Skip to content

Commit ee0a425

Browse files
SamTebbs33SamTebbs33
committed
Remove FunctionComparator changes and add MergeFunc test
1 parent e73271c commit ee0a425

File tree

4 files changed

+214
-10
lines changed

4 files changed

+214
-10
lines changed

‎llvm/lib/IR/Instruction.cpp‎

Lines changed: 0 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -935,8 +935,6 @@ bool Instruction::hasSameSpecialState(const Instruction *I2,
935935
if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I1))
936936
return GEP->getSourceElementType() ==
937937
cast<GetElementPtrInst>(I2)->getSourceElementType();
938-
if (const CastInst *Cast = dyn_cast<CastInst>(I1))
939-
return Cast->getDestTy() == cast<CastInst>(I2)->getDestTy();
940938

941939
return true;
942940
}

‎llvm/lib/Transforms/Utils/FunctionComparator.cpp‎

Lines changed: 0 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -804,10 +804,6 @@ int FunctionComparator::cmpOperations(const Instruction *L,
804804
return Res;
805805
}
806806
}
807-
if (const CastInst *Cast = dyn_cast<CastInst>(L)) {
808-
const CastInst *CastR = cast<CastInst>(R);
809-
return cmpTypes(Cast->getDestTy(), CastR->getDestTy());
810-
}
811807
return 0;
812808
}
813809

‎llvm/test/Transforms/LoopVectorize/cse.ll‎

Lines changed: 155 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -137,8 +137,8 @@ exit: ; preds = %exit.loopexit, %entry
137137
ret i32 %res.0.lcssa
138138
}
139139

140-
define i32 @common_sext_different_types(ptr %a, ptr %b, ptr %c, i32 %N) {
141-
; CHECK-LABEL: define i32 @common_sext_different_types(
140+
define i32 @common_sext_different_src_types(ptr %a, ptr %b, ptr %c, i32 %N) {
141+
; CHECK-LABEL: define i32 @common_sext_different_src_types(
142142
; CHECK-SAME: ptr [[A:%.*]], ptr [[B:%.*]], ptr [[C:%.*]], i32 [[N:%.*]]) {
143143
; CHECK-NEXT: [[ENTRY:.*:]]
144144
; CHECK-NEXT: [[CMP28_NOT:%.*]] = icmp ult i32 [[N]], 2
@@ -222,8 +222,9 @@ exit: ; preds = %exit.loopexit, %entry
222222
%res.0.lcssa = phi i32 [ %add3, %for.body ]
223223
ret i32 %res.0.lcssa
224224
}
225-
define i32 @common_zext_different_types(ptr %a, ptr %b, ptr %c, i32 %N) {
226-
; CHECK-LABEL: define i32 @common_zext_different_types(
225+
226+
define i32 @common_zext_different_src_types(ptr %a, ptr %b, ptr %c, i32 %N) {
227+
; CHECK-LABEL: define i32 @common_zext_different_src_types(
227228
; CHECK-SAME: ptr [[A:%.*]], ptr [[B:%.*]], ptr [[C:%.*]], i32 [[N:%.*]]) {
228229
; CHECK-NEXT: [[ENTRY:.*:]]
229230
; CHECK-NEXT: [[CMP28_NOT:%.*]] = icmp ult i32 [[N]], 2
@@ -307,3 +308,153 @@ exit: ; preds = %exit.loopexit, %entry
307308
%res.0.lcssa = phi i32 [ %add3, %for.body ]
308309
ret i32 %res.0.lcssa
309310
}
311+
312+
define i32 @common_sext_different_dest_types(ptr %a, ptr %b, ptr %c, i32 %N) {
313+
; CHECK-LABEL: define i32 @common_sext_different_dest_types(
314+
; CHECK-SAME: ptr [[A:%.*]], ptr [[B:%.*]], ptr [[C:%.*]], i32 [[N:%.*]]) {
315+
; CHECK-NEXT: [[ENTRY:.*:]]
316+
; CHECK-NEXT: [[CMP28_NOT:%.*]] = icmp ult i32 [[N]], 2
317+
; CHECK-NEXT: [[DIV27:%.*]] = lshr i32 [[N]], 1
318+
; CHECK-NEXT: [[WIDE_TRIP_COUNT:%.*]] = zext nneg i32 [[DIV27]] to i64
319+
; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[WIDE_TRIP_COUNT]], 4
320+
; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
321+
; CHECK: [[VECTOR_PH]]:
322+
; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[WIDE_TRIP_COUNT]], 4
323+
; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[WIDE_TRIP_COUNT]], [[N_MOD_VF]]
324+
; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
325+
; CHECK: [[VECTOR_BODY]]:
326+
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
327+
; CHECK-NEXT: [[TMP0:%.*]] = getelementptr inbounds nuw i8, ptr [[A]], i64 [[INDEX]]
328+
; CHECK-NEXT: [[TMP1:%.*]] = getelementptr inbounds nuw i8, ptr [[B]], i64 [[INDEX]]
329+
; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds nuw i8, ptr [[C]], i64 [[INDEX]]
330+
; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds nuw i8, ptr [[TMP0]], i32 0
331+
; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i8>, ptr [[TMP3]], align 1
332+
; CHECK-NEXT: [[TMP4:%.*]] = getelementptr inbounds nuw i8, ptr [[TMP1]], i32 0
333+
; CHECK-NEXT: [[WIDE_LOAD1:%.*]] = load <4 x i8>, ptr [[TMP4]], align 1
334+
; CHECK-NEXT: [[TMP5:%.*]] = getelementptr inbounds nuw i8, ptr [[TMP2]], i32 0
335+
; CHECK-NEXT: [[WIDE_LOAD2:%.*]] = load <4 x i8>, ptr [[TMP5]], align 1
336+
; CHECK-NEXT: [[TMP6:%.*]] = sext <4 x i8> [[WIDE_LOAD]] to <4 x i32>
337+
; CHECK-NEXT: [[TMP7:%.*]] = sext <4 x i8> [[WIDE_LOAD1]] to <4 x i32>
338+
; CHECK-NEXT: [[TMP8:%.*]] = sext <4 x i8> [[WIDE_LOAD1]] to <4 x i16>
339+
; CHECK-NEXT: [[TMP9:%.*]] = sext <4 x i8> [[WIDE_LOAD2]] to <4 x i16>
340+
; CHECK-NEXT: [[TMP10:%.*]] = add nsw <4 x i32> [[TMP6]], [[TMP7]]
341+
; CHECK-NEXT: [[TMP11:%.*]] = add nsw <4 x i16> [[TMP9]], [[TMP8]]
342+
; CHECK-NEXT: [[TMP12:%.*]] = sext <4 x i16> [[TMP11]] to <4 x i32>
343+
; CHECK-NEXT: [[TMP13:%.*]] = add <4 x i32> [[TMP10]], [[TMP12]]
344+
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
345+
; CHECK-NEXT: [[TMP14:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
346+
; CHECK-NEXT: br i1 [[TMP14]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP10:![0-9]+]]
347+
; CHECK: [[MIDDLE_BLOCK]]:
348+
; CHECK-NEXT: [[TMP15:%.*]] = extractelement <4 x i32> [[TMP13]], i32 3
349+
; CHECK-NEXT: [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i32> [[TMP13]], i32 3
350+
; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[WIDE_TRIP_COUNT]], [[N_VEC]]
351+
; CHECK-NEXT: br i1 [[CMP_N]], [[EXIT:label %.*]], label %[[SCALAR_PH]]
352+
; CHECK: [[SCALAR_PH]]:
353+
;
354+
entry:
355+
%cmp28.not = icmp ult i32 %N, 2
356+
%div27 = lshr i32 %N, 1
357+
%wide.trip.count = zext nneg i32 %div27 to i64
358+
br label %for.body
359+
360+
for.body: ; preds = %for.body.preheader, %for.body
361+
%iv = phi i64 [ 0, %entry ], [ %iv.next, %for.body ]
362+
%res = phi i32 [ 0, %entry ], [ %add3, %for.body ]
363+
%a.ptr = getelementptr inbounds nuw i8, ptr %a, i64 %iv
364+
%b.ptr = getelementptr inbounds nuw i8, ptr %b, i64 %iv
365+
%b.ptr2 = getelementptr inbounds nuw i16, ptr %b, i64 %iv
366+
%c.ptr = getelementptr inbounds nuw i8, ptr %c, i64 %iv
367+
%a.val = load i8, ptr %a.ptr, align 1
368+
%b.val = load i8, ptr %b.ptr, align 1
369+
%c.val = load i8, ptr %c.ptr, align 1
370+
%a.ext = sext i8 %a.val to i32
371+
%b.ext = sext i8 %b.val to i32
372+
%b.ext2 = sext i8 %b.val to i16
373+
%c.ext = sext i8 %c.val to i16
374+
%add = add nsw i32 %a.ext, %b.ext
375+
%add2 = add nsw i16 %c.ext, %b.ext2
376+
%add2.ext = sext i16 %add2 to i32
377+
%add3 = add i32 %add, %add2.ext
378+
%iv.next = add nuw nsw i64 %iv, 1
379+
%exitcond = icmp eq i64 %iv.next, %wide.trip.count
380+
br i1 %exitcond, label %exit, label %for.body
381+
382+
exit: ; preds = %exit.loopexit, %entry
383+
%res.0.lcssa = phi i32 [ %add3, %for.body ]
384+
ret i32 %res.0.lcssa
385+
}
386+
387+
define i32 @common_zext_different_dest_types(ptr %a, ptr %b, ptr %c, i32 %N) {
388+
; CHECK-LABEL: define i32 @common_zext_different_dest_types(
389+
; CHECK-SAME: ptr [[A:%.*]], ptr [[B:%.*]], ptr [[C:%.*]], i32 [[N:%.*]]) {
390+
; CHECK-NEXT: [[ENTRY:.*:]]
391+
; CHECK-NEXT: [[CMP28_NOT:%.*]] = icmp ult i32 [[N]], 2
392+
; CHECK-NEXT: [[DIV27:%.*]] = lshr i32 [[N]], 1
393+
; CHECK-NEXT: [[WIDE_TRIP_COUNT:%.*]] = zext nneg i32 [[DIV27]] to i64
394+
; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[WIDE_TRIP_COUNT]], 4
395+
; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
396+
; CHECK: [[VECTOR_PH]]:
397+
; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[WIDE_TRIP_COUNT]], 4
398+
; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[WIDE_TRIP_COUNT]], [[N_MOD_VF]]
399+
; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
400+
; CHECK: [[VECTOR_BODY]]:
401+
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
402+
; CHECK-NEXT: [[TMP0:%.*]] = getelementptr inbounds nuw i8, ptr [[A]], i64 [[INDEX]]
403+
; CHECK-NEXT: [[TMP1:%.*]] = getelementptr inbounds nuw i8, ptr [[B]], i64 [[INDEX]]
404+
; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds nuw i8, ptr [[C]], i64 [[INDEX]]
405+
; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds nuw i8, ptr [[TMP0]], i32 0
406+
; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i8>, ptr [[TMP3]], align 1
407+
; CHECK-NEXT: [[TMP4:%.*]] = getelementptr inbounds nuw i8, ptr [[TMP1]], i32 0
408+
; CHECK-NEXT: [[WIDE_LOAD1:%.*]] = load <4 x i8>, ptr [[TMP4]], align 1
409+
; CHECK-NEXT: [[TMP5:%.*]] = getelementptr inbounds nuw i8, ptr [[TMP2]], i32 0
410+
; CHECK-NEXT: [[WIDE_LOAD2:%.*]] = load <4 x i8>, ptr [[TMP5]], align 1
411+
; CHECK-NEXT: [[TMP6:%.*]] = zext <4 x i8> [[WIDE_LOAD]] to <4 x i32>
412+
; CHECK-NEXT: [[TMP7:%.*]] = zext <4 x i8> [[WIDE_LOAD1]] to <4 x i32>
413+
; CHECK-NEXT: [[TMP8:%.*]] = zext <4 x i8> [[WIDE_LOAD1]] to <4 x i16>
414+
; CHECK-NEXT: [[TMP9:%.*]] = zext <4 x i8> [[WIDE_LOAD2]] to <4 x i16>
415+
; CHECK-NEXT: [[TMP10:%.*]] = add nsw <4 x i32> [[TMP6]], [[TMP7]]
416+
; CHECK-NEXT: [[TMP11:%.*]] = add nsw <4 x i16> [[TMP9]], [[TMP8]]
417+
; CHECK-NEXT: [[TMP12:%.*]] = zext <4 x i16> [[TMP11]] to <4 x i32>
418+
; CHECK-NEXT: [[TMP13:%.*]] = add <4 x i32> [[TMP10]], [[TMP12]]
419+
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
420+
; CHECK-NEXT: [[TMP14:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
421+
; CHECK-NEXT: br i1 [[TMP14]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP12:![0-9]+]]
422+
; CHECK: [[MIDDLE_BLOCK]]:
423+
; CHECK-NEXT: [[TMP15:%.*]] = extractelement <4 x i32> [[TMP13]], i32 3
424+
; CHECK-NEXT: [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i32> [[TMP13]], i32 3
425+
; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[WIDE_TRIP_COUNT]], [[N_VEC]]
426+
; CHECK-NEXT: br i1 [[CMP_N]], [[EXIT:label %.*]], label %[[SCALAR_PH]]
427+
; CHECK: [[SCALAR_PH]]:
428+
;
429+
entry:
430+
%cmp28.not = icmp ult i32 %N, 2
431+
%div27 = lshr i32 %N, 1
432+
%wide.trip.count = zext nneg i32 %div27 to i64
433+
br label %for.body
434+
435+
for.body: ; preds = %for.body.preheader, %for.body
436+
%iv = phi i64 [ 0, %entry ], [ %iv.next, %for.body ]
437+
%res = phi i32 [ 0, %entry ], [ %add3, %for.body ]
438+
%a.ptr = getelementptr inbounds nuw i8, ptr %a, i64 %iv
439+
%b.ptr = getelementptr inbounds nuw i8, ptr %b, i64 %iv
440+
%b.ptr2 = getelementptr inbounds nuw i16, ptr %b, i64 %iv
441+
%c.ptr = getelementptr inbounds nuw i8, ptr %c, i64 %iv
442+
%a.val = load i8, ptr %a.ptr, align 1
443+
%b.val = load i8, ptr %b.ptr, align 1
444+
%c.val = load i8, ptr %c.ptr, align 1
445+
%a.ext = zext i8 %a.val to i32
446+
%b.ext = zext i8 %b.val to i32
447+
%b.ext2 = zext i8 %b.val to i16
448+
%c.ext = zext i8 %c.val to i16
449+
%add = add nsw i32 %a.ext, %b.ext
450+
%add2 = add nsw i16 %c.ext, %b.ext2
451+
%add2.ext = zext i16 %add2 to i32
452+
%add3 = add i32 %add, %add2.ext
453+
%iv.next = add nuw nsw i64 %iv, 1
454+
%exitcond = icmp eq i64 %iv.next, %wide.trip.count
455+
br i1 %exitcond, label %exit, label %for.body
456+
457+
exit: ; preds = %exit.loopexit, %entry
458+
%res.0.lcssa = phi i32 [ %add3, %for.body ]
459+
ret i32 %res.0.lcssa
460+
}

‎llvm/test/Transforms/MergeFunc/merge-ext.ll‎

Lines changed: 59 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,59 @@
1+
; RUN: opt -S -passes=mergefunc < %s | FileCheck %s
2+
3+
@llvm.compiler.used = appending global [2 x ptr] [ptr @zext_a, ptr @sext_a], section "llvm.metadata"
4+
5+
define internal i32 @sext_a(i16 %a) unnamed_addr {
6+
%b = sext i16 %a to i32
7+
ret i32 %b
8+
}
9+
10+
define i32 @sext_b(i16 %a) unnamed_addr {
11+
%b = sext i16 %a to i32
12+
ret i32 %b
13+
}
14+
15+
define i32 @sext_c(i32 %a) unnamed_addr {
16+
%b = tail call i32 @sext_a(i32 %a)
17+
ret i32 %b
18+
}
19+
20+
define internal i32 @zext_a(i16 %a) unnamed_addr {
21+
%b = zext i16 %a to i32
22+
ret i32 %b
23+
}
24+
25+
define i32 @zext_b(i16 %a) unnamed_addr {
26+
%b = zext i16 %a to i32
27+
ret i32 %b
28+
}
29+
30+
define i32 @zext_c(i32 %a) unnamed_addr {
31+
%b = tail call i32 @zext_a(i32 %a)
32+
ret i32 %b
33+
}
34+
35+
; CHECK-LABEL: @llvm.compiler.used = appending global [2 x ptr] [ptr @zext_a, ptr @sext_a], section "llvm.metadata"
36+
37+
; CHECK-LABEL: define i32 @sext_b(i16 %a) unnamed_addr
38+
; CHECK-NEXT: sext
39+
; CHECK-NEXT: ret
40+
41+
; CHECK-LABEL: define i32 @sext_c(i32 %a) unnamed_addr
42+
; CHECK-NEXT: tail call i32 @sext_b(i32 %a)
43+
; CHECK-NEXT: ret
44+
45+
; CHECK-LABEL: define i32 @zext_b(i16 %a) unnamed_addr
46+
; CHECK-NEXT: zext
47+
; CHECK-NEXT: ret
48+
49+
; CHECK-LABEL: define i32 @zext_c(i32 %a) unnamed_addr
50+
; CHECK-NEXT: tail call i32 @zext_b(i32 %a)
51+
; CHECK-NEXT: ret
52+
53+
; CHECK-LABEL: define internal i32 @sext_a(i16 %0) unnamed_addr
54+
; CHECK-NEXT: tail call i32 @sext_b(i16 %0)
55+
; CHECK-NEXT: ret
56+
57+
; CHECK-LABEL: define internal i32 @zext_a(i16 %0) unnamed_addr
58+
; CHECK-NEXT: tail call i32 @zext_b(i16 %0)
59+
; CHECK-NEXT: ret

0 commit comments

Comments
 (0)