1 ; Disable machine cse to stress the different path of the algorithm. 2 ; Otherwise, we always fall in the simple case, i.e., only one definition. 3 ; RUN: llc < %s -mtriple=arm64-apple-ios7.0 -disable-machine-cse -aarch64-stress-promote-const -mcpu=cyclone | FileCheck -check-prefix=PROMOTED %s 4 ; The REGULAR run just checks that the inputs passed to promote const expose 5 ; the appropriate patterns. 6 ; RUN: llc < %s -mtriple=arm64-apple-ios7.0 -disable-machine-cse -aarch64-promote-const=false -mcpu=cyclone | FileCheck -check-prefix=REGULAR %s 7 8 %struct.uint8x16x4_t = type { [4 x <16 x i8>] } 9 10 ; Constant is a structure 11 define %struct.uint8x16x4_t @test1() { 12 ; PROMOTED-LABEL: test1: 13 ; Promote constant has created a big constant for the whole structure 14 ; PROMOTED: adrp [[PAGEADDR:x[0-9]+]], __PromotedConst@PAGE 15 ; PROMOTED: add [[BASEADDR:x[0-9]+]], [[PAGEADDR]], __PromotedConst@PAGEOFF 16 ; Destination registers are defined by the ABI 17 ; PROMOTED-NEXT: ldp q0, q1, {{\[}}[[BASEADDR]]] 18 ; PROMOTED-NEXT: ldp q2, q3, {{\[}}[[BASEADDR]], #32] 19 ; PROMOTED-NEXT: ret 20 21 ; REGULAR-LABEL: test1: 22 ; Regular access is quite bad, it performs 4 loads, one for each chunk of 23 ; the structure 24 ; REGULAR: adrp [[PAGEADDR:x[0-9]+]], [[CSTLABEL:lCP.*]]@PAGE 25 ; Destination registers are defined by the ABI 26 ; REGULAR: ldr q0, {{\[}}[[PAGEADDR]], [[CSTLABEL]]@PAGEOFF] 27 ; REGULAR: adrp [[PAGEADDR:x[0-9]+]], [[CSTLABEL:lCP.*]]@PAGE 28 ; REGULAR: ldr q1, {{\[}}[[PAGEADDR]], [[CSTLABEL]]@PAGEOFF] 29 ; REGULAR: adrp [[PAGEADDR2:x[0-9]+]], [[CSTLABEL2:lCP.*]]@PAGE 30 ; REGULAR: ldr q2, {{\[}}[[PAGEADDR2]], [[CSTLABEL2]]@PAGEOFF] 31 ; REGULAR: adrp [[PAGEADDR3:x[0-9]+]], [[CSTLABEL3:lCP.*]]@PAGE 32 ; REGULAR: ldr q3, {{\[}}[[PAGEADDR3]], [[CSTLABEL3]]@PAGEOFF] 33 ; REGULAR-NEXT: ret 34 entry: 35 ret %struct.uint8x16x4_t { [4 x <16 x i8>] [<16 x i8> <i8 -40, i8 -93, i8 -118, i8 -99, i8 -75, i8 -105, i8 74, i8 -110, i8 62, i8 -115, i8 -119, i8 -120, i8 34, i8 -124, i8 0, i8 -128>, <16 x i8> <i8 32, i8 124, i8 121, i8 120, i8 8, i8 117, i8 -56, i8 113, i8 -76, i8 110, i8 -53, i8 107, i8 7, i8 105, i8 103, i8 102>, <16 x i8> <i8 -24, i8 99, i8 -121, i8 97, i8 66, i8 95, i8 24, i8 93, i8 6, i8 91, i8 12, i8 89, i8 39, i8 87, i8 86, i8 85>, <16 x i8> <i8 -104, i8 83, i8 -20, i8 81, i8 81, i8 80, i8 -59, i8 78, i8 73, i8 77, i8 -37, i8 75, i8 122, i8 74, i8 37, i8 73>] } 36 } 37 38 ; Two different uses of the same constant in the same basic block 39 define <16 x i8> @test2(<16 x i8> %arg) { 40 entry: 41 ; PROMOTED-LABEL: test2: 42 ; In stress mode, constant vector are promoted 43 ; PROMOTED: adrp [[PAGEADDR:x[0-9]+]], [[CSTV1:__PromotedConst[0-9]+]]@PAGE 44 ; PROMOTED: add [[BASEADDR:x[0-9]+]], [[PAGEADDR]], [[CSTV1]]@PAGEOFF 45 ; PROMOTED: ldr q[[REGNUM:[0-9]+]], {{\[}}[[BASEADDR]]] 46 ; Destination register is defined by ABI 47 ; PROMOTED-NEXT: add.16b v0, v0, v[[REGNUM]] 48 ; PROMOTED-NEXT: mla.16b v0, v0, v[[REGNUM]] 49 ; PROMOTED-NEXT: ret 50 51 ; REGULAR-LABEL: test2: 52 ; Regular access is strickly the same as promoted access. 53 ; The difference is that the address (and thus the space in memory) is not 54 ; shared between constants 55 ; REGULAR: adrp [[PAGEADDR:x[0-9]+]], [[CSTLABEL:lCP.*]]@PAGE 56 ; REGULAR: ldr q[[REGNUM:[0-9]+]], {{\[}}[[PAGEADDR]], [[CSTLABEL]]@PAGEOFF] 57 ; Destination register is defined by ABI 58 ; REGULAR-NEXT: add.16b v0, v0, v[[REGNUM]] 59 ; REGULAR-NEXT: mla.16b v0, v0, v[[REGNUM]] 60 ; REGULAR-NEXT: ret 61 %add.i = add <16 x i8> %arg, <i8 -40, i8 -93, i8 -118, i8 -99, i8 -75, i8 -105, i8 74, i8 -110, i8 62, i8 -115, i8 -119, i8 -120, i8 34, i8 -124, i8 0, i8 -128> 62 %mul.i = mul <16 x i8> %add.i, <i8 -40, i8 -93, i8 -118, i8 -99, i8 -75, i8 -105, i8 74, i8 -110, i8 62, i8 -115, i8 -119, i8 -120, i8 34, i8 -124, i8 0, i8 -128> 63 %add.i9 = add <16 x i8> %add.i, %mul.i 64 ret <16 x i8> %add.i9 65 } 66 67 ; Two different uses of the sane constant in two different basic blocks, 68 ; one dominates the other 69 define <16 x i8> @test3(<16 x i8> %arg, i32 %path) { 70 ; PROMOTED-LABEL: test3: 71 ; In stress mode, constant vector are promoted 72 ; Since, the constant is the same as the previous function, 73 ; the same address must be used 74 ; PROMOTED: adrp [[PAGEADDR:x[0-9]+]], [[CSTV1]]@PAGE 75 ; PROMOTED: add [[BASEADDR:x[0-9]+]], [[PAGEADDR]], [[CSTV1]]@PAGEOFF 76 ; PROMOTED-NEXT: ldr q[[REGNUM:[0-9]+]], {{\[}}[[BASEADDR]]] 77 ; Destination register is defined by ABI 78 ; PROMOTED-NEXT: add.16b v0, v0, v[[REGNUM]] 79 ; PROMOTED-NEXT: cbnz w0, [[LABEL:LBB.*]] 80 ; Next BB 81 ; PROMOTED: adrp [[PAGEADDR:x[0-9]+]], [[CSTV2:__PromotedConst[0-9]+]]@PAGE 82 ; PROMOTED: add [[BASEADDR:x[0-9]+]], [[PAGEADDR]], [[CSTV2]]@PAGEOFF 83 ; PROMOTED-NEXT: ldr q[[REGNUM]], {{\[}}[[BASEADDR]]] 84 ; Next BB 85 ; PROMOTED-NEXT: [[LABEL]]: 86 ; PROMOTED-NEXT: mul.16b [[DESTV:v[0-9]+]], v0, v[[REGNUM]] 87 ; PROMOTED-NEXT: add.16b v0, v0, [[DESTV]] 88 ; PROMOTED-NEXT: ret 89 90 ; REGULAR-LABEL: test3: 91 ; Regular mode does not elimitate common sub expression by its own. 92 ; In other words, the same loads appears several times. 93 ; REGULAR: adrp [[PAGEADDR:x[0-9]+]], [[CSTLABEL1:lCP.*]]@PAGE 94 ; REGULAR-NEXT: ldr q[[REGNUM:[0-9]+]], {{\[}}[[PAGEADDR]], [[CSTLABEL1]]@PAGEOFF] 95 ; Destination register is defined by ABI 96 ; REGULAR-NEXT: add.16b v0, v0, v[[REGNUM]] 97 ; REGULAR-NEXT: cbz w0, [[LABELelse:LBB.*]] 98 ; Next BB 99 ; Redundant load 100 ; REGULAR: adrp [[PAGEADDR:x[0-9]+]], [[CSTLABEL1]]@PAGE 101 ; REGULAR-NEXT: ldr q[[REGNUM]], {{\[}}[[PAGEADDR]], [[CSTLABEL1]]@PAGEOFF] 102 ; REGULAR-NEXT: b [[LABELend:LBB.*]] 103 ; Next BB 104 ; REGULAR-NEXT: [[LABELelse]] 105 ; REGULAR-NEXT: adrp [[PAGEADDR:x[0-9]+]], [[CSTLABEL2:lCP.*]]@PAGE 106 ; REGULAR-NEXT: ldr q[[REGNUM]], {{\[}}[[PAGEADDR]], [[CSTLABEL2]]@PAGEOFF] 107 ; Next BB 108 ; REGULAR-NEXT: [[LABELend]]: 109 ; REGULAR-NEXT: mul.16b [[DESTV:v[0-9]+]], v0, v[[REGNUM]] 110 ; REGULAR-NEXT: add.16b v0, v0, [[DESTV]] 111 ; REGULAR-NEXT: ret 112 entry: 113 %add.i = add <16 x i8> %arg, <i8 -40, i8 -93, i8 -118, i8 -99, i8 -75, i8 -105, i8 74, i8 -110, i8 62, i8 -115, i8 -119, i8 -120, i8 34, i8 -124, i8 0, i8 -128> 114 %tobool = icmp eq i32 %path, 0 115 br i1 %tobool, label %if.else, label %if.then 116 117 if.then: ; preds = %entry 118 %mul.i13 = mul <16 x i8> %add.i, <i8 -40, i8 -93, i8 -118, i8 -99, i8 -75, i8 -105, i8 74, i8 -110, i8 62, i8 -115, i8 -119, i8 -120, i8 34, i8 -124, i8 0, i8 -128> 119 br label %if.end 120 121 if.else: ; preds = %entry 122 %mul.i = mul <16 x i8> %add.i, <i8 -24, i8 99, i8 -121, i8 97, i8 66, i8 95, i8 24, i8 93, i8 6, i8 91, i8 12, i8 89, i8 39, i8 87, i8 86, i8 85> 123 br label %if.end 124 125 if.end: ; preds = %if.else, %if.then 126 %ret2.0 = phi <16 x i8> [ %mul.i13, %if.then ], [ %mul.i, %if.else ] 127 %add.i12 = add <16 x i8> %add.i, %ret2.0 128 ret <16 x i8> %add.i12 129 } 130 131 ; Two different uses of the sane constant in two different basic blocks, 132 ; none dominates the other 133 define <16 x i8> @test4(<16 x i8> %arg, i32 %path) { 134 ; PROMOTED-LABEL: test4: 135 ; In stress mode, constant vector are promoted 136 ; Since, the constant is the same as the previous function, 137 ; the same address must be used 138 ; PROMOTED: adrp [[PAGEADDR:x[0-9]+]], [[CSTV1]]@PAGE 139 ; PROMOTED: add [[BASEADDR:x[0-9]+]], [[PAGEADDR]], [[CSTV1]]@PAGEOFF 140 ; PROMOTED-NEXT: ldr q[[REGNUM:[0-9]+]], {{\[}}[[BASEADDR]]] 141 ; Destination register is defined by ABI 142 ; PROMOTED-NEXT: add.16b v0, v0, v[[REGNUM]] 143 ; PROMOTED-NEXT: cbz w0, [[LABEL:LBB.*]] 144 ; Next BB 145 ; PROMOTED: mul.16b v0, v0, v[[REGNUM]] 146 ; Next BB 147 ; PROMOTED-NEXT: [[LABEL]]: 148 ; PROMOTED-NEXT: ret 149 150 151 ; REGULAR-LABEL: test4: 152 ; REGULAR: adrp [[PAGEADDR:x[0-9]+]], [[CSTLABEL3:lCP.*]]@PAGE 153 ; REGULAR-NEXT: ldr q[[REGNUM:[0-9]+]], {{\[}}[[PAGEADDR]], [[CSTLABEL3]]@PAGEOFF] 154 ; Destination register is defined by ABI 155 ; REGULAR-NEXT: add.16b v0, v0, v[[REGNUM]] 156 ; REGULAR-NEXT: cbz w0, [[LABEL:LBB.*]] 157 ; Next BB 158 ; Redundant expression 159 ; REGULAR: adrp [[PAGEADDR:x[0-9]+]], [[CSTLABEL3]]@PAGE 160 ; REGULAR-NEXT: ldr q[[REGNUM:[0-9]+]], {{\[}}[[PAGEADDR]], [[CSTLABEL3]]@PAGEOFF] 161 ; Destination register is defined by ABI 162 ; REGULAR-NEXT: mul.16b v0, v0, v[[REGNUM]] 163 ; Next BB 164 ; REGULAR-NEXT: [[LABEL]]: 165 ; REGULAR-NEXT: ret 166 entry: 167 %add.i = add <16 x i8> %arg, <i8 -40, i8 -93, i8 -118, i8 -99, i8 -75, i8 -105, i8 74, i8 -110, i8 62, i8 -115, i8 -119, i8 -120, i8 34, i8 -124, i8 0, i8 -128> 168 %tobool = icmp eq i32 %path, 0 169 br i1 %tobool, label %if.end, label %if.then 170 171 if.then: ; preds = %entry 172 %mul.i = mul <16 x i8> %add.i, <i8 -40, i8 -93, i8 -118, i8 -99, i8 -75, i8 -105, i8 74, i8 -110, i8 62, i8 -115, i8 -119, i8 -120, i8 34, i8 -124, i8 0, i8 -128> 173 br label %if.end 174 175 if.end: ; preds = %entry, %if.then 176 %ret.0 = phi <16 x i8> [ %mul.i, %if.then ], [ %add.i, %entry ] 177 ret <16 x i8> %ret.0 178 } 179 180 ; Two different uses of the sane constant in two different basic blocks, 181 ; one is in a phi. 182 define <16 x i8> @test5(<16 x i8> %arg, i32 %path) { 183 ; PROMOTED-LABEL: test5: 184 ; In stress mode, constant vector are promoted 185 ; Since, the constant is the same as the previous function, 186 ; the same address must be used 187 ; PROMOTED: adrp [[PAGEADDR:x[0-9]+]], [[CSTV1]]@PAGE 188 ; PROMOTED: add [[BASEADDR:x[0-9]+]], [[PAGEADDR]], [[CSTV1]]@PAGEOFF 189 ; PROMOTED-NEXT: ldr q[[REGNUM:[0-9]+]], {{\[}}[[BASEADDR]]] 190 ; PROMOTED-NEXT: cbz w0, [[LABEL:LBB.*]] 191 ; Next BB 192 ; PROMOTED: add.16b [[DESTV:v[0-9]+]], v0, v[[REGNUM]] 193 ; PROMOTED-NEXT: mul.16b v[[REGNUM]], [[DESTV]], v[[REGNUM]] 194 ; Next BB 195 ; PROMOTED-NEXT: [[LABEL]]: 196 ; PROMOTED-NEXT: mul.16b [[TMP1:v[0-9]+]], v[[REGNUM]], v[[REGNUM]] 197 ; PROMOTED-NEXT: mul.16b [[TMP2:v[0-9]+]], [[TMP1]], [[TMP1]] 198 ; PROMOTED-NEXT: mul.16b [[TMP3:v[0-9]+]], [[TMP2]], [[TMP2]] 199 ; PROMOTED-NEXT: mul.16b v0, [[TMP3]], [[TMP3]] 200 ; PROMOTED-NEXT: ret 201 202 ; REGULAR-LABEL: test5: 203 ; REGULAR: cbz w0, [[LABELelse:LBB.*]] 204 ; Next BB 205 ; REGULAR: adrp [[PAGEADDR:x[0-9]+]], [[CSTLABEL:lCP.*]]@PAGE 206 ; REGULAR-NEXT: ldr q[[REGNUM:[0-9]+]], {{\[}}[[PAGEADDR]], [[CSTLABEL]]@PAGEOFF] 207 ; REGULAR-NEXT: add.16b [[DESTV:v[0-9]+]], v0, v[[REGNUM]] 208 ; REGULAR-NEXT: mul.16b v[[DESTREGNUM:[0-9]+]], [[DESTV]], v[[REGNUM]] 209 ; REGULAR-NEXT: b [[LABELend:LBB.*]] 210 ; Next BB 211 ; REGULAR-NEXT: [[LABELelse]] 212 ; REGULAR-NEXT: adrp [[PAGEADDR:x[0-9]+]], [[CSTLABEL:lCP.*]]@PAGE 213 ; REGULAR-NEXT: ldr q[[DESTREGNUM]], {{\[}}[[PAGEADDR]], [[CSTLABEL]]@PAGEOFF] 214 ; Next BB 215 ; REGULAR-NEXT: [[LABELend]]: 216 ; REGULAR-NEXT: mul.16b [[TMP1:v[0-9]+]], v[[DESTREGNUM]], v[[DESTREGNUM]] 217 ; REGULAR-NEXT: mul.16b [[TMP2:v[0-9]+]], [[TMP1]], [[TMP1]] 218 ; REGULAR-NEXT: mul.16b [[TMP3:v[0-9]+]], [[TMP2]], [[TMP2]] 219 ; REGULAR-NEXT: mul.16b v0, [[TMP3]], [[TMP3]] 220 ; REGULAR-NEXT: ret 221 entry: 222 %tobool = icmp eq i32 %path, 0 223 br i1 %tobool, label %if.end, label %if.then 224 225 if.then: ; preds = %entry 226 %add.i = add <16 x i8> %arg, <i8 -40, i8 -93, i8 -118, i8 -99, i8 -75, i8 -105, i8 74, i8 -110, i8 62, i8 -115, i8 -119, i8 -120, i8 34, i8 -124, i8 0, i8 -128> 227 %mul.i26 = mul <16 x i8> %add.i, <i8 -40, i8 -93, i8 -118, i8 -99, i8 -75, i8 -105, i8 74, i8 -110, i8 62, i8 -115, i8 -119, i8 -120, i8 34, i8 -124, i8 0, i8 -128> 228 br label %if.end 229 230 if.end: ; preds = %entry, %if.then 231 %ret.0 = phi <16 x i8> [ %mul.i26, %if.then ], [ <i8 -40, i8 -93, i8 -118, i8 -99, i8 -75, i8 -105, i8 74, i8 -110, i8 62, i8 -115, i8 -119, i8 -120, i8 34, i8 -124, i8 0, i8 -128>, %entry ] 232 %mul.i25 = mul <16 x i8> %ret.0, %ret.0 233 %mul.i24 = mul <16 x i8> %mul.i25, %mul.i25 234 %mul.i23 = mul <16 x i8> %mul.i24, %mul.i24 235 %mul.i = mul <16 x i8> %mul.i23, %mul.i23 236 ret <16 x i8> %mul.i 237 } 238 239 define void @accessBig(i64* %storage) { 240 ; PROMOTED-LABEL: accessBig: 241 ; PROMOTED: adrp 242 ; PROMOTED: ret 243 %addr = bitcast i64* %storage to <1 x i80>* 244 store <1 x i80> <i80 483673642326615442599424>, <1 x i80>* %addr 245 ret void 246 } 247 248 define void @asmStatement() { 249 ; PROMOTED-LABEL: asmStatement: 250 ; PROMOTED-NOT: adrp 251 ; PROMOTED: ret 252 call void asm sideeffect "bfxil w0, w0, $0, $1", "i,i"(i32 28, i32 4) 253 ret void 254 } 255 256