This is an archive of the discontinued LLVM Phabricator instance.

Differential D57286

AMDGPU/GlobalISel: Legalize icmp for pointer types
ClosedPublic

Authored by arsenm on Jan 26 2019, 9:41 AM.

Details

Reviewers
tstellar
rampitec

Diff Detail

Event Timeline

arsenm created this revision.Jan 26 2019, 9:41 AM
Herald added subscribers: volkan, t-tye, tpr and 8 others. · View Herald TranscriptJan 26 2019, 9:42 AM
tstellar accepted this revision.Feb 1 2019, 4:06 PM

LGTM.

This revision is now accepted and ready to land.Feb 1 2019, 4:06 PM
Herald added a subscriber: Petar.Avramovic. · View Herald TranscriptFeb 1 2019, 4:06 PM
arsenm closed this revision.Feb 2 2019, 3:35 PM

r352976

Revision Contents

PathSize
lib/
Target/
AMDGPU/
11 lines
test/
CodeGen/
AMDGPU/
GlobalISel/
175 lines

Diff 183709

lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

Show First 20 Lines • Show All 188 Lines • ▼ Show 20 LinesAMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST,
// instruction with scalar types for their pointer operands. In assert builds, // instruction with scalar types for their pointer operands. In assert builds,
// the instruction selector will assert if it sees a generic instruction which // the instruction selector will assert if it sees a generic instruction which
// isn't legal, so we need to tell it that scalar types are legal for pointer // isn't legal, so we need to tell it that scalar types are legal for pointer
// operands // operands
setAction({G_GEP, S64}, Legal); setAction({G_GEP, S64}, Legal);
setAction({G_BLOCK_ADDR, CodePtr}, Legal); setAction({G_BLOCK_ADDR, CodePtr}, Legal);
getActionDefinitionsBuilder({G_ICMP, G_FCMP}) getActionDefinitionsBuilder(G_ICMP)
.legalForCartesianProduct(
{S1}, {S32, S64, GlobalPtr, LocalPtr, ConstantPtr, PrivatePtr, FlatPtr})
.legalFor({{S1, S32}, {S1, S64}})
.widenScalarToNextPow2(1)
.clampScalar(1, S32, S64)
.scalarize(0)
.legalIf(all(typeIs(0, S1), isPointer(1)));
getActionDefinitionsBuilder(G_FCMP)
.legalFor({{S1, S32}, {S1, S64}}) .legalFor({{S1, S32}, {S1, S64}})
.widenScalarToNextPow2(1) .widenScalarToNextPow2(1)
.clampScalar(1, S32, S64) .clampScalar(1, S32, S64)
.scalarize(0); .scalarize(0);
// FIXME: fexp, flog2, flog10 needs to be custom lowered. // FIXME: fexp, flog2, flog10 needs to be custom lowered.
getActionDefinitionsBuilder({G_FPOW, G_FEXP, G_FEXP2, getActionDefinitionsBuilder({G_FPOW, G_FEXP, G_FEXP2,
G_FLOG, G_FLOG2, G_FLOG10}) G_FLOG, G_FLOG2, G_FLOG10})
▲ Show 20 Lines • Show All 222 Lines • Show Last 20 Lines

test/CodeGen/AMDGPU/GlobalISel/legalize-icmp.mir

Show First 20 Lines • Show All 158 Lines • ▼ Show 20 Lines bb.0.entry:
; CHECK: S_NOP 0, implicit [[BUILD_VECTOR]](<4 x s1>) ; CHECK: S_NOP 0, implicit [[BUILD_VECTOR]](<4 x s1>)
%0:_(p1) = G_IMPLICIT_DEF %0:_(p1) = G_IMPLICIT_DEF
%1:_(<4 x s32>) = G_LOAD %0 :: (volatile load 16) %1:_(<4 x s32>) = G_LOAD %0 :: (volatile load 16)
%2:_(<4 x s32>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3 %2:_(<4 x s32>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3
%3:_(<4 x s1>) = G_ICMP intpred(ne), %1, %2 %3:_(<4 x s1>) = G_ICMP intpred(ne), %1, %2
S_NOP 0, implicit %3 S_NOP 0, implicit %3
... ...
---
name: test_icmp_p0
body: |
bb.0.entry:
liveins: $vgpr0_vgpr1, $vgpr2_vgpr3
; CHECK-LABEL: name: test_icmp_p0
; CHECK: [[COPY:%[0-9]+]]:_(p0) = COPY $vgpr0_vgpr1
; CHECK: [[COPY1:%[0-9]+]]:_(p0) = COPY $vgpr0_vgpr1
; CHECK: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(ne), [[COPY]](p0), [[COPY1]]
; CHECK: [[SEXT:%[0-9]+]]:_(s32) = G_SEXT [[ICMP]](s1)
; CHECK: $vgpr0 = COPY [[SEXT]](s32)
%0:_(p0) = COPY $vgpr0_vgpr1
%1:_(p0) = COPY $vgpr0_vgpr1
%2:_(s1) = G_ICMP intpred(ne), %0, %1
%3:_(s32) = G_SEXT %2
$vgpr0 = COPY %3
...
---
name: test_icmp_p1
body: |
bb.0.entry:
liveins: $vgpr0_vgpr1, $vgpr2_vgpr3
; CHECK-LABEL: name: test_icmp_p1
; CHECK: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
; CHECK: [[COPY1:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
; CHECK: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(ne), [[COPY]](p1), [[COPY1]]
; CHECK: [[SEXT:%[0-9]+]]:_(s32) = G_SEXT [[ICMP]](s1)
; CHECK: $vgpr0 = COPY [[SEXT]](s32)
%0:_(p1) = COPY $vgpr0_vgpr1
%1:_(p1) = COPY $vgpr0_vgpr1
%2:_(s1) = G_ICMP intpred(ne), %0, %1
%3:_(s32) = G_SEXT %2
$vgpr0 = COPY %3
...
---
name: test_icmp_p2
body: |
bb.0.entry:
liveins: $vgpr0, $vgpr1
; CHECK-LABEL: name: test_icmp_p2
; CHECK: [[COPY:%[0-9]+]]:_(p2) = COPY $vgpr0
; CHECK: [[COPY1:%[0-9]+]]:_(p2) = COPY $vgpr1
; CHECK: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(ne), [[COPY]](p2), [[COPY1]]
; CHECK: [[SEXT:%[0-9]+]]:_(s32) = G_SEXT [[ICMP]](s1)
; CHECK: $vgpr0 = COPY [[SEXT]](s32)
%0:_(p2) = COPY $vgpr0
%1:_(p2) = COPY $vgpr1
%2:_(s1) = G_ICMP intpred(ne), %0, %1
%3:_(s32) = G_SEXT %2
$vgpr0 = COPY %3
...
---
name: test_icmp_p3
body: |
bb.0.entry:
liveins: $vgpr0, $vgpr1
; CHECK-LABEL: name: test_icmp_p3
; CHECK: [[COPY:%[0-9]+]]:_(p3) = COPY $vgpr0
; CHECK: [[COPY1:%[0-9]+]]:_(p3) = COPY $vgpr1
; CHECK: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(ne), [[COPY]](p3), [[COPY1]]
; CHECK: [[SEXT:%[0-9]+]]:_(s32) = G_SEXT [[ICMP]](s1)
; CHECK: $vgpr0 = COPY [[SEXT]](s32)
%0:_(p3) = COPY $vgpr0
%1:_(p3) = COPY $vgpr1
%2:_(s1) = G_ICMP intpred(ne), %0, %1
%3:_(s32) = G_SEXT %2
$vgpr0 = COPY %3
...
---
name: test_icmp_p4
body: |
bb.0.entry:
liveins: $vgpr0_vgpr1, $vgpr2_vgpr3
; CHECK-LABEL: name: test_icmp_p4
; CHECK: [[COPY:%[0-9]+]]:_(p4) = COPY $vgpr0_vgpr1
; CHECK: [[COPY1:%[0-9]+]]:_(p4) = COPY $vgpr0_vgpr1
; CHECK: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(ne), [[COPY]](p4), [[COPY1]]
; CHECK: [[SEXT:%[0-9]+]]:_(s32) = G_SEXT [[ICMP]](s1)
; CHECK: $vgpr0 = COPY [[SEXT]](s32)
%0:_(p4) = COPY $vgpr0_vgpr1
%1:_(p4) = COPY $vgpr0_vgpr1
%2:_(s1) = G_ICMP intpred(ne), %0, %1
%3:_(s32) = G_SEXT %2
$vgpr0 = COPY %3
...
---
name: test_icmp_p5
body: |
bb.0.entry:
liveins: $vgpr0, $vgpr1
; CHECK-LABEL: name: test_icmp_p5
; CHECK: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0
; CHECK: [[COPY1:%[0-9]+]]:_(p5) = COPY $vgpr1
; CHECK: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(ne), [[COPY]](p5), [[COPY1]]
; CHECK: [[SEXT:%[0-9]+]]:_(s32) = G_SEXT [[ICMP]](s1)
; CHECK: $vgpr0 = COPY [[SEXT]](s32)
%0:_(p5) = COPY $vgpr0
%1:_(p5) = COPY $vgpr1
%2:_(s1) = G_ICMP intpred(ne), %0, %1
%3:_(s32) = G_SEXT %2
$vgpr0 = COPY %3
...
---
name: test_icmp_p999
body: |
bb.0.entry:
liveins: $vgpr0_vgpr1, $vgpr2_vgpr3
; CHECK-LABEL: name: test_icmp_p999
; CHECK: [[COPY:%[0-9]+]]:_(p999) = COPY $vgpr0_vgpr1
; CHECK: [[COPY1:%[0-9]+]]:_(p999) = COPY $vgpr0_vgpr1
; CHECK: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(ne), [[COPY]](p999), [[COPY1]]
; CHECK: [[SEXT:%[0-9]+]]:_(s32) = G_SEXT [[ICMP]](s1)
; CHECK: $vgpr0 = COPY [[SEXT]](s32)
%0:_(p999) = COPY $vgpr0_vgpr1
%1:_(p999) = COPY $vgpr0_vgpr1
%2:_(s1) = G_ICMP intpred(ne), %0, %1
%3:_(s32) = G_SEXT %2
$vgpr0 = COPY %3
...
---
name: test_icmp_v2p3
body: |
bb.0.entry:
liveins: $vgpr0_vgpr1, $vgpr2_vgpr3
; CHECK-LABEL: name: test_icmp_v2p3
; CHECK: [[COPY:%[0-9]+]]:_(<2 x p3>) = COPY $vgpr0_vgpr1
; CHECK: [[COPY1:%[0-9]+]]:_(<2 x p3>) = COPY $vgpr0_vgpr1
; CHECK: [[UV:%[0-9]+]]:_(p3), [[UV1:%[0-9]+]]:_(p3) = G_UNMERGE_VALUES [[COPY]](<2 x p3>)
; CHECK: [[UV2:%[0-9]+]]:_(p3), [[UV3:%[0-9]+]]:_(p3) = G_UNMERGE_VALUES [[COPY1]](<2 x p3>)
; CHECK: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(ne), [[UV]](p3), [[UV2]]
; CHECK: [[ICMP1:%[0-9]+]]:_(s1) = G_ICMP intpred(ne), [[UV1]](p3), [[UV3]]
; CHECK: [[SEXT:%[0-9]+]]:_(s32) = G_SEXT [[ICMP]](s1)
; CHECK: [[SEXT1:%[0-9]+]]:_(s32) = G_SEXT [[ICMP1]](s1)
; CHECK: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[SEXT]](s32), [[SEXT1]](s32)
; CHECK: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
%0:_(<2 x p3>) = COPY $vgpr0_vgpr1
%1:_(<2 x p3>) = COPY $vgpr0_vgpr1
%2:_(<2 x s1>) = G_ICMP intpred(ne), %0, %1
%3:_(<2 x s32>) = G_SEXT %2
$vgpr0_vgpr1 = COPY %3
...
---
name: test_icmp_v2p999
body: |
bb.0.entry:
liveins: $vgpr0_vgpr1_vgpr2_vgpr3, $vgpr4_vgpr5_vgpr6_vgpr7
; CHECK-LABEL: name: test_icmp_v2p999
; CHECK: [[COPY:%[0-9]+]]:_(<2 x p999>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3
; CHECK: [[COPY1:%[0-9]+]]:_(<2 x p999>) = COPY $vgpr4_vgpr5_vgpr6_vgpr7
; CHECK: [[UV:%[0-9]+]]:_(p999), [[UV1:%[0-9]+]]:_(p999) = G_UNMERGE_VALUES [[COPY]](<2 x p999>)
; CHECK: [[UV2:%[0-9]+]]:_(p999), [[UV3:%[0-9]+]]:_(p999) = G_UNMERGE_VALUES [[COPY1]](<2 x p999>)
; CHECK: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(ne), [[UV]](p999), [[UV2]]
; CHECK: [[ICMP1:%[0-9]+]]:_(s1) = G_ICMP intpred(ne), [[UV1]](p999), [[UV3]]
; CHECK: [[SEXT:%[0-9]+]]:_(s32) = G_SEXT [[ICMP]](s1)
; CHECK: [[SEXT1:%[0-9]+]]:_(s32) = G_SEXT [[ICMP1]](s1)
; CHECK: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[SEXT]](s32), [[SEXT1]](s32)
; CHECK: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
%0:_(<2 x p999>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3
%1:_(<2 x p999>) = COPY $vgpr4_vgpr5_vgpr6_vgpr7
%2:_(<2 x s1>) = G_ICMP intpred(ne), %0, %1
%3:_(<2 x s32>) = G_SEXT %2
$vgpr0_vgpr1 = COPY %3
...