Skip to content

Commit 0e4ecfa

Browse files
committed
[X86] Add test for PR38788 missing PTEST fold
1 parent c486b82 commit 0e4ecfa

File tree

1 file changed

+48
-0
lines changed

1 file changed

+48
-0
lines changed

llvm/test/CodeGen/X86/combine-ptest.ll

Lines changed: 48 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -416,6 +416,54 @@ define i32 @ptestz_v2i64_concat(<4 x i64> %c, <4 x i64> %d, i32 %a, i32 %b) {
416416
ret i32 %t9
417417
}
418418

419+
; FIXME: Foldable to ptest(xor(%0,%1),xor(%0,%1))
420+
define i1 @PR38788_0(<4 x i32> %0, <4 x i32> %1) {
421+
; CHECK-LABEL: PR38788_0:
422+
; CHECK: # %bb.0:
423+
; CHECK-NEXT: vpcmpeqd %xmm1, %xmm0, %xmm0
424+
; CHECK-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
425+
; CHECK-NEXT: vptest %xmm1, %xmm0
426+
; CHECK-NEXT: setb %al
427+
; CHECK-NEXT: retq
428+
%3 = icmp eq <4 x i32> %0, %1
429+
%4 = sext <4 x i1> %3 to <4 x i32>
430+
%5 = bitcast <4 x i32> %4 to <2 x i64>
431+
%6 = tail call i32 @llvm.x86.sse41.ptestc(<2 x i64> %5, <2 x i64> <i64 -1, i64 -1>)
432+
%7 = icmp eq i32 %6, 1
433+
ret i1 %7
434+
}
435+
436+
define i1 @PR38788_1(<16 x i16> %0, <16 x i16> %1) {
437+
; AVX1-LABEL: PR38788_1:
438+
; AVX1: # %bb.0:
439+
; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2
440+
; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm3
441+
; AVX1-NEXT: vpcmpeqw %xmm2, %xmm3, %xmm2
442+
; AVX1-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
443+
; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0
444+
; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
445+
; AVX1-NEXT: vcmptrueps %ymm1, %ymm1, %ymm1
446+
; AVX1-NEXT: vptest %ymm1, %ymm0
447+
; AVX1-NEXT: setae %al
448+
; AVX1-NEXT: vzeroupper
449+
; AVX1-NEXT: retq
450+
;
451+
; AVX2-LABEL: PR38788_1:
452+
; AVX2: # %bb.0:
453+
; AVX2-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0
454+
; AVX2-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1
455+
; AVX2-NEXT: vptest %ymm1, %ymm0
456+
; AVX2-NEXT: setae %al
457+
; AVX2-NEXT: vzeroupper
458+
; AVX2-NEXT: retq
459+
%3 = icmp eq <16 x i16> %0, %1
460+
%4 = sext <16 x i1> %3 to <16 x i16>
461+
%5 = bitcast <16 x i16> %4 to <4 x i64>
462+
%6 = tail call i32 @llvm.x86.avx.ptestc.256(<4 x i64> %5, <4 x i64> <i64 -1, i64 -1, i64 -1, i64 -1>)
463+
%7 = icmp eq i32 %6, 0
464+
ret i1 %7
465+
}
466+
419467
declare i32 @llvm.x86.sse41.ptestz(<2 x i64>, <2 x i64>) nounwind readnone
420468
declare i32 @llvm.x86.sse41.ptestc(<2 x i64>, <2 x i64>) nounwind readnone
421469
declare i32 @llvm.x86.sse41.ptestnzc(<2 x i64>, <2 x i64>) nounwind readnone

0 commit comments

Comments
 (0)