236 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -verify-machineinstrs -mcpu=pwr9 -mtriple=powerpc64le-unknown-unknown | FileCheck %s -check-prefix=PWR93; RUN: llc < %s -verify-machineinstrs -mcpu=pwr8 -mtriple=powerpc64le-unknown-unknown | FileCheck %s -check-prefix=PWR84 5@a = internal global fp128 0xL00000000000000000000000000000000, align 166@x = internal global [4 x fp128] zeroinitializer, align 167@y = internal global [4 x fp128] zeroinitializer, align 168 9define void @fmul_ctrloop_fp128() nounwind {10; PWR9-LABEL: fmul_ctrloop_fp128:11; PWR9: # %bb.0: # %entry12; PWR9-NEXT: addis 3, 2, .L_MergedGlobals@toc@ha13; PWR9-NEXT: li 4, 414; PWR9-NEXT: addi 3, 3, .L_MergedGlobals@toc@l15; PWR9-NEXT: lxv 34, 0(3)16; PWR9-NEXT: mtctr 417; PWR9-NEXT: .p2align 518; PWR9-NEXT: .LBB0_1: # %for.body19; PWR9-NEXT: #20; PWR9-NEXT: lxv 35, 16(3)21; PWR9-NEXT: xsmulqp 3, 2, 322; PWR9-NEXT: stxv 35, 80(3)23; PWR9-NEXT: addi 3, 3, 1624; PWR9-NEXT: bdnz .LBB0_125; PWR9-NEXT: # %bb.2: # %for.end26; PWR9-NEXT: blr27;28; PWR8-LABEL: fmul_ctrloop_fp128:29; PWR8: # %bb.0: # %entry30; PWR8-NEXT: mflr 031; PWR8-NEXT: stdu 1, -112(1)32; PWR8-NEXT: li 3, 4833; PWR8-NEXT: std 0, 128(1)34; PWR8-NEXT: std 30, 96(1) # 8-byte Folded Spill35; PWR8-NEXT: std 27, 72(1) # 8-byte Folded Spill36; PWR8-NEXT: std 28, 80(1) # 8-byte Folded Spill37; PWR8-NEXT: li 28, 1638; PWR8-NEXT: li 27, 8039; PWR8-NEXT: std 29, 88(1) # 8-byte Folded Spill40; PWR8-NEXT: stxvd2x 63, 1, 3 # 16-byte Folded Spill41; PWR8-NEXT: addis 3, 2, .L_MergedGlobals@toc@ha42; PWR8-NEXT: li 29, 443; PWR8-NEXT: addi 30, 3, .L_MergedGlobals@toc@l44; PWR8-NEXT: std 26, 64(1) # 8-byte Folded Spill45; PWR8-NEXT: lxvd2x 0, 0, 3046; PWR8-NEXT: xxswapd 63, 047; PWR8-NEXT: .p2align 448; PWR8-NEXT: .LBB0_1: # %for.body49; PWR8-NEXT: #50; PWR8-NEXT: lxvd2x 0, 30, 2851; PWR8-NEXT: vmr 2, 3152; PWR8-NEXT: addi 26, 30, 1653; PWR8-NEXT: xxswapd 35, 054; PWR8-NEXT: bl __mulkf355; PWR8-NEXT: nop56; PWR8-NEXT: addi 29, 29, -157; PWR8-NEXT: xxswapd 0, 3458; PWR8-NEXT: cmpldi 29, 059; PWR8-NEXT: stxvd2x 0, 30, 2760; PWR8-NEXT: mr 30, 2661; PWR8-NEXT: bc 12, 1, .LBB0_162; PWR8-NEXT: # %bb.2: # %for.end63; PWR8-NEXT: li 3, 4864; PWR8-NEXT: ld 30, 96(1) # 8-byte Folded Reload65; PWR8-NEXT: ld 29, 88(1) # 8-byte Folded Reload66; PWR8-NEXT: lxvd2x 63, 1, 3 # 16-byte Folded Reload67; PWR8-NEXT: ld 28, 80(1) # 8-byte Folded Reload68; PWR8-NEXT: ld 27, 72(1) # 8-byte Folded Reload69; PWR8-NEXT: ld 26, 64(1) # 8-byte Folded Reload70; PWR8-NEXT: addi 1, 1, 11271; PWR8-NEXT: ld 0, 16(1)72; PWR8-NEXT: mtlr 073; PWR8-NEXT: blr74entry:75 %0 = load fp128, ptr @a, align 1676 br label %for.body77 78for.body: ; preds = %for.body, %entry79 %i.06 = phi i64 [ 0, %entry ], [ %inc, %for.body ]80 %arrayidx = getelementptr inbounds [4 x fp128], ptr @x, i64 0, i64 %i.0681 %1 = load fp128, ptr %arrayidx, align 1682 %mul = fmul fp128 %0, %183 %arrayidx1 = getelementptr inbounds [4 x fp128], ptr @y, i64 0, i64 %i.0684 store fp128 %mul, ptr %arrayidx1, align 1685 %inc = add nuw nsw i64 %i.06, 186 %exitcond = icmp eq i64 %inc, 487 br i1 %exitcond, label %for.end, label %for.body88 89for.end: ; preds = %for.body90 ret void91}92 93define void @fpext_ctrloop_fp128(ptr %a) nounwind {94; PWR9-LABEL: fpext_ctrloop_fp128:95; PWR9: # %bb.0: # %entry96; PWR9-NEXT: li 4, 497; PWR9-NEXT: addi 3, 3, -898; PWR9-NEXT: mtctr 499; PWR9-NEXT: addis 4, 2, .L_MergedGlobals@toc@ha100; PWR9-NEXT: addi 4, 4, .L_MergedGlobals@toc@l101; PWR9-NEXT: addi 4, 4, 64102; PWR9-NEXT: .p2align 5103; PWR9-NEXT: .LBB1_1: # %for.body104; PWR9-NEXT: #105; PWR9-NEXT: lfdu 0, 8(3)106; PWR9-NEXT: xscpsgndp 34, 0, 0107; PWR9-NEXT: xscvdpqp 2, 2108; PWR9-NEXT: stxv 34, 16(4)109; PWR9-NEXT: addi 4, 4, 16110; PWR9-NEXT: bdnz .LBB1_1111; PWR9-NEXT: # %bb.2: # %for.end112; PWR9-NEXT: blr113;114; PWR8-LABEL: fpext_ctrloop_fp128:115; PWR8: # %bb.0: # %entry116; PWR8-NEXT: mflr 0117; PWR8-NEXT: std 28, -32(1) # 8-byte Folded Spill118; PWR8-NEXT: std 29, -24(1) # 8-byte Folded Spill119; PWR8-NEXT: std 30, -16(1) # 8-byte Folded Spill120; PWR8-NEXT: stdu 1, -64(1)121; PWR8-NEXT: addi 30, 3, -8122; PWR8-NEXT: addis 3, 2, .L_MergedGlobals@toc@ha123; PWR8-NEXT: li 29, 4124; PWR8-NEXT: std 0, 80(1)125; PWR8-NEXT: addi 3, 3, .L_MergedGlobals@toc@l126; PWR8-NEXT: addi 28, 3, 64127; PWR8-NEXT: .p2align 4128; PWR8-NEXT: .LBB1_1: # %for.body129; PWR8-NEXT: #130; PWR8-NEXT: lfdu 1, 8(30)131; PWR8-NEXT: addi 28, 28, 16132; PWR8-NEXT: bl __extenddfkf2133; PWR8-NEXT: nop134; PWR8-NEXT: addi 29, 29, -1135; PWR8-NEXT: xxswapd 0, 34136; PWR8-NEXT: cmpldi 29, 0137; PWR8-NEXT: stxvd2x 0, 0, 28138; PWR8-NEXT: bc 12, 1, .LBB1_1139; PWR8-NEXT: # %bb.2: # %for.end140; PWR8-NEXT: addi 1, 1, 64141; PWR8-NEXT: ld 0, 16(1)142; PWR8-NEXT: ld 30, -16(1) # 8-byte Folded Reload143; PWR8-NEXT: ld 29, -24(1) # 8-byte Folded Reload144; PWR8-NEXT: ld 28, -32(1) # 8-byte Folded Reload145; PWR8-NEXT: mtlr 0146; PWR8-NEXT: blr147entry:148 br label %for.body149 150for.body:151 %i.06 = phi i64 [ 0, %entry ], [ %inc, %for.body ]152 %arrayidx = getelementptr inbounds double, ptr %a, i64 %i.06153 %0 = load double, ptr %arrayidx, align 8154 %ext = fpext double %0 to fp128155 %arrayidx1 = getelementptr inbounds [4 x fp128], ptr @y, i64 0, i64 %i.06156 store fp128 %ext, ptr %arrayidx1, align 16157 %inc = add nuw nsw i64 %i.06, 1158 %exitcond = icmp eq i64 %inc, 4159 br i1 %exitcond, label %for.end, label %for.body160 161for.end:162 ret void163}164 165define void @fptrunc_ctrloop_fp128(ptr %a) nounwind {166; PWR9-LABEL: fptrunc_ctrloop_fp128:167; PWR9: # %bb.0: # %entry168; PWR9-NEXT: li 4, 4169; PWR9-NEXT: addi 3, 3, -8170; PWR9-NEXT: mtctr 4171; PWR9-NEXT: addis 4, 2, .L_MergedGlobals@toc@ha172; PWR9-NEXT: addi 4, 4, .L_MergedGlobals@toc@l173; PWR9-NEXT: .p2align 5174; PWR9-NEXT: .LBB2_1: # %for.body175; PWR9-NEXT: #176; PWR9-NEXT: lxv 34, 16(4)177; PWR9-NEXT: addi 4, 4, 16178; PWR9-NEXT: xscvqpdp 2, 2179; PWR9-NEXT: xscpsgndp 0, 34, 34180; PWR9-NEXT: stfdu 0, 8(3)181; PWR9-NEXT: bdnz .LBB2_1182; PWR9-NEXT: # %bb.2: # %for.end183; PWR9-NEXT: blr184;185; PWR8-LABEL: fptrunc_ctrloop_fp128:186; PWR8: # %bb.0: # %entry187; PWR8-NEXT: mflr 0188; PWR8-NEXT: std 28, -32(1) # 8-byte Folded Spill189; PWR8-NEXT: std 29, -24(1) # 8-byte Folded Spill190; PWR8-NEXT: std 30, -16(1) # 8-byte Folded Spill191; PWR8-NEXT: stdu 1, -64(1)192; PWR8-NEXT: addi 30, 3, -8193; PWR8-NEXT: addis 3, 2, .L_MergedGlobals@toc@ha194; PWR8-NEXT: li 29, 4195; PWR8-NEXT: std 0, 80(1)196; PWR8-NEXT: addi 28, 3, .L_MergedGlobals@toc@l197; PWR8-NEXT: .p2align 4198; PWR8-NEXT: .LBB2_1: # %for.body199; PWR8-NEXT: #200; PWR8-NEXT: addi 28, 28, 16201; PWR8-NEXT: lxvd2x 0, 0, 28202; PWR8-NEXT: xxswapd 34, 0203; PWR8-NEXT: bl __trunckfdf2204; PWR8-NEXT: nop205; PWR8-NEXT: addi 29, 29, -1206; PWR8-NEXT: stfdu 1, 8(30)207; PWR8-NEXT: cmpldi 29, 0208; PWR8-NEXT: bc 12, 1, .LBB2_1209; PWR8-NEXT: # %bb.2: # %for.end210; PWR8-NEXT: addi 1, 1, 64211; PWR8-NEXT: ld 0, 16(1)212; PWR8-NEXT: ld 30, -16(1) # 8-byte Folded Reload213; PWR8-NEXT: ld 29, -24(1) # 8-byte Folded Reload214; PWR8-NEXT: ld 28, -32(1) # 8-byte Folded Reload215; PWR8-NEXT: mtlr 0216; PWR8-NEXT: blr217entry:218 br label %for.body219 220for.body:221 %i.06 = phi i64 [ 0, %entry ], [ %inc, %for.body ]222 %arrayidx = getelementptr inbounds [4 x fp128], ptr @x, i64 0, i64 %i.06223 %0 = load fp128, ptr %arrayidx, align 16224 %trunc = fptrunc fp128 %0 to double225 %arrayidx1 = getelementptr inbounds double, ptr %a, i64 %i.06226 store double %trunc, ptr %arrayidx1, align 16227 %inc = add nuw nsw i64 %i.06, 1228 %exitcond = icmp eq i64 %inc, 4229 br i1 %exitcond, label %for.end, label %for.body230 231for.end:232 ret void233}234 235declare void @obfuscate(ptr, ...) local_unnamed_addr #2236