v v8, (a0)
+; LMULMAX2-RV64D-NEXT: vrsub.vi v9, v8, 0
+; LMULMAX2-RV64D-NEXT: vand.vv v8, v8, v9
+; LMULMAX2-RV64D-NEXT: fsrmi a1, 1
+; LMULMAX2-RV64D-NEXT: vfcvt.f.xu.v v8, v8
+; LMULMAX2-RV64D-NEXT: fsrm a1
+; LMULMAX2-RV64D-NEXT: li a1, 52
+; LMULMAX2-RV64D-NEXT: vsrl.vx v8, v8, a1
+; LMULMAX2-RV64D-NEXT: li a1, 1023
+; LMULMAX2-RV64D-NEXT: vsub.vx v8, v8, a1
+; LMULMAX2-RV64D-NEXT: vse64.v v8, (a0)
+; LMULMAX2-RV64D-NEXT: ret
+;
+; LMULMAX8-RV32-LABEL: cttz_zero_undef_v2i64:
+; LMULMAX8-RV32: # %bb.0:
+; LMULMAX8-RV32-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; LMULMAX8-RV32-NEXT: vle64.v v8, (a0)
+; LMULMAX8-RV32-NEXT: vsetivli zero, 4, e32, m1, ta, ma
+; LMULMAX8-RV32-NEXT: vmv.v.i v9, 0
+; LMULMAX8-RV32-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; LMULMAX8-RV32-NEXT: vsub.vv v9, v9, v8
+; LMULMAX8-RV32-NEXT: vand.vv v8, v8, v9
+; LMULMAX8-RV32-NEXT: fsrmi a1, 1
+; LMULMAX8-RV32-NEXT: vfcvt.f.xu.v v8, v8
+; LMULMAX8-RV32-NEXT: fsrm a1
+; LMULMAX8-RV32-NEXT: li a1, 52
+; LMULMAX8-RV32-NEXT: vsrl.vx v8, v8, a1
+; LMULMAX8-RV32-NEXT: li a1, 1023
+; LMULMAX8-RV32-NEXT: vsub.vx v8, v8, a1
+; LMULMAX8-RV32-NEXT: vse64.v v8, (a0)
+; LMULMAX8-RV32-NEXT: ret
+;
+; LMULMAX8-RV64-LABEL: cttz_zero_undef_v2i64:
+; LMULMAX8-RV64: # %bb.0:
+; LMULMAX8-RV64-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; LMULMAX8-RV64-NEXT: vle64.v v8, (a0)
+; LMULMAX8-RV64-NEXT: vrsub.vi v9, v8, 0
+; LMULMAX8-RV64-NEXT: vand.vv v8, v8, v9
+; LMULMAX8-RV64-NEXT: fsrmi a1, 1
+; LMULMAX8-RV64-NEXT: vfcvt.f.xu.v v8, v8
+; LMULMAX8-RV64-NEXT: fsrm a1
+; LMULMAX8-RV64-NEXT: li a1, 52
+; LMULMAX8-RV64-NEXT: vsrl.vx v8, v8, a1
+; LMULMAX8-RV64-NEXT: li a1, 1023
+; LMULMAX8-RV64-NEXT: vsub.vx v8, v8, a1
+; LMULMAX8-RV64-NEXT: vse64.v v8, (a0)
+; LMULMAX8-RV64-NEXT: ret
+ %a = load <2 x i64>, ptr %x
+ %b = load <2 x i64>, ptr %y
+ %c = call <2 x i64> @llvm.cttz.v2i64(<2 x i64> %a, i1 true)
+ store <2 x i64> %c, ptr %x
+ ret void
+}
+
+define void @cttz_zero_undef_v32i8(ptr %x, ptr %y) nounwind {
+; LMULMAX2-LABEL: cttz_zero_undef_v32i8:
+; LMULMAX2: # %bb.0:
+; LMULMAX2-NEXT: li a1, 32
+; LMULMAX2-NEXT: vsetvli zero, a1, e8, m2, ta, ma
+; LMULMAX2-NEXT: vle8.v v8, (a0)
+; LMULMAX2-NEXT: li a1, 1
+; LMULMAX2-NEXT: vsub.vx v10, v8, a1
+; LMULMAX2-NEXT: vnot.v v8, v8
+; LMULMAX2-NEXT: vand.vv v8, v8, v10
+; LMULMAX2-NEXT: vsrl.vi v10, v8, 1
+; LMULMAX2-NEXT: li a1, 85
+; LMULMAX2-NEXT: vand.vx v10, v10, a1
+; LMULMAX2-NEXT: vsub.vv v8, v8, v10
+; LMULMAX2-NEXT: li a1, 51
+; LMULMAX2-NEXT: vand.vx v10, v8, a1
+; LMULMAX2-NEXT: vsrl.vi v8, v8, 2
+; LMULMAX2-NEXT: vand.vx v8, v8, a1
+; LMULMAX2-NEXT: vadd.vv v8, v10, v8
+; LMULMAX2-NEXT: vsrl.vi v10, v8, 4
+; LMULMAX2-NEXT: vadd.vv v8, v8, v10
+; LMULMAX2-NEXT: vand.vi v8, v8, 15
+; LMULMAX2-NEXT: vse8.v v8, (a0)
+; LMULMAX2-NEXT: ret
+;
+; LMULMAX1-LABEL: cttz_zero_undef_v32i8:
+; LMULMAX1: # %bb.0:
+; LMULMAX1-NEXT: vsetivli zero, 16, e8, m1, ta, ma
+; LMULMAX1-NEXT: addi a1, a0, 16
+; LMULMAX1-NEXT: vle8.v v8, (a1)
+; LMULMAX1-NEXT: vle8.v v9, (a0)
+; LMULMAX1-NEXT: li a2, 1
+; LMULMAX1-NEXT: vsub.vx v10, v8, a2
+; LMULMAX1-NEXT: vnot.v v8, v8
+; LMULMAX1-NEXT: vand.vv v8, v8, v10
+; LMULMAX1-NEXT: vsrl.vi v10, v8, 1
+; LMULMAX1-NEXT: li a3, 85
+; LMULMAX1-NEXT: vand.vx v10, v10, a3
+; LMULMAX1-NEXT: vsub.vv v8, v8, v10
+; LMULMAX1-NEXT: li a4, 51
+; LMULMAX1-NEXT: vand.vx v10, v8, a4
+; LMULMAX1-NEXT: vsrl.vi v8, v8, 2
+; LMULMAX1-NEXT: vand.vx v8, v8, a4
+; LMULMAX1-NEXT: vadd.vv v8, v10, v8
+; LMULMAX1-NEXT: vsrl.vi v10, v8, 4
+; LMULMAX1-NEXT: vadd.vv v8, v8, v10
+; LMULMAX1-NEXT: vand.vi v8, v8, 15
+; LMULMAX1-NEXT: vsub.vx v10, v9, a2
+; LMULMAX1-NEXT: vnot.v v9, v9
+; LMULMAX1-NEXT: vand.vv v9, v9, v10
+; LMULMAX1-NEXT: vsrl.vi v10, v9, 1
+; LMULMAX1-NEXT: vand.vx v10, v10, a3
+; LMULMAX1-NEXT: vsub.vv v9, v9, v10
+; LMULMAX1-NEXT: vand.vx v10, v9, a4
+; LMULMAX1-NEXT: vsrl.vi v9, v9, 2
+; LMULMAX1-NEXT: vand.vx v9, v9, a4
+; LMULMAX1-NEXT: vadd.vv v9, v10, v9
+; LMULMAX1-NEXT: vsrl.vi v10, v9, 4
+; LMULMAX1-NEXT: vadd.vv v9, v9, v10
+; LMULMAX1-NEXT: vand.vi v9, v9, 15
+; LMULMAX1-NEXT: vse8.v v9, (a0)
+; LMULMAX1-NEXT: vse8.v v8, (a1)
+; LMULMAX1-NEXT: ret
+;
+; LMULMAX8-LABEL: cttz_zero_undef_v32i8:
+; LMULMAX8: # %bb.0:
+; LMULMAX8-NEXT: li a1, 32
+; LMULMAX8-NEXT: vsetvli zero, a1, e8, m2, ta, ma
+; LMULMAX8-NEXT: vle8.v v8, (a0)
+; LMULMAX8-NEXT: vrsub.vi v10, v8, 0
+; LMULMAX8-NEXT: vand.vv v8, v8, v10
+; LMULMAX8-NEXT: vsetvli zero, zero, e16, m4, ta, ma
+; LMULMAX8-NEXT: vzext.vf2 v12, v8
+; LMULMAX8-NEXT: vfwcvt.f.xu.v v16, v12
+; LMULMAX8-NEXT: vnsrl.wi v8, v16, 23
+; LMULMAX8-NEXT: vsetvli zero, zero, e8, m2, ta, ma
+; LMULMAX8-NEXT: vnsrl.wi v12, v8, 0
+; LMULMAX8-NEXT: li a1, 127
+; LMULMAX8-NEXT: vsub.vx v8, v12, a1
+; LMULMAX8-NEXT: vse8.v v8, (a0)
+; LMULMAX8-NEXT: ret
+ %a = load <32 x i8>, ptr %x
+ %b = load <32 x i8>, ptr %y
+ %c = call <32 x i8> @llvm.cttz.v32i8(<32 x i8> %a, i1 true)
+ store <32 x i8> %c, ptr %x
+ ret void
+}
+
+define void @cttz_zero_undef_v16i16(ptr %x, ptr %y) nounwind {
+; LMULMAX2-RV32-LABEL: cttz_zero_undef_v16i16:
+; LMULMAX2-RV32: # %bb.0:
+; LMULMAX2-RV32-NEXT: vsetivli zero, 16, e16, m2, ta, ma
+; LMULMAX2-RV32-NEXT: vle16.v v8, (a0)
+; LMULMAX2-RV32-NEXT: li a1, 1
+; LMULMAX2-RV32-NEXT: vsub.vx v10, v8, a1
+; LMULMAX2-RV32-NEXT: vnot.v v8, v8
+; 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_zero_undef_v16i16:
+; LMULMAX2-RV64: # %bb.0:
+; LMULMAX2-RV64-NEXT: vsetivli zero, 16, e16, m2, ta, ma
+; LMULMAX2-RV64-NEXT: vle16.v v8, (a0)
+; LMULMAX2-RV64-NEXT: li a1, 1
+; LMULMAX2-RV64-NEXT: vsub.vx v10, v8, a1
+; LMULMAX2-RV64-NEXT: vnot.v v8, v8
+; 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_zero_undef_v16i16:
+; LMULMAX1-RV32: # %bb.0:
+; LMULMAX1-RV32-NEXT: vsetivli zero, 8, e16, m1, ta, ma
+; 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 a2, 1
+; LMULMAX1-RV32-NEXT: vsub.vx v10, v8, a2
+; LMULMAX1-RV32-NEXT: vnot.v v8, v8
+; 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 a6, 257
+; LMULMAX1-RV32-NEXT: vmul.vx v8, v8, a6
+; LMULMAX1-RV32-NEXT: vsrl.vi v8, v8, 8
+; LMULMAX1-RV32-NEXT: vsub.vx v10, v9, a2
+; LMULMAX1-RV32-NEXT: vnot.v v9, v9
+; 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, a6
+; 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_zero_undef_v16i16:
+; LMULMAX1-RV64: # %bb.0:
+; LMULMAX1-RV64-NEXT: vsetivli zero, 8, e16, m1, ta, ma
+; 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 a2, 1
+; LMULMAX1-RV64-NEXT: vsub.vx v10, v8, a2
+; LMULMAX1-RV64-NEXT: vnot.v v8, v8
+; 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 a6, 257
+; LMULMAX1-RV64-NEXT: vmul.vx v8, v8, a6
+; LMULMAX1-RV64-NEXT: vsrl.vi v8, v8, 8
+; LMULMAX1-RV64-NEXT: vsub.vx v10, v9, a2
+; LMULMAX1-RV64-NEXT: vnot.v v9, v9
+; 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, a6
+; 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-LABEL: cttz_zero_undef_v16i16:
+; LMULMAX8: # %bb.0:
+; LMULMAX8-NEXT: vsetivli zero, 16, e16, m2, ta, ma
+; LMULMAX8-NEXT: vle16.v v8, (a0)
+; LMULMAX8-NEXT: vrsub.vi v10, v8, 0
+; LMULMAX8-NEXT: vand.vv v8, v8, v10
+; LMULMAX8-NEXT: vfwcvt.f.xu.v v12, v8
+; LMULMAX8-NEXT: vnsrl.wi v8, v12, 23
+; LMULMAX8-NEXT: li a1, 127
+; LMULMAX8-NEXT: vsub.vx v8, v8, a1
+; LMULMAX8-NEXT: vse16.v v8, (a0)
+; LMULMAX8-NEXT: ret
+ %a = load <16 x i16>, ptr %x
+ %b = load <16 x i16>, ptr %y
+ %c = call <16 x i16> @llvm.cttz.v16i16(<16 x i16> %a, i1 true)
+ store <16 x i16> %c, ptr %x
+ ret void
+}
+
+define void @cttz_zero_undef_v8i32(ptr %x, ptr %y) nounwind {
+; LMULMAX2-RV32I-LABEL: cttz_zero_undef_v8i32:
+; LMULMAX2-RV32I: # %bb.0:
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; LMULMAX2-RV32I-NEXT: vle32.v v8, (a0)
+; LMULMAX2-RV32I-NEXT: li a1, 1
+; LMULMAX2-RV32I-NEXT: vsub.vx v10, v8, a1
+; LMULMAX2-RV32I-NEXT: vnot.v v8, v8
+; LMULMAX2-RV32I-NEXT: vand.vv v8, v8, v10
+; LMULMAX2-RV32I-NEXT: vsrl.vi v10, v8, 1
+; LMULMAX2-RV32I-NEXT: lui a1, 349525
+; LMULMAX2-RV32I-NEXT: addi a1, a1, 1365
+; LMULMAX2-RV32I-NEXT: vand.vx v10, v10, a1
+; LMULMAX2-RV32I-NEXT: vsub.vv v8, v8, v10
+; LMULMAX2-RV32I-NEXT: lui a1, 209715
+; LMULMAX2-RV32I-NEXT: addi a1, a1, 819
+; LMULMAX2-RV32I-NEXT: vand.vx v10, v8, a1
+; LMULMAX2-RV32I-NEXT: vsrl.vi v8, v8, 2
+; LMULMAX2-RV32I-NEXT: vand.vx v8, v8, a1
+; LMULMAX2-RV32I-NEXT: vadd.vv v8, v10, v8
+; LMULMAX2-RV32I-NEXT: vsrl.vi v10, v8, 4
+; LMULMAX2-RV32I-NEXT: vadd.vv v8, v8, v10
+; 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_zero_undef_v8i32:
+; LMULMAX2-RV64I: # %bb.0:
+; LMULMAX2-RV64I-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; LMULMAX2-RV64I-NEXT: vle32.v v8, (a0)
+; LMULMAX2-RV64I-NEXT: li a1, 1
+; LMULMAX2-RV64I-NEXT: vsub.vx v10, v8, a1
+; LMULMAX2-RV64I-NEXT: vnot.v v8, v8
+; LMULMAX2-RV64I-NEXT: vand.vv v8, v8, v10
+; LMULMAX2-RV64I-NEXT: vsrl.vi v10, v8, 1
+; LMULMAX2-RV64I-NEXT: lui a1, 349525
+; LMULMAX2-RV64I-NEXT: addiw a1, a1, 1365
+; LMULMAX2-RV64I-NEXT: vand.vx v10, v10, a1
+; LMULMAX2-RV64I-NEXT: vsub.vv v8, v8, v10
+; LMULMAX2-RV64I-NEXT: lui a1, 209715
+; LMULMAX2-RV64I-NEXT: addiw a1, a1, 819
+; LMULMAX2-RV64I-NEXT: vand.vx v10, v8, a1
+; LMULMAX2-RV64I-NEXT: vsrl.vi v8, v8, 2
+; LMULMAX2-RV64I-NEXT: vand.vx v8, v8, a1
+; LMULMAX2-RV64I-NEXT: vadd.vv v8, v10, v8
+; LMULMAX2-RV64I-NEXT: vsrl.vi v10, v8, 4
+; LMULMAX2-RV64I-NEXT: vadd.vv v8, v8, v10
+; 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
+;
+; LMULMAX2-RV32F-LABEL: cttz_zero_undef_v8i32:
+; LMULMAX2-RV32F: # %bb.0:
+; LMULMAX2-RV32F-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; LMULMAX2-RV32F-NEXT: vle32.v v8, (a0)
+; LMULMAX2-RV32F-NEXT: vrsub.vi v10, v8, 0
+; LMULMAX2-RV32F-NEXT: vand.vv v8, v8, v10
+; LMULMAX2-RV32F-NEXT: fsrmi a1, 1
+; LMULMAX2-RV32F-NEXT: vfcvt.f.xu.v v8, v8
+; LMULMAX2-RV32F-NEXT: fsrm a1
+; LMULMAX2-RV32F-NEXT: vsrl.vi v8, v8, 23
+; LMULMAX2-RV32F-NEXT: li a1, 127
+; LMULMAX2-RV32F-NEXT: vsub.vx v8, v8, a1
+; LMULMAX2-RV32F-NEXT: vse32.v v8, (a0)
+; LMULMAX2-RV32F-NEXT: ret
+;
+; LMULMAX2-RV64F-LABEL: cttz_zero_undef_v8i32:
+; LMULMAX2-RV64F: # %bb.0:
+; LMULMAX2-RV64F-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; LMULMAX2-RV64F-NEXT: vle32.v v8, (a0)
+; LMULMAX2-RV64F-NEXT: vrsub.vi v10, v8, 0
+; LMULMAX2-RV64F-NEXT: vand.vv v8, v8, v10
+; LMULMAX2-RV64F-NEXT: fsrmi a1, 1
+; LMULMAX2-RV64F-NEXT: vfcvt.f.xu.v v8, v8
+; LMULMAX2-RV64F-NEXT: fsrm a1
+; LMULMAX2-RV64F-NEXT: vsrl.vi v8, v8, 23
+; LMULMAX2-RV64F-NEXT: li a1, 127
+; LMULMAX2-RV64F-NEXT: vsub.vx v8, v8, a1
+; LMULMAX2-RV64F-NEXT: vse32.v v8, (a0)
+; LMULMAX2-RV64F-NEXT: ret
+;
+; LMULMAX2-RV32D-LABEL: cttz_zero_undef_v8i32:
+; LMULMAX2-RV32D: # %bb.0:
+; LMULMAX2-RV32D-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; LMULMAX2-RV32D-NEXT: vle32.v v8, (a0)
+; LMULMAX2-RV32D-NEXT: vrsub.vi v10, v8, 0
+; LMULMAX2-RV32D-NEXT: vand.vv v8, v8, v10
+; LMULMAX2-RV32D-NEXT: fsrmi a1, 1
+; LMULMAX2-RV32D-NEXT: vfcvt.f.xu.v v8, v8
+; LMULMAX2-RV32D-NEXT: fsrm a1
+; LMULMAX2-RV32D-NEXT: vsrl.vi v8, v8, 23
+; LMULMAX2-RV32D-NEXT: li a1, 127
+; LMULMAX2-RV32D-NEXT: vsub.vx v8, v8, a1
+; LMULMAX2-RV32D-NEXT: vse32.v v8, (a0)
+; LMULMAX2-RV32D-NEXT: ret
+;
+; LMULMAX2-RV64D-LABEL: cttz_zero_undef_v8i32:
+; LMULMAX2-RV64D: # %bb.0:
+; LMULMAX2-RV64D-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; LMULMAX2-RV64D-NEXT: vle32.v v8, (a0)
+; LMULMAX2-RV64D-NEXT: vrsub.vi v10, v8, 0
+; LMULMAX2-RV64D-NEXT: vand.vv v8, v8, v10
+; LMULMAX2-RV64D-NEXT: fsrmi a1, 1
+; LMULMAX2-RV64D-NEXT: vfcvt.f.xu.v v8, v8
+; LMULMAX2-RV64D-NEXT: fsrm a1
+; LMULMAX2-RV64D-NEXT: vsrl.vi v8, v8, 23
+; LMULMAX2-RV64D-NEXT: li a1, 127
+; LMULMAX2-RV64D-NEXT: vsub.vx v8, v8, a1
+; LMULMAX2-RV64D-NEXT: vse32.v v8, (a0)
+; LMULMAX2-RV64D-NEXT: ret
+;
+; LMULMAX8-LABEL: cttz_zero_undef_v8i32:
+; LMULMAX8: # %bb.0:
+; LMULMAX8-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; LMULMAX8-NEXT: vle32.v v8, (a0)
+; LMULMAX8-NEXT: vrsub.vi v10, v8, 0
+; LMULMAX8-NEXT: vand.vv v8, v8, v10
+; LMULMAX8-NEXT: vfwcvt.f.xu.v v12, v8
+; LMULMAX8-NEXT: li a1, 52
+; LMULMAX8-NEXT: vnsrl.wx v8, v12, a1
+; LMULMAX8-NEXT: li a1, 1023
+; LMULMAX8-NEXT: vsub.vx v8, v8, a1
+; LMULMAX8-NEXT: vse32.v v8, (a0)
+; LMULMAX8-NEXT: ret
+ %a = load <8 x i32>, ptr %x
+ %b = load <8 x i32>, ptr %y
+ %c = call <8 x i32> @llvm.cttz.v8i32(<8 x i32> %a, i1 true)
+ store <8 x i32> %c, ptr %x
+ ret void
+}
+
+define void @cttz_zero_undef_v4i64(ptr %x, ptr %y) nounwind {
+; LMULMAX2-RV32I-LABEL: cttz_zero_undef_v4i64:
+; LMULMAX2-RV32I: # %bb.0:
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 4, e64, m2, ta, ma
+; LMULMAX2-RV32I-NEXT: vle64.v v8, (a0)
+; LMULMAX2-RV32I-NEXT: li a1, 1
+; LMULMAX2-RV32I-NEXT: vsub.vx v10, v8, a1
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; LMULMAX2-RV32I-NEXT: vmv.v.i v12, -1
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 4, e64, m2, ta, ma
+; LMULMAX2-RV32I-NEXT: vxor.vv v8, v8, v12
+; LMULMAX2-RV32I-NEXT: vand.vv v8, v8, v10
+; LMULMAX2-RV32I-NEXT: vsrl.vi v10, v8, 1
+; LMULMAX2-RV32I-NEXT: lui a1, 349525
+; LMULMAX2-RV32I-NEXT: addi a1, a1, 1365
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; LMULMAX2-RV32I-NEXT: vmv.v.x v12, a1
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 4, e64, m2, ta, ma
+; LMULMAX2-RV32I-NEXT: vand.vv v10, v10, v12
+; LMULMAX2-RV32I-NEXT: vsub.vv v8, v8, v10
+; LMULMAX2-RV32I-NEXT: lui a1, 209715
+; LMULMAX2-RV32I-NEXT: addi a1, a1, 819
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; LMULMAX2-RV32I-NEXT: vmv.v.x v10, a1
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 4, e64, m2, ta, ma
+; LMULMAX2-RV32I-NEXT: vand.vv v12, v8, v10
+; LMULMAX2-RV32I-NEXT: vsrl.vi v8, v8, 2
+; LMULMAX2-RV32I-NEXT: vand.vv v8, v8, v10
+; LMULMAX2-RV32I-NEXT: vadd.vv v8, v12, v8
+; LMULMAX2-RV32I-NEXT: vsrl.vi v10, v8, 4
+; LMULMAX2-RV32I-NEXT: vadd.vv v8, v8, v10
+; LMULMAX2-RV32I-NEXT: lui a1, 61681
+; LMULMAX2-RV32I-NEXT: addi a1, a1, -241
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; LMULMAX2-RV32I-NEXT: vmv.v.x v10, a1
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 4, e64, m2, ta, ma
+; LMULMAX2-RV32I-NEXT: vand.vv v8, v8, v10
+; LMULMAX2-RV32I-NEXT: lui a1, 4112
+; LMULMAX2-RV32I-NEXT: addi a1, a1, 257
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; LMULMAX2-RV32I-NEXT: vmv.v.x v10, a1
+; LMULMAX2-RV32I-NEXT: vsetivli zero, 4, e64, m2, ta, ma
+; LMULMAX2-RV32I-NEXT: vmul.vv v8, v8, v10
+; LMULMAX2-RV32I-NEXT: li a1, 56
+; LMULMAX2-RV32I-NEXT: vsrl.vx v8, v8, a1
+; LMULMAX2-RV32I-NEXT: vse64.v v8, (a0)
+; LMULMAX2-RV32I-NEXT: ret
+;
+; LMULMAX2-RV64I-LABEL: cttz_zero_undef_v4i64:
+; LMULMAX2-RV64I: # %bb.0:
+; LMULMAX2-RV64I-NEXT: vsetivli zero, 4, e64, m2, ta, ma
+; LMULMAX2-RV64I-NEXT: vle64.v v8, (a0)
+; LMULMAX2-RV64I-NEXT: li a1, 1
+; LMULMAX2-RV64I-NEXT: vsub.vx v10, v8, a1
+; LMULMAX2-RV64I-NEXT: vnot.v v8, v8
+; LMULMAX2-RV64I-NEXT: vand.vv v8, v8, v10
+; LMULMAX2-RV64I-NEXT: vsrl.vi v10, v8, 1
+; LMULMAX2-RV64I-NEXT: lui a1, 349525
+; LMULMAX2-RV64I-NEXT: addiw a1, a1, 1365
+; LMULMAX2-RV64I-NEXT: slli a2, a1, 32
+; LMULMAX2-RV64I-NEXT: add a1, a1, a2
+; LMULMAX2-RV64I-NEXT: vand.vx v10, v10, a1
+; LMULMAX2-RV64I-NEXT: vsub.vv v8, v8, v10
+; LMULMAX2-RV64I-NEXT: lui a1, 209715
+; LMULMAX2-RV64I-NEXT: addiw a1, a1, 819
+; LMULMAX2-RV64I-NEXT: slli a2, a1, 32
+; LMULMAX2-RV64I-NEXT: add a1, a1, a2
+; LMULMAX2-RV64I-NEXT: vand.vx v10, v8, a1
+; LMULMAX2-RV64I-NEXT: vsrl.vi v8, v8, 2
+; LMULMAX2-RV64I-NEXT: vand.vx v8, v8, a1
+; LMULMAX2-RV64I-NEXT: vadd.vv v8, v10, v8
+; LMULMAX2-RV64I-NEXT: vsrl.vi v10, v8, 4
+; LMULMAX2-RV64I-NEXT: vadd.vv v8, v8, v10
+; LMULMAX2-RV64I-NEXT: lui a1, 61681
+; LMULMAX2-RV64I-NEXT: addiw a1, a1, -241
+; LMULMAX2-RV64I-NEXT: slli a2, a1, 32
+; LMULMAX2-RV64I-NEXT: add a1, a1, a2
+; LMULMAX2-RV64I-NEXT: vand.vx v8, v8, a1
+; LMULMAX2-RV64I-NEXT: lui a1, 4112
+; LMULMAX2-RV64I-NEXT: addiw a1, a1, 257
+; LMULMAX2-RV64I-NEXT: slli a2, a1, 32
+; LMULMAX2-RV64I-NEXT: add a1, a1, a2
+; LMULMAX2-RV64I-NEXT: vmul.vx v8, v8, a1
+; LMULMAX2-RV64I-NEXT: li a1, 56
+; LMULMAX2-RV64I-NEXT: vsrl.vx v8, v8, a1
+; LMULMAX2-RV64I-NEXT: vse64.v v8, (a0)
+; LMULMAX2-RV64I-NEXT: ret
+;
+; LMULMAX2-RV32F-LABEL: cttz_zero_undef_v4i64:
+; LMULMAX2-RV32F: # %bb.0:
+; LMULMAX2-RV32F-NEXT: vsetivli zero, 4, e64, m2, ta, ma
+; LMULMAX2-RV32F-NEXT: vle64.v v8, (a0)
+; LMULMAX2-RV32F-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; LMULMAX2-RV32F-NEXT: vmv.v.i v10, 0
+; LMULMAX2-RV32F-NEXT: vsetivli zero, 4, e64, m2, ta, ma
+; LMULMAX2-RV32F-NEXT: vsub.vv v10, v10, v8
+; LMULMAX2-RV32F-NEXT: vand.vv v8, v8, v10
+; LMULMAX2-RV32F-NEXT: fsrmi a1, 1
+; LMULMAX2-RV32F-NEXT: vsetvli zero, zero, e32, m1, ta, ma
+; LMULMAX2-RV32F-NEXT: vfncvt.f.xu.w v10, v8
+; LMULMAX2-RV32F-NEXT: fsrm a1
+; LMULMAX2-RV32F-NEXT: vsrl.vi v8, v10, 23
+; LMULMAX2-RV32F-NEXT: vsetvli zero, zero, e64, m2, ta, ma
+; LMULMAX2-RV32F-NEXT: vzext.vf2 v10, v8
+; LMULMAX2-RV32F-NEXT: li a1, 127
+; LMULMAX2-RV32F-NEXT: vsub.vx v8, v10, a1
+; LMULMAX2-RV32F-NEXT: vse64.v v8, (a0)
+; LMULMAX2-RV32F-NEXT: ret
+;
+; LMULMAX2-RV64F-LABEL: cttz_zero_undef_v4i64:
+; LMULMAX2-RV64F: # %bb.0:
+; LMULMAX2-RV64F-NEXT: vsetivli zero, 4, e64, m2, ta, ma
+; LMULMAX2-RV64F-NEXT: vle64.v v8, (a0)
+; LMULMAX2-RV64F-NEXT: vrsub.vi v10, v8, 0
+; LMULMAX2-RV64F-NEXT: vand.vv v8, v8, v10
+; LMULMAX2-RV64F-NEXT: fsrmi a1, 1
+; LMULMAX2-RV64F-NEXT: vsetvli zero, zero, e32, m1, ta, ma
+; LMULMAX2-RV64F-NEXT: vfncvt.f.xu.w v10, v8
+; LMULMAX2-RV64F-NEXT: fsrm a1
+; LMULMAX2-RV64F-NEXT: vsrl.vi v8, v10, 23
+; LMULMAX2-RV64F-NEXT: li a1, 127
+; LMULMAX2-RV64F-NEXT: vwsubu.vx v10, v8, a1
+; LMULMAX2-RV64F-NEXT: vse64.v v10, (a0)
+; LMULMAX2-RV64F-NEXT: ret
+;
+; LMULMAX2-RV32D-LABEL: cttz_zero_undef_v4i64:
+; LMULMAX2-RV32D: # %bb.0:
+; LMULMAX2-RV32D-NEXT: vsetivli zero, 4, e64, m2, ta, ma
+; LMULMAX2-RV32D-NEXT: vle64.v v8, (a0)
+; LMULMAX2-RV32D-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; LMULMAX2-RV32D-NEXT: vmv.v.i v10, 0
+; LMULMAX2-RV32D-NEXT: vsetivli zero, 4, e64, m2, ta, ma
+; LMULMAX2-RV32D-NEXT: vsub.vv v10, v10, v8
+; LMULMAX2-RV32D-NEXT: vand.vv v8, v8, v10
+; LMULMAX2-RV32D-NEXT: fsrmi a1, 1
+; LMULMAX2-RV32D-NEXT: vfcvt.f.xu.v v8, v8
+; LMULMAX2-RV32D-NEXT: fsrm a1
+; LMULMAX2-RV32D-NEXT: li a1, 52
+; LMULMAX2-RV32D-NEXT: vsrl.vx v8, v8, a1
+; LMULMAX2-RV32D-NEXT: li a1, 1023
+; LMULMAX2-RV32D-NEXT: vsub.vx v8, v8, a1
+; LMULMAX2-RV32D-NEXT: vse64.v v8, (a0)
+; LMULMAX2-RV32D-NEXT: ret
+;
+; LMULMAX2-RV64D-LABEL: cttz_zero_undef_v4i64:
+; LMULMAX2-RV64D: # %bb.0:
+; LMULMAX2-RV64D-NEXT: vsetivli zero, 4, e64, m2, ta, ma
+; LMULMAX2-RV64D-NEXT: vle64.v v8, (a0)
+; LMULMAX2-RV64D-NEXT: vrsub.vi v10, v8, 0
+; LMULMAX2-RV64D-NEXT: vand.vv v8, v8, v10
+; LMULMAX2-RV64D-NEXT: fsrmi a1, 1
+; LMULMAX2-RV64D-NEXT: vfcvt.f.xu.v v8, v8
+; LMULMAX2-RV64D-NEXT: fsrm a1
+; LMULMAX2-RV64D-NEXT: li a1, 52
+; LMULMAX2-RV64D-NEXT: vsrl.vx v8, v8, a1
+; LMULMAX2-RV64D-NEXT: li a1, 1023
+; LMULMAX2-RV64D-NEXT: vsub.vx v8, v8, a1
+; LMULMAX2-RV64D-NEXT: vse64.v v8, (a0)
+; LMULMAX2-RV64D-NEXT: ret
+;
+; LMULMAX8-RV32-LABEL: cttz_zero_undef_v4i64:
+; LMULMAX8-RV32: # %bb.0:
+; LMULMAX8-RV32-NEXT: vsetivli zero, 4, e64, m2, ta, ma
+; LMULMAX8-RV32-NEXT: vle64.v v8, (a0)
+; LMULMAX8-RV32-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; LMULMAX8-RV32-NEXT: vmv.v.i v10, 0
+; LMULMAX8-RV32-NEXT: vsetivli zero, 4, e64, m2, ta, ma
+; LMULMAX8-RV32-NEXT: vsub.vv v10, v10, v8
+; LMULMAX8-RV32-NEXT: vand.vv v8, v8, v10
+; LMULMAX8-RV32-NEXT: fsrmi a1, 1
+; LMULMAX8-RV32-NEXT: vfcvt.f.xu.v v8, v8
+; LMULMAX8-RV32-NEXT: fsrm a1
+; LMULMAX8-RV32-NEXT: li a1, 52
+; LMULMAX8-RV32-NEXT: vsrl.vx v8, v8, a1
+; LMULMAX8-RV32-NEXT: li a1, 1023
+; LMULMAX8-RV32-NEXT: vsub.vx v8, v8, a1
+; LMULMAX8-RV32-NEXT: vse64.v v8, (a0)
+; LMULMAX8-RV32-NEXT: ret
+;
+; LMULMAX8-RV64-LABEL: cttz_zero_undef_v4i64:
+; LMULMAX8-RV64: # %bb.0:
+; LMULMAX8-RV64-NEXT: vsetivli zero, 4, e64, m2, ta, ma
+; LMULMAX8-RV64-NEXT: vle64.v v8, (a0)
+; LMULMAX8-RV64-NEXT: vrsub.vi v10, v8, 0
+; LMULMAX8-RV64-NEXT: vand.vv v8, v8, v10
+; LMULMAX8-RV64-NEXT: fsrmi a1, 1
+; LMULMAX8-RV64-NEXT: vfcvt.f.xu.v v8, v8
+; LMULMAX8-RV64-NEXT: fsrm a1
+; LMULMAX8-RV64-NEXT: li a1, 52
+; LMULMAX8-RV64-NEXT: vsrl.vx v8, v8, a1
+; LMULMAX8-RV64-NEXT: li a1, 1023
+; LMULMAX8-RV64-NEXT: vsub.vx v8, v8, a1
+; LMULMAX8-RV64-NEXT: vse64.v v8, (a0)
+; LMULMAX8-RV64-NEXT: ret
+ %a = load <4 x i64>, ptr %x
+ %b = load <4 x i64>, ptr %y
+ %c = call <4 x i64> @llvm.cttz.v4i64(<4 x i64> %a, i1 true)
+ store <4 x i64> %c, ptr %x
+ ret void
+}
More information about the llvm-commits
mailing list