This is an archive of the discontinued LLVM Phabricator instance.

Differential D148590

[InstCombine] Add tests for new eq/ne patterns combining eq/ne by parts; NFC
Needs ReviewPublic

Authored by goldstein.w.n on Apr 17 2023, 5:32 PM.

Details

Reviewers
nikic
spatel
majnemer

Diff Detail

Unit TestsFailed

TimeTest
190 msx64 debian > LLVM.Examples/OrcV2Examples::lljit-with-thinlto-summaries.test
40 msx64 debian > LLVM.Examples/OrcV2Examples::orcv2-cbindings-add-object-file.test
40 msx64 debian > LLVM.Examples/OrcV2Examples::orcv2-cbindings-basic-usage.test
40 msx64 debian > LLVM.Examples/OrcV2Examples::orcv2-cbindings-lazy.test
40 msx64 debian > LLVM.Examples/OrcV2Examples::orcv2-cbindings-removable-code.test

Event Timeline

goldstein.w.n created this revision.Apr 17 2023, 5:32 PM
Herald added a project: Restricted Project. · View Herald TranscriptApr 17 2023, 5:32 PM
Herald added a subscriber: StephenFan. · View Herald Transcript
goldstein.w.n requested review of this revision.Apr 17 2023, 5:32 PM
Herald added a project: Restricted Project. · View Herald TranscriptApr 17 2023, 5:32 PM
Herald added a subscriber: llvm-commits. · View Herald Transcript
Harbormaster completed remote builds in B226274: Diff 514477.Apr 17 2023, 6:47 PM

Revision Contents

PathSize
llvm/
test/
Transforms/
InstCombine/
114 lines

Diff 514477

llvm/test/Transforms/InstCombine/eq-of-parts.ll

Show First 20 Lines • Show All 1,327 Lines • ▼ Show 20 Lines;
%y.1 = trunc i32 %y.321 to i8 %y.1 = trunc i32 %y.321 to i8
%y.32 = lshr i32 %y, 16 %y.32 = lshr i32 %y, 16
%y.2 = trunc i32 %y.32 to i8 %y.2 = trunc i32 %y.32 to i8
%c.1 = icmp eq i8 %x.1, %y.1 %c.1 = icmp eq i8 %x.1, %y.1
%c.2 = icmp eq i8 %x.2, %y.2 %c.2 = icmp eq i8 %x.2, %y.2
%c.210 = or i1 %c.2, %c.1 %c.210 = or i1 %c.2, %c.1
ret i1 %c.210 ret i1 %c.210
} }
define i1 @eq_optimized_highbits_cmp(i32 %x, i32 %y) {
; CHECK-LABEL: @eq_optimized_highbits_cmp(
; CHECK-NEXT: [[XOR:%.*]] = xor i32 [[Y:%.*]], [[X:%.*]]
; CHECK-NEXT: [[CMP_HI:%.*]] = icmp ult i32 [[XOR]], 33554432
; CHECK-NEXT: [[TX:%.*]] = trunc i32 [[X]] to i25
; CHECK-NEXT: [[TY:%.*]] = trunc i32 [[Y]] to i25
; CHECK-NEXT: [[CMP_LO:%.*]] = icmp eq i25 [[TX]], [[TY]]
; CHECK-NEXT: [[R:%.*]] = and i1 [[CMP_HI]], [[CMP_LO]]
; CHECK-NEXT: ret i1 [[R]]
;
%xor = xor i32 %y, %x
%cmp_hi = icmp ult i32 %xor, 33554432
%tx = trunc i32 %x to i25
%ty = trunc i32 %y to i25
%cmp_lo = icmp eq i25 %tx, %ty
%r = and i1 %cmp_hi, %cmp_lo
ret i1 %r
}
define i1 @eq_optimized_highbits_cmp_todo_overlapping(i32 %x, i32 %y) {
; CHECK-LABEL: @eq_optimized_highbits_cmp_todo_overlapping(
; CHECK-NEXT: [[XOR:%.*]] = xor i32 [[Y:%.*]], [[X:%.*]]
; CHECK-NEXT: [[CMP_HI:%.*]] = icmp ult i32 [[XOR]], 16777216
; CHECK-NEXT: [[TX:%.*]] = trunc i32 [[X]] to i25
; CHECK-NEXT: [[TY:%.*]] = trunc i32 [[Y]] to i25
; CHECK-NEXT: [[CMP_LO:%.*]] = icmp eq i25 [[TX]], [[TY]]
; CHECK-NEXT: [[R:%.*]] = and i1 [[CMP_HI]], [[CMP_LO]]
; CHECK-NEXT: ret i1 [[R]]
;
%xor = xor i32 %y, %x
%cmp_hi = icmp ult i32 %xor, 16777216
%tx = trunc i32 %x to i25
%ty = trunc i32 %y to i25
%cmp_lo = icmp eq i25 %tx, %ty
%r = and i1 %cmp_hi, %cmp_lo
ret i1 %r
}
define i1 @eq_optimized_highbits_cmp_fail_not_pow2(i32 %x, i32 %y) {
; CHECK-LABEL: @eq_optimized_highbits_cmp_fail_not_pow2(
; CHECK-NEXT: [[XOR:%.*]] = xor i32 [[Y:%.*]], [[X:%.*]]
; CHECK-NEXT: [[CMP_HI:%.*]] = icmp ult i32 [[XOR]], 16777215
; CHECK-NEXT: [[TX:%.*]] = trunc i32 [[X]] to i24
; CHECK-NEXT: [[TY:%.*]] = trunc i32 [[Y]] to i24
; CHECK-NEXT: [[CMP_LO:%.*]] = icmp eq i24 [[TX]], [[TY]]
; CHECK-NEXT: [[R:%.*]] = and i1 [[CMP_HI]], [[CMP_LO]]
; CHECK-NEXT: ret i1 [[R]]
;
%xor = xor i32 %y, %x
%cmp_hi = icmp ult i32 %xor, 16777215
%tx = trunc i32 %x to i24
%ty = trunc i32 %y to i24
%cmp_lo = icmp eq i24 %tx, %ty
%r = and i1 %cmp_hi, %cmp_lo
ret i1 %r
}
define i1 @ne_optimized_highbits_cmp(i32 %x, i32 %y) {
; CHECK-LABEL: @ne_optimized_highbits_cmp(
; CHECK-NEXT: [[XOR:%.*]] = xor i32 [[Y:%.*]], [[X:%.*]]
; CHECK-NEXT: [[CMP_HI:%.*]] = icmp ugt i32 [[XOR]], 16777215
; CHECK-NEXT: [[TX:%.*]] = trunc i32 [[X]] to i24
; CHECK-NEXT: [[TY:%.*]] = trunc i32 [[Y]] to i24
; CHECK-NEXT: [[CMP_LO:%.*]] = icmp ne i24 [[TX]], [[TY]]
; CHECK-NEXT: [[R:%.*]] = or i1 [[CMP_HI]], [[CMP_LO]]
; CHECK-NEXT: ret i1 [[R]]
;
%xor = xor i32 %y, %x
%cmp_hi = icmp ugt i32 %xor, 16777215
%tx = trunc i32 %x to i24
%ty = trunc i32 %y to i24
%cmp_lo = icmp ne i24 %tx, %ty
%r = or i1 %cmp_hi, %cmp_lo
ret i1 %r
}
define i1 @ne_optimized_highbits_cmp_fail_not_mask(i32 %x, i32 %y) {
; CHECK-LABEL: @ne_optimized_highbits_cmp_fail_not_mask(
; CHECK-NEXT: [[XOR:%.*]] = xor i32 [[Y:%.*]], [[X:%.*]]
; CHECK-NEXT: [[CMP_HI:%.*]] = icmp ugt i32 [[XOR]], 16777216
; CHECK-NEXT: [[TX:%.*]] = trunc i32 [[X]] to i24
; CHECK-NEXT: [[TY:%.*]] = trunc i32 [[Y]] to i24
; CHECK-NEXT: [[CMP_LO:%.*]] = icmp ne i24 [[TX]], [[TY]]
; CHECK-NEXT: [[R:%.*]] = or i1 [[CMP_HI]], [[CMP_LO]]
; CHECK-NEXT: ret i1 [[R]]
;
%xor = xor i32 %y, %x
%cmp_hi = icmp ugt i32 %xor, 16777216
%tx = trunc i32 %x to i24
%ty = trunc i32 %y to i24
%cmp_lo = icmp ne i24 %tx, %ty
%r = or i1 %cmp_hi, %cmp_lo
ret i1 %r
}
define i1 @ne_optimized_highbits_cmp_fail_no_combined_int(i32 %x, i32 %y) {
; CHECK-LABEL: @ne_optimized_highbits_cmp_fail_no_combined_int(
; CHECK-NEXT: [[XOR:%.*]] = xor i32 [[Y:%.*]], [[X:%.*]]
; CHECK-NEXT: [[CMP_HI:%.*]] = icmp ugt i32 [[XOR]], 16777215
; CHECK-NEXT: [[TX:%.*]] = trunc i32 [[X]] to i23
; CHECK-NEXT: [[TY:%.*]] = trunc i32 [[Y]] to i23
; CHECK-NEXT: [[CMP_LO:%.*]] = icmp ne i23 [[TX]], [[TY]]
; CHECK-NEXT: [[R:%.*]] = or i1 [[CMP_HI]], [[CMP_LO]]
; CHECK-NEXT: ret i1 [[R]]
;
%xor = xor i32 %y, %x
%cmp_hi = icmp ugt i32 %xor, 16777215
%tx = trunc i32 %x to i23
%ty = trunc i32 %y to i23
%cmp_lo = icmp ne i23 %tx, %ty
%r = or i1 %cmp_hi, %cmp_lo
ret i1 %r
}