Skip to content

[PreISelIntrinsicLowering] Look through ptrtoint for a constant (e.g. global ptr) when trying to produce memset_pattern16 #129220

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wants to merge 7 commits into from
Closed

[PreISelIntrinsicLowering] Look through ptrtoint for a constant (e.g. global ptr) when trying to produce memset_pattern16 #129220

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
a313322
[PreISelIntrinsicLowering][test] Precommit test for failure to select…
asb Feb 28, 2025
7dea145
[PreISelIntrinsicLowering] Support producing memset_pattern16 when lo…
asb Feb 28, 2025
cb47c53
Merge remote-tracking branch 'origin/main' into 2025q1-preiselintrins…
asb Mar 5, 2025
8fd2d02
Update test case to no longer assume global ptrs arg is loaded from g…
asb Mar 5, 2025
bd31315
Look through a ptrtoint cast for a constant (e.g. global ptr)
asb Mar 5, 2025
237c0bf
Merge remote-tracking branch 'origin/main' into 2025q1-preiselintrins…
asb Mar 12, 2025
2b6d262
Merge remote-tracking branch 'origin/main' into 2025q1-preiselintrins…
asb Mar 12, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
12 changes: 10 additions & 2 deletions llvm/lib/CodeGen/PreISelIntrinsicLowering.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -254,10 +254,18 @@ static Constant *getMemSetPattern16Value(MemSetPatternInst *Inst,
if (!isLibFuncEmittable(M, &TLI, LibFunc_memset_pattern16))
return nullptr;

// Look through a ptrtoint cast for a candidate constant. This could be
// extended to look through other casts, but the assumption is earlier
// passes that introduced memset.pattern intrinsic would have just emitted
// the integer argument directly for CosntantFP or ConstantInt cases.
Copy link
Contributor

@nikic nikic Mar 12, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
// the integer argument directly for CosntantFP or ConstantInt cases.
// the integer argument directly for ConstantFP or ConstantInt cases.

if (auto *PI = dyn_cast<PtrToIntInst>(V))
V = PI->getPointerOperand();

Constant *C = dyn_cast<Constant>(V);

// If the value isn't a constant, we can't promote it to being in a constant
// array. We could theoretically do a store to an alloca or something, but
// that doesn't seem worthwhile.
Constant *C = dyn_cast<Constant>(V);
if (!C || isa<ConstantExpr>(C))
return nullptr;

Expand All @@ -284,7 +292,7 @@ static Constant *getMemSetPattern16Value(MemSetPatternInst *Inst,

// Otherwise, we'll use an array of the constants.
uint64_t ArraySize = 16 / Size;
ArrayType *AT = ArrayType::get(V->getType(), ArraySize);
ArrayType *AT = ArrayType::get(C->getType(), ArraySize);
return ConstantArray::get(AT, std::vector<Constant *>(ArraySize, C));
}

Expand Down
69 changes: 61 additions & 8 deletions llvm/test/Transforms/PreISelIntrinsicLowering/X86/memset-pattern.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,13 +2,16 @@
; RUN: opt -mtriple=x86_64-apple-darwin10.0.0 -passes=pre-isel-intrinsic-lowering -S -o - %s | FileCheck %s

;.
; CHECK: @G = global i32 5
; CHECK: @.memset_pattern = private unnamed_addr constant [2 x i64] [i64 -6148895925951734307, i64 -6148895925951734307], align 16
; CHECK: @.memset_pattern.1 = private unnamed_addr constant [2 x i64] [i64 4614256656552045848, i64 4614256656552045848], align 16
; CHECK: @.memset_pattern.2 = private unnamed_addr constant [8 x i16] [i16 -21555, i16 -21555, i16 -21555, i16 -21555, i16 -21555, i16 -21555, i16 -21555, i16 -21555], align 16
; CHECK: @.memset_pattern.3 = private unnamed_addr constant i128 -113427455635030943652277463699152839203, align 16
; CHECK: @.memset_pattern.4 = private unnamed_addr constant i128 -113427455635030943652277463699152839203, align 16
; CHECK: @.memset_pattern.2 = private unnamed_addr constant [2 x ptr] [ptr @G, ptr @G], align 16
; CHECK: @.memset_pattern.3 = private unnamed_addr constant [2 x ptr] [ptr @G, ptr @G], align 16
; CHECK: @.memset_pattern.4 = private unnamed_addr constant [8 x i16] [i16 -21555, i16 -21555, i16 -21555, i16 -21555, i16 -21555, i16 -21555, i16 -21555, i16 -21555], align 16
; CHECK: @.memset_pattern.5 = private unnamed_addr constant i128 -113427455635030943652277463699152839203, align 16
; CHECK: @.memset_pattern.6 = private unnamed_addr constant i128 -113427455635030943652277463699152839203, align 16
; CHECK: @.memset_pattern.7 = private unnamed_addr constant i128 -113427455635030943652277463699152839203, align 16
; CHECK: @.memset_pattern.8 = private unnamed_addr constant i128 -113427455635030943652277463699152839203, align 16
;.
define void @memset_pattern_i128_1_dynvalue(ptr %a, i128 %value) nounwind {
; CHECK-LABEL: define void @memset_pattern_i128_1_dynvalue(
Expand All @@ -31,7 +34,7 @@ define void @memset_pattern_i128_1_dynvalue(ptr %a, i128 %value) nounwind {
define void @memset_pattern_i128_1(ptr %a, i128 %value) nounwind {
; CHECK-LABEL: define void @memset_pattern_i128_1(
; CHECK-SAME: ptr [[A:%.*]], i128 [[VALUE:%.*]]) #[[ATTR0]] {
; CHECK-NEXT: call void @memset_pattern16(ptr [[A]], ptr @.memset_pattern.3, i64 16)
; CHECK-NEXT: call void @memset_pattern16(ptr [[A]], ptr @.memset_pattern.5, i64 16)
; CHECK-NEXT: ret void
;
tail call void @llvm.experimental.memset.pattern(ptr %a, i128 u0xaaaaaaaabbbbbbbbccccccccdddddddd, i64 1, i1 false)
Expand Down Expand Up @@ -59,7 +62,7 @@ define void @memset_pattern_i128_1_nz_as(ptr addrspace(1) %a, i128 %value) nounw
define void @memset_pattern_i128_1_align_attr(ptr align(16) %a, i128 %value) nounwind {
; CHECK-LABEL: define void @memset_pattern_i128_1_align_attr(
; CHECK-SAME: ptr align 16 [[A:%.*]], i128 [[VALUE:%.*]]) #[[ATTR0]] {
; CHECK-NEXT: call void @memset_pattern16(ptr align 16 [[A]], ptr @.memset_pattern.4, i64 16)
; CHECK-NEXT: call void @memset_pattern16(ptr align 16 [[A]], ptr @.memset_pattern.6, i64 16)
; CHECK-NEXT: ret void
;
tail call void @llvm.experimental.memset.pattern(ptr align(16) %a, i128 u0xaaaaaaaabbbbbbbbccccccccdddddddd, i64 1, i1 false)
Expand All @@ -69,7 +72,7 @@ define void @memset_pattern_i128_1_align_attr(ptr align(16) %a, i128 %value) nou
define void @memset_pattern_i128_16(ptr %a) nounwind {
; CHECK-LABEL: define void @memset_pattern_i128_16(
; CHECK-SAME: ptr [[A:%.*]]) #[[ATTR0]] {
; CHECK-NEXT: call void @memset_pattern16(ptr [[A]], ptr @.memset_pattern.5, i64 256)
; CHECK-NEXT: call void @memset_pattern16(ptr [[A]], ptr @.memset_pattern.7, i64 256)
; CHECK-NEXT: ret void
;
tail call void @llvm.experimental.memset.pattern(ptr %a, i128 u0xaaaaaaaabbbbbbbbccccccccdddddddd, i64 16, i1 false)
Expand All @@ -80,7 +83,7 @@ define void @memset_pattern_i128_x(ptr %a, i64 %x) nounwind {
; CHECK-LABEL: define void @memset_pattern_i128_x(
; CHECK-SAME: ptr [[A:%.*]], i64 [[X:%.*]]) #[[ATTR0]] {
; CHECK-NEXT: [[TMP1:%.*]] = mul i64 16, [[X]]
; CHECK-NEXT: call void @memset_pattern16(ptr [[A]], ptr @.memset_pattern.6, i64 [[TMP1]])
; CHECK-NEXT: call void @memset_pattern16(ptr [[A]], ptr @.memset_pattern.8, i64 [[TMP1]])
; CHECK-NEXT: ret void
;
tail call void @llvm.experimental.memset.pattern(ptr %a, i128 u0xaaaaaaaabbbbbbbbccccccccdddddddd, i64 %x, i1 false)
Expand Down Expand Up @@ -110,7 +113,7 @@ define void @memset_pattern_i16_x(ptr %a, i64 %x) nounwind {
; CHECK-LABEL: define void @memset_pattern_i16_x(
; CHECK-SAME: ptr [[A:%.*]], i64 [[X:%.*]]) #[[ATTR0]] {
; CHECK-NEXT: [[TMP1:%.*]] = mul i64 2, [[X]]
; CHECK-NEXT: call void @memset_pattern16(ptr [[A]], ptr @.memset_pattern.2, i64 [[TMP1]])
; CHECK-NEXT: call void @memset_pattern16(ptr [[A]], ptr @.memset_pattern.4, i64 [[TMP1]])
; CHECK-NEXT: ret void
;
tail call void @llvm.experimental.memset.pattern(ptr %a, i16 u0xabcd, i64 %x, i1 false)
Expand Down Expand Up @@ -144,6 +147,56 @@ define void @memset_pattern_i64_128_tbaa(ptr %a) nounwind {
!7 = !{!"omnipotent char", !8, i64 0}
!8 = !{!"Simple C++ TBAA"}

@G = global i32 5

define void @memset_pattern_i64_16_fromptr(ptr %a) nounwind {
; CHECK-LABEL: define void @memset_pattern_i64_16_fromptr(
; CHECK-SAME: ptr [[A:%.*]]) #[[ATTR0]] {
; CHECK-NEXT: [[TMP1:%.*]] = ptrtoint ptr @G to i64
; CHECK-NEXT: call void @memset_pattern16(ptr [[A]], ptr @.memset_pattern.2, i64 128)
; CHECK-NEXT: ret void
;
%1 = ptrtoint ptr @G to i64
tail call void @llvm.experimental.memset.pattern(ptr %a, i64 %1, i64 16, i1 false)
ret void
}

define void @memset_pattern_i64_x_fromptr(ptr %a, i64 %x) nounwind {
; CHECK-LABEL: define void @memset_pattern_i64_x_fromptr(
; CHECK-SAME: ptr [[A:%.*]], i64 [[X:%.*]]) #[[ATTR0]] {
; CHECK-NEXT: [[TMP1:%.*]] = ptrtoint ptr @G to i64
; CHECK-NEXT: [[TMP2:%.*]] = mul i64 8, [[X]]
; CHECK-NEXT: call void @memset_pattern16(ptr [[A]], ptr @.memset_pattern.3, i64 [[TMP2]])
; CHECK-NEXT: ret void
;
%1 = ptrtoint ptr @G to i64
tail call void @llvm.experimental.memset.pattern(ptr %a, i64 %1, i64 %x, i1 false)
ret void
}

; memset_pattern16 shouldn't be used for this example (at least not by just
; creating a constantarray global at compile time), as the ptr isn't constant.
define void @memset_pattern_i64_x_fromnonconstptr(ptr %a, i64 %x, ptr %p) nounwind {
; CHECK-LABEL: define void @memset_pattern_i64_x_fromnonconstptr(
; CHECK-SAME: ptr [[A:%.*]], i64 [[X:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
; CHECK-NEXT: [[TMP1:%.*]] = ptrtoint ptr [[P]] to i64
; CHECK-NEXT: [[TMP2:%.*]] = icmp eq i64 0, [[X]]
; CHECK-NEXT: br i1 [[TMP2]], label %[[SPLIT:.*]], label %[[LOADSTORELOOP:.*]]
; CHECK: [[LOADSTORELOOP]]:
; CHECK-NEXT: [[TMP3:%.*]] = phi i64 [ 0, [[TMP0:%.*]] ], [ [[TMP5:%.*]], %[[LOADSTORELOOP]] ]
; CHECK-NEXT: [[TMP4:%.*]] = getelementptr inbounds i64, ptr [[A]], i64 [[TMP3]]
; CHECK-NEXT: store i64 [[TMP1]], ptr [[TMP4]], align 1
; CHECK-NEXT: [[TMP5]] = add i64 [[TMP3]], 1
; CHECK-NEXT: [[TMP6:%.*]] = icmp ult i64 [[TMP5]], [[X]]
; CHECK-NEXT: br i1 [[TMP6]], label %[[LOADSTORELOOP]], label %[[SPLIT]]
; CHECK: [[SPLIT]]:
; CHECK-NEXT: ret void
;
%1 = ptrtoint ptr %p to i64
tail call void @llvm.experimental.memset.pattern(ptr %a, i64 %1, i64 %x, i1 false)
ret void
Comment on lines +195 to +197
Copy link
Contributor

@arsenm arsenm Mar 13, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

(finding the new no-required mangling suffix behavior confusing)

Could instcombine fold this to pull the cast in, and adjust the call signature?

}

;.
; CHECK: attributes #[[ATTR0]] = { nounwind }
; CHECK: attributes #[[ATTR1:[0-9]+]] = { nocallback nofree nounwind willreturn memory(argmem: write) }
Expand Down