Skip to content

Commit 98509c7

Browse files
fhahnfhahn
committed
[AArch64] Add vec3 tests with different load/store alignments.
Add extra tests with different load/store alignments for llvm#78637.
1 parent c3e7707 commit 98509c7

File tree

1 file changed

+248
-8
lines changed

1 file changed

+248
-8
lines changed

‎llvm/test/CodeGen/AArch64/vec3-loads-ext-trunc-stores.ll

+248-8
Original file line numberDiff line numberDiff line change
@@ -87,6 +87,76 @@ define <4 x i32> @load_v3i8_to_4xi32(ptr %src) {
8787
ret <4 x i32> %e
8888
}
8989

90+
define <4 x i32> @load_v3i8_to_4xi32_align_2(ptr %src) {
91+
; CHECK-LABEL: load_v3i8_to_4xi32_align_2:
92+
; CHECK: ; %bb.0:
93+
; CHECK-NEXT: sub sp, sp, #16
94+
; CHECK-NEXT: .cfi_def_cfa_offset 16
95+
; CHECK-NEXT: ldrh w8, [x0]
96+
; CHECK-NEXT: movi.2d v1, #0x0000ff000000ff
97+
; CHECK-NEXT: strh w8, [sp, #12]
98+
; CHECK-NEXT: ldr s0, [sp, #12]
99+
; CHECK-NEXT: ldrsb w8, [x0, #2]
100+
; CHECK-NEXT: ushll.8h v0, v0, #0
101+
; CHECK-NEXT: mov.h v0[1], v0[1]
102+
; CHECK-NEXT: mov.h v0[2], w8
103+
; CHECK-NEXT: ushll.4s v0, v0, #0
104+
; CHECK-NEXT: and.16b v0, v0, v1
105+
; CHECK-NEXT: add sp, sp, #16
106+
; CHECK-NEXT: ret
107+
;
108+
; BE-LABEL: load_v3i8_to_4xi32_align_2:
109+
; BE: // %bb.0:
110+
; BE-NEXT: sub sp, sp, #16
111+
; BE-NEXT: .cfi_def_cfa_offset 16
112+
; BE-NEXT: ldrh w8, [x0]
113+
; BE-NEXT: movi v1.2d, #0x0000ff000000ff
114+
; BE-NEXT: strh w8, [sp, #12]
115+
; BE-NEXT: ldr s0, [sp, #12]
116+
; BE-NEXT: ldrsb w8, [x0, #2]
117+
; BE-NEXT: rev32 v0.8b, v0.8b
118+
; BE-NEXT: ushll v0.8h, v0.8b, #0
119+
; BE-NEXT: mov v0.h[1], v0.h[1]
120+
; BE-NEXT: mov v0.h[2], w8
121+
; BE-NEXT: ushll v0.4s, v0.4h, #0
122+
; BE-NEXT: and v0.16b, v0.16b, v1.16b
123+
; BE-NEXT: rev64 v0.4s, v0.4s
124+
; BE-NEXT: ext v0.16b, v0.16b, v0.16b, #8
125+
; BE-NEXT: add sp, sp, #16
126+
; BE-NEXT: ret
127+
%l = load <3 x i8>, ptr %src, align 2
128+
%s = shufflevector <3 x i8> poison, <3 x i8> %l, <4 x i32> <i32 3, i32 4, i32 5, i32 undef>
129+
%e = zext <4 x i8> %s to <4 x i32>
130+
ret <4 x i32> %e
131+
}
132+
133+
define <4 x i32> @load_v3i8_to_4xi32_align_4(ptr %src) {
134+
; CHECK-LABEL: load_v3i8_to_4xi32_align_4:
135+
; CHECK: ; %bb.0:
136+
; CHECK-NEXT: ldr s0, [x0]
137+
; CHECK-NEXT: movi.2d v1, #0x0000ff000000ff
138+
; CHECK-NEXT: zip1.8b v0, v0, v0
139+
; CHECK-NEXT: ushll.4s v0, v0, #0
140+
; CHECK-NEXT: and.16b v0, v0, v1
141+
; CHECK-NEXT: ret
142+
;
143+
; BE-LABEL: load_v3i8_to_4xi32_align_4:
144+
; BE: // %bb.0:
145+
; BE-NEXT: ldr s0, [x0]
146+
; BE-NEXT: movi v1.2d, #0x0000ff000000ff
147+
; BE-NEXT: rev32 v0.8b, v0.8b
148+
; BE-NEXT: zip1 v0.8b, v0.8b, v0.8b
149+
; BE-NEXT: ushll v0.4s, v0.4h, #0
150+
; BE-NEXT: and v0.16b, v0.16b, v1.16b
151+
; BE-NEXT: rev64 v0.4s, v0.4s
152+
; BE-NEXT: ext v0.16b, v0.16b, v0.16b, #8
153+
; BE-NEXT: ret
154+
%l = load <3 x i8>, ptr %src, align 4
155+
%s = shufflevector <3 x i8> poison, <3 x i8> %l, <4 x i32> <i32 3, i32 4, i32 5, i32 undef>
156+
%e = zext <4 x i8> %s to <4 x i32>
157+
ret <4 x i32> %e
158+
}
159+
90160
define <4 x i32> @load_v3i8_to_4xi32_const_offset_1(ptr %src) {
91161
; CHECK-LABEL: load_v3i8_to_4xi32_const_offset_1:
92162
; CHECK: ; %bb.0:
@@ -176,6 +246,42 @@ define <4 x i32> @load_v3i8_to_4xi32_const_offset_3(ptr %src) {
176246
}
177247

178248
define <4 x i32> @volatile_load_v3i8_to_4xi32(ptr %src) {
249+
; check-label: volatile_load_v3i8_to_4xi32:
250+
; check: ; %bb.0:
251+
; check-next: sub sp, sp, #16
252+
; check-next: .cfi_def_cfa_offset 16
253+
; check-next: ldrh w8, [x0]
254+
; check-next: movi.2d v1, #0x0000ff000000ff
255+
; check-next: strh w8, [sp, #12]
256+
; check-next: ldr s0, [sp, #12]
257+
; check-next: ldrsb w8, [x0, #2]
258+
; check-next: ushll.8h v0, v0, #0
259+
; check-next: mov.h v0[1], v0[1]
260+
; check-next: mov.h v0[2], w8
261+
; check-next: ushll.4s v0, v0, #0
262+
; check-next: and.16b v0, v0, v1
263+
; check-next: add sp, sp, #16
264+
; check-next: ret
265+
;
266+
; be-label: volatile_load_v3i8_to_4xi32:
267+
; be: // %bb.0:
268+
; be-next: sub sp, sp, #16
269+
; be-next: .cfi_def_cfa_offset 16
270+
; be-next: ldrh w8, [x0]
271+
; be-next: movi v1.2d, #0x0000ff000000ff
272+
; be-next: strh w8, [sp, #12]
273+
; be-next: ldr s0, [sp, #12]
274+
; be-next: ldrsb w8, [x0, #2]
275+
; be-next: rev32 v0.8b, v0.8b
276+
; be-next: ushll v0.8h, v0.8b, #0
277+
; be-next: mov v0.h[1], v0.h[1]
278+
; be-next: mov v0.h[2], w8
279+
; be-next: ushll v0.4s, v0.4h, #0
280+
; be-next: and v0.16b, v0.16b, v1.16b
281+
; be-next: rev64 v0.4s, v0.4s
282+
; be-next: ext v0.16b, v0.16b, v0.16b, #8
283+
; be-next: add sp, sp, #16
284+
; be-next: ret
179285
; CHECK-LABEL: volatile_load_v3i8_to_4xi32:
180286
; CHECK: ; %bb.0:
181287
; CHECK-NEXT: sub sp, sp, #16
@@ -286,9 +392,9 @@ define void @store_trunc_add_from_64bits(ptr %src, ptr %dst) {
286392
; CHECK-NEXT: ldr s0, [x0]
287393
; CHECK-NEXT: add x9, x0, #4
288394
; CHECK-NEXT: Lloh0:
289-
; CHECK-NEXT: adrp x8, lCPI7_0@PAGE
395+
; CHECK-NEXT: adrp x8, lCPI9_0@PAGE
290396
; CHECK-NEXT: Lloh1:
291-
; CHECK-NEXT: ldr d1, [x8, lCPI7_0@PAGEOFF]
397+
; CHECK-NEXT: ldr d1, [x8, lCPI9_0@PAGEOFF]
292398
; CHECK-NEXT: ld1.h { v0 }[2], [x9]
293399
; CHECK-NEXT: add.4h v0, v0, v1
294400
; CHECK-NEXT: xtn.8b v1, v0
@@ -309,8 +415,8 @@ define void @store_trunc_add_from_64bits(ptr %src, ptr %dst) {
309415
; BE-NEXT: add x8, x0, #4
310416
; BE-NEXT: rev32 v0.4h, v0.4h
311417
; BE-NEXT: ld1 { v0.h }[2], [x8]
312-
; BE-NEXT: adrp x8, .LCPI7_0
313-
; BE-NEXT: add x8, x8, :lo12:.LCPI7_0
418+
; BE-NEXT: adrp x8, .LCPI9_0
419+
; BE-NEXT: add x8, x8, :lo12:.LCPI9_0
314420
; BE-NEXT: ld1 { v1.4h }, [x8]
315421
; BE-NEXT: add v0.4h, v0.4h, v1.4h
316422
; BE-NEXT: xtn v1.8b, v0.8h
@@ -373,16 +479,74 @@ entry:
373479
ret void
374480
}
375481

482+
define void @load_ext_to_64bits_default_align(ptr %src, ptr %dst) {
483+
; CHECK-LABEL: load_ext_to_64bits_default_align:
484+
; CHECK: ; %bb.0: ; %entry
485+
; CHECK-NEXT: ldr s0, [x0]
486+
; CHECK-NEXT: add x8, x1, #4
487+
; CHECK-NEXT: zip1.8b v0, v0, v0
488+
; CHECK-NEXT: bic.4h v0, #255, lsl #8
489+
; CHECK-NEXT: st1.h { v0 }[2], [x8]
490+
; CHECK-NEXT: str s0, [x1]
491+
; CHECK-NEXT: ret
492+
;
493+
; BE-LABEL: load_ext_to_64bits_default_align:
494+
; BE: // %bb.0: // %entry
495+
; BE-NEXT: ldr s0, [x0]
496+
; BE-NEXT: add x8, x1, #4
497+
; BE-NEXT: rev32 v0.8b, v0.8b
498+
; BE-NEXT: zip1 v0.8b, v0.8b, v0.8b
499+
; BE-NEXT: bic v0.4h, #255, lsl #8
500+
; BE-NEXT: rev32 v1.8h, v0.8h
501+
; BE-NEXT: st1 { v0.h }[2], [x8]
502+
; BE-NEXT: str s1, [x1]
503+
; BE-NEXT: ret
504+
entry:
505+
%l = load <3 x i8>, ptr %src
506+
%e = zext <3 x i8> %l to <3 x i16>
507+
store <3 x i16> %e, ptr %dst, align 1
508+
ret void
509+
}
510+
511+
define void @load_ext_to_64bits_align_4(ptr %src, ptr %dst) {
512+
; CHECK-LABEL: load_ext_to_64bits_align_4:
513+
; CHECK: ; %bb.0: ; %entry
514+
; CHECK-NEXT: ldr s0, [x0]
515+
; CHECK-NEXT: add x8, x1, #4
516+
; CHECK-NEXT: zip1.8b v0, v0, v0
517+
; CHECK-NEXT: bic.4h v0, #255, lsl #8
518+
; CHECK-NEXT: st1.h { v0 }[2], [x8]
519+
; CHECK-NEXT: str s0, [x1]
520+
; CHECK-NEXT: ret
521+
;
522+
; BE-LABEL: load_ext_to_64bits_align_4:
523+
; BE: // %bb.0: // %entry
524+
; BE-NEXT: ldr s0, [x0]
525+
; BE-NEXT: add x8, x1, #4
526+
; BE-NEXT: rev32 v0.8b, v0.8b
527+
; BE-NEXT: zip1 v0.8b, v0.8b, v0.8b
528+
; BE-NEXT: bic v0.4h, #255, lsl #8
529+
; BE-NEXT: rev32 v1.8h, v0.8h
530+
; BE-NEXT: st1 { v0.h }[2], [x8]
531+
; BE-NEXT: str s1, [x1]
532+
; BE-NEXT: ret
533+
entry:
534+
%l = load <3 x i8>, ptr %src, align 4
535+
%e = zext <3 x i8> %l to <3 x i16>
536+
store <3 x i16> %e, ptr %dst, align 1
537+
ret void
538+
}
539+
376540
define void @load_ext_add_to_64bits(ptr %src, ptr %dst) {
377541
; CHECK-LABEL: load_ext_add_to_64bits:
378542
; CHECK: ; %bb.0: ; %entry
379543
; CHECK-NEXT: sub sp, sp, #16
380544
; CHECK-NEXT: .cfi_def_cfa_offset 16
381545
; CHECK-NEXT: ldrh w9, [x0]
382546
; CHECK-NEXT: Lloh2:
383-
; CHECK-NEXT: adrp x8, lCPI9_0@PAGE
547+
; CHECK-NEXT: adrp x8, lCPI13_0@PAGE
384548
; CHECK-NEXT: Lloh3:
385-
; CHECK-NEXT: ldr d1, [x8, lCPI9_0@PAGEOFF]
549+
; CHECK-NEXT: ldr d1, [x8, lCPI13_0@PAGEOFF]
386550
; CHECK-NEXT: add x8, x1, #4
387551
; CHECK-NEXT: strh w9, [sp, #12]
388552
; CHECK-NEXT: add x9, x0, #2
@@ -408,8 +572,8 @@ define void @load_ext_add_to_64bits(ptr %src, ptr %dst) {
408572
; BE-NEXT: rev32 v0.8b, v0.8b
409573
; BE-NEXT: ushll v0.8h, v0.8b, #0
410574
; BE-NEXT: ld1 { v0.b }[4], [x8]
411-
; BE-NEXT: adrp x8, .LCPI9_0
412-
; BE-NEXT: add x8, x8, :lo12:.LCPI9_0
575+
; BE-NEXT: adrp x8, .LCPI13_0
576+
; BE-NEXT: add x8, x8, :lo12:.LCPI13_0
413577
; BE-NEXT: ld1 { v1.4h }, [x8]
414578
; BE-NEXT: add x8, x1, #4
415579
; BE-NEXT: bic v0.4h, #255, lsl #8
@@ -465,6 +629,82 @@ define void @shift_trunc_store(ptr %src, ptr %dst) {
465629
ret void
466630
}
467631

632+
define void @shift_trunc_store_default_align(ptr %src, ptr %dst) {
633+
; CHECK-LABEL: shift_trunc_store_default_align:
634+
; CHECK: ; %bb.0:
635+
; CHECK-NEXT: sub sp, sp, #16
636+
; CHECK-NEXT: .cfi_def_cfa_offset 16
637+
; CHECK-NEXT: ldr q0, [x0]
638+
; CHECK-NEXT: shrn.4h v0, v0, #16
639+
; CHECK-NEXT: xtn.8b v1, v0
640+
; CHECK-NEXT: umov.h w8, v0[2]
641+
; CHECK-NEXT: str s1, [sp, #12]
642+
; CHECK-NEXT: ldrh w9, [sp, #12]
643+
; CHECK-NEXT: strb w8, [x1, #2]
644+
; CHECK-NEXT: strh w9, [x1]
645+
; CHECK-NEXT: add sp, sp, #16
646+
; CHECK-NEXT: ret
647+
;
648+
; BE-LABEL: shift_trunc_store_default_align:
649+
; BE: // %bb.0:
650+
; BE-NEXT: sub sp, sp, #16
651+
; BE-NEXT: .cfi_def_cfa_offset 16
652+
; BE-NEXT: ld1 { v0.4s }, [x0]
653+
; BE-NEXT: shrn v0.4h, v0.4s, #16
654+
; BE-NEXT: xtn v1.8b, v0.8h
655+
; BE-NEXT: umov w8, v0.h[2]
656+
; BE-NEXT: rev32 v1.16b, v1.16b
657+
; BE-NEXT: str s1, [sp, #12]
658+
; BE-NEXT: ldrh w9, [sp, #12]
659+
; BE-NEXT: strb w8, [x1, #2]
660+
; BE-NEXT: strh w9, [x1]
661+
; BE-NEXT: add sp, sp, #16
662+
; BE-NEXT: ret
663+
%l = load <3 x i32>, ptr %src
664+
%s = lshr <3 x i32> %l, <i32 16, i32 16, i32 16>
665+
%t = trunc <3 x i32> %s to <3 x i8>
666+
store <3 x i8> %t, ptr %dst
667+
ret void
668+
}
669+
670+
define void @shift_trunc_store_align_4(ptr %src, ptr %dst) {
671+
; CHECK-LABEL: shift_trunc_store_align_4:
672+
; CHECK: ; %bb.0:
673+
; CHECK-NEXT: sub sp, sp, #16
674+
; CHECK-NEXT: .cfi_def_cfa_offset 16
675+
; CHECK-NEXT: ldr q0, [x0]
676+
; CHECK-NEXT: shrn.4h v0, v0, #16
677+
; CHECK-NEXT: xtn.8b v1, v0
678+
; CHECK-NEXT: umov.h w8, v0[2]
679+
; CHECK-NEXT: str s1, [sp, #12]
680+
; CHECK-NEXT: ldrh w9, [sp, #12]
681+
; CHECK-NEXT: strb w8, [x1, #2]
682+
; CHECK-NEXT: strh w9, [x1]
683+
; CHECK-NEXT: add sp, sp, #16
684+
; CHECK-NEXT: ret
685+
;
686+
; BE-LABEL: shift_trunc_store_align_4:
687+
; BE: // %bb.0:
688+
; BE-NEXT: sub sp, sp, #16
689+
; BE-NEXT: .cfi_def_cfa_offset 16
690+
; BE-NEXT: ld1 { v0.4s }, [x0]
691+
; BE-NEXT: shrn v0.4h, v0.4s, #16
692+
; BE-NEXT: xtn v1.8b, v0.8h
693+
; BE-NEXT: umov w8, v0.h[2]
694+
; BE-NEXT: rev32 v1.16b, v1.16b
695+
; BE-NEXT: str s1, [sp, #12]
696+
; BE-NEXT: ldrh w9, [sp, #12]
697+
; BE-NEXT: strb w8, [x1, #2]
698+
; BE-NEXT: strh w9, [x1]
699+
; BE-NEXT: add sp, sp, #16
700+
; BE-NEXT: ret
701+
%l = load <3 x i32>, ptr %src
702+
%s = lshr <3 x i32> %l, <i32 16, i32 16, i32 16>
703+
%t = trunc <3 x i32> %s to <3 x i8>
704+
store <3 x i8> %t, ptr %dst, align 4
705+
ret void
706+
}
707+
468708
define void @shift_trunc_store_const_offset_1(ptr %src, ptr %dst) {
469709
; CHECK-LABEL: shift_trunc_store_const_offset_1:
470710
; CHECK: ; %bb.0:

0 commit comments

Comments
 (0)