| ; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z13 | FileCheck %s | |
| ; | |
| ; Test that sign-extending the result of a shuffle that takes its leading | |
| ; lanes from the second shuffle operand unpacks the second operand. The | |
| ; shuffle mask indexes the concatenation of both operands, so mask values | |
| ; that are at least the lane count of one operand select lanes of the second. | |
| ; | |
| ; Registers other than the incoming arguments (%v24, %v26) are matched with | |
| ; variables, so that the checks do not depend on register allocation. | |
| ; Lanes 0-3 are b[0..3] (mask offset 8), lanes 4-7 are a[4..7]. | |
| define void @fun0(<8 x i16> %a, <8 x i16> %b, ptr %dst) { | |
| ; CHECK-LABEL: fun0: | |
| ; CHECK-DAG: vuphh [[BLANES:%v[0-9]+]], %v26 | |
| ; CHECK-DAG: vuplhw [[ALANES:%v[0-9]+]], %v24 | |
| ; CHECK-DAG: vst [[BLANES]], 0(%r2), 4 | |
| ; CHECK-DAG: vst [[ALANES]], 16(%r2), 4 | |
| ; CHECK: br %r14 | |
| %s = shufflevector <8 x i16> %a, <8 x i16> %b, | |
| <8 x i32> <i32 8, i32 9, i32 10, i32 11, | |
| i32 4, i32 5, i32 6, i32 7> | |
| %z = sext <8 x i16> %s to <8 x i32> | |
| store <8 x i32> %z, ptr %dst | |
| ret void | |
| } | |
| ; Lanes 0-1 are b[0..1] (mask offset 4), lanes 2-3 are a[2..3]. | |
| define void @fun1(<4 x i32> %a, <4 x i32> %b, ptr %dst) { | |
| ; CHECK-LABEL: fun1: | |
| ; CHECK: vuphf [[BLANES:%v[0-9]+]], %v26 | |
| ; CHECK: vst [[BLANES]], 0(%r2), 4 | |
| ; CHECK: br %r14 | |
| %s = shufflevector <4 x i32> %a, <4 x i32> %b, | |
| <4 x i32> <i32 4, i32 5, i32 2, i32 3> | |
| %z = sext <4 x i32> %s to <4 x i64> | |
| store <4 x i64> %z, ptr %dst | |
| ret void | |
| } | |
| ; The reproducer from the bug report: lanes 0-1 are b[6..7] (mask offset 22). | |
| ; Only the unpack chain that produces the first result register is checked. | |
| ; It must start from %v26 (b), not %v24 (a). | |
| define <16 x i64> @fun2(<16 x i8> %a, <16 x i8> %b) { | |
| ; CHECK-LABEL: fun2: | |
| ; CHECK: vuphb [[HW:%v[0-9]+]], %v26 | |
| ; CHECK: vuplhw [[W:%v[0-9]+]], [[HW]] | |
| ; CHECK: vuplf {{%v[0-9]+}}, [[W]] | |
| ; CHECK: br %r14 | |
| %s = shufflevector <16 x i8> %a, <16 x i8> %b, | |
| <16 x i32> <i32 22, i32 23, i32 1, i32 2, | |
| i32 3, i32 4, i32 5, i32 6, | |
| i32 7, i32 8, i32 9, i32 10, | |
| i32 11, i32 12, i32 13, i32 14> | |
| %z = sext <16 x i8> %s to <16 x i64> | |
| ret <16 x i64> %z | |
| } | |
| ; Lanes 0-3 are a[4..7]. The first operand must still be used for the leading | |
| ; lanes when the offset is nonzero, but less than the lane count. | |
| define void @fun3(<8 x i16> %a, <8 x i16> %b, ptr %dst) { | |
| ; CHECK-LABEL: fun3: | |
| ; CHECK-DAG: vuplhw [[ALANES:%v[0-9]+]], %v24 | |
| ; CHECK-DAG: vuphh [[BLANES:%v[0-9]+]], %v26 | |
| ; CHECK-DAG: vst [[ALANES]], 0(%r2), 4 | |
| ; CHECK-DAG: vst [[BLANES]], 16(%r2), 4 | |
| ; CHECK: br %r14 | |
| %s = shufflevector <8 x i16> %a, <8 x i16> %b, | |
| <8 x i32> <i32 4, i32 5, i32 6, i32 7, | |
| i32 8, i32 9, i32 10, i32 11> | |
| %z = sext <8 x i16> %s to <8 x i32> | |
| store <8 x i32> %z, ptr %dst | |
| ret void | |
| } |