@@ -32379,15 +32379,15 @@ define bfloat @v_uitofp_i16_to_bf16(i16 %x) {
32379
32379
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
32380
32380
; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
32381
32381
; GCN-NEXT: v_cvt_f32_u32_e32 v0, v0
32382
- ; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000 , v0
32382
+ ; GCN-NEXT: v_and_b32_e32 v0, 0x7fff0000 , v0
32383
32383
; GCN-NEXT: s_setpc_b64 s[30:31]
32384
32384
;
32385
32385
; GFX7-LABEL: v_uitofp_i16_to_bf16:
32386
32386
; GFX7: ; %bb.0:
32387
32387
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
32388
32388
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff, v0
32389
32389
; GFX7-NEXT: v_cvt_f32_u32_e32 v0, v0
32390
- ; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000 , v0
32390
+ ; GFX7-NEXT: v_and_b32_e32 v0, 0x7fff0000 , v0
32391
32391
; GFX7-NEXT: s_setpc_b64 s[30:31]
32392
32392
;
32393
32393
; GFX8-LABEL: v_uitofp_i16_to_bf16:
@@ -32455,8 +32455,8 @@ define <2 x bfloat> @v_uitofp_v2i16_to_v2bf16(<2 x i16> %x) {
32455
32455
; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
32456
32456
; GCN-NEXT: v_cvt_f32_u32_e32 v1, v1
32457
32457
; GCN-NEXT: v_cvt_f32_u32_e32 v0, v0
32458
- ; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000 , v0
32459
- ; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000 , v1
32458
+ ; GCN-NEXT: v_and_b32_e32 v0, 0x7fff0000 , v0
32459
+ ; GCN-NEXT: v_and_b32_e32 v1, 0x7fff0000 , v1
32460
32460
; GCN-NEXT: s_setpc_b64 s[30:31]
32461
32461
;
32462
32462
; GFX7-LABEL: v_uitofp_v2i16_to_v2bf16:
@@ -32466,8 +32466,8 @@ define <2 x bfloat> @v_uitofp_v2i16_to_v2bf16(<2 x i16> %x) {
32466
32466
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff, v1
32467
32467
; GFX7-NEXT: v_cvt_f32_u32_e32 v0, v0
32468
32468
; GFX7-NEXT: v_cvt_f32_u32_e32 v1, v1
32469
- ; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000 , v0
32470
- ; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000 , v1
32469
+ ; GFX7-NEXT: v_and_b32_e32 v0, 0x7fff0000 , v0
32470
+ ; GFX7-NEXT: v_and_b32_e32 v1, 0x7fff0000 , v1
32471
32471
; GFX7-NEXT: s_setpc_b64 s[30:31]
32472
32472
;
32473
32473
; GFX8-LABEL: v_uitofp_v2i16_to_v2bf16:
@@ -32566,9 +32566,9 @@ define <3 x bfloat> @v_uitofp_v3i16_to_v3bf16(<3 x i16> %x) {
32566
32566
; GCN-NEXT: v_cvt_f32_u32_e32 v2, v2
32567
32567
; GCN-NEXT: v_cvt_f32_u32_e32 v1, v1
32568
32568
; GCN-NEXT: v_cvt_f32_u32_e32 v0, v0
32569
- ; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000 , v0
32570
- ; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000 , v1
32571
- ; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000 , v2
32569
+ ; GCN-NEXT: v_and_b32_e32 v0, 0x7fff0000 , v0
32570
+ ; GCN-NEXT: v_and_b32_e32 v1, 0x7fff0000 , v1
32571
+ ; GCN-NEXT: v_and_b32_e32 v2, 0x7fff0000 , v2
32572
32572
; GCN-NEXT: s_setpc_b64 s[30:31]
32573
32573
;
32574
32574
; GFX7-LABEL: v_uitofp_v3i16_to_v3bf16:
@@ -32580,9 +32580,9 @@ define <3 x bfloat> @v_uitofp_v3i16_to_v3bf16(<3 x i16> %x) {
32580
32580
; GFX7-NEXT: v_cvt_f32_u32_e32 v0, v0
32581
32581
; GFX7-NEXT: v_cvt_f32_u32_e32 v1, v1
32582
32582
; GFX7-NEXT: v_cvt_f32_u32_e32 v2, v2
32583
- ; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000 , v0
32584
- ; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000 , v1
32585
- ; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000 , v2
32583
+ ; GFX7-NEXT: v_and_b32_e32 v0, 0x7fff0000 , v0
32584
+ ; GFX7-NEXT: v_and_b32_e32 v1, 0x7fff0000 , v1
32585
+ ; GFX7-NEXT: v_and_b32_e32 v2, 0x7fff0000 , v2
32586
32586
; GFX7-NEXT: s_setpc_b64 s[30:31]
32587
32587
;
32588
32588
; GFX8-LABEL: v_uitofp_v3i16_to_v3bf16:
@@ -32682,10 +32682,10 @@ define <4 x bfloat> @v_uitofp_v4i16_to_v4bf16(<4 x i16> %x) {
32682
32682
; GCN-NEXT: v_cvt_f32_u32_e32 v2, v2
32683
32683
; GCN-NEXT: v_cvt_f32_u32_e32 v1, v1
32684
32684
; GCN-NEXT: v_cvt_f32_u32_e32 v0, v0
32685
- ; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000 , v0
32686
- ; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000 , v1
32687
- ; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000 , v2
32688
- ; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000 , v3
32685
+ ; GCN-NEXT: v_and_b32_e32 v0, 0x7fff0000 , v0
32686
+ ; GCN-NEXT: v_and_b32_e32 v1, 0x7fff0000 , v1
32687
+ ; GCN-NEXT: v_and_b32_e32 v2, 0x7fff0000 , v2
32688
+ ; GCN-NEXT: v_and_b32_e32 v3, 0x7fff0000 , v3
32689
32689
; GCN-NEXT: s_setpc_b64 s[30:31]
32690
32690
;
32691
32691
; GFX7-LABEL: v_uitofp_v4i16_to_v4bf16:
@@ -32699,10 +32699,10 @@ define <4 x bfloat> @v_uitofp_v4i16_to_v4bf16(<4 x i16> %x) {
32699
32699
; GFX7-NEXT: v_cvt_f32_u32_e32 v1, v1
32700
32700
; GFX7-NEXT: v_cvt_f32_u32_e32 v2, v2
32701
32701
; GFX7-NEXT: v_cvt_f32_u32_e32 v3, v3
32702
- ; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000 , v0
32703
- ; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000 , v1
32704
- ; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000 , v2
32705
- ; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000 , v3
32702
+ ; GFX7-NEXT: v_and_b32_e32 v0, 0x7fff0000 , v0
32703
+ ; GFX7-NEXT: v_and_b32_e32 v1, 0x7fff0000 , v1
32704
+ ; GFX7-NEXT: v_and_b32_e32 v2, 0x7fff0000 , v2
32705
+ ; GFX7-NEXT: v_and_b32_e32 v3, 0x7fff0000 , v3
32706
32706
; GFX7-NEXT: s_setpc_b64 s[30:31]
32707
32707
;
32708
32708
; GFX8-LABEL: v_uitofp_v4i16_to_v4bf16:
@@ -32857,14 +32857,14 @@ define bfloat @v_uitofp_i32_to_bf16(i32 %x) {
32857
32857
; GCN: ; %bb.0:
32858
32858
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
32859
32859
; GCN-NEXT: v_cvt_f32_u32_e32 v0, v0
32860
- ; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000 , v0
32860
+ ; GCN-NEXT: v_and_b32_e32 v0, 0x7fff0000 , v0
32861
32861
; GCN-NEXT: s_setpc_b64 s[30:31]
32862
32862
;
32863
32863
; GFX7-LABEL: v_uitofp_i32_to_bf16:
32864
32864
; GFX7: ; %bb.0:
32865
32865
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
32866
32866
; GFX7-NEXT: v_cvt_f32_u32_e32 v0, v0
32867
- ; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000 , v0
32867
+ ; GFX7-NEXT: v_and_b32_e32 v0, 0x7fff0000 , v0
32868
32868
; GFX7-NEXT: s_setpc_b64 s[30:31]
32869
32869
;
32870
32870
; GFX8-LABEL: v_uitofp_i32_to_bf16:
@@ -32928,17 +32928,17 @@ define <2 x bfloat> @v_uitofp_v2i32_to_v2bf16(<2 x i32> %x) {
32928
32928
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
32929
32929
; GCN-NEXT: v_cvt_f32_u32_e32 v1, v1
32930
32930
; GCN-NEXT: v_cvt_f32_u32_e32 v0, v0
32931
- ; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000 , v0
32932
- ; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000 , v1
32931
+ ; GCN-NEXT: v_and_b32_e32 v0, 0x7fff0000 , v0
32932
+ ; GCN-NEXT: v_and_b32_e32 v1, 0x7fff0000 , v1
32933
32933
; GCN-NEXT: s_setpc_b64 s[30:31]
32934
32934
;
32935
32935
; GFX7-LABEL: v_uitofp_v2i32_to_v2bf16:
32936
32936
; GFX7: ; %bb.0:
32937
32937
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
32938
32938
; GFX7-NEXT: v_cvt_f32_u32_e32 v0, v0
32939
32939
; GFX7-NEXT: v_cvt_f32_u32_e32 v1, v1
32940
- ; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000 , v0
32941
- ; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000 , v1
32940
+ ; GFX7-NEXT: v_and_b32_e32 v0, 0x7fff0000 , v0
32941
+ ; GFX7-NEXT: v_and_b32_e32 v1, 0x7fff0000 , v1
32942
32942
; GFX7-NEXT: s_setpc_b64 s[30:31]
32943
32943
;
32944
32944
; GFX8-LABEL: v_uitofp_v2i32_to_v2bf16:
@@ -33031,9 +33031,9 @@ define <3 x bfloat> @v_uitofp_v3i32_to_v3bf16(<3 x i32> %x) {
33031
33031
; GCN-NEXT: v_cvt_f32_u32_e32 v2, v2
33032
33032
; GCN-NEXT: v_cvt_f32_u32_e32 v1, v1
33033
33033
; GCN-NEXT: v_cvt_f32_u32_e32 v0, v0
33034
- ; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000 , v0
33035
- ; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000 , v1
33036
- ; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000 , v2
33034
+ ; GCN-NEXT: v_and_b32_e32 v0, 0x7fff0000 , v0
33035
+ ; GCN-NEXT: v_and_b32_e32 v1, 0x7fff0000 , v1
33036
+ ; GCN-NEXT: v_and_b32_e32 v2, 0x7fff0000 , v2
33037
33037
; GCN-NEXT: s_setpc_b64 s[30:31]
33038
33038
;
33039
33039
; GFX7-LABEL: v_uitofp_v3i32_to_v3bf16:
@@ -33042,9 +33042,9 @@ define <3 x bfloat> @v_uitofp_v3i32_to_v3bf16(<3 x i32> %x) {
33042
33042
; GFX7-NEXT: v_cvt_f32_u32_e32 v0, v0
33043
33043
; GFX7-NEXT: v_cvt_f32_u32_e32 v1, v1
33044
33044
; GFX7-NEXT: v_cvt_f32_u32_e32 v2, v2
33045
- ; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000 , v0
33046
- ; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000 , v1
33047
- ; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000 , v2
33045
+ ; GFX7-NEXT: v_and_b32_e32 v0, 0x7fff0000 , v0
33046
+ ; GFX7-NEXT: v_and_b32_e32 v1, 0x7fff0000 , v1
33047
+ ; GFX7-NEXT: v_and_b32_e32 v2, 0x7fff0000 , v2
33048
33048
; GFX7-NEXT: s_setpc_b64 s[30:31]
33049
33049
;
33050
33050
; GFX8-LABEL: v_uitofp_v3i32_to_v3bf16:
@@ -33140,10 +33140,10 @@ define <4 x bfloat> @v_uitofp_v4i32_to_v4bf16(<4 x i32> %x) {
33140
33140
; GCN-NEXT: v_cvt_f32_u32_e32 v2, v2
33141
33141
; GCN-NEXT: v_cvt_f32_u32_e32 v1, v1
33142
33142
; GCN-NEXT: v_cvt_f32_u32_e32 v0, v0
33143
- ; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000 , v0
33144
- ; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000 , v1
33145
- ; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000 , v2
33146
- ; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000 , v3
33143
+ ; GCN-NEXT: v_and_b32_e32 v0, 0x7fff0000 , v0
33144
+ ; GCN-NEXT: v_and_b32_e32 v1, 0x7fff0000 , v1
33145
+ ; GCN-NEXT: v_and_b32_e32 v2, 0x7fff0000 , v2
33146
+ ; GCN-NEXT: v_and_b32_e32 v3, 0x7fff0000 , v3
33147
33147
; GCN-NEXT: s_setpc_b64 s[30:31]
33148
33148
;
33149
33149
; GFX7-LABEL: v_uitofp_v4i32_to_v4bf16:
@@ -33153,10 +33153,10 @@ define <4 x bfloat> @v_uitofp_v4i32_to_v4bf16(<4 x i32> %x) {
33153
33153
; GFX7-NEXT: v_cvt_f32_u32_e32 v1, v1
33154
33154
; GFX7-NEXT: v_cvt_f32_u32_e32 v2, v2
33155
33155
; GFX7-NEXT: v_cvt_f32_u32_e32 v3, v3
33156
- ; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000 , v0
33157
- ; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000 , v1
33158
- ; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000 , v2
33159
- ; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000 , v3
33156
+ ; GFX7-NEXT: v_and_b32_e32 v0, 0x7fff0000 , v0
33157
+ ; GFX7-NEXT: v_and_b32_e32 v1, 0x7fff0000 , v1
33158
+ ; GFX7-NEXT: v_and_b32_e32 v2, 0x7fff0000 , v2
33159
+ ; GFX7-NEXT: v_and_b32_e32 v3, 0x7fff0000 , v3
33160
33160
; GFX7-NEXT: s_setpc_b64 s[30:31]
33161
33161
;
33162
33162
; GFX8-LABEL: v_uitofp_v4i32_to_v4bf16:
0 commit comments