[llvm] [NVPTX] Vectorize and lower 256-bit global loads/stores for sm_100+/ptx88+ (PR #139292)
Artem Belevich via llvm-commits
llvm-commits at lists.llvm.org
Fri May 9 11:56:59 PDT 2025
================
@@ -0,0 +1,728 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -mtriple=nvptx64-nvidia-cuda -passes=load-store-vectorizer -mcpu=sm_90 -mattr=+ptx87 -S < %s | FileCheck %s -check-prefixes=CHECK,SM90
+; RUN: opt -mtriple=nvptx64-nvidia-cuda -passes=load-store-vectorizer -mcpu=sm_100 -mattr=+ptx88 -S < %s | FileCheck %s -check-prefixes=CHECK,SM100
+
+; 256 bit loads/stores are only currently supported for:
+; - global
+; - blackwell (sm_100)
+; - ptx 8.8
+; - 32/64-bit types
+
+; Currently, the LSV produces 256 bit loads/stores if the first three conditions
+; are satisfied, as the backend will either upsize or split vectors
+; of smaller elements in the Type Legalization stage.
+
+; In this file, we test i8, i16, i32, i64, f32, f64.
+; The other floating point types are omitted for simplicity.
+; We also test the negative case for non-global i32.
+
+define void @int8x32(ptr addrspace(1) %ptr) {
+; SM90-LABEL: define void @int8x32(
+; SM90-SAME: ptr addrspace(1) [[PTR:%.*]]) #[[ATTR0:[0-9]+]] {
+; SM90-NEXT: [[PTR0:%.*]] = getelementptr i8, ptr addrspace(1) [[PTR]], i64 0
+; SM90-NEXT: [[PTR10:%.*]] = getelementptr i8, ptr addrspace(1) [[PTR]], i64 16
+; SM90-NEXT: [[TMP1:%.*]] = load <16 x i8>, ptr addrspace(1) [[PTR0]], align 32
+; SM90-NEXT: [[L01:%.*]] = extractelement <16 x i8> [[TMP1]], i32 0
+; SM90-NEXT: [[L110:%.*]] = extractelement <16 x i8> [[TMP1]], i32 1
+; SM90-NEXT: [[L211:%.*]] = extractelement <16 x i8> [[TMP1]], i32 2
+; SM90-NEXT: [[L312:%.*]] = extractelement <16 x i8> [[TMP1]], i32 3
+; SM90-NEXT: [[L413:%.*]] = extractelement <16 x i8> [[TMP1]], i32 4
+; SM90-NEXT: [[L514:%.*]] = extractelement <16 x i8> [[TMP1]], i32 5
+; SM90-NEXT: [[L615:%.*]] = extractelement <16 x i8> [[TMP1]], i32 6
+; SM90-NEXT: [[L716:%.*]] = extractelement <16 x i8> [[TMP1]], i32 7
+; SM90-NEXT: [[L817:%.*]] = extractelement <16 x i8> [[TMP1]], i32 8
+; SM90-NEXT: [[L918:%.*]] = extractelement <16 x i8> [[TMP1]], i32 9
+; SM90-NEXT: [[LA19:%.*]] = extractelement <16 x i8> [[TMP1]], i32 10
+; SM90-NEXT: [[LB20:%.*]] = extractelement <16 x i8> [[TMP1]], i32 11
+; SM90-NEXT: [[LC21:%.*]] = extractelement <16 x i8> [[TMP1]], i32 12
+; SM90-NEXT: [[LD22:%.*]] = extractelement <16 x i8> [[TMP1]], i32 13
+; SM90-NEXT: [[LE23:%.*]] = extractelement <16 x i8> [[TMP1]], i32 14
+; SM90-NEXT: [[LF24:%.*]] = extractelement <16 x i8> [[TMP1]], i32 15
+; SM90-NEXT: [[TMP2:%.*]] = load <16 x i8>, ptr addrspace(1) [[PTR10]], align 16
+; SM90-NEXT: [[L1025:%.*]] = extractelement <16 x i8> [[TMP2]], i32 0
+; SM90-NEXT: [[L1126:%.*]] = extractelement <16 x i8> [[TMP2]], i32 1
+; SM90-NEXT: [[L1227:%.*]] = extractelement <16 x i8> [[TMP2]], i32 2
+; SM90-NEXT: [[L1328:%.*]] = extractelement <16 x i8> [[TMP2]], i32 3
+; SM90-NEXT: [[L1429:%.*]] = extractelement <16 x i8> [[TMP2]], i32 4
+; SM90-NEXT: [[L1530:%.*]] = extractelement <16 x i8> [[TMP2]], i32 5
+; SM90-NEXT: [[L1631:%.*]] = extractelement <16 x i8> [[TMP2]], i32 6
+; SM90-NEXT: [[L1732:%.*]] = extractelement <16 x i8> [[TMP2]], i32 7
+; SM90-NEXT: [[L1833:%.*]] = extractelement <16 x i8> [[TMP2]], i32 8
+; SM90-NEXT: [[L1934:%.*]] = extractelement <16 x i8> [[TMP2]], i32 9
+; SM90-NEXT: [[L1A35:%.*]] = extractelement <16 x i8> [[TMP2]], i32 10
+; SM90-NEXT: [[L1B36:%.*]] = extractelement <16 x i8> [[TMP2]], i32 11
+; SM90-NEXT: [[L1C37:%.*]] = extractelement <16 x i8> [[TMP2]], i32 12
+; SM90-NEXT: [[L1D38:%.*]] = extractelement <16 x i8> [[TMP2]], i32 13
+; SM90-NEXT: [[L1E39:%.*]] = extractelement <16 x i8> [[TMP2]], i32 14
+; SM90-NEXT: [[L1F40:%.*]] = extractelement <16 x i8> [[TMP2]], i32 15
+; SM90-NEXT: [[TMP3:%.*]] = insertelement <16 x i8> poison, i8 [[LB20]], i32 0
+; SM90-NEXT: [[TMP4:%.*]] = insertelement <16 x i8> [[TMP3]], i8 [[LA19]], i32 1
+; SM90-NEXT: [[TMP5:%.*]] = insertelement <16 x i8> [[TMP4]], i8 [[L918]], i32 2
+; SM90-NEXT: [[TMP6:%.*]] = insertelement <16 x i8> [[TMP5]], i8 [[L817]], i32 3
+; SM90-NEXT: [[TMP7:%.*]] = insertelement <16 x i8> [[TMP6]], i8 [[L716]], i32 4
+; SM90-NEXT: [[TMP8:%.*]] = insertelement <16 x i8> [[TMP7]], i8 [[L615]], i32 5
+; SM90-NEXT: [[TMP9:%.*]] = insertelement <16 x i8> [[TMP8]], i8 [[L514]], i32 6
+; SM90-NEXT: [[TMP10:%.*]] = insertelement <16 x i8> [[TMP9]], i8 [[L413]], i32 7
+; SM90-NEXT: [[TMP11:%.*]] = insertelement <16 x i8> [[TMP10]], i8 [[L312]], i32 8
+; SM90-NEXT: [[TMP12:%.*]] = insertelement <16 x i8> [[TMP11]], i8 [[L211]], i32 9
+; SM90-NEXT: [[TMP13:%.*]] = insertelement <16 x i8> [[TMP12]], i8 [[L110]], i32 10
+; SM90-NEXT: [[TMP14:%.*]] = insertelement <16 x i8> [[TMP13]], i8 [[L01]], i32 11
+; SM90-NEXT: [[TMP15:%.*]] = insertelement <16 x i8> [[TMP14]], i8 [[LF24]], i32 12
+; SM90-NEXT: [[TMP16:%.*]] = insertelement <16 x i8> [[TMP15]], i8 [[LE23]], i32 13
+; SM90-NEXT: [[TMP17:%.*]] = insertelement <16 x i8> [[TMP16]], i8 [[LD22]], i32 14
+; SM90-NEXT: [[TMP18:%.*]] = insertelement <16 x i8> [[TMP17]], i8 [[LC21]], i32 15
+; SM90-NEXT: store <16 x i8> [[TMP18]], ptr addrspace(1) [[PTR0]], align 32
+; SM90-NEXT: [[TMP19:%.*]] = insertelement <16 x i8> poison, i8 [[L1B36]], i32 0
+; SM90-NEXT: [[TMP20:%.*]] = insertelement <16 x i8> [[TMP19]], i8 [[L1A35]], i32 1
+; SM90-NEXT: [[TMP21:%.*]] = insertelement <16 x i8> [[TMP20]], i8 [[L1934]], i32 2
+; SM90-NEXT: [[TMP22:%.*]] = insertelement <16 x i8> [[TMP21]], i8 [[L1833]], i32 3
+; SM90-NEXT: [[TMP23:%.*]] = insertelement <16 x i8> [[TMP22]], i8 [[L1732]], i32 4
+; SM90-NEXT: [[TMP24:%.*]] = insertelement <16 x i8> [[TMP23]], i8 [[L1631]], i32 5
+; SM90-NEXT: [[TMP25:%.*]] = insertelement <16 x i8> [[TMP24]], i8 [[L1530]], i32 6
+; SM90-NEXT: [[TMP26:%.*]] = insertelement <16 x i8> [[TMP25]], i8 [[L1429]], i32 7
+; SM90-NEXT: [[TMP27:%.*]] = insertelement <16 x i8> [[TMP26]], i8 [[L1328]], i32 8
+; SM90-NEXT: [[TMP28:%.*]] = insertelement <16 x i8> [[TMP27]], i8 [[L1227]], i32 9
+; SM90-NEXT: [[TMP29:%.*]] = insertelement <16 x i8> [[TMP28]], i8 [[L1126]], i32 10
+; SM90-NEXT: [[TMP30:%.*]] = insertelement <16 x i8> [[TMP29]], i8 [[L1025]], i32 11
+; SM90-NEXT: [[TMP31:%.*]] = insertelement <16 x i8> [[TMP30]], i8 [[L1F40]], i32 12
+; SM90-NEXT: [[TMP32:%.*]] = insertelement <16 x i8> [[TMP31]], i8 [[L1E39]], i32 13
+; SM90-NEXT: [[TMP33:%.*]] = insertelement <16 x i8> [[TMP32]], i8 [[L1D38]], i32 14
+; SM90-NEXT: [[TMP34:%.*]] = insertelement <16 x i8> [[TMP33]], i8 [[L1C37]], i32 15
+; SM90-NEXT: store <16 x i8> [[TMP34]], ptr addrspace(1) [[PTR10]], align 16
+; SM90-NEXT: ret void
+;
+; SM100-LABEL: define void @int8x32(
+; SM100-SAME: ptr addrspace(1) [[PTR:%.*]]) #[[ATTR0:[0-9]+]] {
+; SM100-NEXT: [[PTR0:%.*]] = getelementptr i8, ptr addrspace(1) [[PTR]], i64 0
+; SM100-NEXT: [[TMP1:%.*]] = load <32 x i8>, ptr addrspace(1) [[PTR0]], align 32
+; SM100-NEXT: [[L01:%.*]] = extractelement <32 x i8> [[TMP1]], i32 0
+; SM100-NEXT: [[L110:%.*]] = extractelement <32 x i8> [[TMP1]], i32 1
+; SM100-NEXT: [[L211:%.*]] = extractelement <32 x i8> [[TMP1]], i32 2
+; SM100-NEXT: [[L312:%.*]] = extractelement <32 x i8> [[TMP1]], i32 3
+; SM100-NEXT: [[L413:%.*]] = extractelement <32 x i8> [[TMP1]], i32 4
+; SM100-NEXT: [[L514:%.*]] = extractelement <32 x i8> [[TMP1]], i32 5
+; SM100-NEXT: [[L615:%.*]] = extractelement <32 x i8> [[TMP1]], i32 6
+; SM100-NEXT: [[L716:%.*]] = extractelement <32 x i8> [[TMP1]], i32 7
+; SM100-NEXT: [[L817:%.*]] = extractelement <32 x i8> [[TMP1]], i32 8
+; SM100-NEXT: [[L918:%.*]] = extractelement <32 x i8> [[TMP1]], i32 9
+; SM100-NEXT: [[LA19:%.*]] = extractelement <32 x i8> [[TMP1]], i32 10
+; SM100-NEXT: [[LB20:%.*]] = extractelement <32 x i8> [[TMP1]], i32 11
+; SM100-NEXT: [[LC21:%.*]] = extractelement <32 x i8> [[TMP1]], i32 12
+; SM100-NEXT: [[LD22:%.*]] = extractelement <32 x i8> [[TMP1]], i32 13
+; SM100-NEXT: [[LE23:%.*]] = extractelement <32 x i8> [[TMP1]], i32 14
+; SM100-NEXT: [[LF24:%.*]] = extractelement <32 x i8> [[TMP1]], i32 15
+; SM100-NEXT: [[L1025:%.*]] = extractelement <32 x i8> [[TMP1]], i32 16
+; SM100-NEXT: [[L1126:%.*]] = extractelement <32 x i8> [[TMP1]], i32 17
+; SM100-NEXT: [[L1227:%.*]] = extractelement <32 x i8> [[TMP1]], i32 18
+; SM100-NEXT: [[L1328:%.*]] = extractelement <32 x i8> [[TMP1]], i32 19
+; SM100-NEXT: [[L1429:%.*]] = extractelement <32 x i8> [[TMP1]], i32 20
+; SM100-NEXT: [[L1530:%.*]] = extractelement <32 x i8> [[TMP1]], i32 21
+; SM100-NEXT: [[L1631:%.*]] = extractelement <32 x i8> [[TMP1]], i32 22
+; SM100-NEXT: [[L1732:%.*]] = extractelement <32 x i8> [[TMP1]], i32 23
+; SM100-NEXT: [[L1833:%.*]] = extractelement <32 x i8> [[TMP1]], i32 24
+; SM100-NEXT: [[L1934:%.*]] = extractelement <32 x i8> [[TMP1]], i32 25
+; SM100-NEXT: [[L1A35:%.*]] = extractelement <32 x i8> [[TMP1]], i32 26
+; SM100-NEXT: [[L1B36:%.*]] = extractelement <32 x i8> [[TMP1]], i32 27
+; SM100-NEXT: [[L1C37:%.*]] = extractelement <32 x i8> [[TMP1]], i32 28
+; SM100-NEXT: [[L1D38:%.*]] = extractelement <32 x i8> [[TMP1]], i32 29
+; SM100-NEXT: [[L1E39:%.*]] = extractelement <32 x i8> [[TMP1]], i32 30
+; SM100-NEXT: [[L1F40:%.*]] = extractelement <32 x i8> [[TMP1]], i32 31
+; SM100-NEXT: [[TMP2:%.*]] = insertelement <32 x i8> poison, i8 [[LB20]], i32 0
+; SM100-NEXT: [[TMP3:%.*]] = insertelement <32 x i8> [[TMP2]], i8 [[LA19]], i32 1
+; SM100-NEXT: [[TMP4:%.*]] = insertelement <32 x i8> [[TMP3]], i8 [[L918]], i32 2
+; SM100-NEXT: [[TMP5:%.*]] = insertelement <32 x i8> [[TMP4]], i8 [[L817]], i32 3
+; SM100-NEXT: [[TMP6:%.*]] = insertelement <32 x i8> [[TMP5]], i8 [[L716]], i32 4
+; SM100-NEXT: [[TMP7:%.*]] = insertelement <32 x i8> [[TMP6]], i8 [[L615]], i32 5
+; SM100-NEXT: [[TMP8:%.*]] = insertelement <32 x i8> [[TMP7]], i8 [[L514]], i32 6
+; SM100-NEXT: [[TMP9:%.*]] = insertelement <32 x i8> [[TMP8]], i8 [[L413]], i32 7
+; SM100-NEXT: [[TMP10:%.*]] = insertelement <32 x i8> [[TMP9]], i8 [[L312]], i32 8
+; SM100-NEXT: [[TMP11:%.*]] = insertelement <32 x i8> [[TMP10]], i8 [[L211]], i32 9
+; SM100-NEXT: [[TMP12:%.*]] = insertelement <32 x i8> [[TMP11]], i8 [[L110]], i32 10
+; SM100-NEXT: [[TMP13:%.*]] = insertelement <32 x i8> [[TMP12]], i8 [[L01]], i32 11
+; SM100-NEXT: [[TMP14:%.*]] = insertelement <32 x i8> [[TMP13]], i8 [[LF24]], i32 12
+; SM100-NEXT: [[TMP15:%.*]] = insertelement <32 x i8> [[TMP14]], i8 [[LE23]], i32 13
+; SM100-NEXT: [[TMP16:%.*]] = insertelement <32 x i8> [[TMP15]], i8 [[LD22]], i32 14
+; SM100-NEXT: [[TMP17:%.*]] = insertelement <32 x i8> [[TMP16]], i8 [[LC21]], i32 15
+; SM100-NEXT: [[TMP18:%.*]] = insertelement <32 x i8> [[TMP17]], i8 [[L1B36]], i32 16
+; SM100-NEXT: [[TMP19:%.*]] = insertelement <32 x i8> [[TMP18]], i8 [[L1A35]], i32 17
+; SM100-NEXT: [[TMP20:%.*]] = insertelement <32 x i8> [[TMP19]], i8 [[L1934]], i32 18
+; SM100-NEXT: [[TMP21:%.*]] = insertelement <32 x i8> [[TMP20]], i8 [[L1833]], i32 19
+; SM100-NEXT: [[TMP22:%.*]] = insertelement <32 x i8> [[TMP21]], i8 [[L1732]], i32 20
+; SM100-NEXT: [[TMP23:%.*]] = insertelement <32 x i8> [[TMP22]], i8 [[L1631]], i32 21
+; SM100-NEXT: [[TMP24:%.*]] = insertelement <32 x i8> [[TMP23]], i8 [[L1530]], i32 22
+; SM100-NEXT: [[TMP25:%.*]] = insertelement <32 x i8> [[TMP24]], i8 [[L1429]], i32 23
+; SM100-NEXT: [[TMP26:%.*]] = insertelement <32 x i8> [[TMP25]], i8 [[L1328]], i32 24
+; SM100-NEXT: [[TMP27:%.*]] = insertelement <32 x i8> [[TMP26]], i8 [[L1227]], i32 25
+; SM100-NEXT: [[TMP28:%.*]] = insertelement <32 x i8> [[TMP27]], i8 [[L1126]], i32 26
+; SM100-NEXT: [[TMP29:%.*]] = insertelement <32 x i8> [[TMP28]], i8 [[L1025]], i32 27
+; SM100-NEXT: [[TMP30:%.*]] = insertelement <32 x i8> [[TMP29]], i8 [[L1F40]], i32 28
+; SM100-NEXT: [[TMP31:%.*]] = insertelement <32 x i8> [[TMP30]], i8 [[L1E39]], i32 29
+; SM100-NEXT: [[TMP32:%.*]] = insertelement <32 x i8> [[TMP31]], i8 [[L1D38]], i32 30
+; SM100-NEXT: [[TMP33:%.*]] = insertelement <32 x i8> [[TMP32]], i8 [[L1C37]], i32 31
+; SM100-NEXT: store <32 x i8> [[TMP33]], ptr addrspace(1) [[PTR0]], align 32
+; SM100-NEXT: ret void
+;
+
+ %ptr0 = getelementptr i8, ptr addrspace(1) %ptr, i64 0
+ %ptr1 = getelementptr i8, ptr addrspace(1) %ptr, i64 1
+ %ptr2 = getelementptr i8, ptr addrspace(1) %ptr, i64 2
+ %ptr3 = getelementptr i8, ptr addrspace(1) %ptr, i64 3
+ %ptr4 = getelementptr i8, ptr addrspace(1) %ptr, i64 4
+ %ptr5 = getelementptr i8, ptr addrspace(1) %ptr, i64 5
+ %ptr6 = getelementptr i8, ptr addrspace(1) %ptr, i64 6
+ %ptr7 = getelementptr i8, ptr addrspace(1) %ptr, i64 7
+ %ptr8 = getelementptr i8, ptr addrspace(1) %ptr, i64 8
+ %ptr9 = getelementptr i8, ptr addrspace(1) %ptr, i64 9
+ %ptra = getelementptr i8, ptr addrspace(1) %ptr, i64 10
+ %ptrb = getelementptr i8, ptr addrspace(1) %ptr, i64 11
+ %ptrc = getelementptr i8, ptr addrspace(1) %ptr, i64 12
+ %ptrd = getelementptr i8, ptr addrspace(1) %ptr, i64 13
+ %ptre = getelementptr i8, ptr addrspace(1) %ptr, i64 14
+ %ptrf = getelementptr i8, ptr addrspace(1) %ptr, i64 15
+ %ptr10 = getelementptr i8, ptr addrspace(1) %ptr, i64 16
+ %ptr11 = getelementptr i8, ptr addrspace(1) %ptr, i64 17
+ %ptr12 = getelementptr i8, ptr addrspace(1) %ptr, i64 18
+ %ptr13 = getelementptr i8, ptr addrspace(1) %ptr, i64 19
+ %ptr14 = getelementptr i8, ptr addrspace(1) %ptr, i64 20
+ %ptr15 = getelementptr i8, ptr addrspace(1) %ptr, i64 21
+ %ptr16 = getelementptr i8, ptr addrspace(1) %ptr, i64 22
+ %ptr17 = getelementptr i8, ptr addrspace(1) %ptr, i64 23
+ %ptr18 = getelementptr i8, ptr addrspace(1) %ptr, i64 24
+ %ptr19 = getelementptr i8, ptr addrspace(1) %ptr, i64 25
+ %ptr1a = getelementptr i8, ptr addrspace(1) %ptr, i64 26
+ %ptr1b = getelementptr i8, ptr addrspace(1) %ptr, i64 27
+ %ptr1c = getelementptr i8, ptr addrspace(1) %ptr, i64 28
+ %ptr1d = getelementptr i8, ptr addrspace(1) %ptr, i64 29
+ %ptr1e = getelementptr i8, ptr addrspace(1) %ptr, i64 30
+ %ptr1f = getelementptr i8, ptr addrspace(1) %ptr, i64 31
+
+ %l0 = load i8, ptr addrspace(1) %ptr0, align 32
+ %l1 = load i8, ptr addrspace(1) %ptr1, align 1
+ %l2 = load i8, ptr addrspace(1) %ptr2, align 1
+ %l3 = load i8, ptr addrspace(1) %ptr3, align 1
+ %l4 = load i8, ptr addrspace(1) %ptr4, align 1
+ %l5 = load i8, ptr addrspace(1) %ptr5, align 1
+ %l6 = load i8, ptr addrspace(1) %ptr6, align 1
+ %l7 = load i8, ptr addrspace(1) %ptr7, align 1
+ %l8 = load i8, ptr addrspace(1) %ptr8, align 1
+ %l9 = load i8, ptr addrspace(1) %ptr9, align 1
+ %la = load i8, ptr addrspace(1) %ptra, align 1
+ %lb = load i8, ptr addrspace(1) %ptrb, align 1
+ %lc = load i8, ptr addrspace(1) %ptrc, align 1
+ %ld = load i8, ptr addrspace(1) %ptrd, align 1
+ %le = load i8, ptr addrspace(1) %ptre, align 1
+ %lf = load i8, ptr addrspace(1) %ptrf, align 1
+ %l10 = load i8, ptr addrspace(1) %ptr10, align 16
+ %l11 = load i8, ptr addrspace(1) %ptr11, align 1
+ %l12 = load i8, ptr addrspace(1) %ptr12, align 1
+ %l13 = load i8, ptr addrspace(1) %ptr13, align 1
+ %l14 = load i8, ptr addrspace(1) %ptr14, align 1
+ %l15 = load i8, ptr addrspace(1) %ptr15, align 1
+ %l16 = load i8, ptr addrspace(1) %ptr16, align 1
+ %l17 = load i8, ptr addrspace(1) %ptr17, align 1
+ %l18 = load i8, ptr addrspace(1) %ptr18, align 1
+ %l19 = load i8, ptr addrspace(1) %ptr19, align 1
+ %l1a = load i8, ptr addrspace(1) %ptr1a, align 1
+ %l1b = load i8, ptr addrspace(1) %ptr1b, align 1
+ %l1c = load i8, ptr addrspace(1) %ptr1c, align 1
+ %l1d = load i8, ptr addrspace(1) %ptr1d, align 1
+ %l1e = load i8, ptr addrspace(1) %ptr1e, align 1
+ %l1f = load i8, ptr addrspace(1) %ptr1f, align 1
+
+ store i8 %lf, ptr addrspace(1) %ptrc, align 1
+ store i8 %le, ptr addrspace(1) %ptrd, align 1
+ store i8 %ld, ptr addrspace(1) %ptre, align 1
+ store i8 %lc, ptr addrspace(1) %ptrf, align 1
+ store i8 %lb, ptr addrspace(1) %ptr0, align 32
+ store i8 %la, ptr addrspace(1) %ptr1, align 1
+ store i8 %l9, ptr addrspace(1) %ptr2, align 1
+ store i8 %l8, ptr addrspace(1) %ptr3, align 1
+ store i8 %l7, ptr addrspace(1) %ptr4, align 1
+ store i8 %l6, ptr addrspace(1) %ptr5, align 1
+ store i8 %l5, ptr addrspace(1) %ptr6, align 1
+ store i8 %l4, ptr addrspace(1) %ptr7, align 1
+ store i8 %l3, ptr addrspace(1) %ptr8, align 1
+ store i8 %l2, ptr addrspace(1) %ptr9, align 1
+ store i8 %l1, ptr addrspace(1) %ptra, align 1
+ store i8 %l0, ptr addrspace(1) %ptrb, align 1
+ store i8 %l1f, ptr addrspace(1) %ptr1c, align 1
+ store i8 %l1e, ptr addrspace(1) %ptr1d, align 1
+ store i8 %l1d, ptr addrspace(1) %ptr1e, align 1
+ store i8 %l1c, ptr addrspace(1) %ptr1f, align 1
+ store i8 %l1b, ptr addrspace(1) %ptr10, align 16
+ store i8 %l1a, ptr addrspace(1) %ptr11, align 1
+ store i8 %l19, ptr addrspace(1) %ptr12, align 1
+ store i8 %l18, ptr addrspace(1) %ptr13, align 1
+ store i8 %l17, ptr addrspace(1) %ptr14, align 1
+ store i8 %l16, ptr addrspace(1) %ptr15, align 1
+ store i8 %l15, ptr addrspace(1) %ptr16, align 1
+ store i8 %l14, ptr addrspace(1) %ptr17, align 1
+ store i8 %l13, ptr addrspace(1) %ptr18, align 1
+ store i8 %l12, ptr addrspace(1) %ptr19, align 1
+ store i8 %l11, ptr addrspace(1) %ptr1a, align 1
+ store i8 %l10, ptr addrspace(1) %ptr1b, align 1
+
+ ret void
+}
+
+define void @int16x16(ptr addrspace(1) %ptr) {
+; SM90-LABEL: define void @int16x16(
+; SM90-SAME: ptr addrspace(1) [[PTR:%.*]]) #[[ATTR0]] {
+; SM90-NEXT: [[PTR0:%.*]] = getelementptr i16, ptr addrspace(1) [[PTR]], i64 0
+; SM90-NEXT: [[PTR8:%.*]] = getelementptr i16, ptr addrspace(1) [[PTR]], i64 8
+; SM90-NEXT: [[TMP1:%.*]] = load <8 x i16>, ptr addrspace(1) [[PTR0]], align 32
+; SM90-NEXT: [[L01:%.*]] = extractelement <8 x i16> [[TMP1]], i32 0
+; SM90-NEXT: [[L12:%.*]] = extractelement <8 x i16> [[TMP1]], i32 1
+; SM90-NEXT: [[L23:%.*]] = extractelement <8 x i16> [[TMP1]], i32 2
+; SM90-NEXT: [[L34:%.*]] = extractelement <8 x i16> [[TMP1]], i32 3
+; SM90-NEXT: [[L45:%.*]] = extractelement <8 x i16> [[TMP1]], i32 4
+; SM90-NEXT: [[L56:%.*]] = extractelement <8 x i16> [[TMP1]], i32 5
+; SM90-NEXT: [[L67:%.*]] = extractelement <8 x i16> [[TMP1]], i32 6
+; SM90-NEXT: [[L78:%.*]] = extractelement <8 x i16> [[TMP1]], i32 7
+; SM90-NEXT: [[TMP2:%.*]] = load <8 x i16>, ptr addrspace(1) [[PTR8]], align 16
+; SM90-NEXT: [[L89:%.*]] = extractelement <8 x i16> [[TMP2]], i32 0
+; SM90-NEXT: [[L910:%.*]] = extractelement <8 x i16> [[TMP2]], i32 1
+; SM90-NEXT: [[LA11:%.*]] = extractelement <8 x i16> [[TMP2]], i32 2
+; SM90-NEXT: [[LB12:%.*]] = extractelement <8 x i16> [[TMP2]], i32 3
+; SM90-NEXT: [[LC13:%.*]] = extractelement <8 x i16> [[TMP2]], i32 4
+; SM90-NEXT: [[LD14:%.*]] = extractelement <8 x i16> [[TMP2]], i32 5
+; SM90-NEXT: [[LE15:%.*]] = extractelement <8 x i16> [[TMP2]], i32 6
+; SM90-NEXT: [[LF16:%.*]] = extractelement <8 x i16> [[TMP2]], i32 7
+; SM90-NEXT: [[TMP3:%.*]] = insertelement <8 x i16> poison, i16 [[LB12]], i32 0
+; SM90-NEXT: [[TMP4:%.*]] = insertelement <8 x i16> [[TMP3]], i16 [[LA11]], i32 1
+; SM90-NEXT: [[TMP5:%.*]] = insertelement <8 x i16> [[TMP4]], i16 [[L910]], i32 2
+; SM90-NEXT: [[TMP6:%.*]] = insertelement <8 x i16> [[TMP5]], i16 [[L89]], i32 3
+; SM90-NEXT: [[TMP7:%.*]] = insertelement <8 x i16> [[TMP6]], i16 [[L78]], i32 4
+; SM90-NEXT: [[TMP8:%.*]] = insertelement <8 x i16> [[TMP7]], i16 [[L67]], i32 5
+; SM90-NEXT: [[TMP9:%.*]] = insertelement <8 x i16> [[TMP8]], i16 [[L56]], i32 6
+; SM90-NEXT: [[TMP10:%.*]] = insertelement <8 x i16> [[TMP9]], i16 [[L45]], i32 7
+; SM90-NEXT: store <8 x i16> [[TMP10]], ptr addrspace(1) [[PTR0]], align 32
+; SM90-NEXT: [[TMP11:%.*]] = insertelement <8 x i16> poison, i16 [[L34]], i32 0
+; SM90-NEXT: [[TMP12:%.*]] = insertelement <8 x i16> [[TMP11]], i16 [[L23]], i32 1
+; SM90-NEXT: [[TMP13:%.*]] = insertelement <8 x i16> [[TMP12]], i16 [[L12]], i32 2
+; SM90-NEXT: [[TMP14:%.*]] = insertelement <8 x i16> [[TMP13]], i16 [[L01]], i32 3
+; SM90-NEXT: [[TMP15:%.*]] = insertelement <8 x i16> [[TMP14]], i16 [[LF16]], i32 4
+; SM90-NEXT: [[TMP16:%.*]] = insertelement <8 x i16> [[TMP15]], i16 [[LE15]], i32 5
+; SM90-NEXT: [[TMP17:%.*]] = insertelement <8 x i16> [[TMP16]], i16 [[LD14]], i32 6
+; SM90-NEXT: [[TMP18:%.*]] = insertelement <8 x i16> [[TMP17]], i16 [[LC13]], i32 7
+; SM90-NEXT: store <8 x i16> [[TMP18]], ptr addrspace(1) [[PTR8]], align 16
+; SM90-NEXT: ret void
+;
+; SM100-LABEL: define void @int16x16(
+; SM100-SAME: ptr addrspace(1) [[PTR:%.*]]) #[[ATTR0]] {
+; SM100-NEXT: [[PTR0:%.*]] = getelementptr i16, ptr addrspace(1) [[PTR]], i64 0
+; SM100-NEXT: [[TMP1:%.*]] = load <16 x i16>, ptr addrspace(1) [[PTR0]], align 32
+; SM100-NEXT: [[L01:%.*]] = extractelement <16 x i16> [[TMP1]], i32 0
+; SM100-NEXT: [[L12:%.*]] = extractelement <16 x i16> [[TMP1]], i32 1
+; SM100-NEXT: [[L23:%.*]] = extractelement <16 x i16> [[TMP1]], i32 2
+; SM100-NEXT: [[L34:%.*]] = extractelement <16 x i16> [[TMP1]], i32 3
+; SM100-NEXT: [[L45:%.*]] = extractelement <16 x i16> [[TMP1]], i32 4
+; SM100-NEXT: [[L56:%.*]] = extractelement <16 x i16> [[TMP1]], i32 5
+; SM100-NEXT: [[L67:%.*]] = extractelement <16 x i16> [[TMP1]], i32 6
+; SM100-NEXT: [[L78:%.*]] = extractelement <16 x i16> [[TMP1]], i32 7
+; SM100-NEXT: [[L89:%.*]] = extractelement <16 x i16> [[TMP1]], i32 8
+; SM100-NEXT: [[L910:%.*]] = extractelement <16 x i16> [[TMP1]], i32 9
+; SM100-NEXT: [[LA11:%.*]] = extractelement <16 x i16> [[TMP1]], i32 10
+; SM100-NEXT: [[LB12:%.*]] = extractelement <16 x i16> [[TMP1]], i32 11
+; SM100-NEXT: [[LC13:%.*]] = extractelement <16 x i16> [[TMP1]], i32 12
+; SM100-NEXT: [[LD14:%.*]] = extractelement <16 x i16> [[TMP1]], i32 13
+; SM100-NEXT: [[LE15:%.*]] = extractelement <16 x i16> [[TMP1]], i32 14
+; SM100-NEXT: [[LF16:%.*]] = extractelement <16 x i16> [[TMP1]], i32 15
+; SM100-NEXT: [[TMP2:%.*]] = insertelement <16 x i16> poison, i16 [[LB12]], i32 0
+; SM100-NEXT: [[TMP3:%.*]] = insertelement <16 x i16> [[TMP2]], i16 [[LA11]], i32 1
+; SM100-NEXT: [[TMP4:%.*]] = insertelement <16 x i16> [[TMP3]], i16 [[L910]], i32 2
+; SM100-NEXT: [[TMP5:%.*]] = insertelement <16 x i16> [[TMP4]], i16 [[L89]], i32 3
+; SM100-NEXT: [[TMP6:%.*]] = insertelement <16 x i16> [[TMP5]], i16 [[L78]], i32 4
+; SM100-NEXT: [[TMP7:%.*]] = insertelement <16 x i16> [[TMP6]], i16 [[L67]], i32 5
+; SM100-NEXT: [[TMP8:%.*]] = insertelement <16 x i16> [[TMP7]], i16 [[L56]], i32 6
+; SM100-NEXT: [[TMP9:%.*]] = insertelement <16 x i16> [[TMP8]], i16 [[L45]], i32 7
+; SM100-NEXT: [[TMP10:%.*]] = insertelement <16 x i16> [[TMP9]], i16 [[L34]], i32 8
+; SM100-NEXT: [[TMP11:%.*]] = insertelement <16 x i16> [[TMP10]], i16 [[L23]], i32 9
+; SM100-NEXT: [[TMP12:%.*]] = insertelement <16 x i16> [[TMP11]], i16 [[L12]], i32 10
+; SM100-NEXT: [[TMP13:%.*]] = insertelement <16 x i16> [[TMP12]], i16 [[L01]], i32 11
+; SM100-NEXT: [[TMP14:%.*]] = insertelement <16 x i16> [[TMP13]], i16 [[LF16]], i32 12
+; SM100-NEXT: [[TMP15:%.*]] = insertelement <16 x i16> [[TMP14]], i16 [[LE15]], i32 13
+; SM100-NEXT: [[TMP16:%.*]] = insertelement <16 x i16> [[TMP15]], i16 [[LD14]], i32 14
+; SM100-NEXT: [[TMP17:%.*]] = insertelement <16 x i16> [[TMP16]], i16 [[LC13]], i32 15
+; SM100-NEXT: store <16 x i16> [[TMP17]], ptr addrspace(1) [[PTR0]], align 32
+; SM100-NEXT: ret void
+;
+
+ %ptr0 = getelementptr i16, ptr addrspace(1) %ptr, i64 0
+ %ptr1 = getelementptr i16, ptr addrspace(1) %ptr, i64 1
+ %ptr2 = getelementptr i16, ptr addrspace(1) %ptr, i64 2
+ %ptr3 = getelementptr i16, ptr addrspace(1) %ptr, i64 3
+ %ptr4 = getelementptr i16, ptr addrspace(1) %ptr, i64 4
+ %ptr5 = getelementptr i16, ptr addrspace(1) %ptr, i64 5
+ %ptr6 = getelementptr i16, ptr addrspace(1) %ptr, i64 6
+ %ptr7 = getelementptr i16, ptr addrspace(1) %ptr, i64 7
+ %ptr8 = getelementptr i16, ptr addrspace(1) %ptr, i64 8
+ %ptr9 = getelementptr i16, ptr addrspace(1) %ptr, i64 9
+ %ptra = getelementptr i16, ptr addrspace(1) %ptr, i64 10
+ %ptrb = getelementptr i16, ptr addrspace(1) %ptr, i64 11
+ %ptrc = getelementptr i16, ptr addrspace(1) %ptr, i64 12
+ %ptrd = getelementptr i16, ptr addrspace(1) %ptr, i64 13
+ %ptre = getelementptr i16, ptr addrspace(1) %ptr, i64 14
+ %ptrf = getelementptr i16, ptr addrspace(1) %ptr, i64 15
+
+ %l0 = load i16, ptr addrspace(1) %ptr0, align 32
+ %l1 = load i16, ptr addrspace(1) %ptr1, align 1
+ %l2 = load i16, ptr addrspace(1) %ptr2, align 1
+ %l3 = load i16, ptr addrspace(1) %ptr3, align 1
+ %l4 = load i16, ptr addrspace(1) %ptr4, align 1
+ %l5 = load i16, ptr addrspace(1) %ptr5, align 1
+ %l6 = load i16, ptr addrspace(1) %ptr6, align 1
+ %l7 = load i16, ptr addrspace(1) %ptr7, align 1
+ %l8 = load i16, ptr addrspace(1) %ptr8, align 16
+ %l9 = load i16, ptr addrspace(1) %ptr9, align 1
+ %la = load i16, ptr addrspace(1) %ptra, align 1
+ %lb = load i16, ptr addrspace(1) %ptrb, align 1
+ %lc = load i16, ptr addrspace(1) %ptrc, align 1
+ %ld = load i16, ptr addrspace(1) %ptrd, align 1
+ %le = load i16, ptr addrspace(1) %ptre, align 1
+ %lf = load i16, ptr addrspace(1) %ptrf, align 1
+
+ store i16 %lf, ptr addrspace(1) %ptrc, align 1
+ store i16 %le, ptr addrspace(1) %ptrd, align 1
+ store i16 %ld, ptr addrspace(1) %ptre, align 1
+ store i16 %lc, ptr addrspace(1) %ptrf, align 1
+ store i16 %lb, ptr addrspace(1) %ptr0, align 32
+ store i16 %la, ptr addrspace(1) %ptr1, align 1
+ store i16 %l9, ptr addrspace(1) %ptr2, align 1
+ store i16 %l8, ptr addrspace(1) %ptr3, align 1
+ store i16 %l7, ptr addrspace(1) %ptr4, align 1
+ store i16 %l6, ptr addrspace(1) %ptr5, align 1
+ store i16 %l5, ptr addrspace(1) %ptr6, align 1
+ store i16 %l4, ptr addrspace(1) %ptr7, align 1
+ store i16 %l3, ptr addrspace(1) %ptr8, align 16
+ store i16 %l2, ptr addrspace(1) %ptr9, align 1
+ store i16 %l1, ptr addrspace(1) %ptra, align 1
+ store i16 %l0, ptr addrspace(1) %ptrb, align 1
+
+ ret void
+}
+
+define void @int32x8(ptr addrspace(1) %ptr) {
+; SM90-LABEL: define void @int32x8(
+; SM90-SAME: ptr addrspace(1) [[PTR:%.*]]) #[[ATTR0]] {
+; SM90-NEXT: [[PTR0:%.*]] = getelementptr i32, ptr addrspace(1) [[PTR]], i64 0
+; SM90-NEXT: [[PTR4:%.*]] = getelementptr i32, ptr addrspace(1) [[PTR]], i64 4
+; SM90-NEXT: [[TMP1:%.*]] = load <4 x i32>, ptr addrspace(1) [[PTR0]], align 32
+; SM90-NEXT: [[L01:%.*]] = extractelement <4 x i32> [[TMP1]], i32 0
+; SM90-NEXT: [[L12:%.*]] = extractelement <4 x i32> [[TMP1]], i32 1
+; SM90-NEXT: [[L23:%.*]] = extractelement <4 x i32> [[TMP1]], i32 2
+; SM90-NEXT: [[L34:%.*]] = extractelement <4 x i32> [[TMP1]], i32 3
+; SM90-NEXT: [[TMP2:%.*]] = load <4 x i32>, ptr addrspace(1) [[PTR4]], align 16
+; SM90-NEXT: [[L45:%.*]] = extractelement <4 x i32> [[TMP2]], i32 0
+; SM90-NEXT: [[L56:%.*]] = extractelement <4 x i32> [[TMP2]], i32 1
+; SM90-NEXT: [[L67:%.*]] = extractelement <4 x i32> [[TMP2]], i32 2
+; SM90-NEXT: [[L78:%.*]] = extractelement <4 x i32> [[TMP2]], i32 3
+; SM90-NEXT: [[TMP3:%.*]] = insertelement <4 x i32> poison, i32 [[L78]], i32 0
+; SM90-NEXT: [[TMP4:%.*]] = insertelement <4 x i32> [[TMP3]], i32 [[L67]], i32 1
+; SM90-NEXT: [[TMP5:%.*]] = insertelement <4 x i32> [[TMP4]], i32 [[L56]], i32 2
+; SM90-NEXT: [[TMP6:%.*]] = insertelement <4 x i32> [[TMP5]], i32 [[L45]], i32 3
+; SM90-NEXT: store <4 x i32> [[TMP6]], ptr addrspace(1) [[PTR0]], align 32
+; SM90-NEXT: [[TMP7:%.*]] = insertelement <4 x i32> poison, i32 [[L34]], i32 0
+; SM90-NEXT: [[TMP8:%.*]] = insertelement <4 x i32> [[TMP7]], i32 [[L23]], i32 1
+; SM90-NEXT: [[TMP9:%.*]] = insertelement <4 x i32> [[TMP8]], i32 [[L12]], i32 2
+; SM90-NEXT: [[TMP10:%.*]] = insertelement <4 x i32> [[TMP9]], i32 [[L01]], i32 3
+; SM90-NEXT: store <4 x i32> [[TMP10]], ptr addrspace(1) [[PTR4]], align 16
+; SM90-NEXT: ret void
+;
+; SM100-LABEL: define void @int32x8(
+; SM100-SAME: ptr addrspace(1) [[PTR:%.*]]) #[[ATTR0]] {
+; SM100-NEXT: [[PTR0:%.*]] = getelementptr i32, ptr addrspace(1) [[PTR]], i64 0
+; SM100-NEXT: [[TMP1:%.*]] = load <8 x i32>, ptr addrspace(1) [[PTR0]], align 32
+; SM100-NEXT: [[L01:%.*]] = extractelement <8 x i32> [[TMP1]], i32 0
+; SM100-NEXT: [[L12:%.*]] = extractelement <8 x i32> [[TMP1]], i32 1
+; SM100-NEXT: [[L23:%.*]] = extractelement <8 x i32> [[TMP1]], i32 2
+; SM100-NEXT: [[L34:%.*]] = extractelement <8 x i32> [[TMP1]], i32 3
+; SM100-NEXT: [[L45:%.*]] = extractelement <8 x i32> [[TMP1]], i32 4
+; SM100-NEXT: [[L56:%.*]] = extractelement <8 x i32> [[TMP1]], i32 5
+; SM100-NEXT: [[L67:%.*]] = extractelement <8 x i32> [[TMP1]], i32 6
+; SM100-NEXT: [[L78:%.*]] = extractelement <8 x i32> [[TMP1]], i32 7
+; SM100-NEXT: [[TMP2:%.*]] = insertelement <8 x i32> poison, i32 [[L78]], i32 0
+; SM100-NEXT: [[TMP3:%.*]] = insertelement <8 x i32> [[TMP2]], i32 [[L67]], i32 1
+; SM100-NEXT: [[TMP4:%.*]] = insertelement <8 x i32> [[TMP3]], i32 [[L56]], i32 2
+; SM100-NEXT: [[TMP5:%.*]] = insertelement <8 x i32> [[TMP4]], i32 [[L45]], i32 3
+; SM100-NEXT: [[TMP6:%.*]] = insertelement <8 x i32> [[TMP5]], i32 [[L34]], i32 4
+; SM100-NEXT: [[TMP7:%.*]] = insertelement <8 x i32> [[TMP6]], i32 [[L23]], i32 5
+; SM100-NEXT: [[TMP8:%.*]] = insertelement <8 x i32> [[TMP7]], i32 [[L12]], i32 6
+; SM100-NEXT: [[TMP9:%.*]] = insertelement <8 x i32> [[TMP8]], i32 [[L01]], i32 7
+; SM100-NEXT: store <8 x i32> [[TMP9]], ptr addrspace(1) [[PTR0]], align 32
+; SM100-NEXT: ret void
+;
+
+ %ptr0 = getelementptr i32, ptr addrspace(1) %ptr, i64 0
+ %ptr1 = getelementptr i32, ptr addrspace(1) %ptr, i64 1
+ %ptr2 = getelementptr i32, ptr addrspace(1) %ptr, i64 2
+ %ptr3 = getelementptr i32, ptr addrspace(1) %ptr, i64 3
+ %ptr4 = getelementptr i32, ptr addrspace(1) %ptr, i64 4
+ %ptr5 = getelementptr i32, ptr addrspace(1) %ptr, i64 5
+ %ptr6 = getelementptr i32, ptr addrspace(1) %ptr, i64 6
+ %ptr7 = getelementptr i32, ptr addrspace(1) %ptr, i64 7
+
+ %l0 = load i32, ptr addrspace(1) %ptr0, align 32
+ %l1 = load i32, ptr addrspace(1) %ptr1, align 1
+ %l2 = load i32, ptr addrspace(1) %ptr2, align 1
+ %l3 = load i32, ptr addrspace(1) %ptr3, align 1
+ %l4 = load i32, ptr addrspace(1) %ptr4, align 16
+ %l5 = load i32, ptr addrspace(1) %ptr5, align 1
+ %l6 = load i32, ptr addrspace(1) %ptr6, align 1
+ %l7 = load i32, ptr addrspace(1) %ptr7, align 1
+
+ store i32 %l7, ptr addrspace(1) %ptr0, align 32
+ store i32 %l6, ptr addrspace(1) %ptr1, align 1
+ store i32 %l5, ptr addrspace(1) %ptr2, align 1
+ store i32 %l4, ptr addrspace(1) %ptr3, align 1
+ store i32 %l3, ptr addrspace(1) %ptr4, align 16
+ store i32 %l2, ptr addrspace(1) %ptr5, align 1
+ store i32 %l1, ptr addrspace(1) %ptr6, align 1
+ store i32 %l0, ptr addrspace(1) %ptr7, align 1
+
+ ret void
+}
+
+define void @int64x4(ptr addrspace(1) %ptr) {
+; SM90-LABEL: define void @int64x4(
+; SM90-SAME: ptr addrspace(1) [[PTR:%.*]]) #[[ATTR0]] {
+; SM90-NEXT: [[PTR0:%.*]] = getelementptr i64, ptr addrspace(1) [[PTR]], i64 0
+; SM90-NEXT: [[PTR2:%.*]] = getelementptr i64, ptr addrspace(1) [[PTR]], i64 2
+; SM90-NEXT: [[TMP1:%.*]] = load <2 x i64>, ptr addrspace(1) [[PTR0]], align 32
+; SM90-NEXT: [[L01:%.*]] = extractelement <2 x i64> [[TMP1]], i32 0
+; SM90-NEXT: [[L12:%.*]] = extractelement <2 x i64> [[TMP1]], i32 1
+; SM90-NEXT: [[TMP2:%.*]] = load <2 x i64>, ptr addrspace(1) [[PTR2]], align 16
+; SM90-NEXT: [[L23:%.*]] = extractelement <2 x i64> [[TMP2]], i32 0
+; SM90-NEXT: [[L34:%.*]] = extractelement <2 x i64> [[TMP2]], i32 1
+; SM90-NEXT: [[TMP3:%.*]] = insertelement <2 x i64> poison, i64 [[L34]], i32 0
+; SM90-NEXT: [[TMP4:%.*]] = insertelement <2 x i64> [[TMP3]], i64 [[L23]], i32 1
+; SM90-NEXT: store <2 x i64> [[TMP4]], ptr addrspace(1) [[PTR0]], align 32
+; SM90-NEXT: [[TMP5:%.*]] = insertelement <2 x i64> poison, i64 [[L12]], i32 0
+; SM90-NEXT: [[TMP6:%.*]] = insertelement <2 x i64> [[TMP5]], i64 [[L01]], i32 1
+; SM90-NEXT: store <2 x i64> [[TMP6]], ptr addrspace(1) [[PTR2]], align 16
+; SM90-NEXT: ret void
+;
+; SM100-LABEL: define void @int64x4(
+; SM100-SAME: ptr addrspace(1) [[PTR:%.*]]) #[[ATTR0]] {
+; SM100-NEXT: [[PTR0:%.*]] = getelementptr i64, ptr addrspace(1) [[PTR]], i64 0
+; SM100-NEXT: [[TMP1:%.*]] = load <4 x i64>, ptr addrspace(1) [[PTR0]], align 32
+; SM100-NEXT: [[L01:%.*]] = extractelement <4 x i64> [[TMP1]], i32 0
+; SM100-NEXT: [[L12:%.*]] = extractelement <4 x i64> [[TMP1]], i32 1
+; SM100-NEXT: [[L23:%.*]] = extractelement <4 x i64> [[TMP1]], i32 2
+; SM100-NEXT: [[L34:%.*]] = extractelement <4 x i64> [[TMP1]], i32 3
+; SM100-NEXT: [[TMP2:%.*]] = insertelement <4 x i64> poison, i64 [[L34]], i32 0
+; SM100-NEXT: [[TMP3:%.*]] = insertelement <4 x i64> [[TMP2]], i64 [[L23]], i32 1
+; SM100-NEXT: [[TMP4:%.*]] = insertelement <4 x i64> [[TMP3]], i64 [[L12]], i32 2
+; SM100-NEXT: [[TMP5:%.*]] = insertelement <4 x i64> [[TMP4]], i64 [[L01]], i32 3
+; SM100-NEXT: store <4 x i64> [[TMP5]], ptr addrspace(1) [[PTR0]], align 32
+; SM100-NEXT: ret void
+;
+
+ %ptr0 = getelementptr i64, ptr addrspace(1) %ptr, i64 0
+ %ptr1 = getelementptr i64, ptr addrspace(1) %ptr, i64 1
+ %ptr2 = getelementptr i64, ptr addrspace(1) %ptr, i64 2
+ %ptr3 = getelementptr i64, ptr addrspace(1) %ptr, i64 3
+
+ %l0 = load i64, ptr addrspace(1) %ptr0, align 32
+ %l1 = load i64, ptr addrspace(1) %ptr1, align 1
+ %l2 = load i64, ptr addrspace(1) %ptr2, align 16
+ %l3 = load i64, ptr addrspace(1) %ptr3, align 1
+
+ store i64 %l3, ptr addrspace(1) %ptr0, align 32
+ store i64 %l2, ptr addrspace(1) %ptr1, align 1
+ store i64 %l1, ptr addrspace(1) %ptr2, align 16
+ store i64 %l0, ptr addrspace(1) %ptr3, align 1
+
+ ret void
+}
+
+define void @float32x8(ptr addrspace(1) %ptr) {
+; SM90-LABEL: define void @float32x8(
+; SM90-SAME: ptr addrspace(1) [[PTR:%.*]]) #[[ATTR0]] {
+; SM90-NEXT: [[PTR0:%.*]] = getelementptr float, ptr addrspace(1) [[PTR]], i64 0
+; SM90-NEXT: [[PTR4:%.*]] = getelementptr float, ptr addrspace(1) [[PTR]], i64 4
+; SM90-NEXT: [[TMP1:%.*]] = load <4 x float>, ptr addrspace(1) [[PTR0]], align 32
+; SM90-NEXT: [[L01:%.*]] = extractelement <4 x float> [[TMP1]], i32 0
+; SM90-NEXT: [[L12:%.*]] = extractelement <4 x float> [[TMP1]], i32 1
+; SM90-NEXT: [[L23:%.*]] = extractelement <4 x float> [[TMP1]], i32 2
+; SM90-NEXT: [[L34:%.*]] = extractelement <4 x float> [[TMP1]], i32 3
+; SM90-NEXT: [[TMP2:%.*]] = load <4 x float>, ptr addrspace(1) [[PTR4]], align 16
+; SM90-NEXT: [[L45:%.*]] = extractelement <4 x float> [[TMP2]], i32 0
+; SM90-NEXT: [[L56:%.*]] = extractelement <4 x float> [[TMP2]], i32 1
+; SM90-NEXT: [[L67:%.*]] = extractelement <4 x float> [[TMP2]], i32 2
+; SM90-NEXT: [[L78:%.*]] = extractelement <4 x float> [[TMP2]], i32 3
+; SM90-NEXT: [[TMP3:%.*]] = insertelement <4 x float> poison, float [[L78]], i32 0
+; SM90-NEXT: [[TMP4:%.*]] = insertelement <4 x float> [[TMP3]], float [[L67]], i32 1
+; SM90-NEXT: [[TMP5:%.*]] = insertelement <4 x float> [[TMP4]], float [[L56]], i32 2
+; SM90-NEXT: [[TMP6:%.*]] = insertelement <4 x float> [[TMP5]], float [[L45]], i32 3
+; SM90-NEXT: store <4 x float> [[TMP6]], ptr addrspace(1) [[PTR0]], align 32
+; SM90-NEXT: [[TMP7:%.*]] = insertelement <4 x float> poison, float [[L34]], i32 0
+; SM90-NEXT: [[TMP8:%.*]] = insertelement <4 x float> [[TMP7]], float [[L23]], i32 1
+; SM90-NEXT: [[TMP9:%.*]] = insertelement <4 x float> [[TMP8]], float [[L12]], i32 2
+; SM90-NEXT: [[TMP10:%.*]] = insertelement <4 x float> [[TMP9]], float [[L01]], i32 3
+; SM90-NEXT: store <4 x float> [[TMP10]], ptr addrspace(1) [[PTR4]], align 16
+; SM90-NEXT: ret void
+;
+; SM100-LABEL: define void @float32x8(
+; SM100-SAME: ptr addrspace(1) [[PTR:%.*]]) #[[ATTR0]] {
+; SM100-NEXT: [[PTR0:%.*]] = getelementptr float, ptr addrspace(1) [[PTR]], i64 0
+; SM100-NEXT: [[TMP1:%.*]] = load <8 x float>, ptr addrspace(1) [[PTR0]], align 32
+; SM100-NEXT: [[L01:%.*]] = extractelement <8 x float> [[TMP1]], i32 0
+; SM100-NEXT: [[L12:%.*]] = extractelement <8 x float> [[TMP1]], i32 1
+; SM100-NEXT: [[L23:%.*]] = extractelement <8 x float> [[TMP1]], i32 2
+; SM100-NEXT: [[L34:%.*]] = extractelement <8 x float> [[TMP1]], i32 3
+; SM100-NEXT: [[L45:%.*]] = extractelement <8 x float> [[TMP1]], i32 4
+; SM100-NEXT: [[L56:%.*]] = extractelement <8 x float> [[TMP1]], i32 5
+; SM100-NEXT: [[L67:%.*]] = extractelement <8 x float> [[TMP1]], i32 6
+; SM100-NEXT: [[L78:%.*]] = extractelement <8 x float> [[TMP1]], i32 7
+; SM100-NEXT: [[TMP2:%.*]] = insertelement <8 x float> poison, float [[L78]], i32 0
+; SM100-NEXT: [[TMP3:%.*]] = insertelement <8 x float> [[TMP2]], float [[L67]], i32 1
+; SM100-NEXT: [[TMP4:%.*]] = insertelement <8 x float> [[TMP3]], float [[L56]], i32 2
+; SM100-NEXT: [[TMP5:%.*]] = insertelement <8 x float> [[TMP4]], float [[L45]], i32 3
+; SM100-NEXT: [[TMP6:%.*]] = insertelement <8 x float> [[TMP5]], float [[L34]], i32 4
+; SM100-NEXT: [[TMP7:%.*]] = insertelement <8 x float> [[TMP6]], float [[L23]], i32 5
+; SM100-NEXT: [[TMP8:%.*]] = insertelement <8 x float> [[TMP7]], float [[L12]], i32 6
+; SM100-NEXT: [[TMP9:%.*]] = insertelement <8 x float> [[TMP8]], float [[L01]], i32 7
+; SM100-NEXT: store <8 x float> [[TMP9]], ptr addrspace(1) [[PTR0]], align 32
+; SM100-NEXT: ret void
+;
+
+ %ptr0 = getelementptr float, ptr addrspace(1) %ptr, i64 0
+ %ptr1 = getelementptr float, ptr addrspace(1) %ptr, i64 1
+ %ptr2 = getelementptr float, ptr addrspace(1) %ptr, i64 2
+ %ptr3 = getelementptr float, ptr addrspace(1) %ptr, i64 3
+ %ptr4 = getelementptr float, ptr addrspace(1) %ptr, i64 4
+ %ptr5 = getelementptr float, ptr addrspace(1) %ptr, i64 5
+ %ptr6 = getelementptr float, ptr addrspace(1) %ptr, i64 6
+ %ptr7 = getelementptr float, ptr addrspace(1) %ptr, i64 7
+
+ %l0 = load float, ptr addrspace(1) %ptr0, align 32
+ %l1 = load float, ptr addrspace(1) %ptr1, align 1
+ %l2 = load float, ptr addrspace(1) %ptr2, align 1
+ %l3 = load float, ptr addrspace(1) %ptr3, align 1
+ %l4 = load float, ptr addrspace(1) %ptr4, align 16
+ %l5 = load float, ptr addrspace(1) %ptr5, align 1
+ %l6 = load float, ptr addrspace(1) %ptr6, align 1
+ %l7 = load float, ptr addrspace(1) %ptr7, align 1
+
+ store float %l7, ptr addrspace(1) %ptr0, align 32
+ store float %l6, ptr addrspace(1) %ptr1, align 1
+ store float %l5, ptr addrspace(1) %ptr2, align 1
+ store float %l4, ptr addrspace(1) %ptr3, align 1
+ store float %l3, ptr addrspace(1) %ptr4, align 16
+ store float %l2, ptr addrspace(1) %ptr5, align 1
+ store float %l1, ptr addrspace(1) %ptr6, align 1
+ store float %l0, ptr addrspace(1) %ptr7, align 1
+
+ ret void
+}
+
+define void @float64x4(ptr addrspace(1) %ptr) {
+; SM90-LABEL: define void @float64x4(
+; SM90-SAME: ptr addrspace(1) [[PTR:%.*]]) #[[ATTR0]] {
+; SM90-NEXT: [[PTR0:%.*]] = getelementptr double, ptr addrspace(1) [[PTR]], i64 0
+; SM90-NEXT: [[PTR2:%.*]] = getelementptr double, ptr addrspace(1) [[PTR]], i64 2
+; SM90-NEXT: [[TMP1:%.*]] = load <2 x double>, ptr addrspace(1) [[PTR0]], align 32
+; SM90-NEXT: [[L01:%.*]] = extractelement <2 x double> [[TMP1]], i32 0
+; SM90-NEXT: [[L12:%.*]] = extractelement <2 x double> [[TMP1]], i32 1
+; SM90-NEXT: [[TMP2:%.*]] = load <2 x double>, ptr addrspace(1) [[PTR2]], align 16
+; SM90-NEXT: [[L23:%.*]] = extractelement <2 x double> [[TMP2]], i32 0
+; SM90-NEXT: [[L34:%.*]] = extractelement <2 x double> [[TMP2]], i32 1
+; SM90-NEXT: [[TMP3:%.*]] = insertelement <2 x double> poison, double [[L34]], i32 0
+; SM90-NEXT: [[TMP4:%.*]] = insertelement <2 x double> [[TMP3]], double [[L23]], i32 1
+; SM90-NEXT: store <2 x double> [[TMP4]], ptr addrspace(1) [[PTR0]], align 32
+; SM90-NEXT: [[TMP5:%.*]] = insertelement <2 x double> poison, double [[L12]], i32 0
+; SM90-NEXT: [[TMP6:%.*]] = insertelement <2 x double> [[TMP5]], double [[L01]], i32 1
+; SM90-NEXT: store <2 x double> [[TMP6]], ptr addrspace(1) [[PTR2]], align 16
+; SM90-NEXT: ret void
+;
+; SM100-LABEL: define void @float64x4(
+; SM100-SAME: ptr addrspace(1) [[PTR:%.*]]) #[[ATTR0]] {
+; SM100-NEXT: [[PTR0:%.*]] = getelementptr double, ptr addrspace(1) [[PTR]], i64 0
+; SM100-NEXT: [[TMP1:%.*]] = load <4 x double>, ptr addrspace(1) [[PTR0]], align 32
+; SM100-NEXT: [[L01:%.*]] = extractelement <4 x double> [[TMP1]], i32 0
+; SM100-NEXT: [[L12:%.*]] = extractelement <4 x double> [[TMP1]], i32 1
+; SM100-NEXT: [[L23:%.*]] = extractelement <4 x double> [[TMP1]], i32 2
+; SM100-NEXT: [[L34:%.*]] = extractelement <4 x double> [[TMP1]], i32 3
+; SM100-NEXT: [[TMP2:%.*]] = insertelement <4 x double> poison, double [[L34]], i32 0
+; SM100-NEXT: [[TMP3:%.*]] = insertelement <4 x double> [[TMP2]], double [[L23]], i32 1
+; SM100-NEXT: [[TMP4:%.*]] = insertelement <4 x double> [[TMP3]], double [[L12]], i32 2
+; SM100-NEXT: [[TMP5:%.*]] = insertelement <4 x double> [[TMP4]], double [[L01]], i32 3
+; SM100-NEXT: store <4 x double> [[TMP5]], ptr addrspace(1) [[PTR0]], align 32
+; SM100-NEXT: ret void
+;
+
+ %ptr0 = getelementptr double, ptr addrspace(1) %ptr, i64 0
+ %ptr1 = getelementptr double, ptr addrspace(1) %ptr, i64 1
+ %ptr2 = getelementptr double, ptr addrspace(1) %ptr, i64 2
+ %ptr3 = getelementptr double, ptr addrspace(1) %ptr, i64 3
+
+ %l0 = load double, ptr addrspace(1) %ptr0, align 32
+ %l1 = load double, ptr addrspace(1) %ptr1, align 1
+ %l2 = load double, ptr addrspace(1) %ptr2, align 16
+ %l3 = load double, ptr addrspace(1) %ptr3, align 1
+
+ store double %l3, ptr addrspace(1) %ptr0, align 32
+ store double %l2, ptr addrspace(1) %ptr1, align 1
+ store double %l1, ptr addrspace(1) %ptr2, align 16
+ store double %l0, ptr addrspace(1) %ptr3, align 1
+
+ ret void
+}
+
+define void @int32x8_non_global(ptr %ptr) {
+; CHECK-LABEL: define void @int32x8_non_global(
+; CHECK-SAME: ptr [[PTR:%.*]]) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT: [[PTR0:%.*]] = getelementptr i32, ptr [[PTR]], i64 0
+; CHECK-NEXT: [[PTR4:%.*]] = getelementptr i32, ptr [[PTR]], i64 4
+; CHECK-NEXT: [[TMP1:%.*]] = load <4 x i32>, ptr [[PTR0]], align 32
+; CHECK-NEXT: [[L01:%.*]] = extractelement <4 x i32> [[TMP1]], i32 0
+; CHECK-NEXT: [[L12:%.*]] = extractelement <4 x i32> [[TMP1]], i32 1
+; CHECK-NEXT: [[L23:%.*]] = extractelement <4 x i32> [[TMP1]], i32 2
+; CHECK-NEXT: [[L34:%.*]] = extractelement <4 x i32> [[TMP1]], i32 3
+; CHECK-NEXT: [[TMP2:%.*]] = load <4 x i32>, ptr [[PTR4]], align 16
+; CHECK-NEXT: [[L45:%.*]] = extractelement <4 x i32> [[TMP2]], i32 0
+; CHECK-NEXT: [[L56:%.*]] = extractelement <4 x i32> [[TMP2]], i32 1
+; CHECK-NEXT: [[L67:%.*]] = extractelement <4 x i32> [[TMP2]], i32 2
+; CHECK-NEXT: [[L78:%.*]] = extractelement <4 x i32> [[TMP2]], i32 3
+; CHECK-NEXT: [[TMP3:%.*]] = insertelement <4 x i32> poison, i32 [[L78]], i32 0
+; CHECK-NEXT: [[TMP4:%.*]] = insertelement <4 x i32> [[TMP3]], i32 [[L67]], i32 1
+; CHECK-NEXT: [[TMP5:%.*]] = insertelement <4 x i32> [[TMP4]], i32 [[L56]], i32 2
+; CHECK-NEXT: [[TMP6:%.*]] = insertelement <4 x i32> [[TMP5]], i32 [[L45]], i32 3
+; CHECK-NEXT: store <4 x i32> [[TMP6]], ptr [[PTR0]], align 32
+; CHECK-NEXT: [[TMP7:%.*]] = insertelement <4 x i32> poison, i32 [[L34]], i32 0
+; CHECK-NEXT: [[TMP8:%.*]] = insertelement <4 x i32> [[TMP7]], i32 [[L23]], i32 1
+; CHECK-NEXT: [[TMP9:%.*]] = insertelement <4 x i32> [[TMP8]], i32 [[L12]], i32 2
+; CHECK-NEXT: [[TMP10:%.*]] = insertelement <4 x i32> [[TMP9]], i32 [[L01]], i32 3
+; CHECK-NEXT: store <4 x i32> [[TMP10]], ptr [[PTR4]], align 16
+; CHECK-NEXT: ret void
+
+ %ptr0 = getelementptr i32, ptr %ptr, i64 0
+ %ptr1 = getelementptr i32, ptr %ptr, i64 1
+ %ptr2 = getelementptr i32, ptr %ptr, i64 2
+ %ptr3 = getelementptr i32, ptr %ptr, i64 3
+ %ptr4 = getelementptr i32, ptr %ptr, i64 4
+ %ptr5 = getelementptr i32, ptr %ptr, i64 5
+ %ptr6 = getelementptr i32, ptr %ptr, i64 6
+ %ptr7 = getelementptr i32, ptr %ptr, i64 7
+
+ %l0 = load i32, ptr %ptr0, align 32
+ %l1 = load i32, ptr %ptr1, align 1
----------------
Artem-B wrote:
Hmm. `align 1` looks rather odd here. We know that `ptr0` is aligned by 32. And we know that `ptr1` is offset by 4, so correct alignment must be 4. And so on for the other elements.
Correct alignment info should be: 32,4,8,4,16,4,8,4
Looks like vectorize was able to figure it out, and vectorized the loads/stores, but it would be great if we'd provide sensible inputs.
https://github.com/llvm/llvm-project/pull/139292
More information about the llvm-commits
mailing list