230 lines · plain
1; RUN: opt -mtriple=dxil-pc-shadermodel6.2-compute -S -dxil-intrinsic-expansion %s | FileCheck %s --check-prefixes=CHECK,CHECK622; RUN: opt -mtriple=dxil-pc-shadermodel6.3-compute -S -dxil-intrinsic-expansion %s | FileCheck %s --check-prefixes=CHECK,CHECK633 4define void @loadf64(i32 %index) {5 ; check the handle from binding is unchanged6 ; CHECK: [[B:%.*]] = call target("dx.Rawbuffer", double, 0, 0)7 ; CHECK-SAME: @llvm.dx.resource.handlefrombinding.tdx.Rawbuffer_f64_0_0t(8 ; CHECK-SAME: i32 0, i32 1, i32 1, i32 0, ptr null)9 %buffer = call target("dx.Rawbuffer", double, 0, 0)10 @llvm.dx.resource.handlefrombinding.tdx.Rawbuffer_f64_1_0_0t(11 i32 0, i32 1, i32 1, i32 0, ptr null)12 13 ; check we don't modify the code in sm6.3 or later14 ; CHECK63: [[L0:%.*]] = call { double, i1 } @llvm.dx.resource.load.rawbuffer15 ; CHECK63-SAME: target("dx.Rawbuffer", double, 0, 0) [[B]], i32 %index, i32 0)16 17 ; check we load an <2 x i32> instead of a double18 ; CHECK62-NOT: call {double, i1} @llvm.dx.resource.load.rawbuffer19 ; CHECK62: [[L0:%.*]] = call { <2 x i32>, i1 }20 ; CHECK62-SAME: @llvm.dx.resource.load.rawbuffer.v2i32.tdx.Rawbuffer_f64_0_0t(21 ; CHECK62-SAME: target("dx.Rawbuffer", double, 0, 0) [[B]], i32 %index, i32 0) 22 %load0 = call {double, i1} @llvm.dx.resource.load.rawbuffer(23 target("dx.Rawbuffer", double, 0, 0) %buffer, i32 %index, i32 0)24 25 ; CHECK63: extractvalue { double, i1 } [[L0]], 026 27 ; check we extract the two i32 and construct a double28 ; CHECK62: [[D0:%.*]] = extractvalue { <2 x i32>, i1 } [[L0]], 029 ; CHECK62: [[Lo:%.*]] = extractelement <2 x i32> [[D0]], i32 030 ; CHECK62: [[Hi:%.*]] = extractelement <2 x i32> [[D0]], i32 131 ; CHECK62: [[DBL:%.*]] = call double @llvm.dx.asdouble.i32(i32 [[Lo]], i32 [[Hi]])32 ; CHECK62-NOT: extractvalue { double, i1 }33 %data0 = extractvalue {double, i1} %load0, 034 ret void35}36 37define void @loadv2f64(i32 %index) {38 ; check the handle from binding is unchanged39 ; CHECK: [[B:%.*]] = call target("dx.Rawbuffer", <2 x double>, 0, 0)40 ; CHECK-SAME: @llvm.dx.resource.handlefrombinding.tdx.Rawbuffer_v2f64_0_0t(41 ; CHECK-SAME: i32 0, i32 1, i32 1, i32 0, ptr null)42 %buffer = call target("dx.Rawbuffer", <2 x double>, 0, 0)43 @llvm.dx.resource.handlefrombinding.tdx.Rawbuffer_v2f64_1_0_0t(44 i32 0, i32 1, i32 1, i32 0, ptr null)45 46 ; check we don't modify the code in sm6.3 or later47 ; CHECK63: [[L0:%.*]] = call { <2 x double>, i1 } @llvm.dx.resource.load.rawbuffer48 ; CHECK63-SAME: target("dx.Rawbuffer", <2 x double>, 0, 0) [[B]], i32 %index, i32 0)49 50 ; check we load an <4 x i32> instead of a double251 ; CHECK62: [[L0:%.*]] = call { <4 x i32>, i1 }52 ; CHECK62-SAME: @llvm.dx.resource.load.rawbuffer.v4i32.tdx.Rawbuffer_v2f64_0_0t(53 ; CHECK62-SAME: target("dx.Rawbuffer", <2 x double>, 0, 0) [[B]], i32 %index, i32 0)54 %load0 = call { <2 x double>, i1 } @llvm.dx.resource.load.rawbuffer(55 target("dx.Rawbuffer", <2 x double>, 0, 0) %buffer, i32 %index, i32 0)56 57 ; CHECK63: extractvalue { <2 x double>, i1 } [[L0]], 058 59 ; check we extract the 4 i32 and construct a <2 x double>60 ; CHECK62: [[D0:%.*]] = extractvalue { <4 x i32>, i1 } [[L0]], 061 ; CHECK62: [[Lo1:%.*]] = extractelement <4 x i32> [[D0]], i32 062 ; CHECK62: [[Hi1:%.*]] = extractelement <4 x i32> [[D0]], i32 163 ; CHECK62: [[Lo2:%.*]] = extractelement <4 x i32> [[D0]], i32 264 ; CHECK62: [[Hi2:%.*]] = extractelement <4 x i32> [[D0]], i32 365 ; CHECK62: [[Dbl1:%.*]] = call double @llvm.dx.asdouble.i32(i32 [[Lo1]], i32 [[Hi1]])66 ; CHECK62: [[Vec:%.*]] = insertelement <2 x double> poison, double [[Dbl1]], i32 067 ; CHECK62: [[Dbl2:%.*]] = call double @llvm.dx.asdouble.i32(i32 [[Lo2]], i32 [[Hi2]])68 ; CHECK62: [[Vec2:%.*]] = insertelement <2 x double> [[Vec]], double [[Dbl2]], i32 169 ; CHECK62-NOT: extractvalue { <2 x double>, i1 }70 %data0 = extractvalue { <2 x double>, i1 } %load0, 071 ret void72}73 74; show we properly handle extracting the check bit75define void @loadf64WithCheckBit(i32 %index) {76 ; check the handle from binding is unchanged77 ; CHECK: [[B:%.*]] = call target("dx.Rawbuffer", double, 0, 0)78 ; CHECK-SAME: @llvm.dx.resource.handlefrombinding.tdx.Rawbuffer_f64_0_0t(79 ; CHECK-SAME: i32 0, i32 1, i32 1, i32 0, ptr null)80 %buffer = call target("dx.Rawbuffer", double, 0, 0)81 @llvm.dx.resource.handlefrombinding.tdx.Rawbuffer_f64_1_0_0t(82 i32 0, i32 1, i32 1, i32 0, ptr null)83 84 ; check we don't modify the code in sm6.3 or later85 ; CHECK63: [[L0:%.*]] = call { double, i1 } @llvm.dx.resource.load.rawbuffer86 ; CHECK63-SAME: target("dx.Rawbuffer", double, 0, 0) [[B]], i32 %index, i32 0)87 88 ; check we load an <2 x i32> instead of a double89 ; CHECK62-NOT: call {double, i1} @llvm.dx.resource.load.rawbuffer90 ; CHECK62: [[L0:%.*]] = call { <2 x i32>, i1 }91 ; CHECK62-SAME: @llvm.dx.resource.load.rawbuffer.v2i32.tdx.Rawbuffer_f64_0_0t(92 ; CHECK62-SAME: target("dx.Rawbuffer", double, 0, 0) [[B]], i32 %index, i32 0) 93 %load0 = call {double, i1} @llvm.dx.resource.load.rawbuffer(94 target("dx.Rawbuffer", double, 0, 0) %buffer, i32 %index, i32 0)95 96 ; CHECK63: extractvalue { double, i1 } [[L0]], 097 ; CHECK63: extractvalue { double, i1 } [[L0]], 198 99 ; check we extract the two i32 and construct a double100 ; CHECK62: [[D0:%.*]] = extractvalue { <2 x i32>, i1 } [[L0]], 0101 ; CHECK62: [[Lo:%.*]] = extractelement <2 x i32> [[D0]], i32 0102 ; CHECK62: [[Hi:%.*]] = extractelement <2 x i32> [[D0]], i32 1103 ; CHECK62: [[DBL:%.*]] = call double @llvm.dx.asdouble.i32(i32 [[Lo]], i32 [[Hi]])104 %data0 = extractvalue {double, i1} %load0, 0105 ; CHECK62: extractvalue { <2 x i32>, i1 } [[L0]], 1106 ; CHECK62-NOT: extractvalue { double, i1 }107 %cb = extractvalue {double, i1} %load0, 1108 ret void109}110 111; Raw Buffer Load allows for double3 and double4 to be loaded112; In SM6.2 and below, two loads will be performed.113; Show we and the checkbits together114 115define void @loadv3f64(i32 %index) {116 ; check the handle from binding is unchanged117 ; CHECK: [[B:%.*]] = call target("dx.Rawbuffer", <3 x double>, 0, 0)118 ; CHECK-SAME: @llvm.dx.resource.handlefrombinding.tdx.Rawbuffer_v3f64_0_0t(119 ; CHECK-SAME: i32 0, i32 0, i32 1, i32 0, ptr null)120 %buffer = call target("dx.Rawbuffer", <3 x double>, 0, 0)121 @llvm.dx.resource.handlefrombinding.tdx.Rawbuffer_v3f64_0_0t(122 i32 0, i32 0, i32 1, i32 0, ptr null)123 124 ; check we don't modify the code in sm6.3 or later125 ; CHECK63: [[L0:%.*]] = call { <3 x double>, i1 } @llvm.dx.resource.load.rawbuffer126 ; CHECK63-SAME: target("dx.Rawbuffer", <3 x double>, 0, 0) [[B]], i32 %index, i32 0)127 128 ; check we perform two loads129 ; and do 6 extracts and construct 3 doubles130 ; CHECK62-NOT: call {<3 x double>, i1} @llvm.dx.resource.load.rawbuffer131 ; CHECK62: [[L0:%.*]] = call { <4 x i32>, i1 }132 ; CHECK62-SAME: @llvm.dx.resource.load.rawbuffer.v4i32.tdx.Rawbuffer_v3f64_0_0t(133 ; CHECK62-SAME: target("dx.Rawbuffer", <3 x double>, 0, 0) [[B]], i32 %index, i32 0)134 135 ; CHECK62: [[D0:%.*]] = extractvalue { <4 x i32>, i1 } [[L0]], 0136 ; CHECK62: [[Lo1:%.*]] = extractelement <4 x i32> [[D0]], i32 0137 ; CHECK62: [[Hi1:%.*]] = extractelement <4 x i32> [[D0]], i32 1138 ; CHECK62: [[Lo2:%.*]] = extractelement <4 x i32> [[D0]], i32 2139 ; CHECK62: [[Hi2:%.*]] = extractelement <4 x i32> [[D0]], i32 3140 ; CHECK62: [[DBL1:%.*]] = call double @llvm.dx.asdouble.i32(i32 [[Lo1]], i32 [[Hi1]])141 ; CHECK62: [[Vec1:%.*]] = insertelement <3 x double> poison, double [[DBL1]], i32 0142 ; CHECK62: [[DBL2:%.*]] = call double @llvm.dx.asdouble.i32(i32 [[Lo2]], i32 [[Hi2]])143 ; CHECK62: [[Vec2:%.*]] = insertelement <3 x double> [[Vec1]], double [[DBL2]], i32 1144 145 ; 2nd load146 ; CHECK62: [[L2:%.*]] = call { <2 x i32>, i1 }147 ; CHECK62-SAME: @llvm.dx.resource.load.rawbuffer.v2i32.tdx.Rawbuffer_v3f64_0_0t(148 ; CHECK62-SAME: target("dx.Rawbuffer", <3 x double>, 0, 0) [[B]], i32 %index, i32 16)149 150 ; CHECK62: [[D2:%.*]] = extractvalue { <2 x i32>, i1 } [[L2]], 0151 ; CHECK62: [[Lo3:%.*]] = extractelement <2 x i32> [[D2]], i32 0152 ; CHECK62: [[Hi3:%.*]] = extractelement <2 x i32> [[D2]], i32 1153 ; CHECK62: [[DBL3:%.*]] = call double @llvm.dx.asdouble.i32(i32 [[Lo3]], i32 [[Hi3]])154 ; CHECK62: [[Vec3:%.*]] = insertelement <3 x double> [[Vec2]], double [[DBL3]], i32 2155 %load0 = call {<3 x double>, i1} @llvm.dx.resource.load.rawbuffer(156 target("dx.Rawbuffer", <3 x double>, 0, 0) %buffer, i32 %index, i32 0)157 158 ; CHECK63: extractvalue { <3 x double>, i1 } [[L0]], 0159 ; CHECK63: extractvalue { <3 x double>, i1 } [[L0]], 1160 161 ; CHECK62-NOT: extractvalue {<3 x double>, i1 }162 %data0 = extractvalue {<3 x double>, i1} %load0, 0163 ; check we extract checkbit from both loads and and them together164 ; CHECK62: [[B1:%.*]] = extractvalue { <4 x i32>, i1 } [[L0]], 1165 ; CHECK62: [[B2:%.*]] = extractvalue { <2 x i32>, i1 } [[L2]], 1166 ; CHECK62: and i1 [[B1]], [[B2]]167 %cb = extractvalue {<3 x double>, i1} %load0, 1168 ret void169}170 171define void @loadv4f64(i32 %index) {172 ; check the handle from binding is unchanged173 ; CHECK: [[B:%.*]] = call target("dx.Rawbuffer", <4 x double>, 0, 0)174 ; CHECK-SAME: @llvm.dx.resource.handlefrombinding.tdx.Rawbuffer_v4f64_0_0t(175 ; CHECK-SAME: i32 0, i32 0, i32 1, i32 0, ptr null)176 %buffer = call target("dx.Rawbuffer", <4 x double>, 0, 0)177 @llvm.dx.resource.handlefrombinding.tdx.Rawbuffer_v4f64_0_0t(178 i32 0, i32 0, i32 1, i32 0, ptr null)179 180 ; check we don't modify the code in sm6.3 or later181 ; CHECK63: [[L0:%.*]] = call { <4 x double>, i1 } @llvm.dx.resource.load.rawbuffer182 ; CHECK63-SAME: target("dx.Rawbuffer", <4 x double>, 0, 0) [[B]], i32 %index, i32 0)183 184 ; check we perform two loads185 ; and do 8 extracts and construct 4 doubles186 ; CHECK62-NOT: call {<4 x double>, i1} @llvm.dx.resource.load.rawbuffer187 ; CHECK62: [[L0:%.*]] = call { <4 x i32>, i1 }188 ; CHECK62-SAME: @llvm.dx.resource.load.rawbuffer.v4i32.tdx.Rawbuffer_v4f64_0_0t(189 ; CHECK62-SAME: target("dx.Rawbuffer", <4 x double>, 0, 0) [[B]], i32 %index, i32 0)190 191 ; CHECK62: [[D0:%.*]] = extractvalue { <4 x i32>, i1 } [[L0]], 0192 ; CHECK62: [[Lo1:%.*]] = extractelement <4 x i32> [[D0]], i32 0193 ; CHECK62: [[Hi1:%.*]] = extractelement <4 x i32> [[D0]], i32 1194 ; CHECK62: [[Lo2:%.*]] = extractelement <4 x i32> [[D0]], i32 2195 ; CHECK62: [[Hi2:%.*]] = extractelement <4 x i32> [[D0]], i32 3196 ; CHECK62: [[DBL1:%.*]] = call double @llvm.dx.asdouble.i32(i32 [[Lo1]], i32 [[Hi1]])197 ; CHECK62: [[Vec1:%.*]] = insertelement <4 x double> poison, double [[DBL1]], i32 0198 ; CHECK62: [[DBL2:%.*]] = call double @llvm.dx.asdouble.i32(i32 [[Lo2]], i32 [[Hi2]])199 ; CHECK62: [[Vec2:%.*]] = insertelement <4 x double> [[Vec1]], double [[DBL2]], i32 1200 201 ; 2nd load202 ; CHECK62: [[L2:%.*]] = call { <4 x i32>, i1 }203 ; CHECK62-SAME: @llvm.dx.resource.load.rawbuffer.v4i32.tdx.Rawbuffer_v4f64_0_0t(204 ; CHECK62-SAME: target("dx.Rawbuffer", <4 x double>, 0, 0) [[B]], i32 %index, i32 16)205 206 ; CHECK62: [[D2:%.*]] = extractvalue { <4 x i32>, i1 } [[L2]], 0207 ; CHECK62: [[Lo3:%.*]] = extractelement <4 x i32> [[D2]], i32 0208 ; CHECK62: [[Hi3:%.*]] = extractelement <4 x i32> [[D2]], i32 1209 ; CHECK62: [[Lo4:%.*]] = extractelement <4 x i32> [[D2]], i32 2210 ; CHECK62: [[Hi4:%.*]] = extractelement <4 x i32> [[D2]], i32 3211 ; CHECK62: [[DBL3:%.*]] = call double @llvm.dx.asdouble.i32(i32 [[Lo3]], i32 [[Hi3]])212 ; CHECK62: [[Vec3:%.*]] = insertelement <4 x double> [[Vec2]], double [[DBL3]], i32 2213 ; CHECK62: [[DBL4:%.*]] = call double @llvm.dx.asdouble.i32(i32 [[Lo4]], i32 [[Hi4]])214 ; CHECK62: [[Vec4:%.*]] = insertelement <4 x double> [[Vec3]], double [[DBL4]], i32 3215 %load0 = call {<4 x double>, i1} @llvm.dx.resource.load.rawbuffer(216 target("dx.Rawbuffer", <4 x double>, 0, 0) %buffer, i32 %index, i32 0)217 218 ; CHECK63: extractvalue { <4 x double>, i1 } [[L0]], 0219 ; CHECK63: extractvalue { <4 x double>, i1 } [[L0]], 1220 221 ; CHECK62-NOT: extractvalue {<4 x double>, i1 }222 %data0 = extractvalue {<4 x double>, i1} %load0, 0223 ; check we extract checkbit from both loads and and them together224 ; CHECK62: [[B1:%.*]] = extractvalue { <4 x i32>, i1 } [[L0]], 1225 ; CHECK62: [[B2:%.*]] = extractvalue { <4 x i32>, i1 } [[L2]], 1226 ; CHECK62: and i1 [[B1]], [[B2]]227 %cb = extractvalue {<4 x double>, i1} %load0, 1228 ret void229}230