1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py 2; RUN: llc < %s -mtriple=i686-unknown-unknown | FileCheck %s --check-prefix=X86-X87 3; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+sse2 | FileCheck %s --check-prefix=X86-SSE 4; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse2 | FileCheck %s --check-prefix=X64 5 6; Ideally this would compile to 5 multiplies. 7 8define double @pow_wrapper(double %a) nounwind readonly ssp noredzone { 9; X86-X87-LABEL: pow_wrapper: 10; X86-X87: # %bb.0: 11; X86-X87-NEXT: fldl {{[0-9]+}}(%esp) 12; X86-X87-NEXT: fld %st(0) 13; X86-X87-NEXT: fmul %st(1), %st 14; X86-X87-NEXT: fmul %st, %st(1) 15; X86-X87-NEXT: fmul %st, %st(0) 16; X86-X87-NEXT: fmul %st, %st(1) 17; X86-X87-NEXT: fmul %st, %st(0) 18; X86-X87-NEXT: fmulp %st, %st(1) 19; X86-X87-NEXT: retl 20; 21; X86-SSE-LABEL: pow_wrapper: 22; X86-SSE: # %bb.0: 23; X86-SSE-NEXT: pushl %ebp 24; X86-SSE-NEXT: movl %esp, %ebp 25; X86-SSE-NEXT: andl $-8, %esp 26; X86-SSE-NEXT: subl $8, %esp 27; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero 28; X86-SSE-NEXT: movapd %xmm0, %xmm1 29; X86-SSE-NEXT: mulsd %xmm0, %xmm1 30; X86-SSE-NEXT: mulsd %xmm1, %xmm0 31; X86-SSE-NEXT: mulsd %xmm1, %xmm1 32; X86-SSE-NEXT: mulsd %xmm1, %xmm0 33; X86-SSE-NEXT: mulsd %xmm1, %xmm1 34; X86-SSE-NEXT: mulsd %xmm0, %xmm1 35; X86-SSE-NEXT: movsd %xmm1, (%esp) 36; X86-SSE-NEXT: fldl (%esp) 37; X86-SSE-NEXT: movl %ebp, %esp 38; X86-SSE-NEXT: popl %ebp 39; X86-SSE-NEXT: retl 40; 41; X64-LABEL: pow_wrapper: 42; X64: # %bb.0: 43; X64-NEXT: movapd %xmm0, %xmm1 44; X64-NEXT: mulsd %xmm0, %xmm1 45; X64-NEXT: mulsd %xmm1, %xmm0 46; X64-NEXT: mulsd %xmm1, %xmm1 47; X64-NEXT: mulsd %xmm1, %xmm0 48; X64-NEXT: mulsd %xmm1, %xmm1 49; X64-NEXT: mulsd %xmm0, %xmm1 50; X64-NEXT: movapd %xmm1, %xmm0 51; X64-NEXT: retq 52 %ret = tail call double @llvm.powi.f64.i32(double %a, i32 15) nounwind ; <double> [#uses=1] 53 ret double %ret 54} 55 56define double @pow_wrapper_optsize(double %a) optsize { 57; X86-X87-LABEL: pow_wrapper_optsize: 58; X86-X87: # %bb.0: 59; X86-X87-NEXT: subl $12, %esp 60; X86-X87-NEXT: .cfi_def_cfa_offset 16 61; X86-X87-NEXT: fldl {{[0-9]+}}(%esp) 62; X86-X87-NEXT: fstpl (%esp) 63; X86-X87-NEXT: movl $15, {{[0-9]+}}(%esp) 64; X86-X87-NEXT: calll __powidf2 65; X86-X87-NEXT: addl $12, %esp 66; X86-X87-NEXT: .cfi_def_cfa_offset 4 67; X86-X87-NEXT: retl 68; 69; X86-SSE-LABEL: pow_wrapper_optsize: 70; X86-SSE: # %bb.0: 71; X86-SSE-NEXT: subl $12, %esp 72; X86-SSE-NEXT: .cfi_def_cfa_offset 16 73; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero 74; X86-SSE-NEXT: movsd %xmm0, (%esp) 75; X86-SSE-NEXT: movl $15, {{[0-9]+}}(%esp) 76; X86-SSE-NEXT: calll __powidf2 77; X86-SSE-NEXT: addl $12, %esp 78; X86-SSE-NEXT: .cfi_def_cfa_offset 4 79; X86-SSE-NEXT: retl 80; 81; X64-LABEL: pow_wrapper_optsize: 82; X64: # %bb.0: 83; X64-NEXT: movl $15, %edi 84; X64-NEXT: jmp __powidf2@PLT # TAILCALL 85 %ret = tail call double @llvm.powi.f64.i32(double %a, i32 15) nounwind ; <double> [#uses=1] 86 ret double %ret 87} 88 89define double @pow_wrapper_pgso(double %a) !prof !14 { 90; X86-X87-LABEL: pow_wrapper_pgso: 91; X86-X87: # %bb.0: 92; X86-X87-NEXT: subl $12, %esp 93; X86-X87-NEXT: .cfi_def_cfa_offset 16 94; X86-X87-NEXT: fldl {{[0-9]+}}(%esp) 95; X86-X87-NEXT: fstpl (%esp) 96; X86-X87-NEXT: movl $15, {{[0-9]+}}(%esp) 97; X86-X87-NEXT: calll __powidf2 98; X86-X87-NEXT: addl $12, %esp 99; X86-X87-NEXT: .cfi_def_cfa_offset 4 100; X86-X87-NEXT: retl 101; 102; X86-SSE-LABEL: pow_wrapper_pgso: 103; X86-SSE: # %bb.0: 104; X86-SSE-NEXT: subl $12, %esp 105; X86-SSE-NEXT: .cfi_def_cfa_offset 16 106; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero 107; X86-SSE-NEXT: movsd %xmm0, (%esp) 108; X86-SSE-NEXT: movl $15, {{[0-9]+}}(%esp) 109; X86-SSE-NEXT: calll __powidf2 110; X86-SSE-NEXT: addl $12, %esp 111; X86-SSE-NEXT: .cfi_def_cfa_offset 4 112; X86-SSE-NEXT: retl 113; 114; X64-LABEL: pow_wrapper_pgso: 115; X64: # %bb.0: 116; X64-NEXT: movl $15, %edi 117; X64-NEXT: jmp __powidf2@PLT # TAILCALL 118 %ret = tail call double @llvm.powi.f64.i32(double %a, i32 15) nounwind ; <double> [#uses=1] 119 ret double %ret 120} 121 122define double @pow_wrapper_minsize(double %a) minsize { 123; X86-X87-LABEL: pow_wrapper_minsize: 124; X86-X87: # %bb.0: 125; X86-X87-NEXT: subl $12, %esp 126; X86-X87-NEXT: .cfi_def_cfa_offset 16 127; X86-X87-NEXT: fldl {{[0-9]+}}(%esp) 128; X86-X87-NEXT: fstpl (%esp) 129; X86-X87-NEXT: movl $15, {{[0-9]+}}(%esp) 130; X86-X87-NEXT: calll __powidf2 131; X86-X87-NEXT: addl $12, %esp 132; X86-X87-NEXT: .cfi_def_cfa_offset 4 133; X86-X87-NEXT: retl 134; 135; X86-SSE-LABEL: pow_wrapper_minsize: 136; X86-SSE: # %bb.0: 137; X86-SSE-NEXT: subl $12, %esp 138; X86-SSE-NEXT: .cfi_def_cfa_offset 16 139; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero 140; X86-SSE-NEXT: movsd %xmm0, (%esp) 141; X86-SSE-NEXT: movl $15, {{[0-9]+}}(%esp) 142; X86-SSE-NEXT: calll __powidf2 143; X86-SSE-NEXT: addl $12, %esp 144; X86-SSE-NEXT: .cfi_def_cfa_offset 4 145; X86-SSE-NEXT: retl 146; 147; X64-LABEL: pow_wrapper_minsize: 148; X64: # %bb.0: 149; X64-NEXT: pushq $15 150; X64-NEXT: .cfi_adjust_cfa_offset 8 151; X64-NEXT: popq %rdi 152; X64-NEXT: .cfi_adjust_cfa_offset -8 153; X64-NEXT: jmp __powidf2@PLT # TAILCALL 154 %ret = tail call double @llvm.powi.f64.i32(double %a, i32 15) nounwind ; <double> [#uses=1] 155 ret double %ret 156} 157 158declare double @llvm.powi.f64.i32(double, i32) nounwind readonly 159 160!llvm.module.flags = !{!0} 161!0 = !{i32 1, !"ProfileSummary", !1} 162!1 = !{!2, !3, !4, !5, !6, !7, !8, !9} 163!2 = !{!"ProfileFormat", !"InstrProf"} 164!3 = !{!"TotalCount", i64 10000} 165!4 = !{!"MaxCount", i64 10} 166!5 = !{!"MaxInternalCount", i64 1} 167!6 = !{!"MaxFunctionCount", i64 1000} 168!7 = !{!"NumCounts", i64 3} 169!8 = !{!"NumFunctions", i64 3} 170!9 = !{!"DetailedSummary", !10} 171!10 = !{!11, !12, !13} 172!11 = !{i32 10000, i64 100, i32 1} 173!12 = !{i32 999000, i64 100, i32 1} 174!13 = !{i32 999999, i64 1, i32 2} 175!14 = !{!"function_entry_count", i64 0} 176