xref: /llvm-project/llvm/test/CodeGen/DirectX/countbits.ll (revision 011b618644113996e2c0a8e57db40f89d20878e3)
167518a44SSarah Spall; RUN: opt -S -scalarizer -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-library %s | FileCheck %s
267518a44SSarah Spall
367518a44SSarah Spall; Make sure dxil operation function calls for countbits are generated for all integer types.
467518a44SSarah Spall
567518a44SSarah Spalldefine noundef i16 @test_countbits_short(i16 noundef %a) {
667518a44SSarah Spallentry:
7*011b6186SFinn Plummer; CHECK: [[A:%.*]] = call i32 @dx.op.unaryBits.i16(i32 31, i16 %{{.*}}) #[[#ATTR:]]
875e7ba8cSSarah Spall; CHECK-NEXT: [[B:%.*]] = trunc i32 [[A]] to i16
975e7ba8cSSarah Spall; CHECK-NEXT ret i16 [[B]]
1067518a44SSarah Spall  %elt.ctpop = call i16 @llvm.ctpop.i16(i16 %a)
1167518a44SSarah Spall  ret i16 %elt.ctpop
1267518a44SSarah Spall}
1367518a44SSarah Spall
1475e7ba8cSSarah Spalldefine noundef i32 @test_countbits_short2(i16 noundef %a) {
1575e7ba8cSSarah Spallentry:
16*011b6186SFinn Plummer; CHECK: [[A:%.*]] = call i32 @dx.op.unaryBits.i16(i32 31, i16 %{{.*}}) #[[#ATTR]]
1775e7ba8cSSarah Spall; CHECK-NEXT: ret i32 [[A]]
1875e7ba8cSSarah Spall  %elt.ctpop = call i16 @llvm.ctpop.i16(i16 %a)
1975e7ba8cSSarah Spall  %elt.zext = zext i16 %elt.ctpop to i32
2075e7ba8cSSarah Spall  ret i32 %elt.zext
2175e7ba8cSSarah Spall}
2275e7ba8cSSarah Spall
2375e7ba8cSSarah Spalldefine noundef i32 @test_countbits_short3(i16 noundef %a) {
2475e7ba8cSSarah Spallentry:
25*011b6186SFinn Plummer; CHECK: [[A:%.*]] = call i32 @dx.op.unaryBits.i16(i32 31, i16 %{{.*}}) #[[#ATTR]]
2675e7ba8cSSarah Spall; CHECK-NEXT: ret i32 [[A]]
2775e7ba8cSSarah Spall  %elt.ctpop = call i16 @llvm.ctpop.i16(i16 %a)
2875e7ba8cSSarah Spall  %elt.sext = sext i16 %elt.ctpop to i32
2975e7ba8cSSarah Spall  ret i32 %elt.sext
3075e7ba8cSSarah Spall}
3175e7ba8cSSarah Spall
3267518a44SSarah Spalldefine noundef i32 @test_countbits_int(i32 noundef %a) {
3367518a44SSarah Spallentry:
34*011b6186SFinn Plummer; CHECK: [[A:%.*]] = call i32 @dx.op.unaryBits.i32(i32 31, i32 %{{.*}}) #[[#ATTR]]
3575e7ba8cSSarah Spall; CHECK-NEXT: ret i32 [[A]]
3667518a44SSarah Spall  %elt.ctpop = call i32 @llvm.ctpop.i32(i32 %a)
3767518a44SSarah Spall  ret i32 %elt.ctpop
3867518a44SSarah Spall}
3967518a44SSarah Spall
4067518a44SSarah Spalldefine noundef i64 @test_countbits_long(i64 noundef %a) {
4167518a44SSarah Spallentry:
42*011b6186SFinn Plummer; CHECK: [[A:%.*]] = call i32 @dx.op.unaryBits.i64(i32 31, i64 %{{.*}}) #[[#ATTR]]
4375e7ba8cSSarah Spall; CHECK-NEXT: [[B:%.*]] = zext i32 [[A]] to i64
4475e7ba8cSSarah Spall; CHECK-NEXT ret i64 [[B]]
4567518a44SSarah Spall  %elt.ctpop = call i64 @llvm.ctpop.i64(i64 %a)
4667518a44SSarah Spall  ret i64 %elt.ctpop
4767518a44SSarah Spall}
4867518a44SSarah Spall
4975e7ba8cSSarah Spalldefine noundef i32 @test_countbits_long2(i64 noundef %a) {
5075e7ba8cSSarah Spallentry:
51*011b6186SFinn Plummer; CHECK: [[A:%.*]] = call i32 @dx.op.unaryBits.i64(i32 31, i64 %{{.*}}) #[[#ATTR]]
5275e7ba8cSSarah Spall; CHECK-NEXT: ret i32 [[A]]
5375e7ba8cSSarah Spall  %elt.ctpop = call i64 @llvm.ctpop.i64(i64 %a)
5475e7ba8cSSarah Spall  %elt.trunc = trunc i64 %elt.ctpop to i32
5575e7ba8cSSarah Spall  ret i32 %elt.trunc
5675e7ba8cSSarah Spall}
5775e7ba8cSSarah Spall
5867518a44SSarah Spalldefine noundef <4 x i32> @countbits_vec4_i32(<4 x i32> noundef %a)  {
5967518a44SSarah Spallentry:
6067518a44SSarah Spall  ; CHECK: [[ee0:%.*]] = extractelement <4 x i32> %a, i64 0
61*011b6186SFinn Plummer  ; CHECK: [[ie0:%.*]] = call i32 @dx.op.unaryBits.i32(i32 31, i32 [[ee0]]) #[[#ATTR]]
6267518a44SSarah Spall  ; CHECK: [[ee1:%.*]] = extractelement <4 x i32> %a, i64 1
63*011b6186SFinn Plummer  ; CHECK: [[ie1:%.*]] = call i32 @dx.op.unaryBits.i32(i32 31, i32 [[ee1]]) #[[#ATTR]]
6467518a44SSarah Spall  ; CHECK: [[ee2:%.*]] = extractelement <4 x i32> %a, i64 2
65*011b6186SFinn Plummer  ; CHECK: [[ie2:%.*]] = call i32 @dx.op.unaryBits.i32(i32 31, i32 [[ee2]]) #[[#ATTR]]
6667518a44SSarah Spall  ; CHECK: [[ee3:%.*]] = extractelement <4 x i32> %a, i64 3
67*011b6186SFinn Plummer  ; CHECK: [[ie3:%.*]] = call i32 @dx.op.unaryBits.i32(i32 31, i32 [[ee3]]) #[[#ATTR]]
6867518a44SSarah Spall  ; CHECK: insertelement <4 x i32> poison, i32 [[ie0]], i64 0
6967518a44SSarah Spall  ; CHECK: insertelement <4 x i32> %{{.*}}, i32 [[ie1]], i64 1
7067518a44SSarah Spall  ; CHECK: insertelement <4 x i32> %{{.*}}, i32 [[ie2]], i64 2
7167518a44SSarah Spall  ; CHECK: insertelement <4 x i32> %{{.*}}, i32 [[ie3]], i64 3
7267518a44SSarah Spall  %2 = call <4 x i32> @llvm.ctpop.v4i32(<4 x i32> %a)
7367518a44SSarah Spall  ret <4 x i32> %2
7467518a44SSarah Spall}
7567518a44SSarah Spall
76*011b6186SFinn Plummer; CHECK: attributes #[[#ATTR]] = {{{.*}} memory(none) {{.*}}}
77*011b6186SFinn Plummer
7867518a44SSarah Spalldeclare i16 @llvm.ctpop.i16(i16)
7967518a44SSarah Spalldeclare i32 @llvm.ctpop.i32(i32)
8067518a44SSarah Spalldeclare i64 @llvm.ctpop.i64(i64)
8167518a44SSarah Spalldeclare <4 x i32> @llvm.ctpop.v4i32(<4 x i32>)
82