1 ; RUN: llc < %s -verify-machineinstrs -mtriple=aarch64-none-linux-gnu -mattr=+neon,+aes | FileCheck %s
3 ; This test checks that pmull2 instruction is used for vmull_high_p64 intrinsic.
4 ; There are two extraction operations located in different basic blocks:
6 ; %4 = extractelement <2 x i64> %0, i32 1
7 ; %12 = extractelement <2 x i64> %9, i32 1
11 ; @llvm.aarch64.neon.pmull64(i64 %12, i64 %4) #2
13 ; We test that pattern replacing llvm.aarch64.neon.pmull64 with pmull2
16 ; IR for that test was generated from the following .cpp file:
18 ; #include <arm_neon.h>
24 ; void func (SS *g, unsigned int count, const unsigned char *buf, poly128_t* res )
26 ; const uint64x2_t x = g->x;
27 ; const uint64x2_t h = g->h;
28 ; uint64x2_t ci = g->x;
30 ; for (int i = 0; i < count; i+=2, buf += 16) {
31 ; ci = vreinterpretq_u64_u8(veorq_u8(vreinterpretq_u8_u64(ci),
32 ; vrbitq_u8(vld1q_u8(buf))));
33 ; res[i] = vmull_p64((poly64_t)vget_low_p64(vreinterpretq_p64_u64(ci)),
34 ; (poly64_t)vget_low_p64(vreinterpretq_p64_u64(h)));
35 ; res[i+1] = vmull_high_p64(vreinterpretq_p64_u64(ci),
36 ; vreinterpretq_p64_u64(h));
44 %struct.SS = type { <2 x i64>, <2 x i64> }
46 ; Function Attrs: nofree noinline nounwind
47 define dso_local void @func(ptr nocapture readonly %g, i32 %count, ptr nocapture readonly %buf, ptr nocapture %res) local_unnamed_addr #0 {
49 %h2 = getelementptr inbounds %struct.SS, ptr %g, i64 0, i32 1
50 %0 = load <2 x i64>, ptr %h2, align 16
51 %cmp34 = icmp eq i32 %count, 0
52 br i1 %cmp34, label %for.cond.cleanup, label %for.body.lr.ph
54 for.body.lr.ph: ; preds = %entry
55 %1 = load <16 x i8>, ptr %g, align 16
56 %2 = extractelement <2 x i64> %0, i32 0
57 %3 = extractelement <2 x i64> %0, i32 1
58 %4 = zext i32 %count to i64
61 for.cond.cleanup: ; preds = %for.body, %entry
64 for.body: ; preds = %for.body.lr.ph, %for.body
65 %indvars.iv = phi i64 [ 0, %for.body.lr.ph ], [ %indvars.iv.next, %for.body ]
66 %buf.addr.036 = phi ptr [ %buf, %for.body.lr.ph ], [ %add.ptr, %for.body ]
67 %5 = phi <16 x i8> [ %1, %for.body.lr.ph ], [ %xor.i, %for.body ]
68 %6 = load <16 x i8>, ptr %buf.addr.036, align 16
69 %vrbit.i = call <16 x i8> @llvm.aarch64.neon.rbit.v16i8(<16 x i8> %6) #0
70 %xor.i = xor <16 x i8> %vrbit.i, %5
71 %7 = bitcast <16 x i8> %xor.i to <2 x i64>
72 %8 = extractelement <2 x i64> %7, i32 0
73 %vmull_p64.i = call <16 x i8> @llvm.aarch64.neon.pmull64(i64 %8, i64 %2) #0
74 %arrayidx = getelementptr inbounds i128, ptr %res, i64 %indvars.iv
75 store <16 x i8> %vmull_p64.i, ptr %arrayidx, align 16
76 %9 = extractelement <2 x i64> %7, i32 1
77 %vmull_p64.i.i = call <16 x i8> @llvm.aarch64.neon.pmull64(i64 %9, i64 %3) #0
78 %10 = or i64 %indvars.iv, 1
79 %arrayidx16 = getelementptr inbounds i128, ptr %res, i64 %10
80 store <16 x i8> %vmull_p64.i.i, ptr %arrayidx16, align 16
81 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 2
82 %add.ptr = getelementptr inbounds i8, ptr %buf.addr.036, i64 16
83 %cmp = icmp ult i64 %indvars.iv.next, %4
84 br i1 %cmp, label %for.body, label %for.cond.cleanup
87 ; Function Attrs: nounwind readnone
88 declare <16 x i8> @llvm.aarch64.neon.rbit.v16i8(<16 x i8>) #0
90 ; Function Attrs: nounwind readnone
91 declare <16 x i8> @llvm.aarch64.neon.pmull64(i64, i64) #0
93 attributes #0 = { nofree noinline nounwind }