980 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -ppc-asm-full-reg-names -verify-machineinstrs -mtriple=powerpc64le-unknown-linux-gnu -mcpu=pwr9 < %s | FileCheck %s3 4; test_no_prep:5; unsigned long test_no_prep(char *p, int count) {6; unsigned long i=0, res=0;7; int DISP1 = 4001;8; int DISP2 = 4002;9; int DISP3 = 4003;10; int DISP4 = 4004;11; for (; i < count ; i++) {12; unsigned long x1 = *(unsigned long *)(p + i + DISP1);13; unsigned long x2 = *(unsigned long *)(p + i + DISP2);14; unsigned long x3 = *(unsigned long *)(p + i + DISP3);15; unsigned long x4 = *(unsigned long *)(p + i + DISP4);16; res += x1*x2*x3*x4;17; }18; return res + count;19; }20 21define i64 @test_no_prep(ptr %arg, i32 signext %arg1) {22; CHECK-LABEL: test_no_prep:23; CHECK: # %bb.0: # %bb24; CHECK-NEXT: cmplwi r4, 025; CHECK-NEXT: beq cr0, .LBB0_426; CHECK-NEXT: # %bb.1: # %bb3.preheader27; CHECK-NEXT: mtctr r428; CHECK-NEXT: addi r5, r3, 400429; CHECK-NEXT: li r3, 030; CHECK-NEXT: li r6, -331; CHECK-NEXT: li r7, -232; CHECK-NEXT: li r8, -133; CHECK-NEXT: .p2align 434; CHECK-NEXT: .LBB0_2: # %bb335; CHECK-NEXT: #36; CHECK-NEXT: ldx r9, r5, r637; CHECK-NEXT: ldx r10, r5, r738; CHECK-NEXT: ldx r11, r5, r839; CHECK-NEXT: ld r12, 0(r5)40; CHECK-NEXT: addi r5, r5, 141; CHECK-NEXT: mulld r9, r10, r942; CHECK-NEXT: mulld r9, r9, r1143; CHECK-NEXT: maddld r3, r9, r12, r344; CHECK-NEXT: bdnz .LBB0_245; CHECK-NEXT: # %bb.3: # %bb2546; CHECK-NEXT: add r3, r3, r447; CHECK-NEXT: blr48; CHECK-NEXT: .LBB0_4:49; CHECK-NEXT: addi r3, r4, 050; CHECK-NEXT: blr51bb:52 %i = sext i32 %arg1 to i6453 %i2 = icmp eq i32 %arg1, 054 br i1 %i2, label %bb25, label %bb355 56bb3: ; preds = %bb3, %bb57 %i4 = phi i64 [ %i23, %bb3 ], [ 0, %bb ]58 %i5 = phi i64 [ %i22, %bb3 ], [ 0, %bb ]59 %i6 = getelementptr inbounds i8, ptr %arg, i64 %i460 %i7 = getelementptr inbounds i8, ptr %i6, i64 400161 %i9 = load i64, ptr %i7, align 862 %i10 = getelementptr inbounds i8, ptr %i6, i64 400263 %i12 = load i64, ptr %i10, align 864 %i13 = getelementptr inbounds i8, ptr %i6, i64 400365 %i15 = load i64, ptr %i13, align 866 %i16 = getelementptr inbounds i8, ptr %i6, i64 400467 %i18 = load i64, ptr %i16, align 868 %i19 = mul i64 %i12, %i969 %i20 = mul i64 %i19, %i1570 %i21 = mul i64 %i20, %i1871 %i22 = add i64 %i21, %i572 %i23 = add nuw i64 %i4, 173 %i24 = icmp ult i64 %i23, %i74 br i1 %i24, label %bb3, label %bb2575 76bb25: ; preds = %bb3, %bb77 %i26 = phi i64 [ 0, %bb ], [ %i22, %bb3 ]78 %i27 = add i64 %i26, %i79 ret i64 %i2780}81 82; test_ds_prep:83; unsigned long test_ds_prep(char *p, int count) {84; unsigned long i=0, res=0;85; int DISP1 = 4001;86; int DISP2 = 4002;87; int DISP3 = 4003;88; int DISP4 = 4006;89; for (; i < count ; i++) {90; unsigned long x1 = *(unsigned long *)(p + i + DISP1);91; unsigned long x2 = *(unsigned long *)(p + i + DISP2);92; unsigned long x3 = *(unsigned long *)(p + i + DISP3);93; unsigned long x4 = *(unsigned long *)(p + i + DISP4);94; res += x1*x2*x3*x4;95; }96; return res + count;97; }98 99define i64 @test_ds_prep(ptr %arg, i32 signext %arg1) {100; CHECK-LABEL: test_ds_prep:101; CHECK: # %bb.0: # %bb102; CHECK-NEXT: cmplwi r4, 0103; CHECK-NEXT: beq cr0, .LBB1_4104; CHECK-NEXT: # %bb.1: # %bb3.preheader105; CHECK-NEXT: mtctr r4106; CHECK-NEXT: addi r7, r3, 4002107; CHECK-NEXT: li r3, 0108; CHECK-NEXT: li r5, -1109; CHECK-NEXT: li r6, 1110; CHECK-NEXT: .p2align 4111; CHECK-NEXT: .LBB1_2: # %bb3112; CHECK-NEXT: #113; CHECK-NEXT: ldx r8, r7, r5114; CHECK-NEXT: ld r9, 0(r7)115; CHECK-NEXT: ldx r10, r7, r6116; CHECK-NEXT: ld r11, 4(r7)117; CHECK-NEXT: addi r7, r7, 1118; CHECK-NEXT: mulld r8, r9, r8119; CHECK-NEXT: mulld r8, r8, r10120; CHECK-NEXT: maddld r3, r8, r11, r3121; CHECK-NEXT: bdnz .LBB1_2122; CHECK-NEXT: # %bb.3: # %bb25123; CHECK-NEXT: add r3, r3, r4124; CHECK-NEXT: blr125; CHECK-NEXT: .LBB1_4:126; CHECK-NEXT: addi r3, r4, 0127; CHECK-NEXT: blr128bb:129 %i = sext i32 %arg1 to i64130 %i2 = icmp eq i32 %arg1, 0131 br i1 %i2, label %bb25, label %bb3132 133bb3: ; preds = %bb3, %bb134 %i4 = phi i64 [ %i23, %bb3 ], [ 0, %bb ]135 %i5 = phi i64 [ %i22, %bb3 ], [ 0, %bb ]136 %i6 = getelementptr inbounds i8, ptr %arg, i64 %i4137 %i7 = getelementptr inbounds i8, ptr %i6, i64 4001138 %i9 = load i64, ptr %i7, align 8139 %i10 = getelementptr inbounds i8, ptr %i6, i64 4002140 %i12 = load i64, ptr %i10, align 8141 %i13 = getelementptr inbounds i8, ptr %i6, i64 4003142 %i15 = load i64, ptr %i13, align 8143 %i16 = getelementptr inbounds i8, ptr %i6, i64 4006144 %i18 = load i64, ptr %i16, align 8145 %i19 = mul i64 %i12, %i9146 %i20 = mul i64 %i19, %i15147 %i21 = mul i64 %i20, %i18148 %i22 = add i64 %i21, %i5149 %i23 = add nuw i64 %i4, 1150 %i24 = icmp ult i64 %i23, %i151 br i1 %i24, label %bb3, label %bb25152 153bb25: ; preds = %bb3, %bb154 %i26 = phi i64 [ 0, %bb ], [ %i22, %bb3 ]155 %i27 = add i64 %i26, %i156 ret i64 %i27157}158 159; test_max_number_reminder:160; unsigned long test_max_number_reminder(char *p, int count) {161; unsigned long i=0, res=0;162; int DISP1 = 4001;163; int DISP2 = 4002;164; int DISP3 = 4003;165; int DISP4 = 4005;166; int DISP5 = 4006;167; int DISP6 = 4007;168; int DISP7 = 4014;169; int DISP8 = 4010;170; int DISP9 = 4011;171; for (; i < count ; i++) {172; unsigned long x1 = *(unsigned long *)(p + i + DISP1);173; unsigned long x2 = *(unsigned long *)(p + i + DISP2);174; unsigned long x3 = *(unsigned long *)(p + i + DISP3);175; unsigned long x4 = *(unsigned long *)(p + i + DISP4);176; unsigned long x5 = *(unsigned long *)(p + i + DISP5);177; unsigned long x6 = *(unsigned long *)(p + i + DISP6);178; unsigned long x7 = *(unsigned long *)(p + i + DISP7);179; unsigned long x8 = *(unsigned long *)(p + i + DISP8);180; unsigned long x9 = *(unsigned long *)(p + i + DISP9);181; res += x1*x2*x3*x4*x5*x6*x7*x8*x9;182; }183; return res + count;184;}185 186define i64 @test_max_number_reminder(ptr %arg, i32 signext %arg1) {187; CHECK-LABEL: test_max_number_reminder:188; CHECK: # %bb.0: # %bb189; CHECK-NEXT: cmplwi r4, 0190; CHECK-NEXT: beq cr0, .LBB2_4191; CHECK-NEXT: # %bb.1: # %bb3.preheader192; CHECK-NEXT: std r25, -56(r1) # 8-byte Folded Spill193; CHECK-NEXT: std r26, -48(r1) # 8-byte Folded Spill194; CHECK-NEXT: addi r10, r3, 4002195; CHECK-NEXT: li r3, 0196; CHECK-NEXT: li r5, -1197; CHECK-NEXT: li r6, 1198; CHECK-NEXT: li r7, 3199; CHECK-NEXT: li r8, 5200; CHECK-NEXT: li r9, 9201; CHECK-NEXT: std r27, -40(r1) # 8-byte Folded Spill202; CHECK-NEXT: std r28, -32(r1) # 8-byte Folded Spill203; CHECK-NEXT: std r29, -24(r1) # 8-byte Folded Spill204; CHECK-NEXT: std r30, -16(r1) # 8-byte Folded Spill205; CHECK-NEXT: mtctr r4206; CHECK-NEXT: .p2align 4207; CHECK-NEXT: .LBB2_2: # %bb3208; CHECK-NEXT: #209; CHECK-NEXT: ldx r11, r10, r5210; CHECK-NEXT: ld r12, 0(r10)211; CHECK-NEXT: ldx r0, r10, r6212; CHECK-NEXT: ldx r30, r10, r7213; CHECK-NEXT: mulld r11, r12, r11214; CHECK-NEXT: ld r29, 4(r10)215; CHECK-NEXT: ldx r28, r10, r8216; CHECK-NEXT: ld r27, 12(r10)217; CHECK-NEXT: ld r26, 8(r10)218; CHECK-NEXT: ldx r25, r10, r9219; CHECK-NEXT: addi r10, r10, 1220; CHECK-NEXT: mulld r11, r11, r0221; CHECK-NEXT: mulld r11, r11, r30222; CHECK-NEXT: mulld r11, r11, r29223; CHECK-NEXT: mulld r11, r11, r28224; CHECK-NEXT: mulld r11, r11, r27225; CHECK-NEXT: mulld r11, r11, r26226; CHECK-NEXT: maddld r3, r11, r25, r3227; CHECK-NEXT: bdnz .LBB2_2228; CHECK-NEXT: # %bb.3:229; CHECK-NEXT: ld r30, -16(r1) # 8-byte Folded Reload230; CHECK-NEXT: ld r29, -24(r1) # 8-byte Folded Reload231; CHECK-NEXT: ld r28, -32(r1) # 8-byte Folded Reload232; CHECK-NEXT: ld r27, -40(r1) # 8-byte Folded Reload233; CHECK-NEXT: add r3, r3, r4234; CHECK-NEXT: ld r26, -48(r1) # 8-byte Folded Reload235; CHECK-NEXT: ld r25, -56(r1) # 8-byte Folded Reload236; CHECK-NEXT: blr237; CHECK-NEXT: .LBB2_4:238; CHECK-NEXT: addi r3, r4, 0239; CHECK-NEXT: blr240bb:241 %i = sext i32 %arg1 to i64242 %i2 = icmp eq i32 %arg1, 0243 br i1 %i2, label %bb45, label %bb3244 245bb3: ; preds = %bb3, %bb246 %i4 = phi i64 [ %i43, %bb3 ], [ 0, %bb ]247 %i5 = phi i64 [ %i42, %bb3 ], [ 0, %bb ]248 %i6 = getelementptr inbounds i8, ptr %arg, i64 %i4249 %i7 = getelementptr inbounds i8, ptr %i6, i64 4001250 %i9 = load i64, ptr %i7, align 8251 %i10 = getelementptr inbounds i8, ptr %i6, i64 4002252 %i12 = load i64, ptr %i10, align 8253 %i13 = getelementptr inbounds i8, ptr %i6, i64 4003254 %i15 = load i64, ptr %i13, align 8255 %i16 = getelementptr inbounds i8, ptr %i6, i64 4005256 %i18 = load i64, ptr %i16, align 8257 %i19 = getelementptr inbounds i8, ptr %i6, i64 4006258 %i21 = load i64, ptr %i19, align 8259 %i22 = getelementptr inbounds i8, ptr %i6, i64 4007260 %i24 = load i64, ptr %i22, align 8261 %i25 = getelementptr inbounds i8, ptr %i6, i64 4014262 %i27 = load i64, ptr %i25, align 8263 %i28 = getelementptr inbounds i8, ptr %i6, i64 4010264 %i30 = load i64, ptr %i28, align 8265 %i31 = getelementptr inbounds i8, ptr %i6, i64 4011266 %i33 = load i64, ptr %i31, align 8267 %i34 = mul i64 %i12, %i9268 %i35 = mul i64 %i34, %i15269 %i36 = mul i64 %i35, %i18270 %i37 = mul i64 %i36, %i21271 %i38 = mul i64 %i37, %i24272 %i39 = mul i64 %i38, %i27273 %i40 = mul i64 %i39, %i30274 %i41 = mul i64 %i40, %i33275 %i42 = add i64 %i41, %i5276 %i43 = add nuw i64 %i4, 1277 %i44 = icmp ult i64 %i43, %i278 br i1 %i44, label %bb3, label %bb45279 280bb45: ; preds = %bb3, %bb281 %i46 = phi i64 [ 0, %bb ], [ %i42, %bb3 ]282 %i47 = add i64 %i46, %i283 ret i64 %i47284}285 286; test_update_ds_prep_interact:287; unsigned long test_update_ds_prep_interact(char *p, int count) {288; unsigned long i=0, res=0;289; int DISP1 = 4001;290; int DISP2 = 4002;291; int DISP3 = 4003;292; int DISP4 = 4006;293; for (; i < count ; i++) {294; unsigned long x1 = *(unsigned long *)(p + 4 * i + DISP1);295; unsigned long x2 = *(unsigned long *)(p + 4 * i + DISP2);296; unsigned long x3 = *(unsigned long *)(p + 4 * i + DISP3);297; unsigned long x4 = *(unsigned long *)(p + 4 * i + DISP4);298; res += x1*x2*x3*x4;299; }300; return res + count;301; }302 303define dso_local i64 @test_update_ds_prep_interact(ptr %arg, i32 signext %arg1) {304; CHECK-LABEL: test_update_ds_prep_interact:305; CHECK: # %bb.0: # %bb306; CHECK-NEXT: cmplwi r4, 0307; CHECK-NEXT: beq cr0, .LBB3_4308; CHECK-NEXT: # %bb.1: # %bb3.preheader309; CHECK-NEXT: mtctr r4310; CHECK-NEXT: addi r5, r3, 3998311; CHECK-NEXT: li r3, 0312; CHECK-NEXT: li r6, -1313; CHECK-NEXT: li r7, 1314; CHECK-NEXT: .p2align 5315; CHECK-NEXT: .LBB3_2: # %bb3316; CHECK-NEXT: #317; CHECK-NEXT: ldu r8, 4(r5)318; CHECK-NEXT: ldx r9, r5, r6319; CHECK-NEXT: ldx r10, r5, r7320; CHECK-NEXT: ld r11, 4(r5)321; CHECK-NEXT: mulld r8, r8, r9322; CHECK-NEXT: mulld r8, r8, r10323; CHECK-NEXT: maddld r3, r8, r11, r3324; CHECK-NEXT: bdnz .LBB3_2325; CHECK-NEXT: # %bb.3: # %bb26326; CHECK-NEXT: add r3, r3, r4327; CHECK-NEXT: blr328; CHECK-NEXT: .LBB3_4:329; CHECK-NEXT: addi r3, r4, 0330; CHECK-NEXT: blr331bb:332 %i = sext i32 %arg1 to i64333 %i2 = icmp eq i32 %arg1, 0334 br i1 %i2, label %bb26, label %bb3335 336bb3: ; preds = %bb3, %bb337 %i4 = phi i64 [ %i24, %bb3 ], [ 0, %bb ]338 %i5 = phi i64 [ %i23, %bb3 ], [ 0, %bb ]339 %i6 = shl i64 %i4, 2340 %i7 = getelementptr inbounds i8, ptr %arg, i64 %i6341 %i8 = getelementptr inbounds i8, ptr %i7, i64 4001342 %i10 = load i64, ptr %i8, align 8343 %i11 = getelementptr inbounds i8, ptr %i7, i64 4002344 %i13 = load i64, ptr %i11, align 8345 %i14 = getelementptr inbounds i8, ptr %i7, i64 4003346 %i16 = load i64, ptr %i14, align 8347 %i17 = getelementptr inbounds i8, ptr %i7, i64 4006348 %i19 = load i64, ptr %i17, align 8349 %i20 = mul i64 %i13, %i10350 %i21 = mul i64 %i20, %i16351 %i22 = mul i64 %i21, %i19352 %i23 = add i64 %i22, %i5353 %i24 = add nuw i64 %i4, 1354 %i25 = icmp ult i64 %i24, %i355 br i1 %i25, label %bb3, label %bb26356 357bb26: ; preds = %bb3, %bb358 %i27 = phi i64 [ 0, %bb ], [ %i23, %bb3 ]359 %i28 = add i64 %i27, %i360 ret i64 %i28361}362 363; test_update_ds_prep_nointeract:364; unsigned long test_update_ds_prep_nointeract(char *p, int count) {365; unsigned long i=0, res=0;366; int DISP1 = 4001;367; int DISP2 = 4002;368; int DISP3 = 4003;369; int DISP4 = 4007;370; for (; i < count ; i++) {371; char x1 = *(p + i + DISP1);372; unsigned long x2 = *(unsigned long *)(p + i + DISP2);373; unsigned long x3 = *(unsigned long *)(p + i + DISP3);374; unsigned long x4 = *(unsigned long *)(p + i + DISP4);375; res += (unsigned long)x1*x2*x3*x4;376; }377; return res + count;378; }379 380define i64 @test_update_ds_prep_nointeract(ptr %arg, i32 signext %arg1) {381; CHECK-LABEL: test_update_ds_prep_nointeract:382; CHECK: # %bb.0: # %bb383; CHECK-NEXT: cmplwi r4, 0384; CHECK-NEXT: beq cr0, .LBB4_4385; CHECK-NEXT: # %bb.1: # %bb3.preheader386; CHECK-NEXT: mtctr r4387; CHECK-NEXT: addi r5, r3, 4000388; CHECK-NEXT: addi r6, r3, 4003389; CHECK-NEXT: li r3, 0390; CHECK-NEXT: li r7, -1391; CHECK-NEXT: .p2align 4392; CHECK-NEXT: .LBB4_2: # %bb3393; CHECK-NEXT: #394; CHECK-NEXT: lbzu r8, 1(r5)395; CHECK-NEXT: ldx r9, r6, r7396; CHECK-NEXT: ld r10, 0(r6)397; CHECK-NEXT: ld r11, 4(r6)398; CHECK-NEXT: addi r6, r6, 1399; CHECK-NEXT: mulld r8, r9, r8400; CHECK-NEXT: mulld r8, r8, r10401; CHECK-NEXT: maddld r3, r8, r11, r3402; CHECK-NEXT: bdnz .LBB4_2403; CHECK-NEXT: # %bb.3: # %bb25404; CHECK-NEXT: add r3, r3, r4405; CHECK-NEXT: blr406; CHECK-NEXT: .LBB4_4:407; CHECK-NEXT: addi r3, r4, 0408; CHECK-NEXT: blr409bb:410 %i = sext i32 %arg1 to i64411 %i2 = icmp eq i32 %arg1, 0412 br i1 %i2, label %bb25, label %bb3413 414bb3: ; preds = %bb3, %bb415 %i4 = phi i64 [ %i23, %bb3 ], [ 0, %bb ]416 %i5 = phi i64 [ %i22, %bb3 ], [ 0, %bb ]417 %i6 = getelementptr inbounds i8, ptr %arg, i64 %i4418 %i7 = getelementptr inbounds i8, ptr %i6, i64 4001419 %i8 = load i8, ptr %i7, align 1420 %i9 = getelementptr inbounds i8, ptr %i6, i64 4002421 %i11 = load i64, ptr %i9, align 8422 %i12 = getelementptr inbounds i8, ptr %i6, i64 4003423 %i14 = load i64, ptr %i12, align 8424 %i15 = getelementptr inbounds i8, ptr %i6, i64 4007425 %i17 = load i64, ptr %i15, align 8426 %i18 = zext i8 %i8 to i64427 %i19 = mul i64 %i11, %i18428 %i20 = mul i64 %i19, %i14429 %i21 = mul i64 %i20, %i17430 %i22 = add i64 %i21, %i5431 %i23 = add nuw i64 %i4, 1432 %i24 = icmp ult i64 %i23, %i433 br i1 %i24, label %bb3, label %bb25434 435bb25: ; preds = %bb3, %bb436 %i26 = phi i64 [ 0, %bb ], [ %i22, %bb3 ]437 %i27 = add i64 %i26, %i438 ret i64 %i27439}440 441; test_ds_multiple_chains:442; unsigned long test_ds_multiple_chains(char *p, char *q, int count) {443; unsigned long i=0, res=0;444; int DISP1 = 4001;445; int DISP2 = 4010;446; int DISP3 = 4005;447; int DISP4 = 4009;448; for (; i < count ; i++) {449; unsigned long x1 = *(unsigned long *)(p + i + DISP1);450; unsigned long x2 = *(unsigned long *)(p + i + DISP2);451; unsigned long x3 = *(unsigned long *)(p + i + DISP3);452; unsigned long x4 = *(unsigned long *)(p + i + DISP4);453; unsigned long x5 = *(unsigned long *)(q + i + DISP1);454; unsigned long x6 = *(unsigned long *)(q + i + DISP2);455; unsigned long x7 = *(unsigned long *)(q + i + DISP3);456; unsigned long x8 = *(unsigned long *)(q + i + DISP4);457; res += x1*x2*x3*x4*x5*x6*x7*x8;458; }459; return res + count;460; }461 462define dso_local i64 @test_ds_multiple_chains(ptr %arg, ptr %arg1, i32 signext %arg2) {463; CHECK-LABEL: test_ds_multiple_chains:464; CHECK: # %bb.0: # %bb465; CHECK-NEXT: cmplwi r5, 0466; CHECK-NEXT: beq cr0, .LBB5_4467; CHECK-NEXT: # %bb.1: # %bb4.preheader468; CHECK-NEXT: std r30, -16(r1) # 8-byte Folded Spill469; CHECK-NEXT: addi r6, r3, 4001470; CHECK-NEXT: addi r4, r4, 4001471; CHECK-NEXT: li r3, 0472; CHECK-NEXT: li r7, 9473; CHECK-NEXT: mtctr r5474; CHECK-NEXT: .p2align 4475; CHECK-NEXT: .LBB5_2: # %bb4476; CHECK-NEXT: #477; CHECK-NEXT: ld r8, 0(r6)478; CHECK-NEXT: ldx r9, r6, r7479; CHECK-NEXT: ld r10, 4(r6)480; CHECK-NEXT: ld r11, 8(r6)481; CHECK-NEXT: addi r6, r6, 1482; CHECK-NEXT: mulld r8, r9, r8483; CHECK-NEXT: ld r12, 0(r4)484; CHECK-NEXT: ldx r0, r4, r7485; CHECK-NEXT: ld r30, 4(r4)486; CHECK-NEXT: ld r9, 8(r4)487; CHECK-NEXT: addi r4, r4, 1488; CHECK-NEXT: mulld r8, r8, r10489; CHECK-NEXT: mulld r8, r8, r11490; CHECK-NEXT: mulld r8, r8, r12491; CHECK-NEXT: mulld r8, r8, r0492; CHECK-NEXT: mulld r8, r8, r30493; CHECK-NEXT: maddld r3, r8, r9, r3494; CHECK-NEXT: bdnz .LBB5_2495; CHECK-NEXT: # %bb.3:496; CHECK-NEXT: ld r30, -16(r1) # 8-byte Folded Reload497; CHECK-NEXT: add r3, r3, r5498; CHECK-NEXT: blr499; CHECK-NEXT: .LBB5_4:500; CHECK-NEXT: addi r3, r5, 0501; CHECK-NEXT: blr502bb:503 %i = sext i32 %arg2 to i64504 %i3 = icmp eq i32 %arg2, 0505 br i1 %i3, label %bb43, label %bb4506 507bb4: ; preds = %bb4, %bb508 %i5 = phi i64 [ %i41, %bb4 ], [ 0, %bb ]509 %i6 = phi i64 [ %i40, %bb4 ], [ 0, %bb ]510 %i7 = getelementptr inbounds i8, ptr %arg, i64 %i5511 %i8 = getelementptr inbounds i8, ptr %i7, i64 4001512 %i10 = load i64, ptr %i8, align 8513 %i11 = getelementptr inbounds i8, ptr %i7, i64 4010514 %i13 = load i64, ptr %i11, align 8515 %i14 = getelementptr inbounds i8, ptr %i7, i64 4005516 %i16 = load i64, ptr %i14, align 8517 %i17 = getelementptr inbounds i8, ptr %i7, i64 4009518 %i19 = load i64, ptr %i17, align 8519 %i20 = getelementptr inbounds i8, ptr %arg1, i64 %i5520 %i21 = getelementptr inbounds i8, ptr %i20, i64 4001521 %i23 = load i64, ptr %i21, align 8522 %i24 = getelementptr inbounds i8, ptr %i20, i64 4010523 %i26 = load i64, ptr %i24, align 8524 %i27 = getelementptr inbounds i8, ptr %i20, i64 4005525 %i29 = load i64, ptr %i27, align 8526 %i30 = getelementptr inbounds i8, ptr %i20, i64 4009527 %i32 = load i64, ptr %i30, align 8528 %i33 = mul i64 %i13, %i10529 %i34 = mul i64 %i33, %i16530 %i35 = mul i64 %i34, %i19531 %i36 = mul i64 %i35, %i23532 %i37 = mul i64 %i36, %i26533 %i38 = mul i64 %i37, %i29534 %i39 = mul i64 %i38, %i32535 %i40 = add i64 %i39, %i6536 %i41 = add nuw i64 %i5, 1537 %i42 = icmp ult i64 %i41, %i538 br i1 %i42, label %bb4, label %bb43539 540bb43: ; preds = %bb4, %bb541 %i44 = phi i64 [ 0, %bb ], [ %i40, %bb4 ]542 %i45 = add i64 %i44, %i543 ret i64 %i45544}545 546; test_ds_cross_basic_blocks:547;extern char *arr;548;unsigned long foo(char *p, int count)549;{550; unsigned long i=0, res=0;551; int DISP1 = 4000;552; int DISP2 = 4001;553; int DISP3 = 4002;554; int DISP4 = 4003;555; int DISP5 = 4005;556; int DISP6 = 4009;557; unsigned long x1, x2, x3, x4, x5, x6;558; x1=x2=x3=x4=x5=x6=1;559; for (; i < count ; i++) {560; if (arr[i] % 3 == 1) {561; x1 += *(unsigned long *)(p + i + DISP1);562; x2 += *(unsigned long *)(p + i + DISP2);563; }564; else if (arr[i] % 3 == 2) {565; x3 += *(unsigned long *)(p + i + DISP3);566; x4 += *(unsigned long *)(p + i + DISP5);567; }568; else {569; x5 += *(unsigned long *)(p + i + DISP4);570; x6 += *(unsigned long *)(p + i + DISP6);571; }572; res += x1*x2*x3*x4*x5*x6;573; }574; return res;575;}576 577@arr = external local_unnamed_addr global ptr, align 8578 579define i64 @test_ds_cross_basic_blocks(ptr %arg, i32 signext %arg1) {580; CHECK-LABEL: test_ds_cross_basic_blocks:581; CHECK: # %bb.0: # %bb582; CHECK-NEXT: cmplwi r4, 0583; CHECK-NEXT: beq cr0, .LBB6_9584; CHECK-NEXT: # %bb.1: # %bb3585; CHECK-NEXT: addis r5, r2, .LC0@toc@ha586; CHECK-NEXT: std r28, -32(r1) # 8-byte Folded Spill587; CHECK-NEXT: std r29, -24(r1) # 8-byte Folded Spill588; CHECK-NEXT: ld r5, .LC0@toc@l(r5)589; CHECK-NEXT: std r30, -16(r1) # 8-byte Folded Spill590; CHECK-NEXT: addi r6, r3, 4009591; CHECK-NEXT: li r3, 0592; CHECK-NEXT: li r7, -7593; CHECK-NEXT: li r8, -6594; CHECK-NEXT: li r9, 1595; CHECK-NEXT: li r10, 1596; CHECK-NEXT: li r11, 1597; CHECK-NEXT: li r12, 1598; CHECK-NEXT: li r0, 1599; CHECK-NEXT: ld r5, 0(r5)600; CHECK-NEXT: mtctr r4601; CHECK-NEXT: li r4, 1602; CHECK-NEXT: addi r5, r5, -1603; CHECK-NEXT: b .LBB6_4604; CHECK-NEXT: .p2align 4605; CHECK-NEXT: .LBB6_2: # %bb18606; CHECK-NEXT: #607; CHECK-NEXT: addi r30, r6, -9608; CHECK-NEXT: ld r30, 0(r30)609; CHECK-NEXT: add r0, r30, r0610; CHECK-NEXT: ld r30, -8(r6)611; CHECK-NEXT: add r12, r30, r12612; CHECK-NEXT: .LBB6_3: # %bb49613; CHECK-NEXT: #614; CHECK-NEXT: mulld r30, r12, r0615; CHECK-NEXT: addi r6, r6, 1616; CHECK-NEXT: mulld r30, r30, r11617; CHECK-NEXT: mulld r30, r30, r10618; CHECK-NEXT: mulld r30, r30, r9619; CHECK-NEXT: maddld r3, r30, r4, r3620; CHECK-NEXT: bdz .LBB6_8621; CHECK-NEXT: .LBB6_4: # %bb5622; CHECK-NEXT: #623; CHECK-NEXT: lbzu r30, 1(r5)624; CHECK-NEXT: mulli r29, r30, 171625; CHECK-NEXT: rlwinm r28, r29, 24, 8, 30626; CHECK-NEXT: srwi r29, r29, 9627; CHECK-NEXT: add r29, r29, r28628; CHECK-NEXT: sub r30, r30, r29629; CHECK-NEXT: clrlwi r30, r30, 24630; CHECK-NEXT: cmplwi r30, 1631; CHECK-NEXT: beq cr0, .LBB6_2632; CHECK-NEXT: # %bb.5: # %bb28633; CHECK-NEXT: #634; CHECK-NEXT: cmplwi r30, 2635; CHECK-NEXT: bne cr0, .LBB6_7636; CHECK-NEXT: # %bb.6: # %bb31637; CHECK-NEXT: #638; CHECK-NEXT: ldx r30, r6, r7639; CHECK-NEXT: add r11, r30, r11640; CHECK-NEXT: ld r30, -4(r6)641; CHECK-NEXT: add r10, r30, r10642; CHECK-NEXT: b .LBB6_3643; CHECK-NEXT: .p2align 4644; CHECK-NEXT: .LBB6_7: # %bb40645; CHECK-NEXT: #646; CHECK-NEXT: ldx r30, r6, r8647; CHECK-NEXT: add r9, r30, r9648; CHECK-NEXT: ld r30, 0(r6)649; CHECK-NEXT: add r4, r30, r4650; CHECK-NEXT: b .LBB6_3651; CHECK-NEXT: .LBB6_8:652; CHECK-NEXT: ld r30, -16(r1) # 8-byte Folded Reload653; CHECK-NEXT: ld r29, -24(r1) # 8-byte Folded Reload654; CHECK-NEXT: ld r28, -32(r1) # 8-byte Folded Reload655; CHECK-NEXT: blr656; CHECK-NEXT: .LBB6_9:657; CHECK-NEXT: li r3, 0658; CHECK-NEXT: blr659bb:660 %i = sext i32 %arg1 to i64661 %i2 = icmp eq i32 %arg1, 0662 br i1 %i2, label %bb64, label %bb3663 664bb3: ; preds = %bb665 %i4 = load ptr, ptr @arr, align 8666 br label %bb5667 668bb5: ; preds = %bb49, %bb3669 %i6 = phi i64 [ 1, %bb3 ], [ %i55, %bb49 ]670 %i7 = phi i64 [ 1, %bb3 ], [ %i54, %bb49 ]671 %i8 = phi i64 [ 1, %bb3 ], [ %i53, %bb49 ]672 %i9 = phi i64 [ 1, %bb3 ], [ %i52, %bb49 ]673 %i10 = phi i64 [ 1, %bb3 ], [ %i51, %bb49 ]674 %i11 = phi i64 [ 1, %bb3 ], [ %i50, %bb49 ]675 %i12 = phi i64 [ 0, %bb3 ], [ %i62, %bb49 ]676 %i13 = phi i64 [ 0, %bb3 ], [ %i61, %bb49 ]677 %i14 = getelementptr inbounds i8, ptr %i4, i64 %i12678 %i15 = load i8, ptr %i14, align 1679 %i16 = urem i8 %i15, 3680 %i17 = icmp eq i8 %i16, 1681 br i1 %i17, label %bb18, label %bb28682 683bb18: ; preds = %bb5684 %i19 = getelementptr inbounds i8, ptr %arg, i64 %i12685 %i20 = getelementptr inbounds i8, ptr %i19, i64 4000686 %i22 = load i64, ptr %i20, align 8687 %i23 = add i64 %i22, %i11688 %i24 = getelementptr inbounds i8, ptr %i19, i64 4001689 %i26 = load i64, ptr %i24, align 8690 %i27 = add i64 %i26, %i10691 br label %bb49692 693bb28: ; preds = %bb5694 %i29 = icmp eq i8 %i16, 2695 %i30 = getelementptr inbounds i8, ptr %arg, i64 %i12696 br i1 %i29, label %bb31, label %bb40697 698bb31: ; preds = %bb28699 %i32 = getelementptr inbounds i8, ptr %i30, i64 4002700 %i34 = load i64, ptr %i32, align 8701 %i35 = add i64 %i34, %i9702 %i36 = getelementptr inbounds i8, ptr %i30, i64 4005703 %i38 = load i64, ptr %i36, align 8704 %i39 = add i64 %i38, %i8705 br label %bb49706 707bb40: ; preds = %bb28708 %i41 = getelementptr inbounds i8, ptr %i30, i64 4003709 %i43 = load i64, ptr %i41, align 8710 %i44 = add i64 %i43, %i7711 %i45 = getelementptr inbounds i8, ptr %i30, i64 4009712 %i47 = load i64, ptr %i45, align 8713 %i48 = add i64 %i47, %i6714 br label %bb49715 716bb49: ; preds = %bb40, %bb31, %bb18717 %i50 = phi i64 [ %i23, %bb18 ], [ %i11, %bb31 ], [ %i11, %bb40 ]718 %i51 = phi i64 [ %i27, %bb18 ], [ %i10, %bb31 ], [ %i10, %bb40 ]719 %i52 = phi i64 [ %i9, %bb18 ], [ %i35, %bb31 ], [ %i9, %bb40 ]720 %i53 = phi i64 [ %i8, %bb18 ], [ %i39, %bb31 ], [ %i8, %bb40 ]721 %i54 = phi i64 [ %i7, %bb18 ], [ %i7, %bb31 ], [ %i44, %bb40 ]722 %i55 = phi i64 [ %i6, %bb18 ], [ %i6, %bb31 ], [ %i48, %bb40 ]723 %i56 = mul i64 %i51, %i50724 %i57 = mul i64 %i56, %i52725 %i58 = mul i64 %i57, %i53726 %i59 = mul i64 %i58, %i54727 %i60 = mul i64 %i59, %i55728 %i61 = add i64 %i60, %i13729 %i62 = add nuw i64 %i12, 1730 %i63 = icmp ult i64 %i62, %i731 br i1 %i63, label %bb5, label %bb64732 733bb64: ; preds = %bb49, %bb734 %i65 = phi i64 [ 0, %bb ], [ %i61, %bb49 ]735 ret i64 %i65736}737 738; test_ds_float:739;float test_ds_float(char *p, int count) {740; int i=0 ;741; float res=0;742; int DISP1 = 4001;743; int DISP2 = 4002;744; int DISP3 = 4022;745; int DISP4 = 4062;746; for (; i < count ; i++) {747; float x1 = *(float *)(p + i + DISP1);748; float x2 = *(float *)(p + i + DISP2);749; float x3 = *(float *)(p + i + DISP3);750; float x4 = *(float *)(p + i + DISP4);751; res += x1*x2*x3*x4;752; }753; return res;754;}755 756define float @test_ds_float(ptr %arg, i32 signext %arg1) {757; CHECK-LABEL: test_ds_float:758; CHECK: # %bb.0: # %bb759; CHECK-NEXT: cmpwi r4, 0760; CHECK-NEXT: ble cr0, .LBB7_4761; CHECK-NEXT: # %bb.1: # %bb2762; CHECK-NEXT: clrldi r4, r4, 32763; CHECK-NEXT: addi r3, r3, 4002764; CHECK-NEXT: xxlxor f1, f1, f1765; CHECK-NEXT: mtctr r4766; CHECK-NEXT: li r4, -1767; CHECK-NEXT: .p2align 4768; CHECK-NEXT: .LBB7_2: # %bb4769; CHECK-NEXT: #770; CHECK-NEXT: lfsx f0, r3, r4771; CHECK-NEXT: lfs f2, 0(r3)772; CHECK-NEXT: xsmulsp f0, f0, f2773; CHECK-NEXT: lfs f3, 20(r3)774; CHECK-NEXT: xsmulsp f0, f0, f3775; CHECK-NEXT: lfs f4, 60(r3)776; CHECK-NEXT: addi r3, r3, 1777; CHECK-NEXT: xsmulsp f0, f0, f4778; CHECK-NEXT: xsaddsp f1, f1, f0779; CHECK-NEXT: bdnz .LBB7_2780; CHECK-NEXT: # %bb.3: # %bb26781; CHECK-NEXT: blr782; CHECK-NEXT: .LBB7_4:783; CHECK-NEXT: xxlxor f1, f1, f1784; CHECK-NEXT: blr785bb:786 %i = icmp sgt i32 %arg1, 0787 br i1 %i, label %bb2, label %bb26788 789bb2: ; preds = %bb790 %i3 = zext i32 %arg1 to i64791 br label %bb4792 793bb4: ; preds = %bb4, %bb2794 %i5 = phi i64 [ 0, %bb2 ], [ %i24, %bb4 ]795 %i6 = phi float [ 0.000000e+00, %bb2 ], [ %i23, %bb4 ]796 %i7 = getelementptr inbounds i8, ptr %arg, i64 %i5797 %i8 = getelementptr inbounds i8, ptr %i7, i64 4001798 %i10 = load float, ptr %i8, align 4799 %i11 = getelementptr inbounds i8, ptr %i7, i64 4002800 %i13 = load float, ptr %i11, align 4801 %i14 = getelementptr inbounds i8, ptr %i7, i64 4022802 %i16 = load float, ptr %i14, align 4803 %i17 = getelementptr inbounds i8, ptr %i7, i64 4062804 %i19 = load float, ptr %i17, align 4805 %i20 = fmul float %i10, %i13806 %i21 = fmul float %i20, %i16807 %i22 = fmul float %i21, %i19808 %i23 = fadd float %i6, %i22809 %i24 = add nuw nsw i64 %i5, 1810 %i25 = icmp eq i64 %i24, %i3811 br i1 %i25, label %bb26, label %bb4812 813bb26: ; preds = %bb4, %bb814 %i27 = phi float [ 0.000000e+00, %bb ], [ %i23, %bb4 ]815 ret float %i27816}817 818; test_ds_combine_float_int:819;float test_ds_combine_float_int(char *p, int count) {820; int i=0 ;821; float res=0;822; int DISP1 = 4001;823; int DISP2 = 4002;824; int DISP3 = 4022;825; int DISP4 = 4062;826; for (; i < count ; i++) {827; float x1 = *(float *)(p + i + DISP1);828; unsigned long x2 = *(unsigned long*)(p + i + DISP2);829; float x3 = *(float *)(p + i + DISP3);830; float x4 = *(float *)(p + i + DISP4);831; res += x1*x2*x3*x4;832; }833; return res;834;}835 836define float @test_ds_combine_float_int(ptr %arg, i32 signext %arg1) {837; CHECK-LABEL: test_ds_combine_float_int:838; CHECK: # %bb.0: # %bb839; CHECK-NEXT: cmpwi r4, 0840; CHECK-NEXT: ble cr0, .LBB8_4841; CHECK-NEXT: # %bb.1: # %bb2842; CHECK-NEXT: clrldi r4, r4, 32843; CHECK-NEXT: addi r3, r3, 4002844; CHECK-NEXT: xxlxor f1, f1, f1845; CHECK-NEXT: mtctr r4846; CHECK-NEXT: li r4, -1847; CHECK-NEXT: .p2align 4848; CHECK-NEXT: .LBB8_2: # %bb4849; CHECK-NEXT: #850; CHECK-NEXT: lfd f4, 0(r3)851; CHECK-NEXT: lfsx f0, r3, r4852; CHECK-NEXT: xscvuxdsp f4, f4853; CHECK-NEXT: lfs f2, 20(r3)854; CHECK-NEXT: lfs f3, 60(r3)855; CHECK-NEXT: addi r3, r3, 1856; CHECK-NEXT: xsmulsp f0, f0, f4857; CHECK-NEXT: xsmulsp f0, f2, f0858; CHECK-NEXT: xsmulsp f0, f3, f0859; CHECK-NEXT: xsaddsp f1, f1, f0860; CHECK-NEXT: bdnz .LBB8_2861; CHECK-NEXT: # %bb.3: # %bb27862; CHECK-NEXT: blr863; CHECK-NEXT: .LBB8_4:864; CHECK-NEXT: xxlxor f1, f1, f1865; CHECK-NEXT: blr866bb:867 %i = icmp sgt i32 %arg1, 0868 br i1 %i, label %bb2, label %bb27869 870bb2: ; preds = %bb871 %i3 = zext i32 %arg1 to i64872 br label %bb4873 874bb4: ; preds = %bb4, %bb2875 %i5 = phi i64 [ 0, %bb2 ], [ %i25, %bb4 ]876 %i6 = phi float [ 0.000000e+00, %bb2 ], [ %i24, %bb4 ]877 %i7 = getelementptr inbounds i8, ptr %arg, i64 %i5878 %i8 = getelementptr inbounds i8, ptr %i7, i64 4001879 %i10 = load float, ptr %i8, align 4880 %i11 = getelementptr inbounds i8, ptr %i7, i64 4002881 %i13 = load i64, ptr %i11, align 8882 %i14 = getelementptr inbounds i8, ptr %i7, i64 4022883 %i16 = load float, ptr %i14, align 4884 %i17 = getelementptr inbounds i8, ptr %i7, i64 4062885 %i19 = load float, ptr %i17, align 4886 %i20 = uitofp i64 %i13 to float887 %i21 = fmul float %i10, %i20888 %i22 = fmul float %i16, %i21889 %i23 = fmul float %i19, %i22890 %i24 = fadd float %i6, %i23891 %i25 = add nuw nsw i64 %i5, 1892 %i26 = icmp eq i64 %i25, %i3893 br i1 %i26, label %bb27, label %bb4894 895bb27: ; preds = %bb4, %bb896 %i28 = phi float [ 0.000000e+00, %bb ], [ %i24, %bb4 ]897 ret float %i28898}899 900; test_ds_lwa_prep:901; long long test_ds_lwa_prep(char *p, int count) {902; long long i=0, res=0;903; int DISP1 = 4001;904; int DISP2 = 4002;905; int DISP3 = 4006;906; int DISP4 = 4010;907; for (; i < count ; i++) {908; long long x1 = *(int *)(p + i + DISP1);909; long long x2 = *(int *)(p + i + DISP2);910; long long x3 = *(int *)(p + i + DISP3);911; long long x4 = *(int *)(p + i + DISP4);912; res += x1*x2*x3*x4;913; }914; return res + count;915; }916 917define i64 @test_ds_lwa_prep(ptr %arg, i32 signext %arg1) {918; CHECK-LABEL: test_ds_lwa_prep:919; CHECK: # %bb.0: # %bb920; CHECK-NEXT: cmpwi r4, 0921; CHECK-NEXT: ble cr0, .LBB9_4922; CHECK-NEXT: # %bb.1: # %bb3.preheader923; CHECK-NEXT: mtctr r4924; CHECK-NEXT: addi r5, r3, 2925; CHECK-NEXT: li r3, 0926; CHECK-NEXT: li r6, -1927; CHECK-NEXT: .p2align 4928; CHECK-NEXT: .LBB9_2: # %bb3929; CHECK-NEXT: #930; CHECK-NEXT: lwax r7, r5, r6931; CHECK-NEXT: lwa r8, 0(r5)932; CHECK-NEXT: lwa r9, 4(r5)933; CHECK-NEXT: lwa r10, 8(r5)934; CHECK-NEXT: addi r5, r5, 1935; CHECK-NEXT: mulld r7, r8, r7936; CHECK-NEXT: mulld r7, r7, r9937; CHECK-NEXT: maddld r3, r7, r10, r3938; CHECK-NEXT: bdnz .LBB9_2939; CHECK-NEXT: # %bb.3: # %bb29940; CHECK-NEXT: add r3, r3, r4941; CHECK-NEXT: blr942; CHECK-NEXT: .LBB9_4:943; CHECK-NEXT: addi r3, r4, 0944; CHECK-NEXT: blr945bb:946 %i = sext i32 %arg1 to i64947 %i2 = icmp sgt i32 %arg1, 0948 br i1 %i2, label %bb3, label %bb29949 950bb3: ; preds = %bb3, %bb951 %i4 = phi i64 [ %i27, %bb3 ], [ 0, %bb ]952 %i5 = phi i64 [ %i26, %bb3 ], [ 0, %bb ]953 %i6 = getelementptr inbounds i8, ptr %arg, i64 %i4954 %i7 = getelementptr inbounds i8, ptr %i6, i64 1955 %i9 = load i32, ptr %i7, align 4956 %i10 = sext i32 %i9 to i64957 %i11 = getelementptr inbounds i8, ptr %i6, i64 2958 %i13 = load i32, ptr %i11, align 4959 %i14 = sext i32 %i13 to i64960 %i15 = getelementptr inbounds i8, ptr %i6, i64 6961 %i17 = load i32, ptr %i15, align 4962 %i18 = sext i32 %i17 to i64963 %i19 = getelementptr inbounds i8, ptr %i6, i64 10964 %i21 = load i32, ptr %i19, align 4965 %i22 = sext i32 %i21 to i64966 %i23 = mul nsw i64 %i14, %i10967 %i24 = mul nsw i64 %i23, %i18968 %i25 = mul nsw i64 %i24, %i22969 %i26 = add nsw i64 %i25, %i5970 %i27 = add nuw nsw i64 %i4, 1971 %i28 = icmp eq i64 %i27, %i972 br i1 %i28, label %bb29, label %bb3973 974bb29: ; preds = %bb3, %bb975 %i30 = phi i64 [ 0, %bb ], [ %i26, %bb3 ]976 %i31 = add nsw i64 %i30, %i977 ret i64 %i31978}979 980