xref: /llvm-project/llvm/test/CodeGen/X86/pseudo_cmov_lower2.ll (revision b468ed494acde4d1cc496a436ab9109660db5b80)
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mtriple=x86_64-linux-gnu -o - | FileCheck %s
3
4; This test checks that only a single jae gets generated in the final code
5; for lowering the CMOV pseudos that get created for this IR.  The tricky part
6; of this test is that it tests the special PHI operand rewriting code in
7; X86TargetLowering::EmitLoweredSelect.
8;
9define double @foo1(float %p1, double %p2, double %p3) nounwind {
10entry:
11  %c1 = fcmp oge float %p1, 0.000000e+00
12  %d0 = fadd double %p2, 1.25e0
13  %d1 = fadd double %p3, 1.25e0
14  %d2 = select i1 %c1, double %d0, double %d1
15  %d3 = select i1 %c1, double %d2, double %p2
16  %d4 = select i1 %c1, double %d3, double %p3
17  %d5 = fsub double %d2, %d3
18  %d6 = fadd double %d5, %d4
19  ret double %d6
20}
21
22; This test checks that only a single jae gets generated in the final code
23; for lowering the CMOV pseudos that get created for this IR.  The tricky part
24; of this test is that it tests the special PHI operand rewriting code in
25; X86TargetLowering::EmitLoweredSelect.
26;
27define double @foo2(float %p1, double %p2, double %p3) nounwind {
28; CHECK-LABEL: foo2:
29; CHECK:       # %bb.0: # %entry
30; CHECK-NEXT:    xorps %xmm3, %xmm3
31; CHECK-NEXT:    ucomiss %xmm3, %xmm0
32; CHECK-NEXT:    movsd {{.*#+}} xmm0 = [1.25E+0,0.0E+0]
33; CHECK-NEXT:    jae .LBB1_1
34; CHECK-NEXT:  # %bb.2: # %entry
35; CHECK-NEXT:    addsd %xmm0, %xmm2
36; CHECK-NEXT:    movapd %xmm2, %xmm0
37; CHECK-NEXT:    movapd %xmm2, %xmm1
38; CHECK-NEXT:    jmp .LBB1_3
39; CHECK-NEXT:  .LBB1_1:
40; CHECK-NEXT:    addsd %xmm1, %xmm0
41; CHECK-NEXT:  .LBB1_3: # %entry
42; CHECK-NEXT:    subsd %xmm1, %xmm0
43; CHECK-NEXT:    addsd %xmm2, %xmm0
44; CHECK-NEXT:    retq
45entry:
46  %c1 = fcmp oge float %p1, 0.000000e+00
47  %d0 = fadd double %p2, 1.25e0
48  %d1 = fadd double %p3, 1.25e0
49  %d2 = select i1 %c1, double %d0, double %d1
50  %d3 = select i1 %c1, double %p2, double %d2
51  %d4 = select i1 %c1, double %p3, double %d3
52  %d5 = fsub double %d2, %d3
53  %d6 = fadd double %d5, %d4
54  ret double %d6
55}
56
57; This test checks that only a single js gets generated in the final code
58; for lowering the CMOV pseudos that get created for this IR.  The tricky part
59; of this test is that it tests the special PHI operand rewriting code in
60; X86TargetLowering::EmitLoweredSelect.  It also tests to make sure all
61; the operands of the resulting instructions are from the proper places.
62;
63define double @foo3(i32 %p1, double %p2, double %p3,
64; CHECK-LABEL: foo3:
65; CHECK:       # %bb.0: # %entry
66; CHECK-NEXT:    testl %edi, %edi
67; CHECK-NEXT:    js .LBB2_2
68; CHECK-NEXT:  # %bb.1: # %entry
69; CHECK-NEXT:    movapd %xmm2, %xmm1
70; CHECK-NEXT:    movapd %xmm2, %xmm0
71; CHECK-NEXT:  .LBB2_2: # %entry
72; CHECK-NEXT:    divsd %xmm1, %xmm0
73; CHECK-NEXT:    retq
74                             double %p4, double %p5) nounwind {
75entry:
76  %c1 = icmp slt i32 %p1, 0
77  %d2 = select i1 %c1, double %p2, double %p3
78  %d3 = select i1 %c1, double %p3, double %p4
79  %d4 = select i1 %c1, double %d2, double %d3
80  %d5 = fdiv double %d4, %d3
81  ret double %d5
82}
83
84; This test checks that only a single js gets generated in the final code
85; for lowering the CMOV pseudos that get created for this IR.  The tricky part
86; of this test is that it tests the special PHI operand rewriting code in
87; X86TargetLowering::EmitLoweredSelect.  It also tests to make sure all
88; the operands of the resulting instructions are from the proper places
89; when the "opposite condition" handling code in the compiler is used.
90; This should be the same code as foo3 above, because we use the opposite
91; condition code in the second two selects, but we also swap the operands
92; of the selects to give the same actual computation.
93;
94define double @foo4(i32 %p1, double %p2, double %p3,
95; CHECK-LABEL: foo4:
96; CHECK:       # %bb.0: # %entry
97; CHECK-NEXT:    testl %edi, %edi
98; CHECK-NEXT:    js .LBB3_2
99; CHECK-NEXT:  # %bb.1: # %entry
100; CHECK-NEXT:    movapd %xmm2, %xmm1
101; CHECK-NEXT:    movapd %xmm2, %xmm0
102; CHECK-NEXT:  .LBB3_2: # %entry
103; CHECK-NEXT:    divsd %xmm1, %xmm0
104; CHECK-NEXT:    retq
105                             double %p4, double %p5) nounwind {
106entry:
107  %c1 = icmp slt i32 %p1, 0
108  %d2 = select i1 %c1, double %p2, double %p3
109  %c2 = icmp sge i32 %p1, 0
110  %d3 = select i1 %c2, double %p4, double %p3
111  %d4 = select i1 %c2, double %d3, double %d2
112  %d5 = fdiv double %d4, %d3
113  ret double %d5
114}
115
116; This test checks that only a single jae gets generated in the final code
117; for lowering the CMOV pseudos that get created for this IR.  The tricky part
118; of this test is that it tests the special code in CodeGenPrepare.
119;
120define double @foo5(float %p1, double %p2, double %p3) nounwind {
121; CHECK-LABEL: foo5:
122; CHECK:       # %bb.0: # %entry
123; CHECK-NEXT:    xorps %xmm3, %xmm3
124; CHECK-NEXT:    ucomiss %xmm3, %xmm0
125; CHECK-NEXT:    movsd {{.*#+}} xmm0 = [1.25E+0,0.0E+0]
126; CHECK-NEXT:    jae .LBB4_1
127; CHECK-NEXT:  # %bb.2: # %select.false
128; CHECK-NEXT:    addsd %xmm2, %xmm0
129; CHECK-NEXT:  .LBB4_3: # %select.end
130; CHECK-NEXT:    subsd %xmm1, %xmm0
131; CHECK-NEXT:    addsd %xmm2, %xmm0
132; CHECK-NEXT:    retq
133; CHECK-NEXT:  .LBB4_1:
134; CHECK-NEXT:    addsd %xmm0, %xmm1
135; CHECK-NEXT:    movapd %xmm1, %xmm0
136; CHECK-NEXT:    movapd %xmm1, %xmm2
137; CHECK-NEXT:    jmp .LBB4_3
138entry:
139  %c1 = fcmp oge float %p1, 0.000000e+00
140  %d0 = fadd double %p2, 1.25e0
141  %d1 = fadd double %p3, 1.25e0
142  %d2 = select i1 %c1, double %d0, double %d1, !prof !0
143  %d3 = select i1 %c1, double %d2, double %p2, !prof !0
144  %d4 = select i1 %c1, double %d3, double %p3, !prof !0
145  %d5 = fsub double %d2, %d3
146  %d6 = fadd double %d5, %d4
147  ret double %d6
148}
149
150; We should expand select instructions into 3 conditional branches as their
151; condtions are different.
152;
153define double @foo6(float %p1, double %p2, double %p3) nounwind {
154; CHECK-LABEL: foo6:
155; CHECK:       # %bb.0: # %entry
156; CHECK-NEXT:    movaps %xmm0, %xmm3
157; CHECK-NEXT:    xorps %xmm0, %xmm0
158; CHECK-NEXT:    ucomiss %xmm0, %xmm3
159; CHECK-NEXT:    movsd {{.*#+}} xmm0 = [1.25E+0,0.0E+0]
160; CHECK-NEXT:    jae .LBB5_1
161; CHECK-NEXT:  # %bb.2: # %select.false
162; CHECK-NEXT:    addsd %xmm2, %xmm0
163; CHECK-NEXT:  .LBB5_3: # %select.end
164; CHECK-NEXT:    ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
165; CHECK-NEXT:    movapd %xmm0, %xmm4
166; CHECK-NEXT:    jae .LBB5_5
167; CHECK-NEXT:  # %bb.4: # %select.false2
168; CHECK-NEXT:    movapd %xmm1, %xmm4
169; CHECK-NEXT:  .LBB5_5: # %select.end1
170; CHECK-NEXT:    ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
171; CHECK-NEXT:    movapd %xmm4, %xmm1
172; CHECK-NEXT:    jae .LBB5_7
173; CHECK-NEXT:  # %bb.6: # %select.false4
174; CHECK-NEXT:    movapd %xmm2, %xmm1
175; CHECK-NEXT:  .LBB5_7: # %select.end3
176; CHECK-NEXT:    subsd %xmm4, %xmm0
177; CHECK-NEXT:    addsd %xmm1, %xmm0
178; CHECK-NEXT:    retq
179; CHECK-NEXT:  .LBB5_1:
180; CHECK-NEXT:    addsd %xmm1, %xmm0
181; CHECK-NEXT:    jmp .LBB5_3
182entry:
183  %c1 = fcmp oge float %p1, 0.000000e+00
184  %c2 = fcmp oge float %p1, 1.000000e+00
185  %c3 = fcmp oge float %p1, 2.000000e+00
186  %d0 = fadd double %p2, 1.25e0
187  %d1 = fadd double %p3, 1.25e0
188  %d2 = select i1 %c1, double %d0, double %d1, !prof !0
189  %d3 = select i1 %c2, double %d2, double %p2, !prof !0
190  %d4 = select i1 %c3, double %d3, double %p3, !prof !0
191  %d5 = fsub double %d2, %d3
192  %d6 = fadd double %d5, %d4
193  ret double %d6
194}
195
196declare void @llvm.dbg.value(metadata, metadata, metadata)
197
198; Like the test for @foo1, but check that the inserted dbg.value does not
199; affect codegen. The CHECK items below should always be identical to @foo1,
200; minus the DEBUG_VALUE line and changes in labels..
201define double @foo1_g(float %p1, double %p2, double %p3) nounwind !dbg !4 {
202; CHECK-LABEL: foo1_g:
203; CHECK:       # %bb.0: # %entry
204; CHECK-NEXT:    .file   1 "." "test.c"
205; CHECK-NEXT:    .loc 1 3 0 prologue_end
206; CHECK-NEXT:    xorps %xmm3, %xmm3
207; CHECK-NEXT:    ucomiss %xmm3, %xmm0
208; CHECK-NEXT:    movsd {{.*#+}} xmm0 = [1.25E+0,0.0E+0]
209; CHECK-NEXT:    jae .LBB6_1
210; CHECK-NEXT:  # %bb.2: # %entry
211; CHECK-NEXT:    addsd %xmm2, %xmm0
212; CHECK-NEXT:    jmp .LBB6_3
213; CHECK-NEXT:  .LBB6_1:
214; CHECK-NEXT:    addsd %xmm0, %xmm1
215; CHECK-NEXT:    movapd %xmm1, %xmm0
216; CHECK-NEXT:    movapd %xmm1, %xmm2
217; CHECK-NEXT:  .LBB6_3: # %entry
218; CHECK-NEXT:    #DEBUG_VALUE: foobar:xyzzy <- undef
219; CHECK-NEXT:    subsd %xmm1, %xmm0
220; CHECK-NEXT:    addsd %xmm2, %xmm0
221; CHECK-NEXT:    retq
222entry:
223  %c1 = fcmp oge float %p1, 0.000000e+00
224  %d0 = fadd double %p2, 1.25e0
225  %d1 = fadd double %p3, 1.25e0
226  %d2 = select i1 %c1, double %d0, double %d1
227  call void @llvm.dbg.value(metadata float undef, metadata !5, metadata !DIExpression()), !dbg !6
228  %d3 = select i1 %c1, double %d2, double %p2
229  %d4 = select i1 %c1, double %d3, double %p3
230  %d5 = fsub double %d2, %d3
231  %d6 = fadd double %d5, %d4
232  ret double %d6
233}
234
235!llvm.module.flags = !{!1}
236!llvm.dbg.cu = !{!2}
237
238!0 = !{!"branch_weights", i32 1, i32 2000}
239!1 = !{i32 2, !"Debug Info Version", i32 3}
240!2 = distinct !DICompileUnit(language: DW_LANG_C99, file: !3, producer: "clang", isOptimized: true, runtimeVersion: 0, emissionKind: FullDebug, nameTableKind: None)
241!3 = !DIFile(filename: "test.c", directory: ".")
242!4 = distinct !DISubprogram(name: "foobar", scope: !2, file: !3, line: 1, type: !9, scopeLine: 3, flags: DIFlagPrototyped, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !2, retainedNodes: !8)
243!5 = !DILocalVariable(name: "xyzzy", scope: !4, file: !3, line: 2, type: !7)
244!6 = !DILocation(line: 1, column: 1, scope: !4)
245!7 = !DIBasicType(name: "float", size: 32, encoding: DW_ATE_float)
246!8 = !{!5}
247!9 = !DISubroutineType(types: !10)
248!10 = !{!7}
249