82 lines · plain
1; RUN: opt -S -scalarizer -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-library %s | FileCheck %s2 3; Make sure dxil operation function calls for countbits are generated for all integer types.4 5define noundef i16 @test_countbits_short(i16 noundef %a) {6entry:7; CHECK: [[A:%.*]] = call i32 @dx.op.unaryBits.i16(i32 31, i16 %{{.*}}) #[[#ATTR:]]8; CHECK-NEXT: [[B:%.*]] = trunc i32 [[A]] to i169; CHECK-NEXT ret i16 [[B]]10 %elt.ctpop = call i16 @llvm.ctpop.i16(i16 %a)11 ret i16 %elt.ctpop12}13 14define noundef i32 @test_countbits_short2(i16 noundef %a) {15entry:16; CHECK: [[A:%.*]] = call i32 @dx.op.unaryBits.i16(i32 31, i16 %{{.*}}) #[[#ATTR]]17; CHECK-NEXT: ret i32 [[A]]18 %elt.ctpop = call i16 @llvm.ctpop.i16(i16 %a)19 %elt.zext = zext i16 %elt.ctpop to i3220 ret i32 %elt.zext21}22 23define noundef i32 @test_countbits_short3(i16 noundef %a) {24entry:25; CHECK: [[A:%.*]] = call i32 @dx.op.unaryBits.i16(i32 31, i16 %{{.*}}) #[[#ATTR]]26; CHECK-NEXT: ret i32 [[A]]27 %elt.ctpop = call i16 @llvm.ctpop.i16(i16 %a)28 %elt.sext = sext i16 %elt.ctpop to i3229 ret i32 %elt.sext30}31 32define noundef i32 @test_countbits_int(i32 noundef %a) {33entry:34; CHECK: [[A:%.*]] = call i32 @dx.op.unaryBits.i32(i32 31, i32 %{{.*}}) #[[#ATTR]]35; CHECK-NEXT: ret i32 [[A]]36 %elt.ctpop = call i32 @llvm.ctpop.i32(i32 %a)37 ret i32 %elt.ctpop38}39 40define noundef i64 @test_countbits_long(i64 noundef %a) {41entry:42; CHECK: [[A:%.*]] = call i32 @dx.op.unaryBits.i64(i32 31, i64 %{{.*}}) #[[#ATTR]]43; CHECK-NEXT: [[B:%.*]] = zext i32 [[A]] to i6444; CHECK-NEXT ret i64 [[B]]45 %elt.ctpop = call i64 @llvm.ctpop.i64(i64 %a)46 ret i64 %elt.ctpop47}48 49define noundef i32 @test_countbits_long2(i64 noundef %a) {50entry:51; CHECK: [[A:%.*]] = call i32 @dx.op.unaryBits.i64(i32 31, i64 %{{.*}}) #[[#ATTR]]52; CHECK-NEXT: ret i32 [[A]]53 %elt.ctpop = call i64 @llvm.ctpop.i64(i64 %a)54 %elt.trunc = trunc i64 %elt.ctpop to i3255 ret i32 %elt.trunc56}57 58define noundef <4 x i32> @countbits_vec4_i32(<4 x i32> noundef %a) {59entry:60 ; CHECK: [[ee0:%.*]] = extractelement <4 x i32> %a, i64 061 ; CHECK: [[ie0:%.*]] = call i32 @dx.op.unaryBits.i32(i32 31, i32 [[ee0]]) #[[#ATTR]]62 ; CHECK: [[ee1:%.*]] = extractelement <4 x i32> %a, i64 163 ; CHECK: [[ie1:%.*]] = call i32 @dx.op.unaryBits.i32(i32 31, i32 [[ee1]]) #[[#ATTR]]64 ; CHECK: [[ee2:%.*]] = extractelement <4 x i32> %a, i64 265 ; CHECK: [[ie2:%.*]] = call i32 @dx.op.unaryBits.i32(i32 31, i32 [[ee2]]) #[[#ATTR]]66 ; CHECK: [[ee3:%.*]] = extractelement <4 x i32> %a, i64 367 ; CHECK: [[ie3:%.*]] = call i32 @dx.op.unaryBits.i32(i32 31, i32 [[ee3]]) #[[#ATTR]]68 ; CHECK: insertelement <4 x i32> poison, i32 [[ie0]], i64 069 ; CHECK: insertelement <4 x i32> %{{.*}}, i32 [[ie1]], i64 170 ; CHECK: insertelement <4 x i32> %{{.*}}, i32 [[ie2]], i64 271 ; CHECK: insertelement <4 x i32> %{{.*}}, i32 [[ie3]], i64 372 %2 = call <4 x i32> @llvm.ctpop.v4i32(<4 x i32> %a)73 ret <4 x i32> %274}75 76; CHECK: attributes #[[#ATTR]] = {{{.*}} memory(none) {{.*}}}77 78declare i16 @llvm.ctpop.i16(i16)79declare i32 @llvm.ctpop.i32(i32)80declare i64 @llvm.ctpop.i64(i64)81declare <4 x i32> @llvm.ctpop.v4i32(<4 x i32>)82