248 lines · plain
1; Test the saving and restoring of FPRs in large frames.2;3; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z10 | FileCheck -check-prefix=CHECK-NOFP %s4; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z10 -frame-pointer=all | FileCheck -check-prefix=CHECK-FP %s5 6; Test a frame size that requires some FPRs to be saved and loaded using7; the 20-bit STDY and LDY while others can use the 12-bit STD and LD.8; The frame is big enough to require two emergency spill slots at 160(%r15),9; as well as the 8 FPR save slots. Get a frame of size 4128 by allocating10; (4128 - 176 - 8 * 8) / 8 = 486 extra doublewords.11define void @f1(ptr %ptr, i64 %x) {12; CHECK-NOFP-LABEL: f1:13; CHECK-NOFP: aghi %r15, -412814; CHECK-NOFP: .cfi_def_cfa_offset 428815; CHECK-NOFP: stdy %f8, 4120(%r15)16; CHECK-NOFP: stdy %f9, 4112(%r15)17; CHECK-NOFP: stdy %f10, 4104(%r15)18; CHECK-NOFP: stdy %f11, 4096(%r15)19; CHECK-NOFP: std %f12, 4088(%r15)20; CHECK-NOFP: std %f13, 4080(%r15)21; CHECK-NOFP: std %f14, 4072(%r15)22; CHECK-NOFP: std %f15, 4064(%r15)23; CHECK-NOFP: .cfi_offset %f8, -16824; CHECK-NOFP: .cfi_offset %f9, -17625; CHECK-NOFP: .cfi_offset %f10, -18426; CHECK-NOFP: .cfi_offset %f11, -19227; CHECK-NOFP: .cfi_offset %f12, -20028; CHECK-NOFP: .cfi_offset %f13, -20829; CHECK-NOFP: .cfi_offset %f14, -21630; CHECK-NOFP: .cfi_offset %f15, -22431; ...main function body...32; CHECK-NOFP: ldy %f8, 4120(%r15)33; CHECK-NOFP: ldy %f9, 4112(%r15)34; CHECK-NOFP: ldy %f10, 4104(%r15)35; CHECK-NOFP: ldy %f11, 4096(%r15)36; CHECK-NOFP: ld %f12, 4088(%r15)37; CHECK-NOFP: ld %f13, 4080(%r15)38; CHECK-NOFP: ld %f14, 4072(%r15)39; CHECK-NOFP: ld %f15, 4064(%r15)40; CHECK-NOFP: aghi %r15, 412841; CHECK-NOFP: br %r1442;43; CHECK-FP-LABEL: f1:44; CHECK-FP: stmg %r11, %r15, 88(%r15)45; CHECK-FP: aghi %r15, -412846; CHECK-FP: .cfi_def_cfa_offset 428847; CHECK-FP: lgr %r11, %r1548; CHECK-FP: .cfi_def_cfa_register %r1149; CHECK-FP: stdy %f8, 4120(%r11)50; CHECK-FP: stdy %f9, 4112(%r11)51; CHECK-FP: stdy %f10, 4104(%r11)52; CHECK-FP: stdy %f11, 4096(%r11)53; CHECK-FP: std %f12, 4088(%r11)54; CHECK-FP: std %f13, 4080(%r11)55; CHECK-FP: std %f14, 4072(%r11)56; CHECK-FP: std %f15, 4064(%r11)57; ...main function body...58; CHECK-FP: ldy %f8, 4120(%r11)59; CHECK-FP: ldy %f9, 4112(%r11)60; CHECK-FP: ldy %f10, 4104(%r11)61; CHECK-FP: ldy %f11, 4096(%r11)62; CHECK-FP: ld %f12, 4088(%r11)63; CHECK-FP: ld %f13, 4080(%r11)64; CHECK-FP: ld %f14, 4072(%r11)65; CHECK-FP: ld %f15, 4064(%r11)66; CHECK-FP: lmg %r11, %r15, 4216(%r11)67; CHECK-FP: br %r1468 %y = alloca [486 x i64], align 869 store volatile i64 %x, ptr %y70 %l0 = load volatile double, ptr %ptr71 %l1 = load volatile double, ptr %ptr72 %l2 = load volatile double, ptr %ptr73 %l3 = load volatile double, ptr %ptr74 %l4 = load volatile double, ptr %ptr75 %l5 = load volatile double, ptr %ptr76 %l6 = load volatile double, ptr %ptr77 %l7 = load volatile double, ptr %ptr78 %l8 = load volatile double, ptr %ptr79 %l9 = load volatile double, ptr %ptr80 %l10 = load volatile double, ptr %ptr81 %l11 = load volatile double, ptr %ptr82 %l12 = load volatile double, ptr %ptr83 %l13 = load volatile double, ptr %ptr84 %l14 = load volatile double, ptr %ptr85 %l15 = load volatile double, ptr %ptr86 %add0 = fadd double %l0, %l087 %add1 = fadd double %l1, %add088 %add2 = fadd double %l2, %add189 %add3 = fadd double %l3, %add290 %add4 = fadd double %l4, %add391 %add5 = fadd double %l5, %add492 %add6 = fadd double %l6, %add593 %add7 = fadd double %l7, %add694 %add8 = fadd double %l8, %add795 %add9 = fadd double %l9, %add896 %add10 = fadd double %l10, %add997 %add11 = fadd double %l11, %add1098 %add12 = fadd double %l12, %add1199 %add13 = fadd double %l13, %add12100 %add14 = fadd double %l14, %add13101 %add15 = fadd double %l15, %add14102 store volatile double %add0, ptr %ptr103 store volatile double %add1, ptr %ptr104 store volatile double %add2, ptr %ptr105 store volatile double %add3, ptr %ptr106 store volatile double %add4, ptr %ptr107 store volatile double %add5, ptr %ptr108 store volatile double %add6, ptr %ptr109 store volatile double %add7, ptr %ptr110 store volatile double %add8, ptr %ptr111 store volatile double %add9, ptr %ptr112 store volatile double %add10, ptr %ptr113 store volatile double %add11, ptr %ptr114 store volatile double %add12, ptr %ptr115 store volatile double %add13, ptr %ptr116 store volatile double %add14, ptr %ptr117 store volatile double %add15, ptr %ptr118 ret void119}120 121; Test a frame size that requires some FPRs to be saved and loaded using122; an indexed STD and LD while others can use the 20-bit STDY and LDY.123; The index can be any call-clobbered GPR except %r0.124;125; Don't require the accesses to share the same LLILH; that would be a126; good optimisation but is really a different test.127;128; As above, get a frame of size 524320 by allocating129; (524320 - 176 - 8 * 8) / 8 = 65510 extra doublewords.130define void @f2(ptr %ptr, i64 %x) {131; CHECK-NOFP-LABEL: f2:132; CHECK-NOFP: agfi %r15, -524320133; CHECK-NOFP: .cfi_def_cfa_offset 524480134; CHECK-NOFP: llilh [[INDEX:%r[1-5]]], 8135; CHECK-NOFP: std %f8, 24([[INDEX]],%r15)136; CHECK-NOFP: std %f9, 16({{%r[1-5]}},%r15)137; CHECK-NOFP: std %f10, 8({{%r[1-5]}},%r15)138; CHECK-NOFP: std %f11, 0({{%r[1-5]}},%r15)139; CHECK-NOFP: stdy %f12, 524280(%r15)140; CHECK-NOFP: stdy %f13, 524272(%r15)141; CHECK-NOFP: stdy %f14, 524264(%r15)142; CHECK-NOFP: stdy %f15, 524256(%r15)143; CHECK-NOFP: .cfi_offset %f8, -168144; CHECK-NOFP: .cfi_offset %f9, -176145; CHECK-NOFP: .cfi_offset %f10, -184146; CHECK-NOFP: .cfi_offset %f11, -192147; CHECK-NOFP: .cfi_offset %f12, -200148; CHECK-NOFP: .cfi_offset %f13, -208149; CHECK-NOFP: .cfi_offset %f14, -216150; CHECK-NOFP: .cfi_offset %f15, -224151; ...main function body...152; CHECK-NOFP: ld %f8, 24({{%r[1-5]}},%r15)153; CHECK-NOFP: ld %f9, 16({{%r[1-5]}},%r15)154; CHECK-NOFP: ld %f10, 8({{%r[1-5]}},%r15)155; CHECK-NOFP: ld %f11, 0({{%r[1-5]}},%r15)156; CHECK-NOFP: ldy %f12, 524280(%r15)157; CHECK-NOFP: ldy %f13, 524272(%r15)158; CHECK-NOFP: ldy %f14, 524264(%r15)159; CHECK-NOFP: ldy %f15, 524256(%r15)160; CHECK-NOFP: agfi %r15, 524320161; CHECK-NOFP: br %r14162;163; CHECK-FP-LABEL: f2:164; CHECK-FP: stmg %r11, %r15, 88(%r15)165; CHECK-FP: agfi %r15, -524320166; CHECK-FP: .cfi_def_cfa_offset 524480167; CHECK-FP: llilh [[INDEX:%r[1-5]]], 8168; CHECK-FP: std %f8, 24([[INDEX]],%r11)169; CHECK-FP: std %f9, 16({{%r[1-5]}},%r11)170; CHECK-FP: std %f10, 8({{%r[1-5]}},%r11)171; CHECK-FP: std %f11, 0({{%r[1-5]}},%r11)172; CHECK-FP: stdy %f12, 524280(%r11)173; CHECK-FP: stdy %f13, 524272(%r11)174; CHECK-FP: stdy %f14, 524264(%r11)175; CHECK-FP: stdy %f15, 524256(%r11)176; CHECK-FP: .cfi_offset %f8, -168177; CHECK-FP: .cfi_offset %f9, -176178; CHECK-FP: .cfi_offset %f10, -184179; CHECK-FP: .cfi_offset %f11, -192180; CHECK-FP: .cfi_offset %f12, -200181; CHECK-FP: .cfi_offset %f13, -208182; CHECK-FP: .cfi_offset %f14, -216183; CHECK-FP: .cfi_offset %f15, -224184; ...main function body...185; CHECK-FP: ld %f8, 24({{%r[1-5]}},%r11)186; CHECK-FP: ld %f9, 16({{%r[1-5]}},%r11)187; CHECK-FP: ld %f10, 8({{%r[1-5]}},%r11)188; CHECK-FP: ld %f11, 0({{%r[1-5]}},%r11)189; CHECK-FP: ldy %f12, 524280(%r11)190; CHECK-FP: ldy %f13, 524272(%r11)191; CHECK-FP: ldy %f14, 524264(%r11)192; CHECK-FP: ldy %f15, 524256(%r11)193; CHECK-FP: aghi %r11, 128194; CHECK-FP: lmg %r11, %r15, 524280(%r11)195; CHECK-FP: br %r14196 %y = alloca [65510 x i64], align 8197 store volatile i64 %x, ptr %y198 %l0 = load volatile double, ptr %ptr199 %l1 = load volatile double, ptr %ptr200 %l2 = load volatile double, ptr %ptr201 %l3 = load volatile double, ptr %ptr202 %l4 = load volatile double, ptr %ptr203 %l5 = load volatile double, ptr %ptr204 %l6 = load volatile double, ptr %ptr205 %l7 = load volatile double, ptr %ptr206 %l8 = load volatile double, ptr %ptr207 %l9 = load volatile double, ptr %ptr208 %l10 = load volatile double, ptr %ptr209 %l11 = load volatile double, ptr %ptr210 %l12 = load volatile double, ptr %ptr211 %l13 = load volatile double, ptr %ptr212 %l14 = load volatile double, ptr %ptr213 %l15 = load volatile double, ptr %ptr214 %add0 = fadd double %l0, %l0215 %add1 = fadd double %l1, %add0216 %add2 = fadd double %l2, %add1217 %add3 = fadd double %l3, %add2218 %add4 = fadd double %l4, %add3219 %add5 = fadd double %l5, %add4220 %add6 = fadd double %l6, %add5221 %add7 = fadd double %l7, %add6222 %add8 = fadd double %l8, %add7223 %add9 = fadd double %l9, %add8224 %add10 = fadd double %l10, %add9225 %add11 = fadd double %l11, %add10226 %add12 = fadd double %l12, %add11227 %add13 = fadd double %l13, %add12228 %add14 = fadd double %l14, %add13229 %add15 = fadd double %l15, %add14230 store volatile double %add0, ptr %ptr231 store volatile double %add1, ptr %ptr232 store volatile double %add2, ptr %ptr233 store volatile double %add3, ptr %ptr234 store volatile double %add4, ptr %ptr235 store volatile double %add5, ptr %ptr236 store volatile double %add6, ptr %ptr237 store volatile double %add7, ptr %ptr238 store volatile double %add8, ptr %ptr239 store volatile double %add9, ptr %ptr240 store volatile double %add10, ptr %ptr241 store volatile double %add11, ptr %ptr242 store volatile double %add12, ptr %ptr243 store volatile double %add13, ptr %ptr244 store volatile double %add14, ptr %ptr245 store volatile double %add15, ptr %ptr246 ret void247}248