204 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=i686-- -mattr=+sse2 | FileCheck %s3; Increment in loop bb.i28.i adjusted to 2, to prevent loop reversal from4; kicking in.5 6declare fastcc void @rdft(i32, i32, ptr, ptr, ptr)7 8define fastcc void @mp_sqrt(i32 %n, i32 %radix, ptr %in, ptr %out, ptr %tmp1, ptr %tmp2, i32 %nfft, ptr %tmp1fft, ptr %tmp2fft, ptr %ip, ptr %w) nounwind {9; CHECK-LABEL: mp_sqrt:10; CHECK: # %bb.0: # %entry11; CHECK-NEXT: pushl %ebp12; CHECK-NEXT: pushl %ebx13; CHECK-NEXT: pushl %edi14; CHECK-NEXT: pushl %esi15; CHECK-NEXT: pushl %eax16; CHECK-NEXT: movb $1, %al17; CHECK-NEXT: movl $1, %ebx18; CHECK-NEXT: movl {{[0-9]+}}(%esp), %esi19; CHECK-NEXT: .p2align 420; CHECK-NEXT: .LBB0_1: # %bb.i521; CHECK-NEXT: # =>This Inner Loop Header: Depth=122; CHECK-NEXT: movl %eax, %ecx23; CHECK-NEXT: addl %ebx, %ebx24; CHECK-NEXT: xorl %eax, %eax25; CHECK-NEXT: testb $1, %cl26; CHECK-NEXT: jne .LBB0_127; CHECK-NEXT: # %bb.2: # %mp_unexp_mp2d.exit.i28; CHECK-NEXT: je .LBB0_329; CHECK-NEXT: # %bb.5: # %cond_next.i30; CHECK-NEXT: jne .LBB0_331; CHECK-NEXT: # %bb.6: # %cond_next36.i32; CHECK-NEXT: movl $0, 033; CHECK-NEXT: movzbl %cl, %ebp34; CHECK-NEXT: andl $1, %ebp35; CHECK-NEXT: xorpd %xmm0, %xmm036; CHECK-NEXT: xorl %eax, %eax37; CHECK-NEXT: xorl %ecx, %ecx38; CHECK-NEXT: xorpd %xmm1, %xmm139; CHECK-NEXT: .p2align 440; CHECK-NEXT: .LBB0_7: # %bb.i28.i41; CHECK-NEXT: # =>This Inner Loop Header: Depth=142; CHECK-NEXT: cvttsd2si %xmm1, %edi43; CHECK-NEXT: cmpl %edx, %edi44; CHECK-NEXT: cmovgel %eax, %edi45; CHECK-NEXT: addl $2, %ecx46; CHECK-NEXT: xorps %xmm2, %xmm247; CHECK-NEXT: cvtsi2sd %edi, %xmm248; CHECK-NEXT: xorpd %xmm1, %xmm149; CHECK-NEXT: subsd %xmm2, %xmm150; CHECK-NEXT: mulsd %xmm0, %xmm151; CHECK-NEXT: addl $-2, %ebp52; CHECK-NEXT: jne .LBB0_753; CHECK-NEXT: # %bb.8: # %mp_unexp_d2mp.exit29.i54; CHECK-NEXT: movl $0, 055; CHECK-NEXT: je .LBB0_956; CHECK-NEXT: # %bb.10: # %mp_sqrt_init.exit57; CHECK-NEXT: xorl %ecx, %ecx58; CHECK-NEXT: movl %edx, %edi59; CHECK-NEXT: movl %esi, %edx60; CHECK-NEXT: calll mp_mul_csqu@PLT61; CHECK-NEXT: xorl %ecx, %ecx62; CHECK-NEXT: movl $-1, %edx63; CHECK-NEXT: pushl {{[0-9]+}}(%esp)64; CHECK-NEXT: pushl {{[0-9]+}}(%esp)65; CHECK-NEXT: pushl $066; CHECK-NEXT: calll rdft@PLT67; CHECK-NEXT: addl $12, %esp68; CHECK-NEXT: xorl %ecx, %ecx69; CHECK-NEXT: movl %edi, (%esp) # 4-byte Spill70; CHECK-NEXT: movl %edi, %edx71; CHECK-NEXT: movl {{[0-9]+}}(%esp), %edi72; CHECK-NEXT: pushl %edi73; CHECK-NEXT: pushl %esi74; CHECK-NEXT: pushl $075; CHECK-NEXT: calll mp_mul_d2i@PLT76; CHECK-NEXT: addl $12, %esp77; CHECK-NEXT: testl %ebp, %ebp78; CHECK-NEXT: je .LBB0_1179; CHECK-NEXT: .LBB0_3: # %cond_true.i80; CHECK-NEXT: addl $4, %esp81; CHECK-NEXT: .LBB0_4: # %cond_true.i82; CHECK-NEXT: popl %esi83; CHECK-NEXT: popl %edi84; CHECK-NEXT: popl %ebx85; CHECK-NEXT: popl %ebp86; CHECK-NEXT: retl87; CHECK-NEXT: .p2align 488; CHECK-NEXT: .LBB0_9: # %bb.i.i89; CHECK-NEXT: # =>This Inner Loop Header: Depth=190; CHECK-NEXT: jmp .LBB0_991; CHECK-NEXT: .LBB0_11: # %cond_false.i92; CHECK-NEXT: xorl %ecx, %ecx93; CHECK-NEXT: movl (%esp), %esi # 4-byte Reload94; CHECK-NEXT: movl %esi, %edx95; CHECK-NEXT: pushl {{[0-9]+}}(%esp)96; CHECK-NEXT: pushl $097; CHECK-NEXT: calll mp_round@PLT98; CHECK-NEXT: addl $8, %esp99; CHECK-NEXT: xorl %ecx, %ecx100; CHECK-NEXT: movl %esi, %edx101; CHECK-NEXT: movl {{[0-9]+}}(%esp), %ebp102; CHECK-NEXT: pushl %ebp103; CHECK-NEXT: pushl %edi104; CHECK-NEXT: pushl %ebp105; CHECK-NEXT: calll mp_add@PLT106; CHECK-NEXT: addl $12, %esp107; CHECK-NEXT: xorl %ecx, %ecx108; CHECK-NEXT: movl %esi, %edx109; CHECK-NEXT: pushl %edi110; CHECK-NEXT: pushl %edi111; CHECK-NEXT: pushl {{[0-9]+}}(%esp)112; CHECK-NEXT: calll mp_sub@PLT113; CHECK-NEXT: addl $12, %esp114; CHECK-NEXT: xorl %ecx, %ecx115; CHECK-NEXT: movl %esi, %edx116; CHECK-NEXT: pushl %ebp117; CHECK-NEXT: pushl $0118; CHECK-NEXT: calll mp_round@PLT119; CHECK-NEXT: addl $8, %esp120; CHECK-NEXT: xorl %ecx, %ecx121; CHECK-NEXT: movl %esi, %edx122; CHECK-NEXT: pushl %edi123; CHECK-NEXT: pushl {{[0-9]+}}(%esp)124; CHECK-NEXT: pushl %ebx125; CHECK-NEXT: calll mp_mul_d2i@PLT126; CHECK-NEXT: addl $16, %esp127; CHECK-NEXT: jmp .LBB0_4128entry:129 br label %bb.i5130 131bb.i5: ; preds = %bb.i5, %entry132 %nfft_init.0.i = phi i32 [ 1, %entry ], [ %tmp7.i3, %bb.i5 ] ; <i32> [#uses=1]133 %foo = phi i1 [1, %entry], [0, %bb.i5]134 %tmp7.i3 = shl i32 %nfft_init.0.i, 1 ; <i32> [#uses=2]135 br i1 %foo, label %bb.i5, label %mp_unexp_mp2d.exit.i136 137mp_unexp_mp2d.exit.i: ; preds = %bb.i5138 br i1 %foo, label %cond_next.i, label %cond_true.i139 140cond_true.i: ; preds = %mp_unexp_mp2d.exit.i141 ret void142 143cond_next.i: ; preds = %mp_unexp_mp2d.exit.i144 %tmp22.i = sdiv i32 0, 2 ; <i32> [#uses=2]145 br i1 %foo, label %cond_true29.i, label %cond_next36.i146 147cond_true29.i: ; preds = %cond_next.i148 ret void149 150cond_next36.i: ; preds = %cond_next.i151 store i32 %tmp22.i, ptr null, align 4152 %tmp8.i14.i = select i1 %foo, i32 1, i32 0 ; <i32> [#uses=1]153 br label %bb.i28.i154 155bb.i28.i: ; preds = %bb.i28.i, %cond_next36.i156 %j.0.reg2mem.0.i16.i = phi i32 [ 0, %cond_next36.i ], [ %indvar.next39.i, %bb.i28.i ] ; <i32> [#uses=2]157 %din_addr.1.reg2mem.0.i17.i = phi double [ 0.000000e+00, %cond_next36.i ], [ %tmp16.i25.i, %bb.i28.i ] ; <double> [#uses=1]158 %tmp1.i18.i = fptosi double %din_addr.1.reg2mem.0.i17.i to i32 ; <i32> [#uses=2]159 %tmp4.i19.i = icmp slt i32 %tmp1.i18.i, %radix ; <i1> [#uses=1]160 %x.0.i21.i = select i1 %tmp4.i19.i, i32 %tmp1.i18.i, i32 0 ; <i32> [#uses=1]161 %tmp41.sum.i = add i32 %j.0.reg2mem.0.i16.i, 2 ; <i32> [#uses=0]162 %tmp1213.i23.i = sitofp i32 %x.0.i21.i to double ; <double> [#uses=1]163 %tmp15.i24.i = fsub double 0.000000e+00, %tmp1213.i23.i ; <double> [#uses=1]164 %tmp16.i25.i = fmul double 0.000000e+00, %tmp15.i24.i ; <double> [#uses=1]165 %indvar.next39.i = add i32 %j.0.reg2mem.0.i16.i, 2 ; <i32> [#uses=2]166 %exitcond40.i = icmp eq i32 %indvar.next39.i, %tmp8.i14.i ; <i1> [#uses=1]167 br i1 %exitcond40.i, label %mp_unexp_d2mp.exit29.i, label %bb.i28.i168 169mp_unexp_d2mp.exit29.i: ; preds = %bb.i28.i170 %tmp46.i = sub i32 0, %tmp22.i ; <i32> [#uses=1]171 store i32 %tmp46.i, ptr null, align 4172 br i1 %exitcond40.i, label %bb.i.i, label %mp_sqrt_init.exit173 174bb.i.i: ; preds = %bb.i.i, %mp_unexp_d2mp.exit29.i175 br label %bb.i.i176 177mp_sqrt_init.exit: ; preds = %mp_unexp_d2mp.exit29.i178 tail call fastcc void @mp_mul_csqu( i32 0, ptr %tmp1fft )179 tail call fastcc void @rdft( i32 0, i32 -1, ptr null, ptr %ip, ptr %w )180 tail call fastcc void @mp_mul_d2i( i32 0, i32 %radix, i32 0, ptr %tmp1fft, ptr %tmp2 )181 br i1 %exitcond40.i, label %cond_false.i, label %cond_true36.i182 183cond_true36.i: ; preds = %mp_sqrt_init.exit184 ret void185 186cond_false.i: ; preds = %mp_sqrt_init.exit187 tail call fastcc void @mp_round( i32 0, i32 %radix, i32 0, ptr %out )188 tail call fastcc void @mp_add( i32 0, i32 %radix, ptr %tmp1, ptr %tmp2, ptr %tmp1 )189 tail call fastcc void @mp_sub( i32 0, i32 %radix, ptr %in, ptr %tmp2, ptr %tmp2 )190 tail call fastcc void @mp_round( i32 0, i32 %radix, i32 0, ptr %tmp1 )191 tail call fastcc void @mp_mul_d2i( i32 0, i32 %radix, i32 %tmp7.i3, ptr %tmp2fft, ptr %tmp2 )192 ret void193}194 195declare fastcc void @mp_add(i32, i32, ptr, ptr, ptr)196 197declare fastcc void @mp_sub(i32, i32, ptr, ptr, ptr)198 199declare fastcc void @mp_round(i32, i32, i32, ptr)200 201declare fastcc void @mp_mul_csqu(i32, ptr)202 203declare fastcc void @mp_mul_d2i(i32, i32, i32, ptr, ptr)204