@@ -326,26 +326,52 @@ cleanup:
326 326
ret i1 %retval.0
327 327
}
328 328 329 -
; From https://github.com/llvm/llvm-project/issues/139050.
330 -
; FIXME: This should be vectorized.
331 329
define i8 @masked_min_reduction(ptr %data, ptr %mask) {
332 330
; CHECK-LABEL: @masked_min_reduction(
333 331
; CHECK-NEXT: entry:
334 332
; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]
335 -
; CHECK: loop:
333 +
; CHECK: vector.body:
336 334
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
337 -
; CHECK-NEXT: [[ACC:%.*]] = phi i8 [ -1, [[ENTRY]] ], [ [[TMP21:%.*]], [[VECTOR_BODY]] ]
335 +
; CHECK-NEXT: [[VEC_PHI:%.*]] = phi <32 x i8> [ splat (i8 -1), [[ENTRY]] ], [ [[TMP16:%.*]], [[VECTOR_BODY]] ]
336 +
; CHECK-NEXT: [[VEC_PHI1:%.*]] = phi <32 x i8> [ splat (i8 -1), [[ENTRY]] ], [ [[TMP17:%.*]], [[VECTOR_BODY]] ]
337 +
; CHECK-NEXT: [[VEC_PHI2:%.*]] = phi <32 x i8> [ splat (i8 -1), [[ENTRY]] ], [ [[TMP18:%.*]], [[VECTOR_BODY]] ]
338 +
; CHECK-NEXT: [[VEC_PHI3:%.*]] = phi <32 x i8> [ splat (i8 -1), [[ENTRY]] ], [ [[TMP19:%.*]], [[VECTOR_BODY]] ]
338 339
; CHECK-NEXT: [[DATA:%.*]] = getelementptr i8, ptr [[DATA1:%.*]], i64 [[INDEX]]
339 -
; CHECK-NEXT: [[VAL:%.*]] = load i8, ptr [[DATA]], align 1
340 +
; CHECK-NEXT: [[TMP1:%.*]] = getelementptr i8, ptr [[DATA]], i64 32
341 +
; CHECK-NEXT: [[TMP2:%.*]] = getelementptr i8, ptr [[DATA]], i64 64
342 +
; CHECK-NEXT: [[TMP3:%.*]] = getelementptr i8, ptr [[DATA]], i64 96
343 +
; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <32 x i8>, ptr [[DATA]], align 1
344 +
; CHECK-NEXT: [[WIDE_LOAD4:%.*]] = load <32 x i8>, ptr [[TMP1]], align 1
345 +
; CHECK-NEXT: [[WIDE_LOAD5:%.*]] = load <32 x i8>, ptr [[TMP2]], align 1
346 +
; CHECK-NEXT: [[WIDE_LOAD6:%.*]] = load <32 x i8>, ptr [[TMP3]], align 1
340 347
; CHECK-NEXT: [[TMP7:%.*]] = getelementptr i8, ptr [[MASK:%.*]], i64 [[INDEX]]
341 -
; CHECK-NEXT: [[M:%.*]] = load i8, ptr [[TMP7]], align 1
342 -
; CHECK-NEXT: [[COND:%.*]] = icmp eq i8 [[M]], 0
343 -
; CHECK-NEXT: [[TMP0:%.*]] = tail call i8 @llvm.umin.i8(i8 [[ACC]], i8 [[VAL]])
344 -
; CHECK-NEXT: [[TMP21]] = select i1 [[COND]], i8 [[TMP0]], i8 [[ACC]]
345 -
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw nsw i64 [[INDEX]], 1
348 +
; CHECK-NEXT: [[TMP5:%.*]] = getelementptr i8, ptr [[TMP7]], i64 32
349 +
; CHECK-NEXT: [[TMP6:%.*]] = getelementptr i8, ptr [[TMP7]], i64 64
350 +
; CHECK-NEXT: [[TMP22:%.*]] = getelementptr i8, ptr [[TMP7]], i64 96
351 +
; CHECK-NEXT: [[WIDE_LOAD7:%.*]] = load <32 x i8>, ptr [[TMP7]], align 1
352 +
; CHECK-NEXT: [[WIDE_LOAD8:%.*]] = load <32 x i8>, ptr [[TMP5]], align 1
353 +
; CHECK-NEXT: [[WIDE_LOAD9:%.*]] = load <32 x i8>, ptr [[TMP6]], align 1
354 +
; CHECK-NEXT: [[WIDE_LOAD10:%.*]] = load <32 x i8>, ptr [[TMP22]], align 1
355 +
; CHECK-NEXT: [[TMP8:%.*]] = icmp eq <32 x i8> [[WIDE_LOAD7]], zeroinitializer
356 +
; CHECK-NEXT: [[TMP9:%.*]] = icmp eq <32 x i8> [[WIDE_LOAD8]], zeroinitializer
357 +
; CHECK-NEXT: [[TMP10:%.*]] = icmp eq <32 x i8> [[WIDE_LOAD9]], zeroinitializer
358 +
; CHECK-NEXT: [[TMP11:%.*]] = icmp eq <32 x i8> [[WIDE_LOAD10]], zeroinitializer
359 +
; CHECK-NEXT: [[TMP12:%.*]] = select <32 x i1> [[TMP8]], <32 x i8> [[WIDE_LOAD]], <32 x i8> splat (i8 -1)
360 +
; CHECK-NEXT: [[TMP13:%.*]] = select <32 x i1> [[TMP9]], <32 x i8> [[WIDE_LOAD4]], <32 x i8> splat (i8 -1)
361 +
; CHECK-NEXT: [[TMP14:%.*]] = select <32 x i1> [[TMP10]], <32 x i8> [[WIDE_LOAD5]], <32 x i8> splat (i8 -1)
362 +
; CHECK-NEXT: [[TMP15:%.*]] = select <32 x i1> [[TMP11]], <32 x i8> [[WIDE_LOAD6]], <32 x i8> splat (i8 -1)
363 +
; CHECK-NEXT: [[TMP16]] = tail call <32 x i8> @llvm.umin.v32i8(<32 x i8> [[VEC_PHI]], <32 x i8> [[TMP12]])
364 +
; CHECK-NEXT: [[TMP17]] = tail call <32 x i8> @llvm.umin.v32i8(<32 x i8> [[VEC_PHI1]], <32 x i8> [[TMP13]])
365 +
; CHECK-NEXT: [[TMP18]] = tail call <32 x i8> @llvm.umin.v32i8(<32 x i8> [[VEC_PHI2]], <32 x i8> [[TMP14]])
366 +
; CHECK-NEXT: [[TMP19]] = tail call <32 x i8> @llvm.umin.v32i8(<32 x i8> [[VEC_PHI3]], <32 x i8> [[TMP15]])
367 +
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 128
346 368
; CHECK-NEXT: [[TMP20:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1024
347 -
; CHECK-NEXT: br i1 [[TMP20]], label [[EXIT:%.*]], label [[VECTOR_BODY]]
348 -
; CHECK: exit:
369 +
; CHECK-NEXT: br i1 [[TMP20]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
370 +
; CHECK: middle.block:
371 +
; CHECK-NEXT: [[RDX_MINMAX:%.*]] = tail call <32 x i8> @llvm.umin.v32i8(<32 x i8> [[TMP16]], <32 x i8> [[TMP17]])
372 +
; CHECK-NEXT: [[RDX_MINMAX11:%.*]] = tail call <32 x i8> @llvm.umin.v32i8(<32 x i8> [[RDX_MINMAX]], <32 x i8> [[TMP18]])
373 +
; CHECK-NEXT: [[RDX_MINMAX12:%.*]] = tail call <32 x i8> @llvm.umin.v32i8(<32 x i8> [[RDX_MINMAX11]], <32 x i8> [[TMP19]])
374 +
; CHECK-NEXT: [[TMP21:%.*]] = tail call i8 @llvm.vector.reduce.umin.v32i8(<32 x i8> [[RDX_MINMAX12]])
349 375
; CHECK-NEXT: ret i8 [[TMP21]]
350 376
;
351 377
entry:
RetroSearch is an open source project built by @garambo | Open a GitHub Issue
Search and Browse the WWW like it's 1997 | Search results from DuckDuckGo
HTML:
3.2
| Encoding:
UTF-8
| Version:
0.7.4