199 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -O3 -disable-peephole -mtriple=x86_64-unknown-unknown < %s | FileCheck %s3 4target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"5target triple = "x86_64-unknown-unknown"6 7; Stack reload folding tests.8;9; By including a nop call with sideeffects we can force a partial register spill of the10; relevant registers and check that the reload is correctly folded into the instruction.11 12;TODO stack_fold_bsf_i1613declare i16 @llvm.cttz.i16(i16, i1)14 15define i32 @stack_fold_bsf_i32(i32 %a0) {16; CHECK-LABEL: stack_fold_bsf_i32:17; CHECK: # %bb.0:18; CHECK-NEXT: pushq %rbp19; CHECK-NEXT: .cfi_def_cfa_offset 1620; CHECK-NEXT: pushq %r1521; CHECK-NEXT: .cfi_def_cfa_offset 2422; CHECK-NEXT: pushq %r1423; CHECK-NEXT: .cfi_def_cfa_offset 3224; CHECK-NEXT: pushq %r1325; CHECK-NEXT: .cfi_def_cfa_offset 4026; CHECK-NEXT: pushq %r1227; CHECK-NEXT: .cfi_def_cfa_offset 4828; CHECK-NEXT: pushq %rbx29; CHECK-NEXT: .cfi_def_cfa_offset 5630; CHECK-NEXT: .cfi_offset %rbx, -5631; CHECK-NEXT: .cfi_offset %r12, -4832; CHECK-NEXT: .cfi_offset %r13, -4033; CHECK-NEXT: .cfi_offset %r14, -3234; CHECK-NEXT: .cfi_offset %r15, -2435; CHECK-NEXT: .cfi_offset %rbp, -1636; CHECK-NEXT: movl %edi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill37; CHECK-NEXT: #APP38; CHECK-NEXT: nop39; CHECK-NEXT: #NO_APP40; CHECK-NEXT: rep bsfl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 4-byte Folded Reload41; CHECK-NEXT: popq %rbx42; CHECK-NEXT: .cfi_def_cfa_offset 4843; CHECK-NEXT: popq %r1244; CHECK-NEXT: .cfi_def_cfa_offset 4045; CHECK-NEXT: popq %r1346; CHECK-NEXT: .cfi_def_cfa_offset 3247; CHECK-NEXT: popq %r1448; CHECK-NEXT: .cfi_def_cfa_offset 2449; CHECK-NEXT: popq %r1550; CHECK-NEXT: .cfi_def_cfa_offset 1651; CHECK-NEXT: popq %rbp52; CHECK-NEXT: .cfi_def_cfa_offset 853; CHECK-NEXT: retq54 %1 = tail call i64 asm sideeffect "nop", "=x,~{rax},~{rbx},~{rcx},~{rdx},~{rsi},~{rdi},~{rbp},~{r8},~{r9},~{r10},~{r11},~{r12},~{r13},~{r14},~{r15}"()55 %2 = call i32 @llvm.cttz.i32(i32 %a0, i1 -1)56 ret i32 %257}58declare i32 @llvm.cttz.i32(i32, i1)59 60define i64 @stack_fold_bsf_i64(i64 %a0) {61; CHECK-LABEL: stack_fold_bsf_i64:62; CHECK: # %bb.0:63; CHECK-NEXT: pushq %rbp64; CHECK-NEXT: .cfi_def_cfa_offset 1665; CHECK-NEXT: pushq %r1566; CHECK-NEXT: .cfi_def_cfa_offset 2467; CHECK-NEXT: pushq %r1468; CHECK-NEXT: .cfi_def_cfa_offset 3269; CHECK-NEXT: pushq %r1370; CHECK-NEXT: .cfi_def_cfa_offset 4071; CHECK-NEXT: pushq %r1272; CHECK-NEXT: .cfi_def_cfa_offset 4873; CHECK-NEXT: pushq %rbx74; CHECK-NEXT: .cfi_def_cfa_offset 5675; CHECK-NEXT: .cfi_offset %rbx, -5676; CHECK-NEXT: .cfi_offset %r12, -4877; CHECK-NEXT: .cfi_offset %r13, -4078; CHECK-NEXT: .cfi_offset %r14, -3279; CHECK-NEXT: .cfi_offset %r15, -2480; CHECK-NEXT: .cfi_offset %rbp, -1681; CHECK-NEXT: movq %rdi, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill82; CHECK-NEXT: #APP83; CHECK-NEXT: nop84; CHECK-NEXT: #NO_APP85; CHECK-NEXT: rep bsfq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Folded Reload86; CHECK-NEXT: popq %rbx87; CHECK-NEXT: .cfi_def_cfa_offset 4888; CHECK-NEXT: popq %r1289; CHECK-NEXT: .cfi_def_cfa_offset 4090; CHECK-NEXT: popq %r1391; CHECK-NEXT: .cfi_def_cfa_offset 3292; CHECK-NEXT: popq %r1493; CHECK-NEXT: .cfi_def_cfa_offset 2494; CHECK-NEXT: popq %r1595; CHECK-NEXT: .cfi_def_cfa_offset 1696; CHECK-NEXT: popq %rbp97; CHECK-NEXT: .cfi_def_cfa_offset 898; CHECK-NEXT: retq99 %1 = tail call i64 asm sideeffect "nop", "=x,~{rax},~{rbx},~{rcx},~{rdx},~{rsi},~{rdi},~{rbp},~{r8},~{r9},~{r10},~{r11},~{r12},~{r13},~{r14},~{r15}"()100 %2 = call i64 @llvm.cttz.i64(i64 %a0, i1 -1)101 ret i64 %2102}103declare i64 @llvm.cttz.i64(i64, i1)104 105;TODO stack_fold_bsr_i16106declare i16 @llvm.ctlz.i16(i16, i1)107 108define i32 @stack_fold_bsr_i32(i32 %a0) {109; CHECK-LABEL: stack_fold_bsr_i32:110; CHECK: # %bb.0:111; CHECK-NEXT: pushq %rbp112; CHECK-NEXT: .cfi_def_cfa_offset 16113; CHECK-NEXT: pushq %r15114; CHECK-NEXT: .cfi_def_cfa_offset 24115; CHECK-NEXT: pushq %r14116; CHECK-NEXT: .cfi_def_cfa_offset 32117; CHECK-NEXT: pushq %r13118; CHECK-NEXT: .cfi_def_cfa_offset 40119; CHECK-NEXT: pushq %r12120; CHECK-NEXT: .cfi_def_cfa_offset 48121; CHECK-NEXT: pushq %rbx122; CHECK-NEXT: .cfi_def_cfa_offset 56123; CHECK-NEXT: .cfi_offset %rbx, -56124; CHECK-NEXT: .cfi_offset %r12, -48125; CHECK-NEXT: .cfi_offset %r13, -40126; CHECK-NEXT: .cfi_offset %r14, -32127; CHECK-NEXT: .cfi_offset %r15, -24128; CHECK-NEXT: .cfi_offset %rbp, -16129; CHECK-NEXT: movl %edi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill130; CHECK-NEXT: #APP131; CHECK-NEXT: nop132; CHECK-NEXT: #NO_APP133; CHECK-NEXT: bsrl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 4-byte Folded Reload134; CHECK-NEXT: xorl $31, %eax135; CHECK-NEXT: popq %rbx136; CHECK-NEXT: .cfi_def_cfa_offset 48137; CHECK-NEXT: popq %r12138; CHECK-NEXT: .cfi_def_cfa_offset 40139; CHECK-NEXT: popq %r13140; CHECK-NEXT: .cfi_def_cfa_offset 32141; CHECK-NEXT: popq %r14142; CHECK-NEXT: .cfi_def_cfa_offset 24143; CHECK-NEXT: popq %r15144; CHECK-NEXT: .cfi_def_cfa_offset 16145; CHECK-NEXT: popq %rbp146; CHECK-NEXT: .cfi_def_cfa_offset 8147; CHECK-NEXT: retq148 %1 = tail call i64 asm sideeffect "nop", "=x,~{rax},~{rbx},~{rcx},~{rdx},~{rsi},~{rdi},~{rbp},~{r8},~{r9},~{r10},~{r11},~{r12},~{r13},~{r14},~{r15}"()149 %2 = call i32 @llvm.ctlz.i32(i32 %a0, i1 -1)150 ret i32 %2151}152declare i32 @llvm.ctlz.i32(i32, i1)153 154define i64 @stack_fold_bsr_i64(i64 %a0) {155; CHECK-LABEL: stack_fold_bsr_i64:156; CHECK: # %bb.0:157; CHECK-NEXT: pushq %rbp158; CHECK-NEXT: .cfi_def_cfa_offset 16159; CHECK-NEXT: pushq %r15160; CHECK-NEXT: .cfi_def_cfa_offset 24161; CHECK-NEXT: pushq %r14162; CHECK-NEXT: .cfi_def_cfa_offset 32163; CHECK-NEXT: pushq %r13164; CHECK-NEXT: .cfi_def_cfa_offset 40165; CHECK-NEXT: pushq %r12166; CHECK-NEXT: .cfi_def_cfa_offset 48167; CHECK-NEXT: pushq %rbx168; CHECK-NEXT: .cfi_def_cfa_offset 56169; CHECK-NEXT: .cfi_offset %rbx, -56170; CHECK-NEXT: .cfi_offset %r12, -48171; CHECK-NEXT: .cfi_offset %r13, -40172; CHECK-NEXT: .cfi_offset %r14, -32173; CHECK-NEXT: .cfi_offset %r15, -24174; CHECK-NEXT: .cfi_offset %rbp, -16175; CHECK-NEXT: movq %rdi, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill176; CHECK-NEXT: #APP177; CHECK-NEXT: nop178; CHECK-NEXT: #NO_APP179; CHECK-NEXT: bsrq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Folded Reload180; CHECK-NEXT: xorq $63, %rax181; CHECK-NEXT: popq %rbx182; CHECK-NEXT: .cfi_def_cfa_offset 48183; CHECK-NEXT: popq %r12184; CHECK-NEXT: .cfi_def_cfa_offset 40185; CHECK-NEXT: popq %r13186; CHECK-NEXT: .cfi_def_cfa_offset 32187; CHECK-NEXT: popq %r14188; CHECK-NEXT: .cfi_def_cfa_offset 24189; CHECK-NEXT: popq %r15190; CHECK-NEXT: .cfi_def_cfa_offset 16191; CHECK-NEXT: popq %rbp192; CHECK-NEXT: .cfi_def_cfa_offset 8193; CHECK-NEXT: retq194 %1 = tail call i64 asm sideeffect "nop", "=x,~{rax},~{rbx},~{rcx},~{rdx},~{rsi},~{rdi},~{rbp},~{r8},~{r9},~{r10},~{r11},~{r12},~{r13},~{r14},~{r15}"()195 %2 = call i64 @llvm.ctlz.i64(i64 %a0, i1 -1)196 ret i64 %2197}198declare i64 @llvm.ctlz.i64(i64, i1)199