Skip to content

Commit f274230

Browse files
author
Simon Moll
committed
[VE] Enable v256 fcmp true|false tests
The broadcast patterns for all-true|false masks are available now. Enable the true|fast fcmp predicate tests that use them. Reviewed By: kaz7 Differential Revision: https://reviews.llvm.org/D119936
1 parent d46e498 commit f274230

File tree

1 file changed

+66
-44
lines changed

1 file changed

+66
-44
lines changed

‎llvm/test/CodeGen/VE/Vector/vec_fcmp.ll

Lines changed: 66 additions & 44 deletions
Original file line numberDiff line numberDiff line change
@@ -4,28 +4,39 @@
44

55
; <256 x float>
66

7-
;; TODO v256i1 zero-mask isel
8-
;; ; Function Attrs: nounwind
9-
;; define fastcc <256 x i1> @fcmp_false_vv_v256f32(<256 x float> %x, <256 x float> %y) {
10-
;; %z = fcmp false <256 x float> %x, %y
11-
;; ret <256 x i1> %z
12-
;; }
13-
;;
14-
;; ; Function Attrs: nounwind
15-
;; define fastcc <256 x i1> @fcmp_false_sv_v256f32(float %x, <256 x float> %y) {
16-
;; %xins = insertelement <256 x float> undef, float %x, i32 0
17-
;; %vx = shufflevector <256 x float> %xins, <256 x float> undef, <256 x i32> zeroinitializer
18-
;; %z = fcmp false <256 x float> %vx, %y
19-
;; ret <256 x i1> %z
20-
;; }
21-
;;
22-
;; ; Function Attrs: nounwind
23-
;; define fastcc <256 x i1> @fcmp_false_vs_v256f32(<256 x float> %x, float %y) {
24-
;; %yins = insertelement <256 x float> undef, float %y, i32 0
25-
;; %vy = shufflevector <256 x float> %yins, <256 x float> undef, <256 x i32> zeroinitializer
26-
;; %z = fcmp false <256 x float> %x, %vy
27-
;; ret <256 x i1> %z
28-
;; }
7+
; Function Attrs: nounwind
8+
define fastcc <256 x i1> @fcmp_false_vv_v256f32(<256 x float> %x, <256 x float> %y) {
9+
; CHECK-LABEL: fcmp_false_vv_v256f32:
10+
; CHECK: # %bb.0:
11+
; CHECK-NEXT: xorm %vm1, %vm0, %vm0
12+
; CHECK-NEXT: b.l.t (, %s10)
13+
%z = fcmp false <256 x float> %x, %y
14+
ret <256 x i1> %z
15+
}
16+
17+
; Function Attrs: nounwind
18+
define fastcc <256 x i1> @fcmp_false_sv_v256f32(float %x, <256 x float> %y) {
19+
; CHECK-LABEL: fcmp_false_sv_v256f32:
20+
; CHECK: # %bb.0:
21+
; CHECK-NEXT: xorm %vm1, %vm0, %vm0
22+
; CHECK-NEXT: b.l.t (, %s10)
23+
%xins = insertelement <256 x float> undef, float %x, i32 0
24+
%vx = shufflevector <256 x float> %xins, <256 x float> undef, <256 x i32> zeroinitializer
25+
%z = fcmp false <256 x float> %vx, %y
26+
ret <256 x i1> %z
27+
}
28+
29+
; Function Attrs: nounwind
30+
define fastcc <256 x i1> @fcmp_false_vs_v256f32(<256 x float> %x, float %y) {
31+
; CHECK-LABEL: fcmp_false_vs_v256f32:
32+
; CHECK: # %bb.0:
33+
; CHECK-NEXT: xorm %vm1, %vm0, %vm0
34+
; CHECK-NEXT: b.l.t (, %s10)
35+
%yins = insertelement <256 x float> undef, float %y, i32 0
36+
%vy = shufflevector <256 x float> %yins, <256 x float> undef, <256 x i32> zeroinitializer
37+
%z = fcmp false <256 x float> %x, %vy
38+
ret <256 x i1> %z
39+
}
2940

3041
; Function Attrs: nounwind
3142
define fastcc <256 x i1> @fcmp_oeq_vv_v256f32(<256 x float> %x, <256 x float> %y) {
@@ -657,25 +668,36 @@ define fastcc <256 x i1> @fcmp_une_vs_v256f32(<256 x float> %x, float %y) {
657668
ret <256 x i1> %z
658669
}
659670

660-
;; TODO v256i1 all-one mask isel.
661-
;; ; Function Attrs: nounwind
662-
;; define fastcc <256 x i1> @fcmp_true_vv_v256f32(<256 x float> %x, <256 x float> %y) {
663-
;; %z = fcmp true <256 x float> %x, %y
664-
;; ret <256 x i1> %z
665-
;; }
666-
;;
667-
;; ; Function Attrs: nounwind
668-
;; define fastcc <256 x i1> @fcmp_true_sv_v256f32(float %x, <256 x float> %y) {
669-
;; %xins = insertelement <256 x float> undef, float %x, i32 0
670-
;; %vx = shufflevector <256 x float> %xins, <256 x float> undef, <256 x i32> zeroinitializer
671-
;; %z = fcmp true <256 x float> %vx, %y
672-
;; ret <256 x i1> %z
673-
;; }
674-
;;
675-
;; ; Function Attrs: nounwind
676-
;; define fastcc <256 x i1> @fcmp_true_vs_v256f32(<256 x float> %x, float %y) {
677-
;; %yins = insertelement <256 x float> undef, float %y, i32 0
678-
;; %vy = shufflevector <256 x float> %yins, <256 x float> undef, <256 x i32> zeroinitializer
679-
;; %z = fcmp true <256 x float> %x, %vy
680-
;; ret <256 x i1> %z
681-
;; }
671+
; Function Attrs: nounwind
672+
define fastcc <256 x i1> @fcmp_true_vv_v256f32(<256 x float> %x, <256 x float> %y) {
673+
; CHECK-LABEL: fcmp_true_vv_v256f32:
674+
; CHECK: # %bb.0:
675+
; CHECK-NEXT: andm %vm1, %vm0, %vm0
676+
; CHECK-NEXT: b.l.t (, %s10)
677+
%z = fcmp true <256 x float> %x, %y
678+
ret <256 x i1> %z
679+
}
680+
681+
; Function Attrs: nounwind
682+
define fastcc <256 x i1> @fcmp_true_sv_v256f32(float %x, <256 x float> %y) {
683+
; CHECK-LABEL: fcmp_true_sv_v256f32:
684+
; CHECK: # %bb.0:
685+
; CHECK-NEXT: andm %vm1, %vm0, %vm0
686+
; CHECK-NEXT: b.l.t (, %s10)
687+
%xins = insertelement <256 x float> undef, float %x, i32 0
688+
%vx = shufflevector <256 x float> %xins, <256 x float> undef, <256 x i32> zeroinitializer
689+
%z = fcmp true <256 x float> %vx, %y
690+
ret <256 x i1> %z
691+
}
692+
693+
; Function Attrs: nounwind
694+
define fastcc <256 x i1> @fcmp_true_vs_v256f32(<256 x float> %x, float %y) {
695+
; CHECK-LABEL: fcmp_true_vs_v256f32:
696+
; CHECK: # %bb.0:
697+
; CHECK-NEXT: andm %vm1, %vm0, %vm0
698+
; CHECK-NEXT: b.l.t (, %s10)
699+
%yins = insertelement <256 x float> undef, float %y, i32 0
700+
%vy = shufflevector <256 x float> %yins, <256 x float> undef, <256 x i32> zeroinitializer
701+
%z = fcmp true <256 x float> %x, %vy
702+
ret <256 x i1> %z
703+
}

0 commit comments

Comments
 (0)