|
| 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 | + |
| 8 | +define 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 %xmm1, %xmm0 |
| 50 | +; X64-NEXT: retq |
| 51 | + %ret = tail call double @llvm.powi.f64.i32(double %a, i32 15) nounwind ; <double> [#uses=1] |
| 52 | + ret double %ret |
| 53 | +} |
| 54 | + |
| 55 | +define double @pow_wrapper_optsize(double %a) nounwind optsize { |
| 56 | +; X86-X87-LABEL: pow_wrapper_optsize: |
| 57 | +; X86-X87: # %bb.0: |
| 58 | +; X86-X87-NEXT: subl $12, %esp |
| 59 | +; X86-X87-NEXT: fldl {{[0-9]+}}(%esp) |
| 60 | +; X86-X87-NEXT: fstpl (%esp) |
| 61 | +; X86-X87-NEXT: movl $15, {{[0-9]+}}(%esp) |
| 62 | +; X86-X87-NEXT: calll __powidf2 |
| 63 | +; X86-X87-NEXT: addl $12, %esp |
| 64 | +; X86-X87-NEXT: retl |
| 65 | +; |
| 66 | +; X86-SSE-LABEL: pow_wrapper_optsize: |
| 67 | +; X86-SSE: # %bb.0: |
| 68 | +; X86-SSE-NEXT: subl $12, %esp |
| 69 | +; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero |
| 70 | +; X86-SSE-NEXT: movsd %xmm0, (%esp) |
| 71 | +; X86-SSE-NEXT: movl $15, {{[0-9]+}}(%esp) |
| 72 | +; X86-SSE-NEXT: calll __powidf2 |
| 73 | +; X86-SSE-NEXT: addl $12, %esp |
| 74 | +; X86-SSE-NEXT: retl |
| 75 | +; |
| 76 | +; X64-LABEL: pow_wrapper_optsize: |
| 77 | +; X64: # %bb.0: |
| 78 | +; X64-NEXT: movl $15, %edi |
| 79 | +; X64-NEXT: jmp __powidf2@PLT # TAILCALL |
| 80 | + %ret = tail call double @llvm.powi.f64.i32(double %a, i32 15) nounwind ; <double> [#uses=1] |
| 81 | + ret double %ret |
| 82 | +} |
| 83 | + |
| 84 | +define double @pow_wrapper_pgso(double %a) nounwind !prof !14 { |
| 85 | +; X86-X87-LABEL: pow_wrapper_pgso: |
| 86 | +; X86-X87: # %bb.0: |
| 87 | +; X86-X87-NEXT: subl $12, %esp |
| 88 | +; X86-X87-NEXT: fldl {{[0-9]+}}(%esp) |
| 89 | +; X86-X87-NEXT: fstpl (%esp) |
| 90 | +; X86-X87-NEXT: movl $15, {{[0-9]+}}(%esp) |
| 91 | +; X86-X87-NEXT: calll __powidf2 |
| 92 | +; X86-X87-NEXT: addl $12, %esp |
| 93 | +; X86-X87-NEXT: retl |
| 94 | +; |
| 95 | +; X86-SSE-LABEL: pow_wrapper_pgso: |
| 96 | +; X86-SSE: # %bb.0: |
| 97 | +; X86-SSE-NEXT: subl $12, %esp |
| 98 | +; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero |
| 99 | +; X86-SSE-NEXT: movsd %xmm0, (%esp) |
| 100 | +; X86-SSE-NEXT: movl $15, {{[0-9]+}}(%esp) |
| 101 | +; X86-SSE-NEXT: calll __powidf2 |
| 102 | +; X86-SSE-NEXT: addl $12, %esp |
| 103 | +; X86-SSE-NEXT: retl |
| 104 | +; |
| 105 | +; X64-LABEL: pow_wrapper_pgso: |
| 106 | +; X64: # %bb.0: |
| 107 | +; X64-NEXT: movl $15, %edi |
| 108 | +; X64-NEXT: jmp __powidf2@PLT # TAILCALL |
| 109 | + %ret = tail call double @llvm.powi.f64.i32(double %a, i32 15) nounwind ; <double> [#uses=1] |
| 110 | + ret double %ret |
| 111 | +} |
| 112 | + |
| 113 | +define double @pow_wrapper_minsize(double %a) nounwind minsize { |
| 114 | +; X86-X87-LABEL: pow_wrapper_minsize: |
| 115 | +; X86-X87: # %bb.0: |
| 116 | +; X86-X87-NEXT: subl $12, %esp |
| 117 | +; X86-X87-NEXT: fldl {{[0-9]+}}(%esp) |
| 118 | +; X86-X87-NEXT: fstpl (%esp) |
| 119 | +; X86-X87-NEXT: movl $15, {{[0-9]+}}(%esp) |
| 120 | +; X86-X87-NEXT: calll __powidf2 |
| 121 | +; X86-X87-NEXT: addl $12, %esp |
| 122 | +; X86-X87-NEXT: retl |
| 123 | +; |
| 124 | +; X86-SSE-LABEL: pow_wrapper_minsize: |
| 125 | +; X86-SSE: # %bb.0: |
| 126 | +; X86-SSE-NEXT: subl $12, %esp |
| 127 | +; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero |
| 128 | +; X86-SSE-NEXT: movsd %xmm0, (%esp) |
| 129 | +; X86-SSE-NEXT: movl $15, {{[0-9]+}}(%esp) |
| 130 | +; X86-SSE-NEXT: calll __powidf2 |
| 131 | +; X86-SSE-NEXT: addl $12, %esp |
| 132 | +; X86-SSE-NEXT: retl |
| 133 | +; |
| 134 | +; X64-LABEL: pow_wrapper_minsize: |
| 135 | +; X64: # %bb.0: |
| 136 | +; X64-NEXT: pushq $15 |
| 137 | +; X64-NEXT: popq %rdi |
| 138 | +; X64-NEXT: jmp __powidf2@PLT # TAILCALL |
| 139 | + %ret = tail call double @llvm.powi.f64.i32(double %a, i32 15) nounwind ; <double> [#uses=1] |
| 140 | + ret double %ret |
| 141 | +} |
| 142 | + |
| 143 | +define <2 x float> @powi_v2f32(<2 x float> %a) nounwind minsize { |
| 144 | +; X86-X87-LABEL: powi_v2f32: |
| 145 | +; X86-X87: # %bb.0: |
| 146 | +; X86-X87-NEXT: pushl %esi |
| 147 | +; X86-X87-NEXT: subl $16, %esp |
| 148 | +; X86-X87-NEXT: flds {{[0-9]+}}(%esp) |
| 149 | +; X86-X87-NEXT: fstps {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Spill |
| 150 | +; X86-X87-NEXT: flds {{[0-9]+}}(%esp) |
| 151 | +; X86-X87-NEXT: pushl $15 |
| 152 | +; X86-X87-NEXT: popl %esi |
| 153 | +; X86-X87-NEXT: movl %esi, {{[0-9]+}}(%esp) |
| 154 | +; X86-X87-NEXT: fstps (%esp) |
| 155 | +; X86-X87-NEXT: calll __powisf2 |
| 156 | +; X86-X87-NEXT: fstps {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Spill |
| 157 | +; X86-X87-NEXT: movl %esi, {{[0-9]+}}(%esp) |
| 158 | +; X86-X87-NEXT: flds {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Reload |
| 159 | +; X86-X87-NEXT: fstps (%esp) |
| 160 | +; X86-X87-NEXT: calll __powisf2 |
| 161 | +; X86-X87-NEXT: flds {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Reload |
| 162 | +; X86-X87-NEXT: fxch %st(1) |
| 163 | +; X86-X87-NEXT: addl $16, %esp |
| 164 | +; X86-X87-NEXT: popl %esi |
| 165 | +; X86-X87-NEXT: retl |
| 166 | +; |
| 167 | +; X86-SSE-LABEL: powi_v2f32: |
| 168 | +; X86-SSE: # %bb.0: |
| 169 | +; X86-SSE-NEXT: pushl %esi |
| 170 | +; X86-SSE-NEXT: subl $32, %esp |
| 171 | +; X86-SSE-NEXT: movups %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill |
| 172 | +; X86-SSE-NEXT: pushl $15 |
| 173 | +; X86-SSE-NEXT: popl %esi |
| 174 | +; X86-SSE-NEXT: movl %esi, {{[0-9]+}}(%esp) |
| 175 | +; X86-SSE-NEXT: movss %xmm0, (%esp) |
| 176 | +; X86-SSE-NEXT: calll __powisf2 |
| 177 | +; X86-SSE-NEXT: movl %esi, {{[0-9]+}}(%esp) |
| 178 | +; X86-SSE-NEXT: movups {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 # 16-byte Reload |
| 179 | +; X86-SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| 180 | +; X86-SSE-NEXT: movss %xmm0, (%esp) |
| 181 | +; X86-SSE-NEXT: fstps {{[0-9]+}}(%esp) |
| 182 | +; X86-SSE-NEXT: calll __powisf2 |
| 183 | +; X86-SSE-NEXT: fstps {{[0-9]+}}(%esp) |
| 184 | +; X86-SSE-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero |
| 185 | +; X86-SSE-NEXT: movss {{.*#+}} xmm1 = mem[0],zero,zero,zero |
| 186 | +; X86-SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] |
| 187 | +; X86-SSE-NEXT: addl $32, %esp |
| 188 | +; X86-SSE-NEXT: popl %esi |
| 189 | +; X86-SSE-NEXT: retl |
| 190 | +; |
| 191 | +; X64-LABEL: powi_v2f32: |
| 192 | +; X64: # %bb.0: |
| 193 | +; X64-NEXT: pushq %rbx |
| 194 | +; X64-NEXT: subq $32, %rsp |
| 195 | +; X64-NEXT: movaps %xmm0, (%rsp) # 16-byte Spill |
| 196 | +; X64-NEXT: pushq $15 |
| 197 | +; X64-NEXT: popq %rbx |
| 198 | +; X64-NEXT: movl %ebx, %edi |
| 199 | +; X64-NEXT: callq __powisf2@PLT |
| 200 | +; X64-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill |
| 201 | +; X64-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload |
| 202 | +; X64-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,1,1] |
| 203 | +; X64-NEXT: movl %ebx, %edi |
| 204 | +; X64-NEXT: callq __powisf2@PLT |
| 205 | +; X64-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload |
| 206 | +; X64-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] |
| 207 | +; X64-NEXT: movaps %xmm1, %xmm0 |
| 208 | +; X64-NEXT: addq $32, %rsp |
| 209 | +; X64-NEXT: popq %rbx |
| 210 | +; X64-NEXT: retq |
| 211 | + %ret = tail call < 2 x float> @llvm.powi.v2f32.i32(<2 x float> %a, i32 15) nounwind ; |
| 212 | + ret <2 x float> %ret |
| 213 | +} |
| 214 | + |
| 215 | +declare double @llvm.powi.f64.i32(double, i32) nounwind readonly |
| 216 | +declare < 2 x float> @llvm.powi.v2f32.i32(<2 x float>, i32) nounwind readonly |
| 217 | + |
| 218 | +!llvm.module.flags = !{!0} |
| 219 | +!0 = !{i32 1, !"ProfileSummary", !1} |
| 220 | +!1 = !{!2, !3, !4, !5, !6, !7, !8, !9} |
| 221 | +!2 = !{!"ProfileFormat", !"InstrProf"} |
| 222 | +!3 = !{!"TotalCount", i64 10000} |
| 223 | +!4 = !{!"MaxCount", i64 10} |
| 224 | +!5 = !{!"MaxInternalCount", i64 1} |
| 225 | +!6 = !{!"MaxFunctionCount", i64 1000} |
| 226 | +!7 = !{!"NumCounts", i64 3} |
| 227 | +!8 = !{!"NumFunctions", i64 3} |
| 228 | +!9 = !{!"DetailedSummary", !10} |
| 229 | +!10 = !{!11, !12, !13} |
| 230 | +!11 = !{i32 10000, i64 100, i32 1} |
| 231 | +!12 = !{i32 999000, i64 100, i32 1} |
| 232 | +!13 = !{i32 999999, i64 1, i32 2} |
| 233 | +!14 = !{!"function_entry_count", i64 0} |
0 commit comments