Skip to content

Commit f061c10

Browse files
RKSimonRKSimon
committed
[SLP][X86] Add ray_sphere intersection methods from c-ray benchmark
We're failing to vectorize several comparison reduction patterns. Issue llvm#43090 was based off this, but while that simplified test case is now folding, the original still fails due to poor cost model values for vXi1 extractions
1 parent a66ff23 commit f061c10

File tree

1 file changed

+170
-0
lines changed
  • llvm/test/Transforms/SLPVectorizer/X86

1 file changed

+170
-0
lines changed

‎llvm/test/Transforms/SLPVectorizer/X86/c-ray.ll

Lines changed: 170 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,170 @@
1+
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2+
; RUN: opt < %s -mtriple=x86_64-unknown -slp-vectorizer -S | FileCheck %s
3+
; RUN: opt < %s -mtriple=x86_64-unknown -mcpu=corei7-avx -slp-vectorizer -S | FileCheck %s
4+
; RUN: opt < %s -mtriple=x86_64-unknown -mcpu=core-avx2 -slp-vectorizer -S | FileCheck %s
5+
6+
%struct.ray = type { %struct.vec3, %struct.vec3 }
7+
%struct.vec3 = type { double, double, double }
8+
%struct.sphere = type { %struct.vec3, double, %struct.material, ptr }
9+
%struct.material = type { %struct.vec3, double, double }
10+
11+
define i32 @ray_sphere(ptr nocapture noundef readonly %sph, ptr nocapture noundef readonly byval(%struct.ray) align 8 %ray, ptr nocapture noundef readnone %sp) {
12+
; CHECK-LABEL: @ray_sphere(
13+
; CHECK-NEXT: entry:
14+
; CHECK-NEXT: [[DIR:%.*]] = getelementptr inbounds [[STRUCT_RAY:%.*]], ptr [[RAY:%.*]], i64 0, i32 1
15+
; CHECK-NEXT: [[TMP0:%.*]] = load double, ptr [[DIR]], align 8
16+
; CHECK-NEXT: [[Y:%.*]] = getelementptr inbounds [[STRUCT_RAY]], ptr [[RAY]], i64 0, i32 1, i32 1
17+
; CHECK-NEXT: [[TMP1:%.*]] = load double, ptr [[Y]], align 8
18+
; CHECK-NEXT: [[MUL6:%.*]] = fmul double [[TMP1]], [[TMP1]]
19+
; CHECK-NEXT: [[TMP2:%.*]] = tail call double @llvm.fmuladd.f64(double [[TMP0]], double [[TMP0]], double [[MUL6]])
20+
; CHECK-NEXT: [[Z:%.*]] = getelementptr inbounds [[STRUCT_RAY]], ptr [[RAY]], i64 0, i32 1, i32 2
21+
; CHECK-NEXT: [[TMP3:%.*]] = load double, ptr [[Z]], align 8
22+
; CHECK-NEXT: [[TMP4:%.*]] = tail call double @llvm.fmuladd.f64(double [[TMP3]], double [[TMP3]], double [[TMP2]])
23+
; CHECK-NEXT: [[MUL:%.*]] = fmul double [[TMP0]], 2.000000e+00
24+
; CHECK-NEXT: [[TMP5:%.*]] = load double, ptr [[RAY]], align 8
25+
; CHECK-NEXT: [[TMP6:%.*]] = load double, ptr [[SPH:%.*]], align 8
26+
; CHECK-NEXT: [[SUB:%.*]] = fsub double [[TMP5]], [[TMP6]]
27+
; CHECK-NEXT: [[MUL17:%.*]] = fmul double [[TMP1]], 2.000000e+00
28+
; CHECK-NEXT: [[Y19:%.*]] = getelementptr inbounds [[STRUCT_VEC3:%.*]], ptr [[RAY]], i64 0, i32 1
29+
; CHECK-NEXT: [[TMP7:%.*]] = load double, ptr [[Y19]], align 8
30+
; CHECK-NEXT: [[Y21:%.*]] = getelementptr inbounds [[STRUCT_VEC3]], ptr [[SPH]], i64 0, i32 1
31+
; CHECK-NEXT: [[TMP8:%.*]] = load double, ptr [[Y21]], align 8
32+
; CHECK-NEXT: [[SUB22:%.*]] = fsub double [[TMP7]], [[TMP8]]
33+
; CHECK-NEXT: [[MUL23:%.*]] = fmul double [[MUL17]], [[SUB22]]
34+
; CHECK-NEXT: [[TMP9:%.*]] = tail call double @llvm.fmuladd.f64(double [[MUL]], double [[SUB]], double [[MUL23]])
35+
; CHECK-NEXT: [[MUL26:%.*]] = fmul double [[TMP3]], 2.000000e+00
36+
; CHECK-NEXT: [[Z28:%.*]] = getelementptr inbounds [[STRUCT_VEC3]], ptr [[RAY]], i64 0, i32 2
37+
; CHECK-NEXT: [[TMP10:%.*]] = load double, ptr [[Z28]], align 8
38+
; CHECK-NEXT: [[Z30:%.*]] = getelementptr inbounds [[STRUCT_VEC3]], ptr [[SPH]], i64 0, i32 2
39+
; CHECK-NEXT: [[TMP11:%.*]] = load double, ptr [[Z30]], align 8
40+
; CHECK-NEXT: [[SUB31:%.*]] = fsub double [[TMP10]], [[TMP11]]
41+
; CHECK-NEXT: [[TMP12:%.*]] = tail call double @llvm.fmuladd.f64(double [[MUL26]], double [[SUB31]], double [[TMP9]])
42+
; CHECK-NEXT: [[MUL42:%.*]] = fmul double [[TMP8]], [[TMP8]]
43+
; CHECK-NEXT: [[TMP13:%.*]] = tail call double @llvm.fmuladd.f64(double [[TMP6]], double [[TMP6]], double [[MUL42]])
44+
; CHECK-NEXT: [[TMP14:%.*]] = tail call double @llvm.fmuladd.f64(double [[TMP11]], double [[TMP11]], double [[TMP13]])
45+
; CHECK-NEXT: [[TMP15:%.*]] = tail call double @llvm.fmuladd.f64(double [[TMP5]], double [[TMP5]], double [[TMP14]])
46+
; CHECK-NEXT: [[TMP16:%.*]] = tail call double @llvm.fmuladd.f64(double [[TMP7]], double [[TMP7]], double [[TMP15]])
47+
; CHECK-NEXT: [[TMP17:%.*]] = tail call double @llvm.fmuladd.f64(double [[TMP10]], double [[TMP10]], double [[TMP16]])
48+
; CHECK-NEXT: [[FNEG:%.*]] = fneg double [[TMP6]]
49+
; CHECK-NEXT: [[TMP18:%.*]] = fneg double [[TMP8]]
50+
; CHECK-NEXT: [[NEG:%.*]] = fmul double [[TMP7]], [[TMP18]]
51+
; CHECK-NEXT: [[TMP19:%.*]] = tail call double @llvm.fmuladd.f64(double [[FNEG]], double [[TMP5]], double [[NEG]])
52+
; CHECK-NEXT: [[NEG78:%.*]] = fneg double [[TMP11]]
53+
; CHECK-NEXT: [[TMP20:%.*]] = tail call double @llvm.fmuladd.f64(double [[NEG78]], double [[TMP10]], double [[TMP19]])
54+
; CHECK-NEXT: [[TMP21:%.*]] = tail call double @llvm.fmuladd.f64(double [[TMP20]], double 2.000000e+00, double [[TMP17]])
55+
; CHECK-NEXT: [[RAD:%.*]] = getelementptr inbounds [[STRUCT_SPHERE:%.*]], ptr [[SPH]], i64 0, i32 1
56+
; CHECK-NEXT: [[TMP22:%.*]] = load double, ptr [[RAD]], align 8
57+
; CHECK-NEXT: [[NEG82:%.*]] = fneg double [[TMP22]]
58+
; CHECK-NEXT: [[TMP23:%.*]] = tail call double @llvm.fmuladd.f64(double [[NEG82]], double [[TMP22]], double [[TMP21]])
59+
; CHECK-NEXT: [[TMP24:%.*]] = fmul double [[TMP4]], -4.000000e+00
60+
; CHECK-NEXT: [[NEG86:%.*]] = fmul double [[TMP24]], [[TMP23]]
61+
; CHECK-NEXT: [[TMP25:%.*]] = tail call double @llvm.fmuladd.f64(double [[TMP12]], double [[TMP12]], double [[NEG86]])
62+
; CHECK-NEXT: [[CMP:%.*]] = fcmp olt double [[TMP25]], 0.000000e+00
63+
; CHECK-NEXT: br i1 [[CMP]], label [[CLEANUP:%.*]], label [[IF_END:%.*]]
64+
; CHECK: if.end:
65+
; CHECK-NEXT: [[CALL:%.*]] = tail call double @sqrt(double noundef [[TMP25]])
66+
; CHECK-NEXT: [[FNEG87:%.*]] = fneg double [[TMP12]]
67+
; CHECK-NEXT: [[MUL88:%.*]] = fmul double [[TMP4]], 2.000000e+00
68+
; CHECK-NEXT: [[TMP26:%.*]] = insertelement <2 x double> poison, double [[FNEG87]], i32 0
69+
; CHECK-NEXT: [[TMP27:%.*]] = insertelement <2 x double> [[TMP26]], double [[CALL]], i32 1
70+
; CHECK-NEXT: [[TMP28:%.*]] = insertelement <2 x double> poison, double [[CALL]], i32 0
71+
; CHECK-NEXT: [[TMP29:%.*]] = insertelement <2 x double> [[TMP28]], double [[TMP12]], i32 1
72+
; CHECK-NEXT: [[TMP30:%.*]] = fsub <2 x double> [[TMP27]], [[TMP29]]
73+
; CHECK-NEXT: [[TMP31:%.*]] = insertelement <2 x double> poison, double [[MUL88]], i32 0
74+
; CHECK-NEXT: [[TMP32:%.*]] = insertelement <2 x double> [[TMP31]], double [[MUL88]], i32 1
75+
; CHECK-NEXT: [[TMP33:%.*]] = fdiv <2 x double> [[TMP30]], [[TMP32]]
76+
; CHECK-NEXT: [[TMP34:%.*]] = extractelement <2 x double> [[TMP33]], i32 1
77+
; CHECK-NEXT: [[CMP93:%.*]] = fcmp olt double [[TMP34]], 0x3EB0C6F7A0B5ED8D
78+
; CHECK-NEXT: [[TMP35:%.*]] = extractelement <2 x double> [[TMP33]], i32 0
79+
; CHECK-NEXT: [[CMP94:%.*]] = fcmp olt double [[TMP35]], 0x3EB0C6F7A0B5ED8D
80+
; CHECK-NEXT: [[OR_COND:%.*]] = select i1 [[CMP93]], i1 [[CMP94]], i1 false
81+
; CHECK-NEXT: br i1 [[OR_COND]], label [[CLEANUP]], label [[LOR_LHS_FALSE:%.*]]
82+
; CHECK: lor.lhs.false:
83+
; CHECK-NEXT: [[TMP36:%.*]] = fcmp ule <2 x double> [[TMP33]], <double 1.000000e+00, double 1.000000e+00>
84+
; CHECK-NEXT: [[TMP37:%.*]] = extractelement <2 x i1> [[TMP36]], i32 0
85+
; CHECK-NEXT: [[TMP38:%.*]] = extractelement <2 x i1> [[TMP36]], i32 1
86+
; CHECK-NEXT: [[OR_COND106:%.*]] = select i1 [[TMP38]], i1 true, i1 [[TMP37]]
87+
; CHECK-NEXT: [[SPEC_SELECT:%.*]] = zext i1 [[OR_COND106]] to i32
88+
; CHECK-NEXT: br label [[CLEANUP]]
89+
; CHECK: cleanup:
90+
; CHECK-NEXT: [[RETVAL_0:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ 0, [[IF_END]] ], [ [[SPEC_SELECT]], [[LOR_LHS_FALSE]] ]
91+
; CHECK-NEXT: ret i32 [[RETVAL_0]]
92+
;
93+
entry:
94+
%dir = getelementptr inbounds %struct.ray, ptr %ray, i64 0, i32 1
95+
%0 = load double, ptr %dir, align 8
96+
%y = getelementptr inbounds %struct.ray, ptr %ray, i64 0, i32 1, i32 1
97+
%1 = load double, ptr %y, align 8
98+
%mul6 = fmul double %1, %1
99+
%2 = tail call double @llvm.fmuladd.f64(double %0, double %0, double %mul6)
100+
%z = getelementptr inbounds %struct.ray, ptr %ray, i64 0, i32 1, i32 2
101+
%3 = load double, ptr %z, align 8
102+
%4 = tail call double @llvm.fmuladd.f64(double %3, double %3, double %2)
103+
%mul = fmul double %0, 2.000000e+00
104+
%5 = load double, ptr %ray, align 8
105+
%6 = load double, ptr %sph, align 8
106+
%sub = fsub double %5, %6
107+
%mul17 = fmul double %1, 2.000000e+00
108+
%y19 = getelementptr inbounds %struct.vec3, ptr %ray, i64 0, i32 1
109+
%7 = load double, ptr %y19, align 8
110+
%y21 = getelementptr inbounds %struct.vec3, ptr %sph, i64 0, i32 1
111+
%8 = load double, ptr %y21, align 8
112+
%sub22 = fsub double %7, %8
113+
%mul23 = fmul double %mul17, %sub22
114+
%9 = tail call double @llvm.fmuladd.f64(double %mul, double %sub, double %mul23)
115+
%mul26 = fmul double %3, 2.000000e+00
116+
%z28 = getelementptr inbounds %struct.vec3, ptr %ray, i64 0, i32 2
117+
%10 = load double, ptr %z28, align 8
118+
%z30 = getelementptr inbounds %struct.vec3, ptr %sph, i64 0, i32 2
119+
%11 = load double, ptr %z30, align 8
120+
%sub31 = fsub double %10, %11
121+
%12 = tail call double @llvm.fmuladd.f64(double %mul26, double %sub31, double %9)
122+
%mul42 = fmul double %8, %8
123+
%13 = tail call double @llvm.fmuladd.f64(double %6, double %6, double %mul42)
124+
%14 = tail call double @llvm.fmuladd.f64(double %11, double %11, double %13)
125+
%15 = tail call double @llvm.fmuladd.f64(double %5, double %5, double %14)
126+
%16 = tail call double @llvm.fmuladd.f64(double %7, double %7, double %15)
127+
%17 = tail call double @llvm.fmuladd.f64(double %10, double %10, double %16)
128+
%fneg = fneg double %6
129+
%18 = fneg double %8
130+
%neg = fmul double %7, %18
131+
%19 = tail call double @llvm.fmuladd.f64(double %fneg, double %5, double %neg)
132+
%neg78 = fneg double %11
133+
%20 = tail call double @llvm.fmuladd.f64(double %neg78, double %10, double %19)
134+
%21 = tail call double @llvm.fmuladd.f64(double %20, double 2.000000e+00, double %17)
135+
%rad = getelementptr inbounds %struct.sphere, ptr %sph, i64 0, i32 1
136+
%22 = load double, ptr %rad, align 8
137+
%neg82 = fneg double %22
138+
%23 = tail call double @llvm.fmuladd.f64(double %neg82, double %22, double %21)
139+
%24 = fmul double %4, -4.000000e+00
140+
%neg86 = fmul double %24, %23
141+
%25 = tail call double @llvm.fmuladd.f64(double %12, double %12, double %neg86)
142+
%cmp = fcmp olt double %25, 0.000000e+00
143+
br i1 %cmp, label %cleanup, label %if.end
144+
145+
if.end: ; preds = %entry
146+
%call = tail call double @sqrt(double noundef %25) #3
147+
%fneg87 = fneg double %12
148+
%add = fsub double %call, %12
149+
%mul88 = fmul double %4, 2.000000e+00
150+
%div = fdiv double %add, %mul88
151+
%sub90 = fsub double %fneg87, %call
152+
%div92 = fdiv double %sub90, %mul88
153+
%cmp93 = fcmp olt double %div, 0x3EB0C6F7A0B5ED8D
154+
%cmp94 = fcmp olt double %div92, 0x3EB0C6F7A0B5ED8D
155+
%or.cond = select i1 %cmp93, i1 %cmp94, i1 false
156+
br i1 %or.cond, label %cleanup, label %lor.lhs.false
157+
158+
lor.lhs.false: ; preds = %if.end
159+
%cmp95 = fcmp ule double %div, 1.000000e+00
160+
%cmp97 = fcmp ule double %div92, 1.000000e+00
161+
%or.cond106 = select i1 %cmp95, i1 true, i1 %cmp97
162+
%spec.select = zext i1 %or.cond106 to i32
163+
br label %cleanup
164+
165+
cleanup: ; preds = %lor.lhs.false, %if.end, %entry
166+
%retval.0 = phi i32 [ 0, %entry ], [ 0, %if.end ], [ %spec.select, %lor.lhs.false ]
167+
ret i32 %retval.0
168+
}
169+
declare double @sqrt(double)
170+
declare double @llvm.fmuladd.f64(double, double, double)

0 commit comments

Comments
 (0)