Skip to content

Commit 9ce4f37

Browse files
jayfoadjayfoad
committed
[AMDGPU][GlobalISel] Fix legalization of G_UMULH
Scalarize before narrowing because the narrowing implementation does not work on vectors. This matches what we do for regular G_MUL. Differential Revision: https://reviews.llvm.org/D111129
1 parent e463b69 commit 9ce4f37

File tree

3 files changed

+305
-3
lines changed

3 files changed

+305
-3
lines changed

‎llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -606,7 +606,7 @@ AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST_,
606606

607607
auto &Mulh = getActionDefinitionsBuilder({G_UMULH, G_SMULH})
608608
.legalFor({S32})
609-
.maxScalarOrElt(0, S32);
609+
.maxScalar(0, S32);
610610

611611
if (ST.hasVOP3PInsts()) {
612612
Mulh

‎llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-umulh.mir

Lines changed: 124 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -1,6 +1,6 @@
11
# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
2-
# RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=fiji -global-isel-abort=0 -O0 -run-pass=legalizer %s -o - | FileCheck %s --check-prefix=GFX8
3-
# RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx906 -global-isel-abort=0 -O0 -run-pass=legalizer %s -o - | FileCheck %s --check-prefix=GFX9
2+
# RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=fiji -O0 -run-pass=legalizer %s -o - | FileCheck %s --check-prefix=GFX8
3+
# RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx906 -O0 -run-pass=legalizer %s -o - | FileCheck %s --check-prefix=GFX9
44

55
---
66
name: test_umulh_s32
@@ -122,6 +122,128 @@ body: |
122122
$vgpr0_vgpr1 = COPY %2
123123
...
124124

125+
---
126+
name: test_umulh_v2s64
127+
body: |
128+
bb.0:
129+
liveins: $vgpr0_vgpr1_vgpr2_vgpr3, $vgpr4_vgpr5_vgpr6_vgpr7
130+
131+
; GFX8-LABEL: name: test_umulh_v2s64
132+
; GFX8: [[COPY:%[0-9]+]]:_(<2 x s64>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3
133+
; GFX8: [[COPY1:%[0-9]+]]:_(<2 x s64>) = COPY $vgpr4_vgpr5_vgpr6_vgpr7
134+
; GFX8: [[UV:%[0-9]+]]:_(s64), [[UV1:%[0-9]+]]:_(s64) = G_UNMERGE_VALUES [[COPY]](<2 x s64>)
135+
; GFX8: [[UV2:%[0-9]+]]:_(s64), [[UV3:%[0-9]+]]:_(s64) = G_UNMERGE_VALUES [[COPY1]](<2 x s64>)
136+
; GFX8: [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[UV]](s64)
137+
; GFX8: [[UV6:%[0-9]+]]:_(s32), [[UV7:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[UV2]](s64)
138+
; GFX8: [[MUL:%[0-9]+]]:_(s32) = G_MUL [[UV5]], [[UV6]]
139+
; GFX8: [[MUL1:%[0-9]+]]:_(s32) = G_MUL [[UV4]], [[UV7]]
140+
; GFX8: [[UMULH:%[0-9]+]]:_(s32) = G_UMULH [[UV4]], [[UV6]]
141+
; GFX8: [[UADDO:%[0-9]+]]:_(s32), [[UADDO1:%[0-9]+]]:_(s1) = G_UADDO [[MUL]], [[MUL1]]
142+
; GFX8: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[UADDO1]](s1)
143+
; GFX8: [[UADDO2:%[0-9]+]]:_(s32), [[UADDO3:%[0-9]+]]:_(s1) = G_UADDO [[UADDO]], [[UMULH]]
144+
; GFX8: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[UADDO3]](s1)
145+
; GFX8: [[ADD:%[0-9]+]]:_(s32) = G_ADD [[ZEXT]], [[ZEXT1]]
146+
; GFX8: [[MUL2:%[0-9]+]]:_(s32) = G_MUL [[UV5]], [[UV7]]
147+
; GFX8: [[UMULH1:%[0-9]+]]:_(s32) = G_UMULH [[UV5]], [[UV6]]
148+
; GFX8: [[UMULH2:%[0-9]+]]:_(s32) = G_UMULH [[UV4]], [[UV7]]
149+
; GFX8: [[UADDO4:%[0-9]+]]:_(s32), [[UADDO5:%[0-9]+]]:_(s1) = G_UADDO [[MUL2]], [[UMULH1]]
150+
; GFX8: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[UADDO5]](s1)
151+
; GFX8: [[UADDO6:%[0-9]+]]:_(s32), [[UADDO7:%[0-9]+]]:_(s1) = G_UADDO [[UADDO4]], [[UMULH2]]
152+
; GFX8: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[UADDO7]](s1)
153+
; GFX8: [[ADD1:%[0-9]+]]:_(s32) = G_ADD [[ZEXT2]], [[ZEXT3]]
154+
; GFX8: [[UADDO8:%[0-9]+]]:_(s32), [[UADDO9:%[0-9]+]]:_(s1) = G_UADDO [[UADDO6]], [[ADD]]
155+
; GFX8: [[ZEXT4:%[0-9]+]]:_(s32) = G_ZEXT [[UADDO9]](s1)
156+
; GFX8: [[ADD2:%[0-9]+]]:_(s32) = G_ADD [[ADD1]], [[ZEXT4]]
157+
; GFX8: [[UMULH3:%[0-9]+]]:_(s32) = G_UMULH [[UV5]], [[UV7]]
158+
; GFX8: [[ADD3:%[0-9]+]]:_(s32) = G_ADD [[UMULH3]], [[ADD2]]
159+
; GFX8: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[UADDO8]](s32), [[ADD3]](s32)
160+
; GFX8: [[UV8:%[0-9]+]]:_(s32), [[UV9:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[UV1]](s64)
161+
; GFX8: [[UV10:%[0-9]+]]:_(s32), [[UV11:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[UV3]](s64)
162+
; GFX8: [[MUL3:%[0-9]+]]:_(s32) = G_MUL [[UV9]], [[UV10]]
163+
; GFX8: [[MUL4:%[0-9]+]]:_(s32) = G_MUL [[UV8]], [[UV11]]
164+
; GFX8: [[UMULH4:%[0-9]+]]:_(s32) = G_UMULH [[UV8]], [[UV10]]
165+
; GFX8: [[UADDO10:%[0-9]+]]:_(s32), [[UADDO11:%[0-9]+]]:_(s1) = G_UADDO [[MUL3]], [[MUL4]]
166+
; GFX8: [[ZEXT5:%[0-9]+]]:_(s32) = G_ZEXT [[UADDO11]](s1)
167+
; GFX8: [[UADDO12:%[0-9]+]]:_(s32), [[UADDO13:%[0-9]+]]:_(s1) = G_UADDO [[UADDO10]], [[UMULH4]]
168+
; GFX8: [[ZEXT6:%[0-9]+]]:_(s32) = G_ZEXT [[UADDO13]](s1)
169+
; GFX8: [[ADD4:%[0-9]+]]:_(s32) = G_ADD [[ZEXT5]], [[ZEXT6]]
170+
; GFX8: [[MUL5:%[0-9]+]]:_(s32) = G_MUL [[UV9]], [[UV11]]
171+
; GFX8: [[UMULH5:%[0-9]+]]:_(s32) = G_UMULH [[UV9]], [[UV10]]
172+
; GFX8: [[UMULH6:%[0-9]+]]:_(s32) = G_UMULH [[UV8]], [[UV11]]
173+
; GFX8: [[UADDO14:%[0-9]+]]:_(s32), [[UADDO15:%[0-9]+]]:_(s1) = G_UADDO [[MUL5]], [[UMULH5]]
174+
; GFX8: [[ZEXT7:%[0-9]+]]:_(s32) = G_ZEXT [[UADDO15]](s1)
175+
; GFX8: [[UADDO16:%[0-9]+]]:_(s32), [[UADDO17:%[0-9]+]]:_(s1) = G_UADDO [[UADDO14]], [[UMULH6]]
176+
; GFX8: [[ZEXT8:%[0-9]+]]:_(s32) = G_ZEXT [[UADDO17]](s1)
177+
; GFX8: [[ADD5:%[0-9]+]]:_(s32) = G_ADD [[ZEXT7]], [[ZEXT8]]
178+
; GFX8: [[UADDO18:%[0-9]+]]:_(s32), [[UADDO19:%[0-9]+]]:_(s1) = G_UADDO [[UADDO16]], [[ADD4]]
179+
; GFX8: [[ZEXT9:%[0-9]+]]:_(s32) = G_ZEXT [[UADDO19]](s1)
180+
; GFX8: [[ADD6:%[0-9]+]]:_(s32) = G_ADD [[ADD5]], [[ZEXT9]]
181+
; GFX8: [[UMULH7:%[0-9]+]]:_(s32) = G_UMULH [[UV9]], [[UV11]]
182+
; GFX8: [[ADD7:%[0-9]+]]:_(s32) = G_ADD [[UMULH7]], [[ADD6]]
183+
; GFX8: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[UADDO18]](s32), [[ADD7]](s32)
184+
; GFX8: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64)
185+
; GFX8: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
186+
; GFX9-LABEL: name: test_umulh_v2s64
187+
; GFX9: [[COPY:%[0-9]+]]:_(<2 x s64>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3
188+
; GFX9: [[COPY1:%[0-9]+]]:_(<2 x s64>) = COPY $vgpr4_vgpr5_vgpr6_vgpr7
189+
; GFX9: [[UV:%[0-9]+]]:_(s64), [[UV1:%[0-9]+]]:_(s64) = G_UNMERGE_VALUES [[COPY]](<2 x s64>)
190+
; GFX9: [[UV2:%[0-9]+]]:_(s64), [[UV3:%[0-9]+]]:_(s64) = G_UNMERGE_VALUES [[COPY1]](<2 x s64>)
191+
; GFX9: [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[UV]](s64)
192+
; GFX9: [[UV6:%[0-9]+]]:_(s32), [[UV7:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[UV2]](s64)
193+
; GFX9: [[MUL:%[0-9]+]]:_(s32) = G_MUL [[UV5]], [[UV6]]
194+
; GFX9: [[MUL1:%[0-9]+]]:_(s32) = G_MUL [[UV4]], [[UV7]]
195+
; GFX9: [[UMULH:%[0-9]+]]:_(s32) = G_UMULH [[UV4]], [[UV6]]
196+
; GFX9: [[UADDO:%[0-9]+]]:_(s32), [[UADDO1:%[0-9]+]]:_(s1) = G_UADDO [[MUL]], [[MUL1]]
197+
; GFX9: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[UADDO1]](s1)
198+
; GFX9: [[UADDO2:%[0-9]+]]:_(s32), [[UADDO3:%[0-9]+]]:_(s1) = G_UADDO [[UADDO]], [[UMULH]]
199+
; GFX9: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[UADDO3]](s1)
200+
; GFX9: [[ADD:%[0-9]+]]:_(s32) = G_ADD [[ZEXT]], [[ZEXT1]]
201+
; GFX9: [[MUL2:%[0-9]+]]:_(s32) = G_MUL [[UV5]], [[UV7]]
202+
; GFX9: [[UMULH1:%[0-9]+]]:_(s32) = G_UMULH [[UV5]], [[UV6]]
203+
; GFX9: [[UMULH2:%[0-9]+]]:_(s32) = G_UMULH [[UV4]], [[UV7]]
204+
; GFX9: [[UADDO4:%[0-9]+]]:_(s32), [[UADDO5:%[0-9]+]]:_(s1) = G_UADDO [[MUL2]], [[UMULH1]]
205+
; GFX9: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[UADDO5]](s1)
206+
; GFX9: [[UADDO6:%[0-9]+]]:_(s32), [[UADDO7:%[0-9]+]]:_(s1) = G_UADDO [[UADDO4]], [[UMULH2]]
207+
; GFX9: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[UADDO7]](s1)
208+
; GFX9: [[ADD1:%[0-9]+]]:_(s32) = G_ADD [[ZEXT2]], [[ZEXT3]]
209+
; GFX9: [[UADDO8:%[0-9]+]]:_(s32), [[UADDO9:%[0-9]+]]:_(s1) = G_UADDO [[UADDO6]], [[ADD]]
210+
; GFX9: [[ZEXT4:%[0-9]+]]:_(s32) = G_ZEXT [[UADDO9]](s1)
211+
; GFX9: [[ADD2:%[0-9]+]]:_(s32) = G_ADD [[ADD1]], [[ZEXT4]]
212+
; GFX9: [[UMULH3:%[0-9]+]]:_(s32) = G_UMULH [[UV5]], [[UV7]]
213+
; GFX9: [[ADD3:%[0-9]+]]:_(s32) = G_ADD [[UMULH3]], [[ADD2]]
214+
; GFX9: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[UADDO8]](s32), [[ADD3]](s32)
215+
; GFX9: [[UV8:%[0-9]+]]:_(s32), [[UV9:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[UV1]](s64)
216+
; GFX9: [[UV10:%[0-9]+]]:_(s32), [[UV11:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[UV3]](s64)
217+
; GFX9: [[MUL3:%[0-9]+]]:_(s32) = G_MUL [[UV9]], [[UV10]]
218+
; GFX9: [[MUL4:%[0-9]+]]:_(s32) = G_MUL [[UV8]], [[UV11]]
219+
; GFX9: [[UMULH4:%[0-9]+]]:_(s32) = G_UMULH [[UV8]], [[UV10]]
220+
; GFX9: [[UADDO10:%[0-9]+]]:_(s32), [[UADDO11:%[0-9]+]]:_(s1) = G_UADDO [[MUL3]], [[MUL4]]
221+
; GFX9: [[ZEXT5:%[0-9]+]]:_(s32) = G_ZEXT [[UADDO11]](s1)
222+
; GFX9: [[UADDO12:%[0-9]+]]:_(s32), [[UADDO13:%[0-9]+]]:_(s1) = G_UADDO [[UADDO10]], [[UMULH4]]
223+
; GFX9: [[ZEXT6:%[0-9]+]]:_(s32) = G_ZEXT [[UADDO13]](s1)
224+
; GFX9: [[ADD4:%[0-9]+]]:_(s32) = G_ADD [[ZEXT5]], [[ZEXT6]]
225+
; GFX9: [[MUL5:%[0-9]+]]:_(s32) = G_MUL [[UV9]], [[UV11]]
226+
; GFX9: [[UMULH5:%[0-9]+]]:_(s32) = G_UMULH [[UV9]], [[UV10]]
227+
; GFX9: [[UMULH6:%[0-9]+]]:_(s32) = G_UMULH [[UV8]], [[UV11]]
228+
; GFX9: [[UADDO14:%[0-9]+]]:_(s32), [[UADDO15:%[0-9]+]]:_(s1) = G_UADDO [[MUL5]], [[UMULH5]]
229+
; GFX9: [[ZEXT7:%[0-9]+]]:_(s32) = G_ZEXT [[UADDO15]](s1)
230+
; GFX9: [[UADDO16:%[0-9]+]]:_(s32), [[UADDO17:%[0-9]+]]:_(s1) = G_UADDO [[UADDO14]], [[UMULH6]]
231+
; GFX9: [[ZEXT8:%[0-9]+]]:_(s32) = G_ZEXT [[UADDO17]](s1)
232+
; GFX9: [[ADD5:%[0-9]+]]:_(s32) = G_ADD [[ZEXT7]], [[ZEXT8]]
233+
; GFX9: [[UADDO18:%[0-9]+]]:_(s32), [[UADDO19:%[0-9]+]]:_(s1) = G_UADDO [[UADDO16]], [[ADD4]]
234+
; GFX9: [[ZEXT9:%[0-9]+]]:_(s32) = G_ZEXT [[UADDO19]](s1)
235+
; GFX9: [[ADD6:%[0-9]+]]:_(s32) = G_ADD [[ADD5]], [[ZEXT9]]
236+
; GFX9: [[UMULH7:%[0-9]+]]:_(s32) = G_UMULH [[UV9]], [[UV11]]
237+
; GFX9: [[ADD7:%[0-9]+]]:_(s32) = G_ADD [[UMULH7]], [[ADD6]]
238+
; GFX9: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[UADDO18]](s32), [[ADD7]](s32)
239+
; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64)
240+
; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
241+
%0:_(<2 x s64>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3
242+
%1:_(<2 x s64>) = COPY $vgpr4_vgpr5_vgpr6_vgpr7
243+
%2:_(<2 x s64>) = G_UMULH %0, %1
244+
$vgpr0_vgpr1_vgpr2_vgpr3 = COPY %2
245+
...
246+
125247
---
126248
name: test_umulh_s16
127249
body: |

0 commit comments

Comments
 (0)