@@ -4765,6 +4765,39 @@ define void @scaleidx_scatter_outofrange(<8 x float> %value, ptr %base, <8 x i32
47654765}
47664766declare void @llvm.masked.scatter.v8f32.v8p0 (<8 x float >, <8 x ptr >, i32 immarg, <8 x i1 >)
47674767
4768+ define <16 x i32 > @pr163023 (ptr %a0 , <16 x i32 > %a1 ) {
4769+ ; X64-LABEL: pr163023:
4770+ ; X64: # %bb.0:
4771+ ; X64-NEXT: vpmovsxdq %ymm0, %zmm1
4772+ ; X64-NEXT: vextracti64x4 $1, %zmm0, %ymm0
4773+ ; X64-NEXT: vpmovsxdq %ymm0, %zmm0
4774+ ; X64-NEXT: kxnorw %k0, %k0, %k1
4775+ ; X64-NEXT: vpxor %xmm2, %xmm2, %xmm2
4776+ ; X64-NEXT: vpxor %xmm3, %xmm3, %xmm3
4777+ ; X64-NEXT: kxnorw %k0, %k0, %k2
4778+ ; X64-NEXT: vpgatherqd (%rdi,%zmm0), %ymm3 {%k2}
4779+ ; X64-NEXT: vpgatherqd (%rdi,%zmm1), %ymm2 {%k1}
4780+ ; X64-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm0
4781+ ; X64-NEXT: retq
4782+ ;
4783+ ; X86-LABEL: pr163023:
4784+ ; X86: # %bb.0:
4785+ ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
4786+ ; X86-NEXT: kxnorw %k0, %k0, %k1
4787+ ; X86-NEXT: vpxor %xmm1, %xmm1, %xmm1
4788+ ; X86-NEXT: vpgatherdd (%eax,%zmm0), %zmm1 {%k1}
4789+ ; X86-NEXT: vmovdqa64 %zmm1, %zmm0
4790+ ; X86-NEXT: retl
4791+ %addr.p = ptrtoint ptr %a0 to i64
4792+ %addr.v = insertelement <1 x i64 > poison, i64 %addr.p , i64 0
4793+ %addr.splat = shufflevector <1 x i64 > %addr.v , <1 x i64 > poison, <16 x i32 > zeroinitializer
4794+ %ofs = sext <16 x i32 > %a1 to <16 x i64 >
4795+ %addr = add nuw <16 x i64 > %addr.splat , %ofs
4796+ %ptr = inttoptr <16 x i64 > %addr to <16 x ptr >
4797+ %gather = tail call fastcc <16 x i32 > @llvm.masked.gather.v16i32.v16p0 (<16 x ptr > %ptr , i32 4 , <16 x i1 > splat (i1 true ), <16 x i32 > poison)
4798+ ret <16 x i32 > %gather
4799+ }
4800+
47684801;
47694802; PR45906
47704803; This used to cause fast-isel to generate bad copy instructions that would
0 commit comments