1183 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -ppc-asm-full-reg-names -verify-machineinstrs -ppc-formprep-chain-commoning \3; RUN: -mtriple=powerpc64le-unknown-linux-gnu -mcpu=pwr9 < %s | FileCheck %s4 5; addresses:6; 1: base1 + offset7; 2: + offset8; 3: + offset9; 4: + offset10;11; chains:12; 1: base: base1 + offset, offsets: (0, offset)13; 2: base: base1 + 3*offset, offsets: (0, offset)14;15; long long two_chain_same_offset_succ(char *p, long long offset, long long base1, long long n) {16; long long o1 = base1 + offset;17; long long o2 = base1 + 2 * offset;18; long long o3 = base1 + 3 * offset;19; long long o4 = base1 + 4 * offset;20; char *p1 = p + o1;21; char *p2 = p + o2;22; char *p3 = p + o3;23; char *p4 = p + o4;24; long long sum = 0;25; for (long long i = 0; i < n; ++i) {26; unsigned long x1 = *(unsigned long *)(p1 + i);27; unsigned long x2 = *(unsigned long *)(p2 + i);28; unsigned long x3 = *(unsigned long *)(p3 + i);29; unsigned long x4 = *(unsigned long *)(p4 + i);30; sum += x1 * x2 * x3 * x4;31; }32; return sum;33; }34;35define i64 @two_chain_same_offset_succ(ptr %p, i64 %offset, i64 %base1, i64 %n) {36; CHECK-LABEL: two_chain_same_offset_succ:37; CHECK: # %bb.0: # %entry38; CHECK-NEXT: cmpdi r6, 039; CHECK-NEXT: ble cr0, .LBB0_440; CHECK-NEXT: # %bb.1: # %for.body.preheader41; CHECK-NEXT: sldi r7, r4, 142; CHECK-NEXT: mtctr r643; CHECK-NEXT: add r8, r4, r744; CHECK-NEXT: add r7, r5, r445; CHECK-NEXT: add r5, r5, r846; CHECK-NEXT: add r7, r3, r747; CHECK-NEXT: add r5, r3, r548; CHECK-NEXT: li r3, 049; CHECK-NEXT: .p2align 450; CHECK-NEXT: .LBB0_2: # %for.body51; CHECK-NEXT: #52; CHECK-NEXT: ld r6, 0(r7)53; CHECK-NEXT: ldx r8, r7, r454; CHECK-NEXT: ld r9, 0(r5)55; CHECK-NEXT: ldx r10, r5, r456; CHECK-NEXT: addi r7, r7, 157; CHECK-NEXT: addi r5, r5, 158; CHECK-NEXT: mulld r6, r8, r659; CHECK-NEXT: mulld r6, r6, r960; CHECK-NEXT: maddld r3, r6, r10, r361; CHECK-NEXT: bdnz .LBB0_262; CHECK-NEXT: # %bb.3: # %for.cond.cleanup63; CHECK-NEXT: blr64; CHECK-NEXT: .LBB0_4:65; CHECK-NEXT: li r3, 066; CHECK-NEXT: blr67entry:68 %mul = shl nsw i64 %offset, 169 %mul2 = mul nsw i64 %offset, 370 %mul4 = shl nsw i64 %offset, 271 %cmp46 = icmp sgt i64 %n, 072 br i1 %cmp46, label %for.body, label %for.cond.cleanup73 74for.cond.cleanup: ; preds = %for.body, %entry75 %sum.0.lcssa = phi i64 [ 0, %entry ], [ %add16, %for.body ]76 ret i64 %sum.0.lcssa77 78for.body: ; preds = %entry, %for.body79 %sum.048 = phi i64 [ %add16, %for.body ], [ 0, %entry ]80 %i.047 = phi i64 [ %inc, %for.body ], [ 0, %entry ]81 %add = add i64 %i.047, %base182 %add.ptr9.idx = add i64 %add, %offset83 %add.ptr9 = getelementptr inbounds i8, ptr %p, i64 %add.ptr9.idx84 %0 = load i64, ptr %add.ptr9, align 885 %add.ptr10.idx = add i64 %add, %mul86 %add.ptr10 = getelementptr inbounds i8, ptr %p, i64 %add.ptr10.idx87 %1 = load i64, ptr %add.ptr10, align 888 %add.ptr11.idx = add i64 %add, %mul289 %add.ptr11 = getelementptr inbounds i8, ptr %p, i64 %add.ptr11.idx90 %2 = load i64, ptr %add.ptr11, align 891 %add.ptr12.idx = add i64 %add, %mul492 %add.ptr12 = getelementptr inbounds i8, ptr %p, i64 %add.ptr12.idx93 %3 = load i64, ptr %add.ptr12, align 894 %mul13 = mul i64 %1, %095 %mul14 = mul i64 %mul13, %296 %mul15 = mul i64 %mul14, %397 %add16 = add i64 %mul15, %sum.04898 %inc = add nuw nsw i64 %i.047, 199 %exitcond.not = icmp eq i64 %inc, %n100 br i1 %exitcond.not, label %for.cond.cleanup, label %for.body101}102 103; addresses:104; 1: base1 + offset105; 2: + offset106; 3: + offset107; 4: + offset108; 5: + offset109;110; It can not be commoned to chains because we need a chain for a single address.111; It is not profitable to common chains if not all addresses are in chains.112;113; long long not_perfect_chain_all_same_offset_fail(char *p, long long offset, long long base1, long long n) {114; long long o1 = base1 + offset;115; long long o2 = base1 + 2 * offset;116; long long o3 = base1 + 3 * offset;117; long long o4 = base1 + 4 * offset;118; long long o5 = base1 + 5 * offset;119; char *p1 = p + o1;120; char *p2 = p + o2;121; char *p3 = p + o3;122; char *p4 = p + o4;123; char *p5 = p + o5;124; long long sum = 0;125; for (long long i = 0; i < n; ++i) {126; unsigned long x1 = *(unsigned long *)(p1 + i);127; unsigned long x2 = *(unsigned long *)(p2 + i);128; unsigned long x3 = *(unsigned long *)(p3 + i);129; unsigned long x4 = *(unsigned long *)(p4 + i);130; unsigned long x5 = *(unsigned long *)(p5 + i);131; sum += x1 * x2 * x3 * x4 * x5;132; }133; return sum;134; }135;136define i64 @not_perfect_chain_all_same_offset_fail(ptr %p, i64 %offset, i64 %base1, i64 %n) {137; CHECK-LABEL: not_perfect_chain_all_same_offset_fail:138; CHECK: # %bb.0: # %entry139; CHECK-NEXT: cmpdi r6, 0140; CHECK-NEXT: ble cr0, .LBB1_4141; CHECK-NEXT: # %bb.1: # %for.body.preheader142; CHECK-NEXT: std r30, -16(r1) # 8-byte Folded Spill143; CHECK-NEXT: sldi r7, r4, 1144; CHECK-NEXT: add r5, r3, r5145; CHECK-NEXT: li r3, 0146; CHECK-NEXT: add r8, r4, r7147; CHECK-NEXT: sldi r9, r4, 2148; CHECK-NEXT: mtctr r6149; CHECK-NEXT: add r10, r4, r9150; CHECK-NEXT: .p2align 4151; CHECK-NEXT: .LBB1_2: # %for.body152; CHECK-NEXT: #153; CHECK-NEXT: ldx r6, r5, r4154; CHECK-NEXT: ldx r11, r5, r7155; CHECK-NEXT: ldx r12, r5, r8156; CHECK-NEXT: ldx r0, r5, r9157; CHECK-NEXT: mulld r6, r11, r6158; CHECK-NEXT: ldx r30, r5, r10159; CHECK-NEXT: addi r5, r5, 1160; CHECK-NEXT: mulld r6, r6, r12161; CHECK-NEXT: mulld r6, r6, r0162; CHECK-NEXT: maddld r3, r6, r30, r3163; CHECK-NEXT: bdnz .LBB1_2164; CHECK-NEXT: # %bb.3:165; CHECK-NEXT: ld r30, -16(r1) # 8-byte Folded Reload166; CHECK-NEXT: blr167; CHECK-NEXT: .LBB1_4:168; CHECK-NEXT: li r3, 0169; CHECK-NEXT: blr170entry:171 %mul = shl nsw i64 %offset, 1172 %mul2 = mul nsw i64 %offset, 3173 %mul4 = shl nsw i64 %offset, 2174 %mul6 = mul nsw i64 %offset, 5175 %cmp58 = icmp sgt i64 %n, 0176 br i1 %cmp58, label %for.body, label %for.cond.cleanup177 178for.cond.cleanup: ; preds = %for.body, %entry179 %sum.0.lcssa = phi i64 [ 0, %entry ], [ %add21, %for.body ]180 ret i64 %sum.0.lcssa181 182for.body: ; preds = %entry, %for.body183 %sum.060 = phi i64 [ %add21, %for.body ], [ 0, %entry ]184 %i.059 = phi i64 [ %inc, %for.body ], [ 0, %entry ]185 %add = add i64 %i.059, %base1186 %add.ptr12.idx = add i64 %add, %offset187 %add.ptr12 = getelementptr inbounds i8, ptr %p, i64 %add.ptr12.idx188 %0 = load i64, ptr %add.ptr12, align 8189 %add.ptr13.idx = add i64 %add, %mul190 %add.ptr13 = getelementptr inbounds i8, ptr %p, i64 %add.ptr13.idx191 %1 = load i64, ptr %add.ptr13, align 8192 %add.ptr14.idx = add i64 %add, %mul2193 %add.ptr14 = getelementptr inbounds i8, ptr %p, i64 %add.ptr14.idx194 %2 = load i64, ptr %add.ptr14, align 8195 %add.ptr15.idx = add i64 %add, %mul4196 %add.ptr15 = getelementptr inbounds i8, ptr %p, i64 %add.ptr15.idx197 %3 = load i64, ptr %add.ptr15, align 8198 %add.ptr16.idx = add i64 %add, %mul6199 %add.ptr16 = getelementptr inbounds i8, ptr %p, i64 %add.ptr16.idx200 %4 = load i64, ptr %add.ptr16, align 8201 %mul17 = mul i64 %1, %0202 %mul18 = mul i64 %mul17, %2203 %mul19 = mul i64 %mul18, %3204 %mul20 = mul i64 %mul19, %4205 %add21 = add i64 %mul20, %sum.060206 %inc = add nuw nsw i64 %i.059, 1207 %exitcond.not = icmp eq i64 %inc, %n208 br i1 %exitcond.not, label %for.cond.cleanup, label %for.body209}210 211; addresses:212; 1: base1213; 2: + 2*offset214; 3: + offset215;216; We need at least 4 addresses to common 2 chains to reuse at least 1 offset.217;218; long long no_enough_elements_fail(char *p, long long offset, long long base1, long long n) {219; long long o1 = base1;220; long long o2 = base1 + 2 * offset;221; long long o3 = base1 + 3 * offset;222; char *p1 = p + o1;223; char *p2 = p + o2;224; char *p3 = p + o3;225; long long sum = 0;226; for (long long i = 0; i < n; ++i) {227; unsigned long x1 = *(unsigned long *)(p1 + i);228; unsigned long x2 = *(unsigned long *)(p2 + i);229; unsigned long x3 = *(unsigned long *)(p3 + i);230; sum += x1 * x2 * x3;231; }232; return sum;233; }234;235define i64 @no_enough_elements_fail(ptr %p, i64 %offset, i64 %base1, i64 %n) {236; CHECK-LABEL: no_enough_elements_fail:237; CHECK: # %bb.0: # %entry238; CHECK-NEXT: cmpdi r6, 0239; CHECK-NEXT: ble cr0, .LBB2_4240; CHECK-NEXT: # %bb.1: # %for.body.preheader241; CHECK-NEXT: sldi r7, r4, 1242; CHECK-NEXT: mtctr r6243; CHECK-NEXT: add r5, r3, r5244; CHECK-NEXT: li r3, 0245; CHECK-NEXT: add r4, r4, r7246; CHECK-NEXT: .p2align 5247; CHECK-NEXT: .LBB2_2: # %for.body248; CHECK-NEXT: #249; CHECK-NEXT: ld r6, 0(r5)250; CHECK-NEXT: ldx r8, r5, r7251; CHECK-NEXT: ldx r9, r5, r4252; CHECK-NEXT: addi r5, r5, 1253; CHECK-NEXT: mulld r6, r8, r6254; CHECK-NEXT: maddld r3, r6, r9, r3255; CHECK-NEXT: bdnz .LBB2_2256; CHECK-NEXT: # %bb.3: # %for.cond.cleanup257; CHECK-NEXT: blr258; CHECK-NEXT: .LBB2_4:259; CHECK-NEXT: li r3, 0260; CHECK-NEXT: blr261entry:262 %mul = shl nsw i64 %offset, 1263 %mul1 = mul nsw i64 %offset, 3264 %cmp32 = icmp sgt i64 %n, 0265 br i1 %cmp32, label %for.body, label %for.cond.cleanup266 267for.cond.cleanup: ; preds = %for.body, %entry268 %sum.0.lcssa = phi i64 [ 0, %entry ], [ %add10, %for.body ]269 ret i64 %sum.0.lcssa270 271for.body: ; preds = %entry, %for.body272 %sum.034 = phi i64 [ %add10, %for.body ], [ 0, %entry ]273 %i.033 = phi i64 [ %inc, %for.body ], [ 0, %entry ]274 %add.ptr5.idx = add i64 %i.033, %base1275 %add.ptr5 = getelementptr inbounds i8, ptr %p, i64 %add.ptr5.idx276 %0 = load i64, ptr %add.ptr5, align 8277 %add.ptr6.idx = add i64 %add.ptr5.idx, %mul278 %add.ptr6 = getelementptr inbounds i8, ptr %p, i64 %add.ptr6.idx279 %1 = load i64, ptr %add.ptr6, align 8280 %add.ptr7.idx = add i64 %add.ptr5.idx, %mul1281 %add.ptr7 = getelementptr inbounds i8, ptr %p, i64 %add.ptr7.idx282 %2 = load i64, ptr %add.ptr7, align 8283 %mul8 = mul i64 %1, %0284 %mul9 = mul i64 %mul8, %2285 %add10 = add i64 %mul9, %sum.034286 %inc = add nuw nsw i64 %i.033, 1287 %exitcond.not = icmp eq i64 %inc, %n288 br i1 %exitcond.not, label %for.cond.cleanup, label %for.body289}290 291; addresses:292; 1: base1293; 2: + 2*offset294; 3: + 2*offset295; 4: + 3*offset296;297; The diff between address 2 and address 1 is 2*offset, and this offset is not reused among other chains,298; so we can not common any chains.299;300; long long no_reuseable_offset_fail(char *p, long long offset, long long base1, long long n) {301; long long o1 = base1;302; long long o2 = base1 + 2 * offset;303; long long o3 = base1 + 4 * offset;304; long long o4 = base1 + 7 * offset;305; char *p1 = p + o1;306; char *p2 = p + o2;307; char *p3 = p + o3;308; char *p4 = p + o4;309; long long sum = 0;310; for (long long i = 0; i < n; ++i) {311; unsigned long x1 = *(unsigned long *)(p1 + i);312; unsigned long x2 = *(unsigned long *)(p2 + i);313; unsigned long x3 = *(unsigned long *)(p3 + i);314; unsigned long x4 = *(unsigned long *)(p4 + i);315; sum += x1 * x2 * x3 * x4;316; }317; return sum;318; }319;320define i64 @no_reuseable_offset_fail(ptr %p, i64 %offset, i64 %base1, i64 %n) {321; CHECK-LABEL: no_reuseable_offset_fail:322; CHECK: # %bb.0: # %entry323; CHECK-NEXT: cmpdi r6, 0324; CHECK-NEXT: ble cr0, .LBB3_4325; CHECK-NEXT: # %bb.1: # %for.body.preheader326; CHECK-NEXT: sldi r9, r4, 3327; CHECK-NEXT: mtctr r6328; CHECK-NEXT: add r5, r3, r5329; CHECK-NEXT: li r3, 0330; CHECK-NEXT: sldi r7, r4, 1331; CHECK-NEXT: sldi r8, r4, 2332; CHECK-NEXT: sub r4, r9, r4333; CHECK-NEXT: .p2align 4334; CHECK-NEXT: .LBB3_2: # %for.body335; CHECK-NEXT: #336; CHECK-NEXT: ld r6, 0(r5)337; CHECK-NEXT: ldx r9, r5, r7338; CHECK-NEXT: ldx r10, r5, r8339; CHECK-NEXT: ldx r11, r5, r4340; CHECK-NEXT: addi r5, r5, 1341; CHECK-NEXT: mulld r6, r9, r6342; CHECK-NEXT: mulld r6, r6, r10343; CHECK-NEXT: maddld r3, r6, r11, r3344; CHECK-NEXT: bdnz .LBB3_2345; CHECK-NEXT: # %bb.3: # %for.cond.cleanup346; CHECK-NEXT: blr347; CHECK-NEXT: .LBB3_4:348; CHECK-NEXT: li r3, 0349; CHECK-NEXT: blr350entry:351 %mul = shl nsw i64 %offset, 1352 %mul1 = shl nsw i64 %offset, 2353 %mul3 = mul nsw i64 %offset, 7354 %cmp44 = icmp sgt i64 %n, 0355 br i1 %cmp44, label %for.body, label %for.cond.cleanup356 357for.cond.cleanup: ; preds = %for.body, %entry358 %sum.0.lcssa = phi i64 [ 0, %entry ], [ %add15, %for.body ]359 ret i64 %sum.0.lcssa360 361for.body: ; preds = %entry, %for.body362 %sum.046 = phi i64 [ %add15, %for.body ], [ 0, %entry ]363 %i.045 = phi i64 [ %inc, %for.body ], [ 0, %entry ]364 %add.ptr8.idx = add i64 %i.045, %base1365 %add.ptr8 = getelementptr inbounds i8, ptr %p, i64 %add.ptr8.idx366 %0 = load i64, ptr %add.ptr8, align 8367 %add.ptr9.idx = add i64 %add.ptr8.idx, %mul368 %add.ptr9 = getelementptr inbounds i8, ptr %p, i64 %add.ptr9.idx369 %1 = load i64, ptr %add.ptr9, align 8370 %add.ptr10.idx = add i64 %add.ptr8.idx, %mul1371 %add.ptr10 = getelementptr inbounds i8, ptr %p, i64 %add.ptr10.idx372 %2 = load i64, ptr %add.ptr10, align 8373 %add.ptr11.idx = add i64 %add.ptr8.idx, %mul3374 %add.ptr11 = getelementptr inbounds i8, ptr %p, i64 %add.ptr11.idx375 %3 = load i64, ptr %add.ptr11, align 8376 %mul12 = mul i64 %1, %0377 %mul13 = mul i64 %mul12, %2378 %mul14 = mul i64 %mul13, %3379 %add15 = add i64 %mul14, %sum.046380 %inc = add nuw nsw i64 %i.045, 1381 %exitcond.not = icmp eq i64 %inc, %n382 br i1 %exitcond.not, label %for.cond.cleanup, label %for.body383}384 385; addresses:386; 1: base1 + offset387; 2: + offset388; 3: + 3*offset389; 4: + 2*offset390; 5: + 1*offset391; 6: + 2*offset392;393; The diff between address 2 and address 1 is 1*offset, and this offset is reused between address 4 and address 5.394; but the diff between address 3 and address 2 (3*offset) is not the same with the diff between address 6395; and address 5(2*offset), so we can not common chains for these addresses.396;397; long long not_same_offset_fail(char *p, long long offset, long long base1, long long n) {398; long long o1 = base1 + offset;399; long long o2 = base1 + 2 * offset;400; long long o3 = base1 + 5 * offset;401; long long o4 = base1 + 7 * offset;402; long long o5 = base1 + 8 * offset;403; long long o6 = base1 + 10 * offset;404; char *p1 = p + o1;405; char *p2 = p + o2;406; char *p3 = p + o3;407; char *p4 = p + o4;408; char *p5 = p + o5;409; char *p6 = p + o6;410; long long sum = 0;411; for (long long i = 0; i < n; ++i) {412; unsigned long x1 = *(unsigned long *)(p1 + i);413; unsigned long x2 = *(unsigned long *)(p2 + i);414; unsigned long x3 = *(unsigned long *)(p3 + i);415; unsigned long x4 = *(unsigned long *)(p4 + i);416; unsigned long x5 = *(unsigned long *)(p5 + i);417; unsigned long x6 = *(unsigned long *)(p6 + i);418; sum += x1 * x2 * x3 * x4 * x5 * x6;419; }420; return sum;421; }422;423define i64 @not_same_offset_fail(ptr %p, i64 %offset, i64 %base1, i64 %n) {424; CHECK-LABEL: not_same_offset_fail:425; CHECK: # %bb.0: # %entry426; CHECK-NEXT: cmpdi r6, 0427; CHECK-NEXT: ble cr0, .LBB4_4428; CHECK-NEXT: # %bb.1: # %for.body.preheader429; CHECK-NEXT: std r28, -32(r1) # 8-byte Folded Spill430; CHECK-NEXT: std r29, -24(r1) # 8-byte Folded Spill431; CHECK-NEXT: add r5, r3, r5432; CHECK-NEXT: li r3, 0433; CHECK-NEXT: std r30, -16(r1) # 8-byte Folded Spill434; CHECK-NEXT: mtctr r6435; CHECK-NEXT: mulli r11, r4, 10436; CHECK-NEXT: sldi r8, r4, 2437; CHECK-NEXT: add r8, r4, r8438; CHECK-NEXT: sldi r9, r4, 3439; CHECK-NEXT: sub r10, r9, r4440; CHECK-NEXT: sldi r7, r4, 1441; CHECK-NEXT: .p2align 4442; CHECK-NEXT: .LBB4_2: # %for.body443; CHECK-NEXT: #444; CHECK-NEXT: ldx r6, r5, r4445; CHECK-NEXT: ldx r12, r5, r7446; CHECK-NEXT: ldx r0, r5, r8447; CHECK-NEXT: ldx r30, r5, r10448; CHECK-NEXT: mulld r6, r12, r6449; CHECK-NEXT: ldx r29, r5, r9450; CHECK-NEXT: ldx r28, r5, r11451; CHECK-NEXT: addi r5, r5, 1452; CHECK-NEXT: mulld r6, r6, r0453; CHECK-NEXT: mulld r6, r6, r30454; CHECK-NEXT: mulld r6, r6, r29455; CHECK-NEXT: maddld r3, r6, r28, r3456; CHECK-NEXT: bdnz .LBB4_2457; CHECK-NEXT: # %bb.3:458; CHECK-NEXT: ld r30, -16(r1) # 8-byte Folded Reload459; CHECK-NEXT: ld r29, -24(r1) # 8-byte Folded Reload460; CHECK-NEXT: ld r28, -32(r1) # 8-byte Folded Reload461; CHECK-NEXT: blr462; CHECK-NEXT: .LBB4_4:463; CHECK-NEXT: li r3, 0464; CHECK-NEXT: blr465entry:466 %mul = shl nsw i64 %offset, 1467 %mul2 = mul nsw i64 %offset, 5468 %mul4 = mul nsw i64 %offset, 7469 %mul6 = shl nsw i64 %offset, 3470 %mul8 = mul nsw i64 %offset, 10471 %cmp70 = icmp sgt i64 %n, 0472 br i1 %cmp70, label %for.body, label %for.cond.cleanup473 474for.cond.cleanup: ; preds = %for.body, %entry475 %sum.0.lcssa = phi i64 [ 0, %entry ], [ %add26, %for.body ]476 ret i64 %sum.0.lcssa477 478for.body: ; preds = %entry, %for.body479 %sum.072 = phi i64 [ %add26, %for.body ], [ 0, %entry ]480 %i.071 = phi i64 [ %inc, %for.body ], [ 0, %entry ]481 %add = add i64 %i.071, %base1482 %add.ptr15.idx = add i64 %add, %offset483 %add.ptr15 = getelementptr inbounds i8, ptr %p, i64 %add.ptr15.idx484 %0 = load i64, ptr %add.ptr15, align 8485 %add.ptr16.idx = add i64 %add, %mul486 %add.ptr16 = getelementptr inbounds i8, ptr %p, i64 %add.ptr16.idx487 %1 = load i64, ptr %add.ptr16, align 8488 %add.ptr17.idx = add i64 %add, %mul2489 %add.ptr17 = getelementptr inbounds i8, ptr %p, i64 %add.ptr17.idx490 %2 = load i64, ptr %add.ptr17, align 8491 %add.ptr18.idx = add i64 %add, %mul4492 %add.ptr18 = getelementptr inbounds i8, ptr %p, i64 %add.ptr18.idx493 %3 = load i64, ptr %add.ptr18, align 8494 %add.ptr19.idx = add i64 %add, %mul6495 %add.ptr19 = getelementptr inbounds i8, ptr %p, i64 %add.ptr19.idx496 %4 = load i64, ptr %add.ptr19, align 8497 %add.ptr20.idx = add i64 %add, %mul8498 %add.ptr20 = getelementptr inbounds i8, ptr %p, i64 %add.ptr20.idx499 %5 = load i64, ptr %add.ptr20, align 8500 %mul21 = mul i64 %1, %0501 %mul22 = mul i64 %mul21, %2502 %mul23 = mul i64 %mul22, %3503 %mul24 = mul i64 %mul23, %4504 %mul25 = mul i64 %mul24, %5505 %add26 = add i64 %mul25, %sum.072506 %inc = add nuw nsw i64 %i.071, 1507 %exitcond.not = icmp eq i64 %inc, %n508 br i1 %exitcond.not, label %for.cond.cleanup, label %for.body509}510 511; addresses:512; 1: base1 + offset513; 2: + offset514; 3: + 3*offset515; 4: + 2*offset516;517; chains:518; 1: base1 + offset, offsets: (0, 2*offset)519; 2: base1 + 4*offset, offsets: (0, 2*offset)520;521; long long two_chain_different_offsets_succ(char *p, long long offset, long long base1, long long n) {522; long long o1 = base1 + offset;523; long long o2 = base1 + 3 * offset;524; long long o3 = base1 + 4 * offset;525; long long o4 = base1 + 6 * offset;526; char *p1 = p + o1;527; char *p2 = p + o2;528; char *p3 = p + o3;529; char *p4 = p + o4;530; long long sum = 0;531; for (long long i = 0; i < n; ++i) {532; unsigned long x1 = *(unsigned long *)(p1 + i);533; unsigned long x2 = *(unsigned long *)(p2 + i);534; unsigned long x3 = *(unsigned long *)(p3 + i);535; unsigned long x4 = *(unsigned long *)(p4 + i);536; sum += x1 * x2 * x3 * x4;537; }538; return sum;539; }540;541define i64 @two_chain_different_offsets_succ(ptr %p, i64 %offset, i64 %base1, i64 %n) {542; CHECK-LABEL: two_chain_different_offsets_succ:543; CHECK: # %bb.0: # %entry544; CHECK-NEXT: cmpdi r6, 0545; CHECK-NEXT: ble cr0, .LBB5_4546; CHECK-NEXT: # %bb.1: # %for.body.preheader547; CHECK-NEXT: sldi r8, r4, 2548; CHECK-NEXT: add r7, r5, r4549; CHECK-NEXT: mtctr r6550; CHECK-NEXT: add r5, r5, r8551; CHECK-NEXT: add r7, r3, r7552; CHECK-NEXT: sldi r4, r4, 1553; CHECK-NEXT: add r5, r3, r5554; CHECK-NEXT: li r3, 0555; CHECK-NEXT: .p2align 4556; CHECK-NEXT: .LBB5_2: # %for.body557; CHECK-NEXT: #558; CHECK-NEXT: ld r6, 0(r7)559; CHECK-NEXT: ldx r8, r7, r4560; CHECK-NEXT: ld r9, 0(r5)561; CHECK-NEXT: ldx r10, r5, r4562; CHECK-NEXT: addi r7, r7, 1563; CHECK-NEXT: addi r5, r5, 1564; CHECK-NEXT: mulld r6, r8, r6565; CHECK-NEXT: mulld r6, r6, r9566; CHECK-NEXT: maddld r3, r6, r10, r3567; CHECK-NEXT: bdnz .LBB5_2568; CHECK-NEXT: # %bb.3: # %for.cond.cleanup569; CHECK-NEXT: blr570; CHECK-NEXT: .LBB5_4:571; CHECK-NEXT: li r3, 0572; CHECK-NEXT: blr573entry:574 %mul = mul nsw i64 %offset, 3575 %mul2 = shl nsw i64 %offset, 2576 %mul4 = mul nsw i64 %offset, 6577 %cmp46 = icmp sgt i64 %n, 0578 br i1 %cmp46, label %for.body, label %for.cond.cleanup579 580for.cond.cleanup: ; preds = %for.body, %entry581 %sum.0.lcssa = phi i64 [ 0, %entry ], [ %add16, %for.body ]582 ret i64 %sum.0.lcssa583 584for.body: ; preds = %entry, %for.body585 %sum.048 = phi i64 [ %add16, %for.body ], [ 0, %entry ]586 %i.047 = phi i64 [ %inc, %for.body ], [ 0, %entry ]587 %add = add i64 %i.047, %base1588 %add.ptr9.idx = add i64 %add, %offset589 %add.ptr9 = getelementptr inbounds i8, ptr %p, i64 %add.ptr9.idx590 %0 = load i64, ptr %add.ptr9, align 8591 %add.ptr10.idx = add i64 %add, %mul592 %add.ptr10 = getelementptr inbounds i8, ptr %p, i64 %add.ptr10.idx593 %1 = load i64, ptr %add.ptr10, align 8594 %add.ptr11.idx = add i64 %add, %mul2595 %add.ptr11 = getelementptr inbounds i8, ptr %p, i64 %add.ptr11.idx596 %2 = load i64, ptr %add.ptr11, align 8597 %add.ptr12.idx = add i64 %add, %mul4598 %add.ptr12 = getelementptr inbounds i8, ptr %p, i64 %add.ptr12.idx599 %3 = load i64, ptr %add.ptr12, align 8600 %mul13 = mul i64 %1, %0601 %mul14 = mul i64 %mul13, %2602 %mul15 = mul i64 %mul14, %3603 %add16 = add i64 %mul15, %sum.048604 %inc = add nuw nsw i64 %i.047, 1605 %exitcond.not = icmp eq i64 %inc, %n606 br i1 %exitcond.not, label %for.cond.cleanup, label %for.body607}608 609; addresses:610; 1: base1 + offset611; 2: + 2*offset612; 3: + base2 - base1 - 2*offset613; 4: + 2*offset614;615; chains:616; 1: base1 + offset, offsets: (0, 2*offset)617; 2: base2 + offset, offsets: (0, 2*offset)618;619; long long two_chain_two_bases_succ(char *p, long long offset, long long base1, long long base2, long long n) {620; long long o1 = base1 + offset;621; long long o2 = base1 + 3 * offset;622; long long o3 = base2 + offset;623; long long o4 = base2 + 3 * offset;624; char *p1 = p + o1;625; char *p2 = p + o2;626; char *p3 = p + o3;627; char *p4 = p + o4;628; long long sum = 0;629; for (long long i = 0; i < n; ++i) {630; unsigned long x1 = *(unsigned long *)(p1 + i);631; unsigned long x2 = *(unsigned long *)(p2 + i);632; unsigned long x3 = *(unsigned long *)(p3 + i);633; unsigned long x4 = *(unsigned long *)(p4 + i);634; sum += x1 * x2 * x3 * x4;635; }636; return sum;637; }638;639define i64 @two_chain_two_bases_succ(ptr %p, i64 %offset, i64 %base1, i64 %base2, i64 %n) {640; CHECK-LABEL: two_chain_two_bases_succ:641; CHECK: # %bb.0: # %entry642; CHECK-NEXT: cmpdi r7, 0643; CHECK-NEXT: ble cr0, .LBB6_4644; CHECK-NEXT: # %bb.1: # %for.body.preheader645; CHECK-NEXT: add r5, r5, r4646; CHECK-NEXT: add r6, r6, r4647; CHECK-NEXT: mtctr r7648; CHECK-NEXT: sldi r4, r4, 1649; CHECK-NEXT: add r5, r3, r5650; CHECK-NEXT: add r6, r3, r6651; CHECK-NEXT: li r3, 0652; CHECK-NEXT: .p2align 4653; CHECK-NEXT: .LBB6_2: # %for.body654; CHECK-NEXT: #655; CHECK-NEXT: ld r7, 0(r5)656; CHECK-NEXT: ldx r8, r5, r4657; CHECK-NEXT: ld r9, 0(r6)658; CHECK-NEXT: ldx r10, r6, r4659; CHECK-NEXT: addi r5, r5, 1660; CHECK-NEXT: addi r6, r6, 1661; CHECK-NEXT: mulld r7, r8, r7662; CHECK-NEXT: mulld r7, r7, r9663; CHECK-NEXT: maddld r3, r7, r10, r3664; CHECK-NEXT: bdnz .LBB6_2665; CHECK-NEXT: # %bb.3: # %for.cond.cleanup666; CHECK-NEXT: blr667; CHECK-NEXT: .LBB6_4:668; CHECK-NEXT: li r3, 0669; CHECK-NEXT: blr670entry:671 %mul = mul nsw i64 %offset, 3672 %cmp44 = icmp sgt i64 %n, 0673 br i1 %cmp44, label %for.body, label %for.cond.cleanup674 675for.cond.cleanup: ; preds = %for.body, %entry676 %sum.0.lcssa = phi i64 [ 0, %entry ], [ %add15, %for.body ]677 ret i64 %sum.0.lcssa678 679for.body: ; preds = %entry, %for.body680 %sum.046 = phi i64 [ %add15, %for.body ], [ 0, %entry ]681 %i.045 = phi i64 [ %inc, %for.body ], [ 0, %entry ]682 %add = add i64 %i.045, %base1683 %add.ptr8.idx = add i64 %add, %offset684 %add.ptr8 = getelementptr inbounds i8, ptr %p, i64 %add.ptr8.idx685 %0 = load i64, ptr %add.ptr8, align 8686 %add1 = add i64 %i.045, %mul687 %add.ptr9.idx = add i64 %add1, %base1688 %add.ptr9 = getelementptr inbounds i8, ptr %p, i64 %add.ptr9.idx689 %1 = load i64, ptr %add.ptr9, align 8690 %add2 = add i64 %i.045, %base2691 %add.ptr10.idx = add i64 %add2, %offset692 %add.ptr10 = getelementptr inbounds i8, ptr %p, i64 %add.ptr10.idx693 %2 = load i64, ptr %add.ptr10, align 8694 %add.ptr11.idx = add i64 %add2, %mul695 %add.ptr11 = getelementptr inbounds i8, ptr %p, i64 %add.ptr11.idx696 %3 = load i64, ptr %add.ptr11, align 8697 %mul12 = mul i64 %1, %0698 %mul13 = mul i64 %mul12, %2699 %mul14 = mul i64 %mul13, %3700 %add15 = add i64 %mul14, %sum.046701 %inc = add nuw nsw i64 %i.045, 1702 %exitcond.not = icmp eq i64 %inc, %n703 br i1 %exitcond.not, label %for.cond.cleanup, label %for.body704}705;706; Check chain commoning can reduce register pressure to save register spill/reload.707;708; int spill_reduce_succ(double *input1, double *input2, double *output, long long m, long long inc1, long long inc2, long long inc3, long long inc4, long long inc) {709; inc = inc4;710; #pragma unroll 4711; for (long long i = 0; i < 4 * m; i++) {712; output[inc + inc1] += input1[inc + inc1] * input2[inc + inc1];713; output[inc + inc2] += input1[inc + inc2] * input2[inc + inc2];714; output[inc + inc3] += input1[inc + inc3] * input2[inc + inc3];715; inc = inc + inc4;716; }717; return 0;718; }719;720define signext i32 @spill_reduce_succ(ptr %input1, ptr %input2, ptr %output, i64 %m, i64 %inc1, i64 %inc2, i64 %inc3, i64 %inc4, i64 %inc) {721; CHECK-LABEL: spill_reduce_succ:722; CHECK: # %bb.0: # %entry723; CHECK-NEXT: cmpdi r6, 0724; CHECK-NEXT: ble cr0, .LBB7_9725; CHECK-NEXT: # %bb.1: # %for.body.preheader726; CHECK-NEXT: sldi r6, r6, 2727; CHECK-NEXT: li r11, 1728; CHECK-NEXT: std r26, -48(r1) # 8-byte Folded Spill729; CHECK-NEXT: mr r26, r10730; CHECK-NEXT: cmpdi r6, 1731; CHECK-NEXT: std r14, -144(r1) # 8-byte Folded Spill732; CHECK-NEXT: std r15, -136(r1) # 8-byte Folded Spill733; CHECK-NEXT: std r16, -128(r1) # 8-byte Folded Spill734; CHECK-NEXT: std r17, -120(r1) # 8-byte Folded Spill735; CHECK-NEXT: std r18, -112(r1) # 8-byte Folded Spill736; CHECK-NEXT: std r19, -104(r1) # 8-byte Folded Spill737; CHECK-NEXT: std r20, -96(r1) # 8-byte Folded Spill738; CHECK-NEXT: std r21, -88(r1) # 8-byte Folded Spill739; CHECK-NEXT: std r22, -80(r1) # 8-byte Folded Spill740; CHECK-NEXT: std r23, -72(r1) # 8-byte Folded Spill741; CHECK-NEXT: std r24, -64(r1) # 8-byte Folded Spill742; CHECK-NEXT: std r25, -56(r1) # 8-byte Folded Spill743; CHECK-NEXT: std r27, -40(r1) # 8-byte Folded Spill744; CHECK-NEXT: std r28, -32(r1) # 8-byte Folded Spill745; CHECK-NEXT: std r29, -24(r1) # 8-byte Folded Spill746; CHECK-NEXT: std r30, -16(r1) # 8-byte Folded Spill747; CHECK-NEXT: std r31, -8(r1) # 8-byte Folded Spill748; CHECK-NEXT: std r2, -152(r1) # 8-byte Folded Spill749; CHECK-NEXT: iselgt r11, r6, r11750; CHECK-NEXT: addi r12, r11, -1751; CHECK-NEXT: cmpldi r12, 3752; CHECK-NEXT: clrldi r6, r11, 63753; CHECK-NEXT: blt cr0, .LBB7_5754; CHECK-NEXT: # %bb.2: # %for.body.preheader.new755; CHECK-NEXT: rldicl r11, r11, 62, 2756; CHECK-NEXT: sldi r20, r8, 3757; CHECK-NEXT: mr r14, r7758; CHECK-NEXT: sldi r7, r7, 3759; CHECK-NEXT: sldi r21, r9, 3760; CHECK-NEXT: std r3, -160(r1) # 8-byte Folded Spill761; CHECK-NEXT: std r9, -208(r1) # 8-byte Folded Spill762; CHECK-NEXT: std r8, -184(r1) # 8-byte Folded Spill763; CHECK-NEXT: std r5, -200(r1) # 8-byte Folded Spill764; CHECK-NEXT: std r4, -168(r1) # 8-byte Folded Spill765; CHECK-NEXT: std r11, -192(r1) # 8-byte Folded Spill766; CHECK-NEXT: sldi r11, r10, 5767; CHECK-NEXT: add r0, r11, r20768; CHECK-NEXT: add r12, r11, r21769; CHECK-NEXT: add r30, r5, r0770; CHECK-NEXT: add r0, r11, r7771; CHECK-NEXT: std r21, -216(r1) # 8-byte Folded Spill772; CHECK-NEXT: std r20, -224(r1) # 8-byte Folded Spill773; CHECK-NEXT: add r12, r5, r12774; CHECK-NEXT: add r29, r5, r0775; CHECK-NEXT: add r28, r4, r0776; CHECK-NEXT: add r27, r3, r0777; CHECK-NEXT: mulli r0, r10, 24778; CHECK-NEXT: std r14, -176(r1) # 8-byte Folded Spill779; CHECK-NEXT: add r26, r0, r21780; CHECK-NEXT: add r25, r0, r20781; CHECK-NEXT: add r0, r0, r7782; CHECK-NEXT: add r24, r5, r0783; CHECK-NEXT: add r23, r4, r0784; CHECK-NEXT: add r22, r3, r0785; CHECK-NEXT: sldi r0, r10, 4786; CHECK-NEXT: add r26, r5, r26787; CHECK-NEXT: add r25, r5, r25788; CHECK-NEXT: add r21, r0, r21789; CHECK-NEXT: add r20, r0, r20790; CHECK-NEXT: add r0, r0, r7791; CHECK-NEXT: add r19, r5, r0792; CHECK-NEXT: add r18, r4, r0793; CHECK-NEXT: add r17, r3, r0794; CHECK-NEXT: add r0, r10, r9795; CHECK-NEXT: add r21, r5, r21796; CHECK-NEXT: add r20, r5, r20797; CHECK-NEXT: sldi r0, r0, 3798; CHECK-NEXT: add r16, r5, r0799; CHECK-NEXT: add r0, r10, r8800; CHECK-NEXT: sldi r0, r0, 3801; CHECK-NEXT: add r15, r5, r0802; CHECK-NEXT: add r0, r10, r14803; CHECK-NEXT: sldi r0, r0, 3804; CHECK-NEXT: add r2, r3, r0805; CHECK-NEXT: ld r3, -224(r1) # 8-byte Folded Reload806; CHECK-NEXT: add r14, r5, r0807; CHECK-NEXT: add r31, r4, r0808; CHECK-NEXT: sub r0, r3, r7809; CHECK-NEXT: ld r3, -192(r1) # 8-byte Folded Reload810; CHECK-NEXT: rldicl r9, r3, 2, 1811; CHECK-NEXT: ld r3, -216(r1) # 8-byte Folded Reload812; CHECK-NEXT: addi r8, r9, -4813; CHECK-NEXT: rldicl r8, r8, 62, 2814; CHECK-NEXT: sub r7, r3, r7815; CHECK-NEXT: ori r3, r9, 1816; CHECK-NEXT: addi r8, r8, 1817; CHECK-NEXT: mulld r3, r10, r3818; CHECK-NEXT: mtctr r8819; CHECK-NEXT: li r8, 0820; CHECK-NEXT: std r10, -192(r1) # 8-byte Folded Spill821; CHECK-NEXT: std r3, -216(r1) # 8-byte Folded Spill822; CHECK-NEXT: .p2align 4823; CHECK-NEXT: .LBB7_3: # %for.body824; CHECK-NEXT: #825; CHECK-NEXT: lfd f0, 0(r2)826; CHECK-NEXT: lfd f1, 0(r31)827; CHECK-NEXT: xsmuldp f0, f0, f1828; CHECK-NEXT: lfd f1, 0(r14)829; CHECK-NEXT: xsadddp f0, f1, f0830; CHECK-NEXT: stfd f0, 0(r14)831; CHECK-NEXT: add r14, r14, r11832; CHECK-NEXT: lfdx f0, r2, r0833; CHECK-NEXT: lfdx f1, r31, r0834; CHECK-NEXT: xsmuldp f0, f0, f1835; CHECK-NEXT: lfdx f1, r15, r8836; CHECK-NEXT: xsadddp f0, f1, f0837; CHECK-NEXT: stfdx f0, r15, r8838; CHECK-NEXT: lfdx f0, r2, r7839; CHECK-NEXT: lfdx f1, r31, r7840; CHECK-NEXT: add r2, r2, r11841; CHECK-NEXT: add r31, r31, r11842; CHECK-NEXT: xsmuldp f0, f0, f1843; CHECK-NEXT: lfdx f1, r16, r8844; CHECK-NEXT: xsadddp f0, f1, f0845; CHECK-NEXT: stfdx f0, r16, r8846; CHECK-NEXT: lfd f0, 0(r17)847; CHECK-NEXT: lfd f1, 0(r18)848; CHECK-NEXT: xsmuldp f0, f0, f1849; CHECK-NEXT: lfdx f1, r19, r8850; CHECK-NEXT: xsadddp f0, f1, f0851; CHECK-NEXT: stfdx f0, r19, r8852; CHECK-NEXT: lfdx f0, r17, r0853; CHECK-NEXT: lfdx f1, r18, r0854; CHECK-NEXT: xsmuldp f0, f0, f1855; CHECK-NEXT: lfdx f1, r20, r8856; CHECK-NEXT: xsadddp f0, f1, f0857; CHECK-NEXT: stfdx f0, r20, r8858; CHECK-NEXT: lfdx f0, r17, r7859; CHECK-NEXT: lfdx f1, r18, r7860; CHECK-NEXT: add r17, r17, r11861; CHECK-NEXT: add r18, r18, r11862; CHECK-NEXT: xsmuldp f0, f0, f1863; CHECK-NEXT: lfdx f1, r21, r8864; CHECK-NEXT: xsadddp f0, f1, f0865; CHECK-NEXT: stfdx f0, r21, r8866; CHECK-NEXT: lfd f0, 0(r22)867; CHECK-NEXT: lfd f1, 0(r23)868; CHECK-NEXT: xsmuldp f0, f0, f1869; CHECK-NEXT: lfdx f1, r24, r8870; CHECK-NEXT: xsadddp f0, f1, f0871; CHECK-NEXT: stfdx f0, r24, r8872; CHECK-NEXT: lfdx f0, r22, r0873; CHECK-NEXT: lfdx f1, r23, r0874; CHECK-NEXT: xsmuldp f0, f0, f1875; CHECK-NEXT: lfdx f1, r25, r8876; CHECK-NEXT: xsadddp f0, f1, f0877; CHECK-NEXT: stfdx f0, r25, r8878; CHECK-NEXT: lfdx f0, r22, r7879; CHECK-NEXT: lfdx f1, r23, r7880; CHECK-NEXT: add r22, r22, r11881; CHECK-NEXT: add r23, r23, r11882; CHECK-NEXT: xsmuldp f0, f0, f1883; CHECK-NEXT: lfdx f1, r26, r8884; CHECK-NEXT: xsadddp f0, f1, f0885; CHECK-NEXT: stfdx f0, r26, r8886; CHECK-NEXT: lfd f0, 0(r27)887; CHECK-NEXT: lfd f1, 0(r28)888; CHECK-NEXT: xsmuldp f0, f0, f1889; CHECK-NEXT: lfdx f1, r29, r8890; CHECK-NEXT: xsadddp f0, f1, f0891; CHECK-NEXT: stfdx f0, r29, r8892; CHECK-NEXT: lfdx f0, r27, r0893; CHECK-NEXT: lfdx f1, r28, r0894; CHECK-NEXT: xsmuldp f0, f0, f1895; CHECK-NEXT: lfdx f1, r30, r8896; CHECK-NEXT: xsadddp f0, f1, f0897; CHECK-NEXT: stfdx f0, r30, r8898; CHECK-NEXT: lfdx f0, r27, r7899; CHECK-NEXT: lfdx f1, r28, r7900; CHECK-NEXT: add r27, r27, r11901; CHECK-NEXT: add r28, r28, r11902; CHECK-NEXT: xsmuldp f0, f0, f1903; CHECK-NEXT: lfdx f1, r12, r8904; CHECK-NEXT: xsadddp f0, f1, f0905; CHECK-NEXT: stfdx f0, r12, r8906; CHECK-NEXT: add r8, r8, r11907; CHECK-NEXT: bdnz .LBB7_3908; CHECK-NEXT: # %bb.4:909; CHECK-NEXT: ld r3, -160(r1) # 8-byte Folded Reload910; CHECK-NEXT: ld r4, -168(r1) # 8-byte Folded Reload911; CHECK-NEXT: ld r7, -176(r1) # 8-byte Folded Reload912; CHECK-NEXT: ld r8, -184(r1) # 8-byte Folded Reload913; CHECK-NEXT: ld r10, -192(r1) # 8-byte Folded Reload914; CHECK-NEXT: ld r5, -200(r1) # 8-byte Folded Reload915; CHECK-NEXT: ld r9, -208(r1) # 8-byte Folded Reload916; CHECK-NEXT: ld r26, -216(r1) # 8-byte Folded Reload917; CHECK-NEXT: .LBB7_5: # %for.cond.cleanup.loopexit.unr-lcssa918; CHECK-NEXT: cmpldi r6, 0919; CHECK-NEXT: beq cr0, .LBB7_8920; CHECK-NEXT: # %bb.6: # %for.body.epil.preheader921; CHECK-NEXT: add r11, r26, r9922; CHECK-NEXT: add r12, r26, r8923; CHECK-NEXT: add r9, r26, r7924; CHECK-NEXT: sldi r27, r10, 3925; CHECK-NEXT: sldi r11, r11, 3926; CHECK-NEXT: sldi r0, r12, 3927; CHECK-NEXT: sldi r9, r9, 3928; CHECK-NEXT: add r28, r5, r11929; CHECK-NEXT: add r10, r4, r11930; CHECK-NEXT: add r11, r3, r11931; CHECK-NEXT: add r12, r5, r0932; CHECK-NEXT: add r30, r4, r0933; CHECK-NEXT: add r29, r3, r0934; CHECK-NEXT: add r5, r5, r9935; CHECK-NEXT: add r4, r4, r9936; CHECK-NEXT: add r3, r3, r9937; CHECK-NEXT: li r9, 0938; CHECK-NEXT: .p2align 4939; CHECK-NEXT: .LBB7_7: # %for.body.epil940; CHECK-NEXT: #941; CHECK-NEXT: lfdx f0, r3, r9942; CHECK-NEXT: lfdx f1, r4, r9943; CHECK-NEXT: addi r6, r6, -1944; CHECK-NEXT: cmpldi r6, 0945; CHECK-NEXT: xsmuldp f0, f0, f1946; CHECK-NEXT: lfd f1, 0(r5)947; CHECK-NEXT: xsadddp f0, f1, f0948; CHECK-NEXT: stfd f0, 0(r5)949; CHECK-NEXT: add r5, r5, r27950; CHECK-NEXT: lfdx f0, r29, r9951; CHECK-NEXT: lfdx f1, r30, r9952; CHECK-NEXT: xsmuldp f0, f0, f1953; CHECK-NEXT: lfdx f1, r12, r9954; CHECK-NEXT: xsadddp f0, f1, f0955; CHECK-NEXT: stfdx f0, r12, r9956; CHECK-NEXT: lfdx f0, r11, r9957; CHECK-NEXT: lfdx f1, r10, r9958; CHECK-NEXT: xsmuldp f0, f0, f1959; CHECK-NEXT: lfdx f1, r28, r9960; CHECK-NEXT: xsadddp f0, f1, f0961; CHECK-NEXT: stfdx f0, r28, r9962; CHECK-NEXT: add r9, r9, r27963; CHECK-NEXT: bne cr0, .LBB7_7964; CHECK-NEXT: .LBB7_8:965; CHECK-NEXT: ld r2, -152(r1) # 8-byte Folded Reload966; CHECK-NEXT: ld r31, -8(r1) # 8-byte Folded Reload967; CHECK-NEXT: ld r30, -16(r1) # 8-byte Folded Reload968; CHECK-NEXT: ld r29, -24(r1) # 8-byte Folded Reload969; CHECK-NEXT: ld r28, -32(r1) # 8-byte Folded Reload970; CHECK-NEXT: ld r27, -40(r1) # 8-byte Folded Reload971; CHECK-NEXT: ld r26, -48(r1) # 8-byte Folded Reload972; CHECK-NEXT: ld r25, -56(r1) # 8-byte Folded Reload973; CHECK-NEXT: ld r24, -64(r1) # 8-byte Folded Reload974; CHECK-NEXT: ld r23, -72(r1) # 8-byte Folded Reload975; CHECK-NEXT: ld r22, -80(r1) # 8-byte Folded Reload976; CHECK-NEXT: ld r21, -88(r1) # 8-byte Folded Reload977; CHECK-NEXT: ld r20, -96(r1) # 8-byte Folded Reload978; CHECK-NEXT: ld r19, -104(r1) # 8-byte Folded Reload979; CHECK-NEXT: ld r18, -112(r1) # 8-byte Folded Reload980; CHECK-NEXT: ld r17, -120(r1) # 8-byte Folded Reload981; CHECK-NEXT: ld r16, -128(r1) # 8-byte Folded Reload982; CHECK-NEXT: ld r15, -136(r1) # 8-byte Folded Reload983; CHECK-NEXT: ld r14, -144(r1) # 8-byte Folded Reload984; CHECK-NEXT: .LBB7_9: # %for.cond.cleanup985; CHECK-NEXT: li r3, 0986; CHECK-NEXT: blr987entry:988 %cmp49 = icmp sgt i64 %m, 0989 br i1 %cmp49, label %for.body.preheader, label %for.cond.cleanup990 991for.body.preheader: ; preds = %entry992 %0 = shl i64 %m, 2993 %smax52 = call i64 @llvm.smax.i64(i64 %0, i64 1)994 %1 = add nsw i64 %smax52, -1995 %xtraiter = and i64 %smax52, 1996 %2 = icmp ult i64 %1, 3997 br i1 %2, label %for.cond.cleanup.loopexit.unr-lcssa, label %for.body.preheader.new998 999for.body.preheader.new: ; preds = %for.body.preheader1000 %unroll_iter = and i64 %smax52, 92233720368547758041001 br label %for.body1002 1003for.cond.cleanup.loopexit.unr-lcssa: ; preds = %for.body, %for.body.preheader1004 %inc.addr.050.unr = phi i64 [ %inc4, %for.body.preheader ], [ %add23.3, %for.body ]1005 %lcmp.mod.not = icmp eq i64 %xtraiter, 01006 br i1 %lcmp.mod.not, label %for.cond.cleanup, label %for.body.epil1007 1008for.body.epil: ; preds = %for.cond.cleanup.loopexit.unr-lcssa, %for.body.epil1009 %inc.addr.050.epil = phi i64 [ %add23.epil, %for.body.epil ], [ %inc.addr.050.unr, %for.cond.cleanup.loopexit.unr-lcssa ]1010 %epil.iter = phi i64 [ %epil.iter.sub, %for.body.epil ], [ %xtraiter, %for.cond.cleanup.loopexit.unr-lcssa ]1011 %add.epil = add nsw i64 %inc.addr.050.epil, %inc11012 %arrayidx.epil = getelementptr inbounds double, ptr %input1, i64 %add.epil1013 %3 = load double, ptr %arrayidx.epil, align 81014 %arrayidx2.epil = getelementptr inbounds double, ptr %input2, i64 %add.epil1015 %4 = load double, ptr %arrayidx2.epil, align 81016 %mul3.epil = fmul double %3, %41017 %arrayidx5.epil = getelementptr inbounds double, ptr %output, i64 %add.epil1018 %5 = load double, ptr %arrayidx5.epil, align 81019 %add6.epil = fadd double %5, %mul3.epil1020 store double %add6.epil, ptr %arrayidx5.epil, align 81021 %add7.epil = add nsw i64 %inc.addr.050.epil, %inc21022 %arrayidx8.epil = getelementptr inbounds double, ptr %input1, i64 %add7.epil1023 %6 = load double, ptr %arrayidx8.epil, align 81024 %arrayidx10.epil = getelementptr inbounds double, ptr %input2, i64 %add7.epil1025 %7 = load double, ptr %arrayidx10.epil, align 81026 %mul11.epil = fmul double %6, %71027 %arrayidx13.epil = getelementptr inbounds double, ptr %output, i64 %add7.epil1028 %8 = load double, ptr %arrayidx13.epil, align 81029 %add14.epil = fadd double %8, %mul11.epil1030 store double %add14.epil, ptr %arrayidx13.epil, align 81031 %add15.epil = add nsw i64 %inc.addr.050.epil, %inc31032 %arrayidx16.epil = getelementptr inbounds double, ptr %input1, i64 %add15.epil1033 %9 = load double, ptr %arrayidx16.epil, align 81034 %arrayidx18.epil = getelementptr inbounds double, ptr %input2, i64 %add15.epil1035 %10 = load double, ptr %arrayidx18.epil, align 81036 %mul19.epil = fmul double %9, %101037 %arrayidx21.epil = getelementptr inbounds double, ptr %output, i64 %add15.epil1038 %11 = load double, ptr %arrayidx21.epil, align 81039 %add22.epil = fadd double %11, %mul19.epil1040 store double %add22.epil, ptr %arrayidx21.epil, align 81041 %add23.epil = add nsw i64 %inc.addr.050.epil, %inc41042 %epil.iter.sub = add nsw i64 %epil.iter, -11043 %epil.iter.cmp.not = icmp eq i64 %epil.iter.sub, 01044 br i1 %epil.iter.cmp.not, label %for.cond.cleanup, label %for.body.epil1045 1046for.cond.cleanup: ; preds = %for.body.epil, %for.cond.cleanup.loopexit.unr-lcssa, %entry1047 ret i32 01048 1049for.body: ; preds = %for.body, %for.body.preheader.new1050 %inc.addr.050 = phi i64 [ %inc4, %for.body.preheader.new ], [ %add23.3, %for.body ]1051 %niter = phi i64 [ %unroll_iter, %for.body.preheader.new ], [ %niter.nsub.3, %for.body ]1052 %add = add nsw i64 %inc.addr.050, %inc11053 %arrayidx = getelementptr inbounds double, ptr %input1, i64 %add1054 %12 = load double, ptr %arrayidx, align 81055 %arrayidx2 = getelementptr inbounds double, ptr %input2, i64 %add1056 %13 = load double, ptr %arrayidx2, align 81057 %mul3 = fmul double %12, %131058 %arrayidx5 = getelementptr inbounds double, ptr %output, i64 %add1059 %14 = load double, ptr %arrayidx5, align 81060 %add6 = fadd double %14, %mul31061 store double %add6, ptr %arrayidx5, align 81062 %add7 = add nsw i64 %inc.addr.050, %inc21063 %arrayidx8 = getelementptr inbounds double, ptr %input1, i64 %add71064 %15 = load double, ptr %arrayidx8, align 81065 %arrayidx10 = getelementptr inbounds double, ptr %input2, i64 %add71066 %16 = load double, ptr %arrayidx10, align 81067 %mul11 = fmul double %15, %161068 %arrayidx13 = getelementptr inbounds double, ptr %output, i64 %add71069 %17 = load double, ptr %arrayidx13, align 81070 %add14 = fadd double %17, %mul111071 store double %add14, ptr %arrayidx13, align 81072 %add15 = add nsw i64 %inc.addr.050, %inc31073 %arrayidx16 = getelementptr inbounds double, ptr %input1, i64 %add151074 %18 = load double, ptr %arrayidx16, align 81075 %arrayidx18 = getelementptr inbounds double, ptr %input2, i64 %add151076 %19 = load double, ptr %arrayidx18, align 81077 %mul19 = fmul double %18, %191078 %arrayidx21 = getelementptr inbounds double, ptr %output, i64 %add151079 %20 = load double, ptr %arrayidx21, align 81080 %add22 = fadd double %20, %mul191081 store double %add22, ptr %arrayidx21, align 81082 %add23 = add nsw i64 %inc.addr.050, %inc41083 %add.1 = add nsw i64 %add23, %inc11084 %arrayidx.1 = getelementptr inbounds double, ptr %input1, i64 %add.11085 %21 = load double, ptr %arrayidx.1, align 81086 %arrayidx2.1 = getelementptr inbounds double, ptr %input2, i64 %add.11087 %22 = load double, ptr %arrayidx2.1, align 81088 %mul3.1 = fmul double %21, %221089 %arrayidx5.1 = getelementptr inbounds double, ptr %output, i64 %add.11090 %23 = load double, ptr %arrayidx5.1, align 81091 %add6.1 = fadd double %23, %mul3.11092 store double %add6.1, ptr %arrayidx5.1, align 81093 %add7.1 = add nsw i64 %add23, %inc21094 %arrayidx8.1 = getelementptr inbounds double, ptr %input1, i64 %add7.11095 %24 = load double, ptr %arrayidx8.1, align 81096 %arrayidx10.1 = getelementptr inbounds double, ptr %input2, i64 %add7.11097 %25 = load double, ptr %arrayidx10.1, align 81098 %mul11.1 = fmul double %24, %251099 %arrayidx13.1 = getelementptr inbounds double, ptr %output, i64 %add7.11100 %26 = load double, ptr %arrayidx13.1, align 81101 %add14.1 = fadd double %26, %mul11.11102 store double %add14.1, ptr %arrayidx13.1, align 81103 %add15.1 = add nsw i64 %add23, %inc31104 %arrayidx16.1 = getelementptr inbounds double, ptr %input1, i64 %add15.11105 %27 = load double, ptr %arrayidx16.1, align 81106 %arrayidx18.1 = getelementptr inbounds double, ptr %input2, i64 %add15.11107 %28 = load double, ptr %arrayidx18.1, align 81108 %mul19.1 = fmul double %27, %281109 %arrayidx21.1 = getelementptr inbounds double, ptr %output, i64 %add15.11110 %29 = load double, ptr %arrayidx21.1, align 81111 %add22.1 = fadd double %29, %mul19.11112 store double %add22.1, ptr %arrayidx21.1, align 81113 %add23.1 = add nsw i64 %add23, %inc41114 %add.2 = add nsw i64 %add23.1, %inc11115 %arrayidx.2 = getelementptr inbounds double, ptr %input1, i64 %add.21116 %30 = load double, ptr %arrayidx.2, align 81117 %arrayidx2.2 = getelementptr inbounds double, ptr %input2, i64 %add.21118 %31 = load double, ptr %arrayidx2.2, align 81119 %mul3.2 = fmul double %30, %311120 %arrayidx5.2 = getelementptr inbounds double, ptr %output, i64 %add.21121 %32 = load double, ptr %arrayidx5.2, align 81122 %add6.2 = fadd double %32, %mul3.21123 store double %add6.2, ptr %arrayidx5.2, align 81124 %add7.2 = add nsw i64 %add23.1, %inc21125 %arrayidx8.2 = getelementptr inbounds double, ptr %input1, i64 %add7.21126 %33 = load double, ptr %arrayidx8.2, align 81127 %arrayidx10.2 = getelementptr inbounds double, ptr %input2, i64 %add7.21128 %34 = load double, ptr %arrayidx10.2, align 81129 %mul11.2 = fmul double %33, %341130 %arrayidx13.2 = getelementptr inbounds double, ptr %output, i64 %add7.21131 %35 = load double, ptr %arrayidx13.2, align 81132 %add14.2 = fadd double %35, %mul11.21133 store double %add14.2, ptr %arrayidx13.2, align 81134 %add15.2 = add nsw i64 %add23.1, %inc31135 %arrayidx16.2 = getelementptr inbounds double, ptr %input1, i64 %add15.21136 %36 = load double, ptr %arrayidx16.2, align 81137 %arrayidx18.2 = getelementptr inbounds double, ptr %input2, i64 %add15.21138 %37 = load double, ptr %arrayidx18.2, align 81139 %mul19.2 = fmul double %36, %371140 %arrayidx21.2 = getelementptr inbounds double, ptr %output, i64 %add15.21141 %38 = load double, ptr %arrayidx21.2, align 81142 %add22.2 = fadd double %38, %mul19.21143 store double %add22.2, ptr %arrayidx21.2, align 81144 %add23.2 = add nsw i64 %add23.1, %inc41145 %add.3 = add nsw i64 %add23.2, %inc11146 %arrayidx.3 = getelementptr inbounds double, ptr %input1, i64 %add.31147 %39 = load double, ptr %arrayidx.3, align 81148 %arrayidx2.3 = getelementptr inbounds double, ptr %input2, i64 %add.31149 %40 = load double, ptr %arrayidx2.3, align 81150 %mul3.3 = fmul double %39, %401151 %arrayidx5.3 = getelementptr inbounds double, ptr %output, i64 %add.31152 %41 = load double, ptr %arrayidx5.3, align 81153 %add6.3 = fadd double %41, %mul3.31154 store double %add6.3, ptr %arrayidx5.3, align 81155 %add7.3 = add nsw i64 %add23.2, %inc21156 %arrayidx8.3 = getelementptr inbounds double, ptr %input1, i64 %add7.31157 %42 = load double, ptr %arrayidx8.3, align 81158 %arrayidx10.3 = getelementptr inbounds double, ptr %input2, i64 %add7.31159 %43 = load double, ptr %arrayidx10.3, align 81160 %mul11.3 = fmul double %42, %431161 %arrayidx13.3 = getelementptr inbounds double, ptr %output, i64 %add7.31162 %44 = load double, ptr %arrayidx13.3, align 81163 %add14.3 = fadd double %44, %mul11.31164 store double %add14.3, ptr %arrayidx13.3, align 81165 %add15.3 = add nsw i64 %add23.2, %inc31166 %arrayidx16.3 = getelementptr inbounds double, ptr %input1, i64 %add15.31167 %45 = load double, ptr %arrayidx16.3, align 81168 %arrayidx18.3 = getelementptr inbounds double, ptr %input2, i64 %add15.31169 %46 = load double, ptr %arrayidx18.3, align 81170 %mul19.3 = fmul double %45, %461171 %arrayidx21.3 = getelementptr inbounds double, ptr %output, i64 %add15.31172 %47 = load double, ptr %arrayidx21.3, align 81173 %add22.3 = fadd double %47, %mul19.31174 store double %add22.3, ptr %arrayidx21.3, align 81175 %add23.3 = add nsw i64 %add23.2, %inc41176 %niter.nsub.3 = add i64 %niter, -41177 %niter.ncmp.3.not = icmp eq i64 %niter.nsub.3, 01178 br i1 %niter.ncmp.3.not, label %for.cond.cleanup.loopexit.unr-lcssa, label %for.body1179}1180 1181declare i64 @llvm.smax.i64(i64, i64)1182 1183