xref: /llvm-project/llvm/test/CodeGen/X86/inline-asm-fpstack.ll (revision 2f448bf509432c1a19ec46ab8cbc7353c03c6280)
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mcpu=generic -mtriple=i386-apple-darwin -verify-machineinstrs -no-integrated-as | FileCheck %s
3
4; There should be no stack manipulations between the inline asm and ret.
5define x86_fp80 @test1() nounwind {
6; CHECK-LABEL: test1:
7; CHECK:       ## %bb.0:
8; CHECK-NEXT:    ## InlineAsm Start
9; CHECK-NEXT:    fld0
10; CHECK-NEXT:    ## InlineAsm End
11; CHECK-NEXT:    retl
12  %tmp85 = call x86_fp80 asm sideeffect "fld0", "={st(0)}"()
13  ret x86_fp80 %tmp85
14}
15
16define double @test2() nounwind {
17; CHECK-LABEL: test2:
18; CHECK:       ## %bb.0:
19; CHECK-NEXT:    ## InlineAsm Start
20; CHECK-NEXT:    fld0
21; CHECK-NEXT:    ## InlineAsm End
22; CHECK-NEXT:    retl
23  %tmp85 = call double asm sideeffect "fld0", "={st(0)}"()
24  ret double %tmp85
25}
26
27; Setting up argument in st(0) should be a single fld.
28; Asm consumes stack, nothing should be popped.
29define void @test3(x86_fp80 %X) nounwind {
30; CHECK-LABEL: test3:
31; CHECK:       ## %bb.0:
32; CHECK-NEXT:    subl $12, %esp
33; CHECK-NEXT:    fldt {{[0-9]+}}(%esp)
34; CHECK-NEXT:    ## InlineAsm Start
35; CHECK-NEXT:    frob
36; CHECK-NEXT:    ## InlineAsm End
37; CHECK-NEXT:    addl $12, %esp
38; CHECK-NEXT:    retl
39  call void asm sideeffect "frob ", "{st(0)},~{st},~{dirflag},~{fpsr},~{flags}"( x86_fp80 %X)
40  ret void
41}
42
43define void @test4(double %X) nounwind {
44; CHECK-LABEL: test4:
45; CHECK:       ## %bb.0:
46; CHECK-NEXT:    fldl {{[0-9]+}}(%esp)
47; CHECK-NEXT:    ## InlineAsm Start
48; CHECK-NEXT:    frob
49; CHECK-NEXT:    ## InlineAsm End
50; CHECK-NEXT:    retl
51  call void asm sideeffect "frob ", "{st(0)},~{st},~{dirflag},~{fpsr},~{flags}"( double %X)
52  ret void
53}
54
55; Same as test3/4, but using value from fadd.
56; The fadd can be done in xmm or x87 regs - we don't test that.
57define void @test5(double %X) nounwind {
58; CHECK-LABEL: test5:
59; CHECK:       ## %bb.0:
60; CHECK-NEXT:    fldl {{[0-9]+}}(%esp)
61; CHECK-NEXT:    fadds {{\.?LCPI[0-9]+_[0-9]+}}
62; CHECK-NEXT:    ## InlineAsm Start
63; CHECK-NEXT:    frob
64; CHECK-NEXT:    ## InlineAsm End
65; CHECK-NEXT:    retl
66  %Y = fadd double %X, 123.0
67  call void asm sideeffect "frob ", "{st(0)},~{st},~{dirflag},~{fpsr},~{flags}"( double %Y)
68  ret void
69}
70
71define void @test6(double %A, double %B, double %C, double %D, double %E) nounwind {
72; CHECK-LABEL: test6:
73; CHECK:       ## %bb.0: ## %entry
74; CHECK-NEXT:    fldl {{[0-9]+}}(%esp)
75; CHECK-NEXT:    fldl {{[0-9]+}}(%esp)
76; CHECK-NEXT:    fldl {{[0-9]+}}(%esp)
77; CHECK-NEXT:    fldl {{[0-9]+}}(%esp)
78; CHECK-NEXT:    fldl {{[0-9]+}}(%esp)
79; CHECK-NEXT:    ## InlineAsm Start
80; CHECK-NEXT:    foo %st %st
81; CHECK-NEXT:    ## InlineAsm End
82; CHECK-NEXT:    fstp %st(0)
83; CHECK-NEXT:    ## InlineAsm Start
84; CHECK-NEXT:    bar %st(1) %st
85; CHECK-NEXT:    ## InlineAsm End
86; CHECK-NEXT:    fstp %st(1)
87; CHECK-NEXT:    fstp %st(0)
88; CHECK-NEXT:    ## InlineAsm Start
89; CHECK-NEXT:    baz %st(1) %st
90; CHECK-NEXT:    ## InlineAsm End
91; CHECK-NEXT:    fstp %st(0)
92; CHECK-NEXT:    ## InlineAsm Start
93; CHECK-NEXT:    baz %st
94; CHECK-NEXT:    ## InlineAsm End
95; CHECK-NEXT:    fstp %st(0)
96; CHECK-NEXT:    retl
97entry:
98; Uses the same value twice, should have one fstp after the asm.
99  tail call void asm sideeffect "foo $0 $1", "f,f,~{dirflag},~{fpsr},~{flags}"( double %A, double %A ) nounwind
100; Uses two different values, should be in st(0)/st(1) and both be popped.
101  tail call void asm sideeffect "bar $0 $1", "f,f,~{dirflag},~{fpsr},~{flags}"( double %B, double %C ) nounwind
102; Uses two different values, one of which isn't killed in this asm, it should not be popped after the asm.
103  tail call void asm sideeffect "baz $0 $1", "f,f,~{dirflag},~{fpsr},~{flags}"( double %D, double %E ) nounwind
104; This is the last use of %D, so it should be popped after.
105  tail call void asm sideeffect "baz $0", "f,~{dirflag},~{fpsr},~{flags}"( double %D ) nounwind
106  ret void
107}
108
109; PR4185
110; Passing a non-killed value to asm in {st}.
111; Make sure it is duped before.
112; asm kills st(0), so we shouldn't pop anything
113; A valid alternative would be to remat the constant pool load before each
114; inline asm.
115define void @testPR4185() nounwind {
116; CHECK-LABEL: testPR4185:
117; CHECK:       ## %bb.0: ## %return
118; CHECK-NEXT:    flds {{\.?LCPI[0-9]+_[0-9]+}}
119; CHECK-NEXT:    fld %st(0)
120; CHECK-NEXT:    ## InlineAsm Start
121; CHECK-NEXT:    fistpl %st
122; CHECK-NEXT:    ## InlineAsm End
123; CHECK-NEXT:    ## InlineAsm Start
124; CHECK-NEXT:    fistpl %st
125; CHECK-NEXT:    ## InlineAsm End
126; CHECK-NEXT:    retl
127return:
128  call void asm sideeffect "fistpl $0", "{st},~{st}"(double 1.000000e+06)
129  call void asm sideeffect "fistpl $0", "{st},~{st}"(double 1.000000e+06)
130  ret void
131}
132
133; Passing a non-killed value through asm in {st}.
134; Make sure it is not duped before.
135; Second asm kills st(0), so we shouldn't pop anything
136; A valid alternative would be to remat the constant pool load before each inline asm.
137define void @testPR4185b() nounwind {
138; CHECK-LABEL: testPR4185b:
139; CHECK:       ## %bb.0: ## %return
140; CHECK-NEXT:    flds {{\.?LCPI[0-9]+_[0-9]+}}
141; CHECK-NEXT:    ## InlineAsm Start
142; CHECK-NEXT:    fistl %st
143; CHECK-NEXT:    ## InlineAsm End
144; CHECK-NEXT:    ## InlineAsm Start
145; CHECK-NEXT:    fistpl %st
146; CHECK-NEXT:    ## InlineAsm End
147; CHECK-NEXT:    retl
148return:
149	call void asm sideeffect "fistl $0", "{st}"(double 1.000000e+06)
150	call void asm sideeffect "fistpl $0", "{st},~{st}"(double 1.000000e+06)
151	ret void
152}
153
154; PR4459
155; The return value from ceil must be duped before being consumed by asm.
156define void @testPR4459(x86_fp80 %a) nounwind {
157; CHECK-LABEL: testPR4459:
158; CHECK:       ## %bb.0: ## %entry
159; CHECK-NEXT:    subl $28, %esp
160; CHECK-NEXT:    fldt {{[0-9]+}}(%esp)
161; CHECK-NEXT:    fstpt (%esp)
162; CHECK-NEXT:    calll _ceil
163; CHECK-NEXT:    fld %st(0)
164; CHECK-NEXT:    fxch %st(1)
165; CHECK-NEXT:    ## InlineAsm Start
166; CHECK-NEXT:    fistpl %st
167; CHECK-NEXT:    ## InlineAsm End
168; CHECK-NEXT:    fstpt (%esp)
169; CHECK-NEXT:    calll _test3
170; CHECK-NEXT:    addl $28, %esp
171; CHECK-NEXT:    retl
172entry:
173  %0 = call x86_fp80 @ceil(x86_fp80 %a)
174  call void asm sideeffect "fistpl $0", "{st},~{st}"( x86_fp80 %0)
175  call void @test3(x86_fp80 %0 )
176  ret void
177}
178declare x86_fp80 @ceil(x86_fp80)
179
180; PR4484
181; test1 leaves a value on the stack that is needed after the asm.
182; Load %a from stack after ceil
183; Set up call to test.
184define void @testPR4484(x86_fp80 %a) nounwind {
185; CHECK-LABEL: testPR4484:
186; CHECK:       ## %bb.0: ## %entry
187; CHECK-NEXT:    subl $28, %esp
188; CHECK-NEXT:    fldt {{[0-9]+}}(%esp)
189; CHECK-NEXT:    fstpt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Spill
190; CHECK-NEXT:    calll _test1
191; CHECK-NEXT:    fldt {{[-0-9]+}}(%e{{[sb]}}p) ## 10-byte Folded Reload
192; CHECK-NEXT:    ## InlineAsm Start
193; CHECK-NEXT:    fistpl %st
194; CHECK-NEXT:    ## InlineAsm End
195; CHECK-NEXT:    fstpt (%esp)
196; CHECK-NEXT:    calll _test3
197; CHECK-NEXT:    addl $28, %esp
198; CHECK-NEXT:    retl
199entry:
200  %0 = call x86_fp80 @test1()
201  call void asm sideeffect "fistpl $0", "{st},~{st}"(x86_fp80 %a)
202  call void @test3(x86_fp80 %0)
203  ret void
204}
205
206; PR4485
207define void @testPR4485(ptr %a) nounwind {
208; CHECK-LABEL: testPR4485:
209; CHECK:       ## %bb.0: ## %entry
210; CHECK-NEXT:    movl {{[0-9]+}}(%esp), %eax
211; CHECK-NEXT:    fldt (%eax)
212; CHECK-NEXT:    flds {{\.?LCPI[0-9]+_[0-9]+}}
213; CHECK-NEXT:    fmul %st, %st(1)
214; CHECK-NEXT:    flds {{\.?LCPI[0-9]+_[0-9]+}}
215; CHECK-NEXT:    fmul %st, %st(2)
216; CHECK-NEXT:    fxch %st(2)
217; CHECK-NEXT:    ## InlineAsm Start
218; CHECK-NEXT:    fistpl %st
219; CHECK-NEXT:    ## InlineAsm End
220; CHECK-NEXT:    fldt (%eax)
221; CHECK-NEXT:    fmulp %st, %st(1)
222; CHECK-NEXT:    fmulp %st, %st(1)
223; CHECK-NEXT:    ## InlineAsm Start
224; CHECK-NEXT:    fistpl %st
225; CHECK-NEXT:    ## InlineAsm End
226; CHECK-NEXT:    retl
227entry:
228  %0 = load x86_fp80, ptr %a, align 16
229  %1 = fmul x86_fp80 %0, 0xK4006B400000000000000
230  %2 = fmul x86_fp80 %1, 0xK4012F424000000000000
231  tail call void asm sideeffect "fistpl $0", "{st},~{st}"(x86_fp80 %2)
232  %3 = load x86_fp80, ptr %a, align 16
233  %4 = fmul x86_fp80 %3, 0xK4006B400000000000000
234  %5 = fmul x86_fp80 %4, 0xK4012F424000000000000
235  tail call void asm sideeffect "fistpl $0", "{st},~{st}"(x86_fp80 %5)
236  ret void
237}
238
239; An input argument in a fixed position is implicitly popped by the asm only if
240; the input argument is tied to an output register, or it is in the clobber list.
241; The clobber list case is tested above.
242;
243; This doesn't implicitly pop the stack:
244;
245;   void fist1(long double x, int *p) {
246;     asm volatile ("fistl %1" : : "t"(x), "m"(*p));
247;   }
248define void @fist1(x86_fp80 %x, ptr %p) nounwind ssp {
249; CHECK-LABEL: fist1:
250; CHECK:       ## %bb.0: ## %entry
251; CHECK-NEXT:    subl $12, %esp
252; CHECK-NEXT:    fldt {{[0-9]+}}(%esp)
253; CHECK-NEXT:    movl {{[0-9]+}}(%esp), %eax
254; CHECK-NEXT:    ## InlineAsm Start
255; CHECK-NEXT:    fistl (%eax)
256; CHECK-NEXT:    ## InlineAsm End
257; CHECK-NEXT:    fstp %st(0)
258; CHECK-NEXT:    addl $12, %esp
259; CHECK-NEXT:    retl
260entry:
261  tail call void asm sideeffect "fistl $1", "{st},*m,~{memory},~{dirflag},~{fpsr},~{flags}"(x86_fp80 %x, ptr elementtype(i32) %p) nounwind
262  ret void
263}
264
265; Here, the input operand is tied to an output which means that is is
266; implicitly popped (and then the output is implicitly pushed).
267;
268;   long double fist2(long double x, int *p) {
269;     long double y;
270;     asm ("fistl %1" : "=&t"(y) : "0"(x), "m"(*p) : "memory");
271;     return y;
272;   }
273define x86_fp80 @fist2(x86_fp80 %x, ptr %p) nounwind ssp {
274; CHECK-LABEL: fist2:
275; CHECK:       ## %bb.0: ## %entry
276; CHECK-NEXT:    subl $12, %esp
277; CHECK-NEXT:    fldt {{[0-9]+}}(%esp)
278; CHECK-NEXT:    movl {{[0-9]+}}(%esp), %eax
279; CHECK-NEXT:    ## InlineAsm Start
280; CHECK-NEXT:    fistl (%eax)
281; CHECK-NEXT:    ## InlineAsm End
282; CHECK-NEXT:    addl $12, %esp
283; CHECK-NEXT:    retl
284entry:
285  %0 = tail call x86_fp80 asm "fistl $2", "=&{st},0,*m,~{memory},~{dirflag},~{fpsr},~{flags}"(x86_fp80 %x, ptr elementtype(i32) %p) nounwind
286  ret x86_fp80 %0
287}
288
289; An 'f' constraint is never implicitly popped:
290;
291;   void fucomp1(long double x, long double y) {
292;     asm volatile ("fucomp %1" : : "t"(x), "f"(y) : "st");
293;   }
294define void @fucomp1(x86_fp80 %x, x86_fp80 %y) nounwind ssp {
295; CHECK-LABEL: fucomp1:
296; CHECK:       ## %bb.0: ## %entry
297; CHECK-NEXT:    subl $12, %esp
298; CHECK-NEXT:    fldt {{[0-9]+}}(%esp)
299; CHECK-NEXT:    fldt {{[0-9]+}}(%esp)
300; CHECK-NEXT:    fxch %st(1)
301; CHECK-NEXT:    ## InlineAsm Start
302; CHECK-NEXT:    fucomp %st(1)
303; CHECK-NEXT:    ## InlineAsm End
304; CHECK-NEXT:    fstp %st(0)
305; CHECK-NEXT:    addl $12, %esp
306; CHECK-NEXT:    retl
307entry:
308  tail call void asm sideeffect "fucomp $1", "{st},f,~{st},~{dirflag},~{fpsr},~{flags}"(x86_fp80 %x, x86_fp80 %y) nounwind
309  ret void
310}
311
312; The 'u' constraint is only popped implicitly when clobbered:
313;
314;   void fucomp2(long double x, long double y) {
315;     asm volatile ("fucomp %1" : : "t"(x), "u"(y) : "st");
316;   }
317;
318;   void fucomp3(long double x, long double y) {
319;     asm volatile ("fucompp %1" : : "t"(x), "u"(y) : "st", "st(1)");
320;   }
321;
322define void @fucomp2(x86_fp80 %x, x86_fp80 %y) nounwind ssp {
323; CHECK-LABEL: fucomp2:
324; CHECK:       ## %bb.0: ## %entry
325; CHECK-NEXT:    subl $12, %esp
326; CHECK-NEXT:    fldt {{[0-9]+}}(%esp)
327; CHECK-NEXT:    fldt {{[0-9]+}}(%esp)
328; CHECK-NEXT:    fxch %st(1)
329; CHECK-NEXT:    ## InlineAsm Start
330; CHECK-NEXT:    fucomp %st(1)
331; CHECK-NEXT:    ## InlineAsm End
332; CHECK-NEXT:    fstp %st(0)
333; CHECK-NEXT:    addl $12, %esp
334; CHECK-NEXT:    retl
335entry:
336  tail call void asm sideeffect "fucomp $1", "{st},{st(1)},~{st},~{dirflag},~{fpsr},~{flags}"(x86_fp80 %x, x86_fp80 %y) nounwind
337  ret void
338}
339
340define void @fucomp3(x86_fp80 %x, x86_fp80 %y) nounwind ssp {
341; CHECK-LABEL: fucomp3:
342; CHECK:       ## %bb.0: ## %entry
343; CHECK-NEXT:    subl $12, %esp
344; CHECK-NEXT:    fldt {{[0-9]+}}(%esp)
345; CHECK-NEXT:    fldt {{[0-9]+}}(%esp)
346; CHECK-NEXT:    fxch %st(1)
347; CHECK-NEXT:    ## InlineAsm Start
348; CHECK-NEXT:    fucompp %st(1)
349; CHECK-NEXT:    ## InlineAsm End
350; CHECK-NEXT:    addl $12, %esp
351; CHECK-NEXT:    retl
352entry:
353  tail call void asm sideeffect "fucompp $1", "{st},{st(1)},~{st},~{st(1)},~{dirflag},~{fpsr},~{flags}"(x86_fp80 %x, x86_fp80 %y) nounwind
354  ret void
355}
356
357; One input, two outputs, one dead output.
358%complex = type { float, float }
359define float @sincos1(float %x) nounwind ssp {
360; CHECK-LABEL: sincos1:
361; CHECK:       ## %bb.0: ## %entry
362; CHECK-NEXT:    flds {{[0-9]+}}(%esp)
363; CHECK-NEXT:    ## InlineAsm Start
364; CHECK-NEXT:    sincos
365; CHECK-NEXT:    ## InlineAsm End
366; CHECK-NEXT:    fstp %st(1)
367; CHECK-NEXT:    retl
368entry:
369  %0 = tail call %complex asm "sincos", "={st},={st(1)},0,~{dirflag},~{fpsr},~{flags}"(float %x) nounwind
370  %asmresult = extractvalue %complex %0, 0
371  ret float %asmresult
372}
373
374; Same thing, swapped output operands.
375define float @sincos2(float %x) nounwind ssp {
376; CHECK-LABEL: sincos2:
377; CHECK:       ## %bb.0: ## %entry
378; CHECK-NEXT:    flds {{[0-9]+}}(%esp)
379; CHECK-NEXT:    ## InlineAsm Start
380; CHECK-NEXT:    sincos
381; CHECK-NEXT:    ## InlineAsm End
382; CHECK-NEXT:    fstp %st(1)
383; CHECK-NEXT:    retl
384entry:
385  %0 = tail call %complex asm "sincos", "={st(1)},={st},1,~{dirflag},~{fpsr},~{flags}"(float %x) nounwind
386  %asmresult = extractvalue %complex %0, 1
387  ret float %asmresult
388}
389
390; Clobber st(0) after it was live-out/dead from the previous asm.
391; Load x, make a copy for the second asm.
392; Discard dead result in st(0), bring x to the top.
393; x is now in st(0) for the second asm
394; Discard both results.
395define float @sincos3(float %x) nounwind ssp {
396; CHECK-LABEL: sincos3:
397; CHECK:       ## %bb.0: ## %entry
398; CHECK-NEXT:    flds {{[0-9]+}}(%esp)
399; CHECK-NEXT:    fld %st(0)
400; CHECK-NEXT:    ## InlineAsm Start
401; CHECK-NEXT:    sincos
402; CHECK-NEXT:    ## InlineAsm End
403; CHECK-NEXT:    fstp %st(0)
404; CHECK-NEXT:    fxch %st(1)
405; CHECK-NEXT:    ## InlineAsm Start
406; CHECK-NEXT:    sincos
407; CHECK-NEXT:    ## InlineAsm End
408; CHECK-NEXT:    fstp %st(1)
409; CHECK-NEXT:    fstp %st(0)
410; CHECK-NEXT:    retl
411entry:
412  %0 = tail call %complex asm sideeffect "sincos", "={st(1)},={st},1,~{dirflag},~{fpsr},~{flags}"(float %x) nounwind
413  %1 = tail call %complex asm sideeffect "sincos", "={st(1)},={st},1,~{dirflag},~{fpsr},~{flags}"(float %x) nounwind
414  %asmresult = extractvalue %complex %0, 0
415  ret float %asmresult
416}
417
418; Pass the same value in two fixed stack slots.
419define i32 @PR10602() nounwind ssp {
420; CHECK-LABEL: PR10602:
421; CHECK:       ## %bb.0: ## %entry
422; CHECK-NEXT:    flds {{\.?LCPI[0-9]+_[0-9]+}}
423; CHECK-NEXT:    fld %st(0)
424; CHECK-NEXT:    fxch %st(1)
425; CHECK-NEXT:    ## InlineAsm Start
426; CHECK-NEXT:    fcomi %st(1), %st; pushf; pop %eax
427; CHECK-NEXT:    ## InlineAsm End
428; CHECK-NEXT:    fstp %st(0)
429; CHECK-NEXT:    fstp %st(0)
430; CHECK-NEXT:    retl
431entry:
432  %0 = tail call i32 asm "fcomi $2, $1; pushf; pop $0", "=r,{st},{st(1)},~{dirflag},~{fpsr},~{flags}"(double 2.000000e+00, double 2.000000e+00) nounwind
433  ret i32 %0
434}
435
436; <rdar://problem/16952634>
437; X87 stackifier asserted when there was an ST register defined by an
438; inline-asm instruction and the ST register was live across another
439; inline-asm instruction.
440;
441; INLINEASM $frndint [sideeffect] [attdialect], $0:[regdef], %st0<imp-def,tied5>, $1:[reguse tiedto:$0], %st0<tied3>, $2:[clobber], early-clobber implicit dead %eflags
442; INLINEASM $fldcw $0 [sideeffect] [mayload] [attdialect], $0:[mem], undef %eax, 1, %noreg, 0, %noreg, $1:[clobber], early-clobber implicit dead %eflags
443; %fp0 = COPY %st0
444
445%struct.fpu_t = type { [8 x x86_fp80], x86_fp80, %struct.anon1, %struct.anon2, i32, i8, [15 x i8] }
446%struct.anon1 = type { i32, i32, i32 }
447%struct.anon2 = type { i32, i32, i32, i32 }
448
449@fpu = external global %struct.fpu_t, align 16
450
451; Function Attrs: ssp
452define void @test_live_st(i32 %a1) nounwind {
453; CHECK-LABEL: test_live_st:
454; CHECK:       ## %bb.0: ## %entry
455; CHECK-NEXT:    subl $12, %esp
456; CHECK-NEXT:    fldt (%eax)
457; CHECK-NEXT:    cmpl $1, {{[0-9]+}}(%esp)
458; CHECK-NEXT:    jne LBB20_2
459; CHECK-NEXT:  ## %bb.1: ## %sw.bb4.i
460; CHECK-NEXT:    ## InlineAsm Start
461; CHECK-NEXT:    frndint
462; CHECK-NEXT:    ## InlineAsm End
463; CHECK-NEXT:    ## InlineAsm Start
464; CHECK-NEXT:    fldcw (%eax)
465; CHECK-NEXT:    ## InlineAsm End
466; CHECK-NEXT:  LBB20_2: ## %_Z5tointRKe.exit
467; CHECK-NEXT:    fnstcw (%esp)
468; CHECK-NEXT:    movzwl (%esp), %eax
469; CHECK-NEXT:    orl $3072, %eax ## imm = 0xC00
470; CHECK-NEXT:    movw %ax, {{[0-9]+}}(%esp)
471; CHECK-NEXT:    fldcw {{[0-9]+}}(%esp)
472; CHECK-NEXT:    fistpl {{[0-9]+}}(%esp)
473; CHECK-NEXT:    fldcw (%esp)
474; CHECK-NEXT:    movl {{[0-9]+}}(%esp), %eax
475; CHECK-NEXT:    movl %eax, {{[0-9]+}}(%esp)
476; CHECK-NEXT:    fildl {{[0-9]+}}(%esp)
477; CHECK-NEXT:    movl L_fpu$non_lazy_ptr, %eax
478; CHECK-NEXT:    fstpt 128(%eax)
479; CHECK-NEXT:    addl $12, %esp
480; CHECK-NEXT:    retl
481entry:
482  %0 = load x86_fp80, ptr undef, align 16
483  %cond = icmp eq i32 %a1, 1
484  br i1 %cond, label %sw.bb4.i, label %_Z5tointRKe.exit
485
486sw.bb4.i:
487  %1 = call x86_fp80 asm sideeffect "frndint", "={st},0,~{dirflag},~{fpsr},~{flags}"(x86_fp80 %0)
488  call void asm sideeffect "fldcw $0", "*m,~{dirflag},~{fpsr},~{flags}"(ptr elementtype(i32) undef)
489  br label %_Z5tointRKe.exit
490
491_Z5tointRKe.exit:
492  %result.0.i = phi x86_fp80 [ %1, %sw.bb4.i ], [ %0, %entry ]
493  %conv.i1814 = fptosi x86_fp80 %result.0.i to i32
494  %conv626 = sitofp i32 %conv.i1814 to x86_fp80
495  store x86_fp80 %conv626, ptr getelementptr inbounds (%struct.fpu_t, ptr @fpu, i32 0, i32 1)
496  br label %return
497
498return:
499  ret void
500}
501
502; Check that x87 stackifier is correctly rewriting FP registers to ST registers.
503define double @test_operand_rewrite() nounwind {
504; CHECK-LABEL: test_operand_rewrite:
505; CHECK:       ## %bb.0: ## %entry
506; CHECK-NEXT:    ## InlineAsm Start
507; CHECK-NEXT:    foo %st, %st(1)
508; CHECK-NEXT:    ## InlineAsm End
509; CHECK-NEXT:    fsubp %st, %st(1)
510; CHECK-NEXT:    retl
511entry:
512  %0 = tail call { double, double } asm sideeffect "foo $0, $1", "={st},={st(1)},~{dirflag},~{fpsr},~{flags}"()
513  %asmresult = extractvalue { double, double } %0, 0
514  %asmresult1 = extractvalue { double, double } %0, 1
515  %sub = fsub double %asmresult, %asmresult1
516  ret double %sub
517}
518