1; RUN: opt -S -dxil-intrinsic-expansion -mtriple=dxil-pc-shadermodel6.3-library %s | FileCheck %s --check-prefixes=CHECK,EXPCHECK 2; RUN: opt -S -dxil-intrinsic-expansion -scalarizer -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-library %s | FileCheck %s --check-prefixes=CHECK,DOPCHECK 3 4; Make sure correct dxil expansions for atan2 are generated for float and half. 5 6define noundef float @atan2_float(float noundef %y, float noundef %x) { 7entry: 8; CHECK: [[DIV:%.+]] = fdiv float %y, %x 9; EXPCHECK: [[ATAN:%.+]] = call float @llvm.atan.f32(float [[DIV]]) 10; DOPCHECK: [[ATAN:%.+]] = call float @dx.op.unary.f32(i32 17, float [[DIV]]) 11; CHECK-DAG: [[ADD_PI:%.+]] = fadd float [[ATAN]], 0x400921FB60000000 12; CHECK-DAG: [[SUB_PI:%.+]] = fsub float [[ATAN]], 0x400921FB60000000 13; CHECK-DAG: [[X_LT_0:%.+]] = fcmp olt float %x, 0.000000e+00 14; CHECK-DAG: [[X_EQ_0:%.+]] = fcmp oeq float %x, 0.000000e+00 15; CHECK-DAG: [[Y_GE_0:%.+]] = fcmp oge float %y, 0.000000e+00 16; CHECK-DAG: [[Y_LT_0:%.+]] = fcmp olt float %y, 0.000000e+00 17; CHECK: [[XLT0_AND_YGE0:%.+]] = and i1 [[X_LT_0]], [[Y_GE_0]] 18; CHECK: [[SELECT_ADD_PI:%.+]] = select i1 [[XLT0_AND_YGE0]], float [[ADD_PI]], float [[ATAN]] 19; CHECK: [[XLT0_AND_YLT0:%.+]] = and i1 [[X_LT_0]], [[Y_LT_0]] 20; CHECK: [[SELECT_SUB_PI:%.+]] = select i1 [[XLT0_AND_YLT0]], float [[SUB_PI]], float [[SELECT_ADD_PI]] 21; CHECK: [[XEQ0_AND_YLT0:%.+]] = and i1 [[X_EQ_0]], [[Y_LT_0]] 22; CHECK: [[SELECT_NEGHPI:%.+]] = select i1 [[XEQ0_AND_YLT0]], float 0xBFF921FB60000000, float [[SELECT_SUB_PI]] 23; CHECK: [[XEQ0_AND_YGE0:%.+]] = and i1 [[X_EQ_0]], [[Y_GE_0]] 24; CHECK: [[SELECT_HPI:%.+]] = select i1 [[XEQ0_AND_YGE0]], float 0x3FF921FB60000000, float [[SELECT_NEGHPI]] 25; CHECK: ret float [[SELECT_HPI]] 26 %elt.atan2 = call float @llvm.atan2.f32(float %y, float %x) 27 ret float %elt.atan2 28} 29 30define noundef half @atan2_half(half noundef %y, half noundef %x) { 31entry: 32; CHECK: [[DIV:%.+]] = fdiv half %y, %x 33; EXPCHECK: [[ATAN:%.+]] = call half @llvm.atan.f16(half [[DIV]]) 34; DOPCHECK: [[ATAN:%.+]] = call half @dx.op.unary.f16(i32 17, half [[DIV]]) 35; CHECK-DAG: [[ADD_PI:%.+]] = fadd half [[ATAN]], 0xH4248 36; CHECK-DAG: [[SUB_PI:%.+]] = fsub half [[ATAN]], 0xH4248 37; CHECK-DAG: [[X_LT_0:%.+]] = fcmp olt half %x, 0xH0000 38; CHECK-DAG: [[X_EQ_0:%.+]] = fcmp oeq half %x, 0xH0000 39; CHECK-DAG: [[Y_GE_0:%.+]] = fcmp oge half %y, 0xH0000 40; CHECK-DAG: [[Y_LT_0:%.+]] = fcmp olt half %y, 0xH0000 41; CHECK: [[XLT0_AND_YGE0:%.+]] = and i1 [[X_LT_0]], [[Y_GE_0]] 42; CHECK: [[SELECT_ADD_PI:%.+]] = select i1 [[XLT0_AND_YGE0]], half [[ADD_PI]], half [[ATAN]] 43; CHECK: [[XLT0_AND_YLT0:%.+]] = and i1 [[X_LT_0]], [[Y_LT_0]] 44; CHECK: [[SELECT_SUB_PI:%.+]] = select i1 [[XLT0_AND_YLT0]], half [[SUB_PI]], half [[SELECT_ADD_PI]] 45; CHECK: [[XEQ0_AND_YLT0:%.+]] = and i1 [[X_EQ_0]], [[Y_LT_0]] 46; CHECK: [[SELECT_NEGHPI:%.+]] = select i1 [[XEQ0_AND_YLT0]], half 0xHBE48, half [[SELECT_SUB_PI]] 47; CHECK: [[XEQ0_AND_YGE0:%.+]] = and i1 [[X_EQ_0]], [[Y_GE_0]] 48; CHECK: [[SELECT_HPI:%.+]] = select i1 [[XEQ0_AND_YGE0]], half 0xH3E48, half [[SELECT_NEGHPI]] 49; CHECK: ret half [[SELECT_HPI]] 50 %elt.atan2 = call half @llvm.atan2.f16(half %y, half %x) 51 ret half %elt.atan2 52} 53 54define noundef <4 x float> @atan2_float4(<4 x float> noundef %y, <4 x float> noundef %x) { 55entry: 56; Just Expansion, no scalarization or lowering: 57; EXPCHECK: [[DIV:%.+]] = fdiv <4 x float> %y, %x 58; EXPCHECK: [[ATAN:%.+]] = call <4 x float> @llvm.atan.v4f32(<4 x float> [[DIV]]) 59; EXPCHECK-DAG: [[ADD_PI:%.+]] = fadd <4 x float> [[ATAN]], splat (float 0x400921FB60000000) 60; EXPCHECK-DAG: [[SUB_PI:%.+]] = fsub <4 x float> [[ATAN]], splat (float 0x400921FB60000000) 61; EXPCHECK-DAG: [[X_LT_0:%.+]] = fcmp olt <4 x float> %x, zeroinitializer 62; EXPCHECK-DAG: [[X_EQ_0:%.+]] = fcmp oeq <4 x float> %x, zeroinitializer 63; EXPCHECK-DAG: [[Y_GE_0:%.+]] = fcmp oge <4 x float> %y, zeroinitializer 64; EXPCHECK-DAG: [[Y_LT_0:%.+]] = fcmp olt <4 x float> %y, zeroinitializer 65; EXPCHECK: [[XLT0_AND_YGE0:%.+]] = and <4 x i1> [[X_LT_0]], [[Y_GE_0]] 66; EXPCHECK: [[SELECT_ADD_PI:%.+]] = select <4 x i1> [[XLT0_AND_YGE0]], <4 x float> [[ADD_PI]], <4 x float> [[ATAN]] 67; EXPCHECK: [[XLT0_AND_YLT0:%.+]] = and <4 x i1> [[X_LT_0]], [[Y_LT_0]] 68; EXPCHECK: [[SELECT_SUB_PI:%.+]] = select <4 x i1> [[XLT0_AND_YLT0]], <4 x float> [[SUB_PI]], <4 x float> [[SELECT_ADD_PI]] 69; EXPCHECK: [[XEQ0_AND_YLT0:%.+]] = and <4 x i1> [[X_EQ_0]], [[Y_LT_0]] 70; EXPCHECK: [[SELECT_NEGHPI:%.+]] = select <4 x i1> [[XEQ0_AND_YLT0]], <4 x float> splat (float 0xBFF921FB60000000), <4 x float> [[SELECT_SUB_PI]] 71; EXPCHECK: [[XEQ0_AND_YGE0:%.+]] = and <4 x i1> [[X_EQ_0]], [[Y_GE_0]] 72; EXPCHECK: [[SELECT_HPI:%.+]] = select <4 x i1> [[XEQ0_AND_YGE0]], <4 x float> splat (float 0x3FF921FB60000000), <4 x float> [[SELECT_NEGHPI]] 73; EXPCHECK: ret <4 x float> [[SELECT_HPI]] 74 75; Scalarization occurs after expansion, so atan scalarization is tested separately. 76; Expansion, scalarization and lowering: 77; Just make sure this expands to exactly 4 scalar DXIL atan (OpCode=17) calls. 78; DOPCHECK-COUNT-4: call float @dx.op.unary.f32(i32 17, float %{{.*}}) 79; DOPCHECK-NOT: call float @dx.op.unary.f32(i32 17, 80 81 %elt.atan2 = call <4 x float> @llvm.atan2.v4f32(<4 x float> %y, <4 x float> %x) 82 ret <4 x float> %elt.atan2 83} 84 85declare half @llvm.atan2.f16(half, half) 86declare float @llvm.atan2.f32(float, float) 87declare <4 x float> @llvm.atan2.v4f32(<4 x float>, <4 x float>) 88