1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc < %s -mtriple=i686-- -mattr=+sse4.1 | FileCheck %s
4 ; Test that when we don't -enable-unsafe-fp-math, we don't do the optimization
5 ; -0 - (A - B) to (B - A) because A==B, -0 != 0
7 define float @negfp(float %a, float %b) nounwind {
9 ; CHECK: # %bb.0: # %entry
10 ; CHECK-NEXT: pushl %eax
11 ; CHECK-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
12 ; CHECK-NEXT: subss {{[0-9]+}}(%esp), %xmm0
13 ; CHECK-NEXT: xorps {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
14 ; CHECK-NEXT: movss %xmm0, (%esp)
15 ; CHECK-NEXT: flds (%esp)
16 ; CHECK-NEXT: popl %eax
19 %sub = fsub float %a, %b ; <float> [#uses=1]
20 %neg = fsub float -0.000000e+00, %sub ; <float> [#uses=1]
24 ; This may infinite loop if isNegatibleForFree and getNegatedExpression are conflicted.
26 define double @negation_propagation(ptr %arg, double %arg1, double %arg2) nounwind {
27 ; CHECK-LABEL: negation_propagation:
29 ; CHECK-NEXT: pushl %ebp
30 ; CHECK-NEXT: movl %esp, %ebp
31 ; CHECK-NEXT: andl $-8, %esp
32 ; CHECK-NEXT: subl $8, %esp
33 ; CHECK-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
34 ; CHECK-NEXT: divsd 12(%ebp), %xmm0
35 ; CHECK-NEXT: movsd {{.*#+}} xmm1 = mem[0],zero
36 ; CHECK-NEXT: mulsd %xmm0, %xmm1
37 ; CHECK-NEXT: movapd %xmm0, %xmm2
38 ; CHECK-NEXT: mulsd %xmm0, %xmm2
39 ; CHECK-NEXT: mulsd %xmm0, %xmm2
40 ; CHECK-NEXT: subsd %xmm2, %xmm1
41 ; CHECK-NEXT: movsd %xmm1, (%esp)
42 ; CHECK-NEXT: fldl (%esp)
43 ; CHECK-NEXT: movl %ebp, %esp
44 ; CHECK-NEXT: popl %ebp
46 %t = fdiv double 1.0, %arg1
47 %t7 = fmul double %t, %arg2
48 %t10 = fneg double %t7
49 %t11 = fmul double %t, %t
50 %t13 = fsub double %t11, %t
52 %t15 = fmul double %t, %t14
53 %t16 = fmul double %t, %t15
54 %t18 = fadd double %t16, %t7
58 ; This would crash because the negated expression for %sub4
59 ; creates a new use of %sub1 and that alters the negated cost
61 define float @fdiv_extra_use_changes_cost(float %a0, float %a1, float %a2) nounwind {
62 ; CHECK-LABEL: fdiv_extra_use_changes_cost:
64 ; CHECK-NEXT: pushl %eax
65 ; CHECK-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
66 ; CHECK-NEXT: movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
67 ; CHECK-NEXT: subss {{[0-9]+}}(%esp), %xmm1
68 ; CHECK-NEXT: movaps %xmm1, %xmm2
69 ; CHECK-NEXT: mulss %xmm0, %xmm2
70 ; CHECK-NEXT: subss %xmm1, %xmm0
71 ; CHECK-NEXT: divss %xmm2, %xmm0
72 ; CHECK-NEXT: movss %xmm0, (%esp)
73 ; CHECK-NEXT: flds (%esp)
74 ; CHECK-NEXT: popl %eax
76 %sub1 = fsub fast float %a0, %a1
77 %mul2 = fmul fast float %sub1, %a2
78 %neg = fneg fast float %a0
79 %add3 = fadd fast float %a1, %neg
80 %sub4 = fadd fast float %add3, %a2
81 %div5 = fdiv fast float %sub4, %mul2
85 ; PR55758 - this is not -(-X)
87 define <2 x i64> @fneg_mismatched_sizes(<4 x float> %x) {
88 ; CHECK-LABEL: fneg_mismatched_sizes:
90 ; CHECK-NEXT: xorps {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
91 ; CHECK-NEXT: xorps {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
93 %n = fneg <4 x float> %x
94 %b = bitcast <4 x float> %n to <2 x i64>
95 %r = xor <2 x i64> %b, <i64 -9223372036854775808, i64 -9223372036854775808>