[llvm] [LV] fix logical error in trunc cost (PR #91136)

David Sherwood via llvm-commits llvm-commits at lists.llvm.org
Tue May 14 03:43:05 PDT 2024


================
@@ -0,0 +1,148 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 4
+; RUN: opt -passes=loop-vectorize,dce,simplifycfg -mtriple=s390x -mcpu=z14 -S %s | FileCheck %s --check-prefix=NOVEC
+; RUN: opt -passes=loop-vectorize,dce,simplifycfg -force-vector-width=2 -mtriple=s390x -mcpu=z14 -S %s | FileCheck %s
+
+define void @test(ptr %p) {
+; NOVEC-LABEL: define void @test(
+; NOVEC-SAME: ptr [[P:%.*]]) #[[ATTR0:[0-9]+]] {
+; NOVEC-NEXT:  entry:
+; NOVEC-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <16 x i32> poison, i32 0, i64 0
+; NOVEC-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <16 x i32> [[BROADCAST_SPLATINSERT]], <16 x i32> poison, <16 x i32> zeroinitializer
+; NOVEC-NEXT:    [[VEC_IV:%.*]] = add <16 x i32> [[BROADCAST_SPLAT]], <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+; NOVEC-NEXT:    [[TMP0:%.*]] = icmp ule <16 x i32> [[VEC_IV]], <i32 9, i32 9, i32 9, i32 9, i32 9, i32 9, i32 9, i32 9, i32 9, i32 9, i32 9, i32 9, i32 9, i32 9, i32 9, i32 9>
+; NOVEC-NEXT:    [[TMP1:%.*]] = extractelement <16 x i1> [[TMP0]], i32 0
+; NOVEC-NEXT:    [[TMP2:%.*]] = extractelement <16 x i1> [[TMP0]], i32 1
+; NOVEC-NEXT:    [[TMP3:%.*]] = xor i1 [[TMP1]], true
+; NOVEC-NEXT:    [[TMP4:%.*]] = xor i1 [[TMP2]], true
+; NOVEC-NEXT:    [[TMP5:%.*]] = xor i1 [[TMP3]], true
+; NOVEC-NEXT:    [[TMP6:%.*]] = xor i1 [[TMP4]], true
+; NOVEC-NEXT:    [[TMP7:%.*]] = or i1 [[TMP5]], [[TMP6]]
+; NOVEC-NEXT:    [[TMP8:%.*]] = extractelement <16 x i1> [[TMP0]], i32 2
+; NOVEC-NEXT:    [[TMP9:%.*]] = xor i1 [[TMP7]], true
+; NOVEC-NEXT:    [[TMP10:%.*]] = xor i1 [[TMP8]], true
+; NOVEC-NEXT:    [[TMP11:%.*]] = xor i1 [[TMP9]], true
+; NOVEC-NEXT:    [[TMP12:%.*]] = xor i1 [[TMP10]], true
+; NOVEC-NEXT:    [[TMP13:%.*]] = or i1 [[TMP11]], [[TMP12]]
+; NOVEC-NEXT:    [[TMP14:%.*]] = extractelement <16 x i1> [[TMP0]], i32 3
+; NOVEC-NEXT:    [[TMP15:%.*]] = xor i1 [[TMP13]], true
+; NOVEC-NEXT:    [[TMP16:%.*]] = xor i1 [[TMP14]], true
+; NOVEC-NEXT:    [[TMP17:%.*]] = xor i1 [[TMP15]], true
+; NOVEC-NEXT:    [[TMP18:%.*]] = xor i1 [[TMP16]], true
+; NOVEC-NEXT:    [[TMP19:%.*]] = or i1 [[TMP17]], [[TMP18]]
+; NOVEC-NEXT:    [[TMP20:%.*]] = extractelement <16 x i1> [[TMP0]], i32 4
+; NOVEC-NEXT:    [[TMP21:%.*]] = xor i1 [[TMP19]], true
+; NOVEC-NEXT:    [[TMP22:%.*]] = xor i1 [[TMP20]], true
+; NOVEC-NEXT:    [[TMP23:%.*]] = xor i1 [[TMP21]], true
+; NOVEC-NEXT:    [[TMP24:%.*]] = xor i1 [[TMP22]], true
+; NOVEC-NEXT:    [[TMP25:%.*]] = or i1 [[TMP23]], [[TMP24]]
+; NOVEC-NEXT:    [[TMP26:%.*]] = extractelement <16 x i1> [[TMP0]], i32 5
+; NOVEC-NEXT:    [[TMP27:%.*]] = xor i1 [[TMP25]], true
+; NOVEC-NEXT:    [[TMP28:%.*]] = xor i1 [[TMP26]], true
+; NOVEC-NEXT:    [[TMP29:%.*]] = xor i1 [[TMP27]], true
+; NOVEC-NEXT:    [[TMP30:%.*]] = xor i1 [[TMP28]], true
+; NOVEC-NEXT:    [[TMP31:%.*]] = or i1 [[TMP29]], [[TMP30]]
+; NOVEC-NEXT:    [[TMP32:%.*]] = extractelement <16 x i1> [[TMP0]], i32 6
+; NOVEC-NEXT:    [[TMP33:%.*]] = xor i1 [[TMP31]], true
+; NOVEC-NEXT:    [[TMP34:%.*]] = xor i1 [[TMP32]], true
+; NOVEC-NEXT:    [[TMP35:%.*]] = xor i1 [[TMP33]], true
+; NOVEC-NEXT:    [[TMP36:%.*]] = xor i1 [[TMP34]], true
+; NOVEC-NEXT:    [[TMP37:%.*]] = or i1 [[TMP35]], [[TMP36]]
+; NOVEC-NEXT:    [[TMP38:%.*]] = extractelement <16 x i1> [[TMP0]], i32 7
+; NOVEC-NEXT:    [[TMP39:%.*]] = xor i1 [[TMP37]], true
+; NOVEC-NEXT:    [[TMP40:%.*]] = xor i1 [[TMP38]], true
+; NOVEC-NEXT:    [[TMP41:%.*]] = xor i1 [[TMP39]], true
+; NOVEC-NEXT:    [[TMP42:%.*]] = xor i1 [[TMP40]], true
+; NOVEC-NEXT:    [[TMP43:%.*]] = or i1 [[TMP41]], [[TMP42]]
+; NOVEC-NEXT:    [[TMP44:%.*]] = extractelement <16 x i1> [[TMP0]], i32 8
+; NOVEC-NEXT:    [[TMP45:%.*]] = xor i1 [[TMP43]], true
+; NOVEC-NEXT:    [[TMP46:%.*]] = xor i1 [[TMP44]], true
+; NOVEC-NEXT:    [[TMP47:%.*]] = xor i1 [[TMP45]], true
+; NOVEC-NEXT:    [[TMP48:%.*]] = xor i1 [[TMP46]], true
+; NOVEC-NEXT:    [[TMP49:%.*]] = or i1 [[TMP47]], [[TMP48]]
+; NOVEC-NEXT:    [[TMP50:%.*]] = extractelement <16 x i1> [[TMP0]], i32 9
+; NOVEC-NEXT:    [[TMP51:%.*]] = xor i1 [[TMP49]], true
+; NOVEC-NEXT:    [[TMP52:%.*]] = xor i1 [[TMP50]], true
+; NOVEC-NEXT:    [[TMP53:%.*]] = xor i1 [[TMP51]], true
+; NOVEC-NEXT:    [[TMP54:%.*]] = xor i1 [[TMP52]], true
+; NOVEC-NEXT:    [[TMP55:%.*]] = or i1 [[TMP53]], [[TMP54]]
+; NOVEC-NEXT:    [[TMP56:%.*]] = extractelement <16 x i1> [[TMP0]], i32 10
+; NOVEC-NEXT:    [[TMP57:%.*]] = xor i1 [[TMP55]], true
+; NOVEC-NEXT:    [[TMP58:%.*]] = xor i1 [[TMP56]], true
+; NOVEC-NEXT:    [[TMP59:%.*]] = xor i1 [[TMP57]], true
+; NOVEC-NEXT:    [[TMP60:%.*]] = xor i1 [[TMP58]], true
+; NOVEC-NEXT:    [[TMP61:%.*]] = or i1 [[TMP59]], [[TMP60]]
+; NOVEC-NEXT:    [[TMP62:%.*]] = extractelement <16 x i1> [[TMP0]], i32 11
+; NOVEC-NEXT:    [[TMP63:%.*]] = xor i1 [[TMP61]], true
+; NOVEC-NEXT:    [[TMP64:%.*]] = xor i1 [[TMP62]], true
+; NOVEC-NEXT:    [[TMP65:%.*]] = xor i1 [[TMP63]], true
+; NOVEC-NEXT:    [[TMP66:%.*]] = xor i1 [[TMP64]], true
+; NOVEC-NEXT:    [[TMP67:%.*]] = or i1 [[TMP65]], [[TMP66]]
+; NOVEC-NEXT:    [[TMP68:%.*]] = extractelement <16 x i1> [[TMP0]], i32 12
+; NOVEC-NEXT:    [[TMP69:%.*]] = xor i1 [[TMP67]], true
+; NOVEC-NEXT:    [[TMP70:%.*]] = xor i1 [[TMP68]], true
+; NOVEC-NEXT:    [[TMP71:%.*]] = xor i1 [[TMP69]], true
+; NOVEC-NEXT:    [[TMP72:%.*]] = xor i1 [[TMP70]], true
+; NOVEC-NEXT:    [[TMP73:%.*]] = or i1 [[TMP71]], [[TMP72]]
+; NOVEC-NEXT:    [[TMP74:%.*]] = extractelement <16 x i1> [[TMP0]], i32 13
+; NOVEC-NEXT:    [[TMP75:%.*]] = xor i1 [[TMP73]], true
+; NOVEC-NEXT:    [[TMP76:%.*]] = xor i1 [[TMP74]], true
+; NOVEC-NEXT:    [[TMP77:%.*]] = xor i1 [[TMP75]], true
+; NOVEC-NEXT:    [[TMP78:%.*]] = xor i1 [[TMP76]], true
+; NOVEC-NEXT:    [[TMP79:%.*]] = or i1 [[TMP77]], [[TMP78]]
+; NOVEC-NEXT:    [[TMP80:%.*]] = extractelement <16 x i1> [[TMP0]], i32 14
+; NOVEC-NEXT:    [[TMP81:%.*]] = xor i1 [[TMP79]], true
+; NOVEC-NEXT:    [[TMP82:%.*]] = xor i1 [[TMP80]], true
+; NOVEC-NEXT:    [[TMP83:%.*]] = xor i1 [[TMP81]], true
+; NOVEC-NEXT:    [[TMP84:%.*]] = xor i1 [[TMP82]], true
+; NOVEC-NEXT:    [[TMP85:%.*]] = or i1 [[TMP83]], [[TMP84]]
+; NOVEC-NEXT:    [[TMP86:%.*]] = extractelement <16 x i1> [[TMP0]], i32 15
+; NOVEC-NEXT:    [[TMP87:%.*]] = xor i1 [[TMP85]], true
+; NOVEC-NEXT:    [[TMP88:%.*]] = xor i1 [[TMP86]], true
+; NOVEC-NEXT:    [[TMP89:%.*]] = xor i1 [[TMP87]], true
+; NOVEC-NEXT:    [[TMP90:%.*]] = xor i1 [[TMP88]], true
+; NOVEC-NEXT:    [[TMP91:%.*]] = or i1 [[TMP89]], [[TMP90]]
+; NOVEC-NEXT:    br i1 [[TMP91]], label [[TMP92:%.*]], label [[MIDDLE_BLOCK:%.*]]
+; NOVEC:       92:
+; NOVEC-NEXT:    store i1 false, ptr [[P]], align 1
+; NOVEC-NEXT:    br label [[MIDDLE_BLOCK]]
+; NOVEC:       middle.block:
+; NOVEC-NEXT:    [[INDEX_NEXT:%.*]] = add i32 0, 16
+; NOVEC-NEXT:    ret void
+;
+; CHECK-LABEL: define void @test(
+; CHECK-SAME: ptr [[P:%.*]]) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    br label [[PRED_STORE_CONTINUE30:%.*]]
+; CHECK:       vector.body:
+; CHECK-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[INDEX_NEXT:%.*]], [[PRED_STORE_CONTINUE30]] ]
+; CHECK-NEXT:    store i1 false, ptr [[P]], align 1
+; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2
+; CHECK-NEXT:    [[TMP0:%.*]] = icmp eq i32 [[INDEX_NEXT]], 10
+; CHECK-NEXT:    br i1 [[TMP0]], label [[EXIT:%.*]], label [[PRED_STORE_CONTINUE30]], !llvm.loop [[LOOP0:![0-9]+]]
+; CHECK:       exit:
+; CHECK-NEXT:    ret void
+;
+entry:
+  br label %for.body
+
+for.body:                                         ; preds = %for.body, %entry
+  %iv = phi i32 [ 0, %entry ], [ %iv.next, %for.body ]
+  %trunc = trunc i40 0 to i32
----------------
david-arm wrote:

I worry these tests might be a bit fragile since you have a trunc of a constant 0 in the loop. I can imagine in future this might be treated as a constant rather than a trunc. If there are particular instructions in this loop that are critical to defending your code change it's probably worth making them more robust. One way to do that would be to pass in a i40 argument to the function and truncate that instead.

https://github.com/llvm/llvm-project/pull/91136


More information about the llvm-commits mailing list