370 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -O1 -mtriple=riscv64 -mattr=+d,+zfh,+zfbfmin -verify-machineinstrs < %s | FileCheck %s3 4@a = common global i32 0, align 45@l = common global i32 0, align 46@b = common global i32 0, align 47@c = common global i32 0, align 48@d = common global i32 0, align 49@e = common global i32 0, align 410@k = common global i32 0, align 411@f = common global i32 0, align 412@j = common global i32 0, align 413@g = common global i32 0, align 414@i = common global i32 0, align 415@h = common global i32 0, align 416 17; This test case benefits from codegen recognising that some values are18; trivially rematerialisable, meaning they are recreated rather than saved to19; the stack and restored. It creates high register pressure to force this20; situation.21 22define i32 @test() nounwind {23; CHECK-LABEL: test:24; CHECK: # %bb.0: # %entry25; CHECK-NEXT: addi sp, sp, -11226; CHECK-NEXT: sd ra, 104(sp) # 8-byte Folded Spill27; CHECK-NEXT: sd s0, 96(sp) # 8-byte Folded Spill28; CHECK-NEXT: sd s1, 88(sp) # 8-byte Folded Spill29; CHECK-NEXT: sd s2, 80(sp) # 8-byte Folded Spill30; CHECK-NEXT: sd s3, 72(sp) # 8-byte Folded Spill31; CHECK-NEXT: sd s4, 64(sp) # 8-byte Folded Spill32; CHECK-NEXT: sd s5, 56(sp) # 8-byte Folded Spill33; CHECK-NEXT: sd s6, 48(sp) # 8-byte Folded Spill34; CHECK-NEXT: sd s7, 40(sp) # 8-byte Folded Spill35; CHECK-NEXT: sd s8, 32(sp) # 8-byte Folded Spill36; CHECK-NEXT: sd s9, 24(sp) # 8-byte Folded Spill37; CHECK-NEXT: sd s10, 16(sp) # 8-byte Folded Spill38; CHECK-NEXT: sd s11, 8(sp) # 8-byte Folded Spill39; CHECK-NEXT: lui s0, %hi(a)40; CHECK-NEXT: lw a0, %lo(a)(s0)41; CHECK-NEXT: beqz a0, .LBB0_1142; CHECK-NEXT: # %bb.1: # %for.body.preheader43; CHECK-NEXT: lui s1, %hi(l)44; CHECK-NEXT: lui s2, %hi(k)45; CHECK-NEXT: lui s3, %hi(j)46; CHECK-NEXT: lui s4, %hi(i)47; CHECK-NEXT: lui s5, %hi(d)48; CHECK-NEXT: lui s6, %hi(e)49; CHECK-NEXT: lui s7, %hi(f)50; CHECK-NEXT: lui s8, %hi(g)51; CHECK-NEXT: lui s9, %hi(h)52; CHECK-NEXT: lui s10, %hi(c)53; CHECK-NEXT: lui s11, %hi(b)54; CHECK-NEXT: j .LBB0_355; CHECK-NEXT: .LBB0_2: # %for.inc56; CHECK-NEXT: # in Loop: Header=BB0_3 Depth=157; CHECK-NEXT: lw a0, %lo(a)(s0)58; CHECK-NEXT: addiw a0, a0, -159; CHECK-NEXT: sw a0, %lo(a)(s0)60; CHECK-NEXT: beqz a0, .LBB0_1161; CHECK-NEXT: .LBB0_3: # %for.body62; CHECK-NEXT: # =>This Inner Loop Header: Depth=163; CHECK-NEXT: lw a1, %lo(l)(s1)64; CHECK-NEXT: beqz a1, .LBB0_565; CHECK-NEXT: # %bb.4: # %if.then66; CHECK-NEXT: # in Loop: Header=BB0_3 Depth=167; CHECK-NEXT: lw a4, %lo(e)(s6)68; CHECK-NEXT: lw a3, %lo(d)(s5)69; CHECK-NEXT: lw a2, %lo(c)(s10)70; CHECK-NEXT: lw a1, %lo(b)(s11)71; CHECK-NEXT: li a5, 3272; CHECK-NEXT: call foo73; CHECK-NEXT: .LBB0_5: # %if.end74; CHECK-NEXT: # in Loop: Header=BB0_3 Depth=175; CHECK-NEXT: lw a0, %lo(k)(s2)76; CHECK-NEXT: beqz a0, .LBB0_777; CHECK-NEXT: # %bb.6: # %if.then378; CHECK-NEXT: # in Loop: Header=BB0_3 Depth=179; CHECK-NEXT: lw a4, %lo(f)(s7)80; CHECK-NEXT: lw a3, %lo(e)(s6)81; CHECK-NEXT: lw a2, %lo(d)(s5)82; CHECK-NEXT: lw a1, %lo(c)(s10)83; CHECK-NEXT: lw a0, %lo(b)(s11)84; CHECK-NEXT: li a5, 6485; CHECK-NEXT: call foo86; CHECK-NEXT: .LBB0_7: # %if.end587; CHECK-NEXT: # in Loop: Header=BB0_3 Depth=188; CHECK-NEXT: lw a0, %lo(j)(s3)89; CHECK-NEXT: beqz a0, .LBB0_990; CHECK-NEXT: # %bb.8: # %if.then791; CHECK-NEXT: # in Loop: Header=BB0_3 Depth=192; CHECK-NEXT: lw a4, %lo(g)(s8)93; CHECK-NEXT: lw a3, %lo(f)(s7)94; CHECK-NEXT: lw a2, %lo(e)(s6)95; CHECK-NEXT: lw a1, %lo(d)(s5)96; CHECK-NEXT: lw a0, %lo(c)(s10)97; CHECK-NEXT: li a5, 3298; CHECK-NEXT: call foo99; CHECK-NEXT: .LBB0_9: # %if.end9100; CHECK-NEXT: # in Loop: Header=BB0_3 Depth=1101; CHECK-NEXT: lw a0, %lo(i)(s4)102; CHECK-NEXT: beqz a0, .LBB0_2103; CHECK-NEXT: # %bb.10: # %if.then11104; CHECK-NEXT: # in Loop: Header=BB0_3 Depth=1105; CHECK-NEXT: lw a4, %lo(h)(s9)106; CHECK-NEXT: lw a3, %lo(g)(s8)107; CHECK-NEXT: lw a2, %lo(f)(s7)108; CHECK-NEXT: lw a1, %lo(e)(s6)109; CHECK-NEXT: lw a0, %lo(d)(s5)110; CHECK-NEXT: li a5, 32111; CHECK-NEXT: call foo112; CHECK-NEXT: j .LBB0_2113; CHECK-NEXT: .LBB0_11: # %for.end114; CHECK-NEXT: li a0, 1115; CHECK-NEXT: ld ra, 104(sp) # 8-byte Folded Reload116; CHECK-NEXT: ld s0, 96(sp) # 8-byte Folded Reload117; CHECK-NEXT: ld s1, 88(sp) # 8-byte Folded Reload118; CHECK-NEXT: ld s2, 80(sp) # 8-byte Folded Reload119; CHECK-NEXT: ld s3, 72(sp) # 8-byte Folded Reload120; CHECK-NEXT: ld s4, 64(sp) # 8-byte Folded Reload121; CHECK-NEXT: ld s5, 56(sp) # 8-byte Folded Reload122; CHECK-NEXT: ld s6, 48(sp) # 8-byte Folded Reload123; CHECK-NEXT: ld s7, 40(sp) # 8-byte Folded Reload124; CHECK-NEXT: ld s8, 32(sp) # 8-byte Folded Reload125; CHECK-NEXT: ld s9, 24(sp) # 8-byte Folded Reload126; CHECK-NEXT: ld s10, 16(sp) # 8-byte Folded Reload127; CHECK-NEXT: ld s11, 8(sp) # 8-byte Folded Reload128; CHECK-NEXT: addi sp, sp, 112129; CHECK-NEXT: ret130entry:131 %.pr = load i32, ptr @a, align 4132 %tobool14 = icmp eq i32 %.pr, 0133 br i1 %tobool14, label %for.end, label %for.body134 135for.body: ; preds = %entry, %for.inc136 %0 = phi i32 [ %dec, %for.inc ], [ %.pr, %entry ]137 %1 = load i32, ptr @l, align 4138 %tobool1 = icmp eq i32 %1, 0139 br i1 %tobool1, label %if.end, label %if.then140 141if.then: ; preds = %for.body142 %2 = load i32, ptr @b, align 4143 %3 = load i32, ptr @c, align 4144 %4 = load i32, ptr @d, align 4145 %5 = load i32, ptr @e, align 4146 %call = tail call i32 @foo(i32 %0, i32 %2, i32 %3, i32 %4, i32 %5, i32 32)147 br label %if.end148 149if.end: ; preds = %for.body, %if.then150 %6 = load i32, ptr @k, align 4151 %tobool2 = icmp eq i32 %6, 0152 br i1 %tobool2, label %if.end5, label %if.then3153 154if.then3: ; preds = %if.end155 %7 = load i32, ptr @b, align 4156 %8 = load i32, ptr @c, align 4157 %9 = load i32, ptr @d, align 4158 %10 = load i32, ptr @e, align 4159 %11 = load i32, ptr @f, align 4160 %call4 = tail call i32 @foo(i32 %7, i32 %8, i32 %9, i32 %10, i32 %11, i32 64)161 br label %if.end5162 163if.end5: ; preds = %if.end, %if.then3164 %12 = load i32, ptr @j, align 4165 %tobool6 = icmp eq i32 %12, 0166 br i1 %tobool6, label %if.end9, label %if.then7167 168if.then7: ; preds = %if.end5169 %13 = load i32, ptr @c, align 4170 %14 = load i32, ptr @d, align 4171 %15 = load i32, ptr @e, align 4172 %16 = load i32, ptr @f, align 4173 %17 = load i32, ptr @g, align 4174 %call8 = tail call i32 @foo(i32 %13, i32 %14, i32 %15, i32 %16, i32 %17, i32 32)175 br label %if.end9176 177if.end9: ; preds = %if.end5, %if.then7178 %18 = load i32, ptr @i, align 4179 %tobool10 = icmp eq i32 %18, 0180 br i1 %tobool10, label %for.inc, label %if.then11181 182if.then11: ; preds = %if.end9183 %19 = load i32, ptr @d, align 4184 %20 = load i32, ptr @e, align 4185 %21 = load i32, ptr @f, align 4186 %22 = load i32, ptr @g, align 4187 %23 = load i32, ptr @h, align 4188 %call12 = tail call i32 @foo(i32 %19, i32 %20, i32 %21, i32 %22, i32 %23, i32 32)189 br label %for.inc190 191for.inc: ; preds = %if.end9, %if.then11192 %24 = load i32, ptr @a, align 4193 %dec = add nsw i32 %24, -1194 store i32 %dec, ptr @a, align 4195 %tobool = icmp eq i32 %dec, 0196 br i1 %tobool, label %for.end, label %for.body197 198for.end: ; preds = %for.inc, %entry199 ret i32 1200}201 202declare i32 @foo(i32, i32, i32, i32, i32, i32)203 204define void @remat_load(i32 %0, i32 %1, i32 %2, i32 %3, i32 %4, i32 %5, i32 %6, i32 %7, double %8, double %9, double %10, double %11, double %12, double %13, double %14, double %15, i8 %stackarg0, i16 %stackarg1, i32 %stackarg2, i64 %stackarg3, half %stackarg4, bfloat %stackarg5, float %stackarg6, double %stackarg7, ptr %p) nounwind {205; CHECK-LABEL: remat_load:206; CHECK: # %bb.0: # %entry207; CHECK-NEXT: addi sp, sp, -208208; CHECK-NEXT: sd ra, 200(sp) # 8-byte Folded Spill209; CHECK-NEXT: sd s0, 192(sp) # 8-byte Folded Spill210; CHECK-NEXT: sd s1, 184(sp) # 8-byte Folded Spill211; CHECK-NEXT: sd s2, 176(sp) # 8-byte Folded Spill212; CHECK-NEXT: sd s3, 168(sp) # 8-byte Folded Spill213; CHECK-NEXT: sd s4, 160(sp) # 8-byte Folded Spill214; CHECK-NEXT: sd s5, 152(sp) # 8-byte Folded Spill215; CHECK-NEXT: sd s6, 144(sp) # 8-byte Folded Spill216; CHECK-NEXT: sd s7, 136(sp) # 8-byte Folded Spill217; CHECK-NEXT: sd s8, 128(sp) # 8-byte Folded Spill218; CHECK-NEXT: sd s9, 120(sp) # 8-byte Folded Spill219; CHECK-NEXT: sd s10, 112(sp) # 8-byte Folded Spill220; CHECK-NEXT: sd s11, 104(sp) # 8-byte Folded Spill221; CHECK-NEXT: fsd fs0, 96(sp) # 8-byte Folded Spill222; CHECK-NEXT: fsd fs1, 88(sp) # 8-byte Folded Spill223; CHECK-NEXT: fsd fs2, 80(sp) # 8-byte Folded Spill224; CHECK-NEXT: fsd fs3, 72(sp) # 8-byte Folded Spill225; CHECK-NEXT: fsd fs4, 64(sp) # 8-byte Folded Spill226; CHECK-NEXT: fsd fs5, 56(sp) # 8-byte Folded Spill227; CHECK-NEXT: fsd fs6, 48(sp) # 8-byte Folded Spill228; CHECK-NEXT: fsd fs7, 40(sp) # 8-byte Folded Spill229; CHECK-NEXT: fsd fs8, 32(sp) # 8-byte Folded Spill230; CHECK-NEXT: fsd fs9, 24(sp) # 8-byte Folded Spill231; CHECK-NEXT: fsd fs10, 16(sp) # 8-byte Folded Spill232; CHECK-NEXT: fsd fs11, 8(sp) # 8-byte Folded Spill233; CHECK-NEXT: fld fa5, 264(sp)234; CHECK-NEXT: flw fa4, 256(sp)235; CHECK-NEXT: flh fa3, 248(sp)236; CHECK-NEXT: flh fa2, 240(sp)237; CHECK-NEXT: ld a0, 272(sp)238; CHECK-NEXT: lbu a4, 208(sp)239; CHECK-NEXT: lh a3, 216(sp)240; CHECK-NEXT: lw a2, 224(sp)241; CHECK-NEXT: ld a1, 232(sp)242; CHECK-NEXT: sb a4, 0(a0)243; CHECK-NEXT: sh a3, 0(a0)244; CHECK-NEXT: sw a2, 0(a0)245; CHECK-NEXT: sd a1, 0(a0)246; CHECK-NEXT: fsh fa2, 0(a0)247; CHECK-NEXT: fsh fa3, 0(a0)248; CHECK-NEXT: fsw fa4, 0(a0)249; CHECK-NEXT: fsd fa5, 0(a0)250; CHECK-NEXT: #APP251; CHECK-NEXT: #NO_APP252; CHECK-NEXT: ld a0, 272(sp)253; CHECK-NEXT: lbu a1, 208(sp)254; CHECK-NEXT: sb a1, 0(a0)255; CHECK-NEXT: lh a1, 216(sp)256; CHECK-NEXT: sh a1, 0(a0)257; CHECK-NEXT: lw a1, 224(sp)258; CHECK-NEXT: sw a1, 0(a0)259; CHECK-NEXT: ld a1, 232(sp)260; CHECK-NEXT: sd a1, 0(a0)261; CHECK-NEXT: flh fa5, 240(sp)262; CHECK-NEXT: fsh fa5, 0(a0)263; CHECK-NEXT: flh fa5, 248(sp)264; CHECK-NEXT: fsh fa5, 0(a0)265; CHECK-NEXT: flw fa5, 256(sp)266; CHECK-NEXT: fsw fa5, 0(a0)267; CHECK-NEXT: fld fa5, 264(sp)268; CHECK-NEXT: fsd fa5, 0(a0)269; CHECK-NEXT: ld ra, 200(sp) # 8-byte Folded Reload270; CHECK-NEXT: ld s0, 192(sp) # 8-byte Folded Reload271; CHECK-NEXT: ld s1, 184(sp) # 8-byte Folded Reload272; CHECK-NEXT: ld s2, 176(sp) # 8-byte Folded Reload273; CHECK-NEXT: ld s3, 168(sp) # 8-byte Folded Reload274; CHECK-NEXT: ld s4, 160(sp) # 8-byte Folded Reload275; CHECK-NEXT: ld s5, 152(sp) # 8-byte Folded Reload276; CHECK-NEXT: ld s6, 144(sp) # 8-byte Folded Reload277; CHECK-NEXT: ld s7, 136(sp) # 8-byte Folded Reload278; CHECK-NEXT: ld s8, 128(sp) # 8-byte Folded Reload279; CHECK-NEXT: ld s9, 120(sp) # 8-byte Folded Reload280; CHECK-NEXT: ld s10, 112(sp) # 8-byte Folded Reload281; CHECK-NEXT: ld s11, 104(sp) # 8-byte Folded Reload282; CHECK-NEXT: fld fs0, 96(sp) # 8-byte Folded Reload283; CHECK-NEXT: fld fs1, 88(sp) # 8-byte Folded Reload284; CHECK-NEXT: fld fs2, 80(sp) # 8-byte Folded Reload285; CHECK-NEXT: fld fs3, 72(sp) # 8-byte Folded Reload286; CHECK-NEXT: fld fs4, 64(sp) # 8-byte Folded Reload287; CHECK-NEXT: fld fs5, 56(sp) # 8-byte Folded Reload288; CHECK-NEXT: fld fs6, 48(sp) # 8-byte Folded Reload289; CHECK-NEXT: fld fs7, 40(sp) # 8-byte Folded Reload290; CHECK-NEXT: fld fs8, 32(sp) # 8-byte Folded Reload291; CHECK-NEXT: fld fs9, 24(sp) # 8-byte Folded Reload292; CHECK-NEXT: fld fs10, 16(sp) # 8-byte Folded Reload293; CHECK-NEXT: fld fs11, 8(sp) # 8-byte Folded Reload294; CHECK-NEXT: addi sp, sp, 208295; CHECK-NEXT: ret296entry:297 ; Add a use of the stack arguments here so that we will have to load them from298 ; the stack before the inline asm. Otherwise we would be exercising the299 ; machine scheduler, not rematerialization.300 store volatile i8 %stackarg0, ptr %p301 store volatile i16 %stackarg1, ptr %p302 store volatile i32 %stackarg2, ptr %p303 store volatile i64 %stackarg3, ptr %p304 store volatile half %stackarg4, ptr %p305 store volatile bfloat %stackarg5, ptr %p306 store volatile float %stackarg6, ptr %p307 store volatile double %stackarg7, ptr %p308 tail call void asm sideeffect "", "~{x1},~{x3},~{x4},~{x5},~{x6},~{x7},~{x8},~{x9},~{x10},~{x11},~{x12},~{x13},~{x14},~{x15},~{x16},~{x17},~{x18},~{x19},~{x20},~{x21},~{x22},~{x23},~{x24},~{x25},~{x26},~{x27},~{x28},~{x29},~{x30},~{x31},~{f0},~{f1},~{f2},~{f3},~{f4},~{f5},~{f6},~{f7},~{f8},~{f9},~{f10},~{f11},~{f12},~{f13},~{f14},~{f15},~{f16},~{f17},~{f18},~{f19},~{f20},~{f21},~{f22},~{f23},~{f24},~{f25},~{f26},~{f27},~{f28},~{f29},~{f30},~{f31}"()309 ; Now use them after spilling everything to force rematerialization310 store volatile i8 %stackarg0, ptr %p311 store volatile i16 %stackarg1, ptr %p312 store volatile i32 %stackarg2, ptr %p313 store volatile i64 %stackarg3, ptr %p314 store volatile half %stackarg4, ptr %p315 store volatile bfloat %stackarg5, ptr %p316 store volatile float %stackarg6, ptr %p317 store volatile double %stackarg7, ptr %p318 ret void319}320 321; We could remat the load of the constant global if we extended the live322; interval of the high bits of the address.323 324@const = external constant i32325define i32 @constglobal_load() nounwind {326; CHECK-LABEL: constglobal_load:327; CHECK: # %bb.0: # %entry328; CHECK-NEXT: addi sp, sp, -112329; CHECK-NEXT: sd ra, 104(sp) # 8-byte Folded Spill330; CHECK-NEXT: sd s0, 96(sp) # 8-byte Folded Spill331; CHECK-NEXT: sd s1, 88(sp) # 8-byte Folded Spill332; CHECK-NEXT: sd s2, 80(sp) # 8-byte Folded Spill333; CHECK-NEXT: sd s3, 72(sp) # 8-byte Folded Spill334; CHECK-NEXT: sd s4, 64(sp) # 8-byte Folded Spill335; CHECK-NEXT: sd s5, 56(sp) # 8-byte Folded Spill336; CHECK-NEXT: sd s6, 48(sp) # 8-byte Folded Spill337; CHECK-NEXT: sd s7, 40(sp) # 8-byte Folded Spill338; CHECK-NEXT: sd s8, 32(sp) # 8-byte Folded Spill339; CHECK-NEXT: sd s9, 24(sp) # 8-byte Folded Spill340; CHECK-NEXT: sd s10, 16(sp) # 8-byte Folded Spill341; CHECK-NEXT: sd s11, 8(sp) # 8-byte Folded Spill342; CHECK-NEXT: lui a0, %hi(const)343; CHECK-NEXT: lw a0, %lo(const)(a0)344; CHECK-NEXT: sd a0, 0(sp) # 8-byte Folded Spill345; CHECK-NEXT: #APP346; CHECK-NEXT: #NO_APP347; CHECK-NEXT: ld a0, 0(sp) # 8-byte Folded Reload348; CHECK-NEXT: addiw a0, a0, 1349; CHECK-NEXT: ld ra, 104(sp) # 8-byte Folded Reload350; CHECK-NEXT: ld s0, 96(sp) # 8-byte Folded Reload351; CHECK-NEXT: ld s1, 88(sp) # 8-byte Folded Reload352; CHECK-NEXT: ld s2, 80(sp) # 8-byte Folded Reload353; CHECK-NEXT: ld s3, 72(sp) # 8-byte Folded Reload354; CHECK-NEXT: ld s4, 64(sp) # 8-byte Folded Reload355; CHECK-NEXT: ld s5, 56(sp) # 8-byte Folded Reload356; CHECK-NEXT: ld s6, 48(sp) # 8-byte Folded Reload357; CHECK-NEXT: ld s7, 40(sp) # 8-byte Folded Reload358; CHECK-NEXT: ld s8, 32(sp) # 8-byte Folded Reload359; CHECK-NEXT: ld s9, 24(sp) # 8-byte Folded Reload360; CHECK-NEXT: ld s10, 16(sp) # 8-byte Folded Reload361; CHECK-NEXT: ld s11, 8(sp) # 8-byte Folded Reload362; CHECK-NEXT: addi sp, sp, 112363; CHECK-NEXT: ret364entry:365 %global = load i32, ptr @const366 tail call void asm sideeffect "", "~{x1},~{x3},~{x4},~{x5},~{x6},~{x7},~{x8},~{x9},~{x10},~{x11},~{x12},~{x13},~{x14},~{x15},~{x16},~{x17},~{x18},~{x19},~{x20},~{x21},~{x22},~{x23},~{x24},~{x25},~{x26},~{x27},~{x28},~{x29},~{x30},~{x31}"()367 %a = add i32 %global, 1368 ret i32 %a369}370