; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
; RUN: opt -passes=loop-load-elim -S < %s | \
; RUN: FileCheck %s -check-prefix=DEFAULT
; RUN: opt -passes=loop-load-elim -S -enable-mem-access-versioning=0 < %s | \
; RUN: FileCheck %s -check-prefix=NO-VERSION
; RUN: opt -passes=loop-load-elim -S -loop-load-elimination-scev-check-threshold=1 < %s | \
; RUN: FileCheck %s -check-prefix=THRESHOLD
; Forwarding in the presence of symbolic strides:
;
; for (unsigned i = 0; i < 100; i++)
; A[i + 1] = A[Stride * i] + B[i];
target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
define void @f(ptr noalias nocapture %A, ptr noalias nocapture readonly %B, i64 %N,
;
;
;
;
;
; DEFAULT-LABEL: @f(
; DEFAULT-NEXT: for.body.lver.check:
; DEFAULT-NEXT: [[IDENT_CHECK:%.*]] = icmp ne i64 [[STRIDE:%.*]], 1
; DEFAULT-NEXT: br i1 [[IDENT_CHECK]], label [[FOR_BODY_PH_LVER_ORIG:%.*]], label [[FOR_BODY_PH:%.*]]
; DEFAULT: for.body.ph.lver.orig:
; DEFAULT-NEXT: br label [[FOR_BODY_LVER_ORIG:%.*]]
; DEFAULT: for.body.lver.orig:
; DEFAULT-NEXT: [[INDVARS_IV_LVER_ORIG:%.*]] = phi i64 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[INDVARS_IV_NEXT_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ]
; DEFAULT-NEXT: [[MUL_LVER_ORIG:%.*]] = mul i64 [[INDVARS_IV_LVER_ORIG]], [[STRIDE]]
; DEFAULT-NEXT: [[ARRAYIDX_LVER_ORIG:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[MUL_LVER_ORIG]]
; DEFAULT-NEXT: [[LOAD_LVER_ORIG:%.*]] = load i32, ptr [[ARRAYIDX_LVER_ORIG]], align 4
; DEFAULT-NEXT: [[ARRAYIDX2_LVER_ORIG:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[INDVARS_IV_LVER_ORIG]]
; DEFAULT-NEXT: [[LOAD_1_LVER_ORIG:%.*]] = load i32, ptr [[ARRAYIDX2_LVER_ORIG]], align 4
; DEFAULT-NEXT: [[ADD_LVER_ORIG:%.*]] = add i32 [[LOAD_1_LVER_ORIG]], [[LOAD_LVER_ORIG]]
; DEFAULT-NEXT: [[INDVARS_IV_NEXT_LVER_ORIG]] = add nuw nsw i64 [[INDVARS_IV_LVER_ORIG]], 1
; DEFAULT-NEXT: [[ARRAYIDX_NEXT_LVER_ORIG:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[INDVARS_IV_NEXT_LVER_ORIG]]
; DEFAULT-NEXT: store i32 [[ADD_LVER_ORIG]], ptr [[ARRAYIDX_NEXT_LVER_ORIG]], align 4
; DEFAULT-NEXT: [[EXITCOND_LVER_ORIG:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT_LVER_ORIG]], [[N:%.*]]
; DEFAULT-NEXT: br i1 [[EXITCOND_LVER_ORIG]], label [[FOR_END_LOOPEXIT:%.*]], label [[FOR_BODY_LVER_ORIG]]
; DEFAULT: for.body.ph:
; DEFAULT-NEXT: [[LOAD_INITIAL:%.*]] = load i32, ptr [[A]], align 4
; DEFAULT-NEXT: br label [[FOR_BODY:%.*]]
; DEFAULT: for.body:
; DEFAULT-NEXT: [[STORE_FORWARDED:%.*]] = phi i32 [ [[LOAD_INITIAL]], [[FOR_BODY_PH]] ], [ [[ADD:%.*]], [[FOR_BODY]] ]
; DEFAULT-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ 0, [[FOR_BODY_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
; DEFAULT-NEXT: [[MUL:%.*]] = mul i64 [[INDVARS_IV]], [[STRIDE]]
; DEFAULT-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[MUL]]
; DEFAULT-NEXT: [[LOAD:%.*]] = load i32, ptr [[ARRAYIDX]], align 4
; DEFAULT-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[INDVARS_IV]]
; DEFAULT-NEXT: [[LOAD_1:%.*]] = load i32, ptr [[ARRAYIDX2]], align 4
; DEFAULT-NEXT: [[ADD]] = add i32 [[LOAD_1]], [[STORE_FORWARDED]]
; DEFAULT-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
; DEFAULT-NEXT: [[ARRAYIDX_NEXT:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[INDVARS_IV_NEXT]]
; DEFAULT-NEXT: store i32 [[ADD]], ptr [[ARRAYIDX_NEXT]], align 4
; DEFAULT-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[N]]
; DEFAULT-NEXT: br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT1:%.*]], label [[FOR_BODY]]
; DEFAULT: for.end.loopexit:
; DEFAULT-NEXT: br label [[FOR_END:%.*]]
; DEFAULT: for.end.loopexit1:
; DEFAULT-NEXT: br label [[FOR_END]]
; DEFAULT: for.end:
; DEFAULT-NEXT: ret void
;
; NO-VERSION-LABEL: @f(
; NO-VERSION-NEXT: entry:
; NO-VERSION-NEXT: br label [[FOR_BODY:%.*]]
; NO-VERSION: for.body:
; NO-VERSION-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
; NO-VERSION-NEXT: [[MUL:%.*]] = mul i64 [[INDVARS_IV]], [[STRIDE:%.*]]
; NO-VERSION-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[MUL]]
; NO-VERSION-NEXT: [[LOAD:%.*]] = load i32, ptr [[ARRAYIDX]], align 4
; NO-VERSION-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[INDVARS_IV]]
; NO-VERSION-NEXT: [[LOAD_1:%.*]] = load i32, ptr [[ARRAYIDX2]], align 4
; NO-VERSION-NEXT: [[ADD:%.*]] = add i32 [[LOAD_1]], [[LOAD]]
; NO-VERSION-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
; NO-VERSION-NEXT: [[ARRAYIDX_NEXT:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[INDVARS_IV_NEXT]]
; NO-VERSION-NEXT: store i32 [[ADD]], ptr [[ARRAYIDX_NEXT]], align 4
; NO-VERSION-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[N:%.*]]
; NO-VERSION-NEXT: br i1 [[EXITCOND]], label [[FOR_END:%.*]], label [[FOR_BODY]]
; NO-VERSION: for.end:
; NO-VERSION-NEXT: ret void
;
; THRESHOLD-LABEL: @f(
; THRESHOLD-NEXT: for.body.lver.check:
; THRESHOLD-NEXT: [[IDENT_CHECK:%.*]] = icmp ne i64 [[STRIDE:%.*]], 1
; THRESHOLD-NEXT: br i1 [[IDENT_CHECK]], label [[FOR_BODY_PH_LVER_ORIG:%.*]], label [[FOR_BODY_PH:%.*]]
; THRESHOLD: for.body.ph.lver.orig:
; THRESHOLD-NEXT: br label [[FOR_BODY_LVER_ORIG:%.*]]
; THRESHOLD: for.body.lver.orig:
; THRESHOLD-NEXT: [[INDVARS_IV_LVER_ORIG:%.*]] = phi i64 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[INDVARS_IV_NEXT_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ]
; THRESHOLD-NEXT: [[MUL_LVER_ORIG:%.*]] = mul i64 [[INDVARS_IV_LVER_ORIG]], [[STRIDE]]
; THRESHOLD-NEXT: [[ARRAYIDX_LVER_ORIG:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[MUL_LVER_ORIG]]
; THRESHOLD-NEXT: [[LOAD_LVER_ORIG:%.*]] = load i32, ptr [[ARRAYIDX_LVER_ORIG]], align 4
; THRESHOLD-NEXT: [[ARRAYIDX2_LVER_ORIG:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[INDVARS_IV_LVER_ORIG]]
; THRESHOLD-NEXT: [[LOAD_1_LVER_ORIG:%.*]] = load i32, ptr [[ARRAYIDX2_LVER_ORIG]], align 4
; THRESHOLD-NEXT: [[ADD_LVER_ORIG:%.*]] = add i32 [[LOAD_1_LVER_ORIG]], [[LOAD_LVER_ORIG]]
; THRESHOLD-NEXT: [[INDVARS_IV_NEXT_LVER_ORIG]] = add nuw nsw i64 [[INDVARS_IV_LVER_ORIG]], 1
; THRESHOLD-NEXT: [[ARRAYIDX_NEXT_LVER_ORIG:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[INDVARS_IV_NEXT_LVER_ORIG]]
; THRESHOLD-NEXT: store i32 [[ADD_LVER_ORIG]], ptr [[ARRAYIDX_NEXT_LVER_ORIG]], align 4
; THRESHOLD-NEXT: [[EXITCOND_LVER_ORIG:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT_LVER_ORIG]], [[N:%.*]]
; THRESHOLD-NEXT: br i1 [[EXITCOND_LVER_ORIG]], label [[FOR_END_LOOPEXIT:%.*]], label [[FOR_BODY_LVER_ORIG]]
; THRESHOLD: for.body.ph:
; THRESHOLD-NEXT: [[LOAD_INITIAL:%.*]] = load i32, ptr [[A]], align 4
; THRESHOLD-NEXT: br label [[FOR_BODY:%.*]]
; THRESHOLD: for.body:
; THRESHOLD-NEXT: [[STORE_FORWARDED:%.*]] = phi i32 [ [[LOAD_INITIAL]], [[FOR_BODY_PH]] ], [ [[ADD:%.*]], [[FOR_BODY]] ]
; THRESHOLD-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ 0, [[FOR_BODY_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
; THRESHOLD-NEXT: [[MUL:%.*]] = mul i64 [[INDVARS_IV]], [[STRIDE]]
; THRESHOLD-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[MUL]]
; THRESHOLD-NEXT: [[LOAD:%.*]] = load i32, ptr [[ARRAYIDX]], align 4
; THRESHOLD-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[INDVARS_IV]]
; THRESHOLD-NEXT: [[LOAD_1:%.*]] = load i32, ptr [[ARRAYIDX2]], align 4
; THRESHOLD-NEXT: [[ADD]] = add i32 [[LOAD_1]], [[STORE_FORWARDED]]
; THRESHOLD-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
; THRESHOLD-NEXT: [[ARRAYIDX_NEXT:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[INDVARS_IV_NEXT]]
; THRESHOLD-NEXT: store i32 [[ADD]], ptr [[ARRAYIDX_NEXT]], align 4
; THRESHOLD-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[N]]
; THRESHOLD-NEXT: br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT1:%.*]], label [[FOR_BODY]]
; THRESHOLD: for.end.loopexit:
; THRESHOLD-NEXT: br label [[FOR_END:%.*]]
; THRESHOLD: for.end.loopexit1:
; THRESHOLD-NEXT: br label [[FOR_END]]
; THRESHOLD: for.end:
; THRESHOLD-NEXT: ret void
;
i64 %stride) {
entry:
br label %for.body
for.body: ; preds = %for.body, %entry
%indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
%mul = mul i64 %indvars.iv, %stride
%arrayidx = getelementptr inbounds i32, ptr %A, i64 %mul
%load = load i32, ptr %arrayidx, align 4
%arrayidx2 = getelementptr inbounds i32, ptr %B, i64 %indvars.iv
%load_1 = load i32, ptr %arrayidx2, align 4
%add = add i32 %load_1, %load
%indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
%arrayidx_next = getelementptr inbounds i32, ptr %A, i64 %indvars.iv.next
store i32 %add, ptr %arrayidx_next, align 4
%exitcond = icmp eq i64 %indvars.iv.next, %N
br i1 %exitcond, label %for.end, label %for.body
for.end: ; preds = %for.body
ret void
}
; Similar to @f(), but with a struct type.
define void @f_struct(ptr noalias nocapture %A, ptr noalias nocapture readonly %B, i64 %N,
;
;
;
;
;
; DEFAULT-LABEL: @f_struct(
; DEFAULT-NEXT: for.body.lver.check:
; DEFAULT-NEXT: [[IDENT_CHECK:%.*]] = icmp ne i64 [[STRIDE:%.*]], 1
; DEFAULT-NEXT: br i1 [[IDENT_CHECK]], label [[FOR_BODY_PH_LVER_ORIG:%.*]], label [[FOR_BODY_PH:%.*]]
; DEFAULT: for.body.ph.lver.orig:
; DEFAULT-NEXT: br label [[FOR_BODY_LVER_ORIG:%.*]]
; DEFAULT: for.body.lver.orig:
; DEFAULT-NEXT: [[INDVARS_IV_LVER_ORIG:%.*]] = phi i64 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[INDVARS_IV_NEXT_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ]
; DEFAULT-NEXT: [[MUL_LVER_ORIG:%.*]] = mul i64 [[INDVARS_IV_LVER_ORIG]], [[STRIDE]]
; DEFAULT-NEXT: [[ARRAYIDX_LVER_ORIG:%.*]] = getelementptr inbounds { i32, i8 }, ptr [[A:%.*]], i64 [[MUL_LVER_ORIG]]
; DEFAULT-NEXT: [[LOAD_LVER_ORIG:%.*]] = load { i32, i8 }, ptr [[ARRAYIDX_LVER_ORIG]], align 4
; DEFAULT-NEXT: [[ARRAYIDX2_LVER_ORIG:%.*]] = getelementptr inbounds { i32, i8 }, ptr [[B:%.*]], i64 [[INDVARS_IV_LVER_ORIG]]
; DEFAULT-NEXT: [[LOAD_1_LVER_ORIG:%.*]] = load { i32, i8 }, ptr [[ARRAYIDX2_LVER_ORIG]], align 4
; DEFAULT-NEXT: [[V1_LVER_ORIG:%.*]] = extractvalue { i32, i8 } [[LOAD_LVER_ORIG]], 0
; DEFAULT-NEXT: [[V2_LVER_ORIG:%.*]] = extractvalue { i32, i8 } [[LOAD_1_LVER_ORIG]], 0
; DEFAULT-NEXT: [[ADD_LVER_ORIG:%.*]] = add i32 [[V1_LVER_ORIG]], [[V2_LVER_ORIG]]
; DEFAULT-NEXT: [[INS_LVER_ORIG:%.*]] = insertvalue { i32, i8 } undef, i32 [[ADD_LVER_ORIG]], 0
; DEFAULT-NEXT: [[INDVARS_IV_NEXT_LVER_ORIG]] = add nuw nsw i64 [[INDVARS_IV_LVER_ORIG]], 1
; DEFAULT-NEXT: [[ARRAYIDX_NEXT_LVER_ORIG:%.*]] = getelementptr inbounds { i32, i8 }, ptr [[A]], i64 [[INDVARS_IV_NEXT_LVER_ORIG]]
; DEFAULT-NEXT: store { i32, i8 } [[INS_LVER_ORIG]], ptr [[ARRAYIDX_NEXT_LVER_ORIG]], align 4
; DEFAULT-NEXT: [[EXITCOND_LVER_ORIG:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT_LVER_ORIG]], [[N:%.*]]
; DEFAULT-NEXT: br i1 [[EXITCOND_LVER_ORIG]], label [[FOR_END_LOOPEXIT:%.*]], label [[FOR_BODY_LVER_ORIG]]
; DEFAULT: for.body.ph:
; DEFAULT-NEXT: [[LOAD_INITIAL:%.*]] = load { i32, i8 }, ptr [[A]], align 4
; DEFAULT-NEXT: br label [[FOR_BODY:%.*]]
; DEFAULT: for.body:
; DEFAULT-NEXT: [[STORE_FORWARDED:%.*]] = phi { i32, i8 } [ [[LOAD_INITIAL]], [[FOR_BODY_PH]] ], [ [[INS:%.*]], [[FOR_BODY]] ]
; DEFAULT-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ 0, [[FOR_BODY_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
; DEFAULT-NEXT: [[MUL:%.*]] = mul i64 [[INDVARS_IV]], [[STRIDE]]
; DEFAULT-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds { i32, i8 }, ptr [[A]], i64 [[MUL]]
; DEFAULT-NEXT: [[LOAD:%.*]] = load { i32, i8 }, ptr [[ARRAYIDX]], align 4
; DEFAULT-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds { i32, i8 }, ptr [[B]], i64 [[INDVARS_IV]]
; DEFAULT-NEXT: [[LOAD_1:%.*]] = load { i32, i8 }, ptr [[ARRAYIDX2]], align 4
; DEFAULT-NEXT: [[V1:%.*]] = extractvalue { i32, i8 } [[STORE_FORWARDED]], 0
; DEFAULT-NEXT: [[V2:%.*]] = extractvalue { i32, i8 } [[LOAD_1]], 0
; DEFAULT-NEXT: [[ADD:%.*]] = add i32 [[V1]], [[V2]]
; DEFAULT-NEXT: [[INS]] = insertvalue { i32, i8 } undef, i32 [[ADD]], 0
; DEFAULT-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
; DEFAULT-NEXT: [[ARRAYIDX_NEXT:%.*]] = getelementptr inbounds { i32, i8 }, ptr [[A]], i64 [[INDVARS_IV_NEXT]]
; DEFAULT-NEXT: store { i32, i8 } [[INS]], ptr [[ARRAYIDX_NEXT]], align 4
; DEFAULT-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[N]]
; DEFAULT-NEXT: br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT1:%.*]], label [[FOR_BODY]]
; DEFAULT: for.end.loopexit:
; DEFAULT-NEXT: br label [[FOR_END:%.*]]
; DEFAULT: for.end.loopexit1:
; DEFAULT-NEXT: br label [[FOR_END]]
; DEFAULT: for.end:
; DEFAULT-NEXT: ret void
;
; NO-VERSION-LABEL: @f_struct(
; NO-VERSION-NEXT: entry:
; NO-VERSION-NEXT: br label [[FOR_BODY:%.*]]
; NO-VERSION: for.body:
; NO-VERSION-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
; NO-VERSION-NEXT: [[MUL:%.*]] = mul i64 [[INDVARS_IV]], [[STRIDE:%.*]]
; NO-VERSION-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds { i32, i8 }, ptr [[A:%.*]], i64 [[MUL]]
; NO-VERSION-NEXT: [[LOAD:%.*]] = load { i32, i8 }, ptr [[ARRAYIDX]], align 4
; NO-VERSION-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds { i32, i8 }, ptr [[B:%.*]], i64 [[INDVARS_IV]]
; NO-VERSION-NEXT: [[LOAD_1:%.*]] = load { i32, i8 }, ptr [[ARRAYIDX2]], align 4
; NO-VERSION-NEXT: [[V1:%.*]] = extractvalue { i32, i8 } [[LOAD]], 0
; NO-VERSION-NEXT: [[V2:%.*]] = extractvalue { i32, i8 } [[LOAD_1]], 0
; NO-VERSION-NEXT: [[ADD:%.*]] = add i32 [[V1]], [[V2]]
; NO-VERSION-NEXT: [[INS:%.*]] = insertvalue { i32, i8 } undef, i32 [[ADD]], 0
; NO-VERSION-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
; NO-VERSION-NEXT: [[ARRAYIDX_NEXT:%.*]] = getelementptr inbounds { i32, i8 }, ptr [[A]], i64 [[INDVARS_IV_NEXT]]
; NO-VERSION-NEXT: store { i32, i8 } [[INS]], ptr [[ARRAYIDX_NEXT]], align 4
; NO-VERSION-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[N:%.*]]
; NO-VERSION-NEXT: br i1 [[EXITCOND]], label [[FOR_END:%.*]], label [[FOR_BODY]]
; NO-VERSION: for.end:
; NO-VERSION-NEXT: ret void
;
; THRESHOLD-LABEL: @f_struct(
; THRESHOLD-NEXT: for.body.lver.check:
; THRESHOLD-NEXT: [[IDENT_CHECK:%.*]] = icmp ne i64 [[STRIDE:%.*]], 1
; THRESHOLD-NEXT: br i1 [[IDENT_CHECK]], label [[FOR_BODY_PH_LVER_ORIG:%.*]], label [[FOR_BODY_PH:%.*]]
; THRESHOLD: for.body.ph.lver.orig:
; THRESHOLD-NEXT: br label [[FOR_BODY_LVER_ORIG:%.*]]
; THRESHOLD: for.body.lver.orig:
; THRESHOLD-NEXT: [[INDVARS_IV_LVER_ORIG:%.*]] = phi i64 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[INDVARS_IV_NEXT_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ]
; THRESHOLD-NEXT: [[MUL_LVER_ORIG:%.*]] = mul i64 [[INDVARS_IV_LVER_ORIG]], [[STRIDE]]
; THRESHOLD-NEXT: [[ARRAYIDX_LVER_ORIG:%.*]] = getelementptr inbounds { i32, i8 }, ptr [[A:%.*]], i64 [[MUL_LVER_ORIG]]
; THRESHOLD-NEXT: [[LOAD_LVER_ORIG:%.*]] = load { i32, i8 }, ptr [[ARRAYIDX_LVER_ORIG]], align 4
; THRESHOLD-NEXT: [[ARRAYIDX2_LVER_ORIG:%.*]] = getelementptr inbounds { i32, i8 }, ptr [[B:%.*]], i64 [[INDVARS_IV_LVER_ORIG]]
; THRESHOLD-NEXT: [[LOAD_1_LVER_ORIG:%.*]] = load { i32, i8 }, ptr [[ARRAYIDX2_LVER_ORIG]], align 4
; THRESHOLD-NEXT: [[V1_LVER_ORIG:%.*]] = extractvalue { i32, i8 } [[LOAD_LVER_ORIG]], 0
; THRESHOLD-NEXT: [[V2_LVER_ORIG:%.*]] = extractvalue { i32, i8 } [[LOAD_1_LVER_ORIG]], 0
; THRESHOLD-NEXT: [[ADD_LVER_ORIG:%.*]] = add i32 [[V1_LVER_ORIG]], [[V2_LVER_ORIG]]
; THRESHOLD-NEXT: [[INS_LVER_ORIG:%.*]] = insertvalue { i32, i8 } undef, i32 [[ADD_LVER_ORIG]], 0
; THRESHOLD-NEXT: [[INDVARS_IV_NEXT_LVER_ORIG]] = add nuw nsw i64 [[INDVARS_IV_LVER_ORIG]], 1
; THRESHOLD-NEXT: [[ARRAYIDX_NEXT_LVER_ORIG:%.*]] = getelementptr inbounds { i32, i8 }, ptr [[A]], i64 [[INDVARS_IV_NEXT_LVER_ORIG]]
; THRESHOLD-NEXT: store { i32, i8 } [[INS_LVER_ORIG]], ptr [[ARRAYIDX_NEXT_LVER_ORIG]], align 4
; THRESHOLD-NEXT: [[EXITCOND_LVER_ORIG:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT_LVER_ORIG]], [[N:%.*]]
; THRESHOLD-NEXT: br i1 [[EXITCOND_LVER_ORIG]], label [[FOR_END_LOOPEXIT:%.*]], label [[FOR_BODY_LVER_ORIG]]
; THRESHOLD: for.body.ph:
; THRESHOLD-NEXT: [[LOAD_INITIAL:%.*]] = load { i32, i8 }, ptr [[A]], align 4
; THRESHOLD-NEXT: br label [[FOR_BODY:%.*]]
; THRESHOLD: for.body:
; THRESHOLD-NEXT: [[STORE_FORWARDED:%.*]] = phi { i32, i8 } [ [[LOAD_INITIAL]], [[FOR_BODY_PH]] ], [ [[INS:%.*]], [[FOR_BODY]] ]
; THRESHOLD-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ 0, [[FOR_BODY_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
; THRESHOLD-NEXT: [[MUL:%.*]] = mul i64 [[INDVARS_IV]], [[STRIDE]]
; THRESHOLD-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds { i32, i8 }, ptr [[A]], i64 [[MUL]]
; THRESHOLD-NEXT: [[LOAD:%.*]] = load { i32, i8 }, ptr [[ARRAYIDX]], align 4
; THRESHOLD-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds { i32, i8 }, ptr [[B]], i64 [[INDVARS_IV]]
; THRESHOLD-NEXT: [[LOAD_1:%.*]] = load { i32, i8 }, ptr [[ARRAYIDX2]], align 4
; THRESHOLD-NEXT: [[V1:%.*]] = extractvalue { i32, i8 } [[STORE_FORWARDED]], 0
; THRESHOLD-NEXT: [[V2:%.*]] = extractvalue { i32, i8 } [[LOAD_1]], 0
; THRESHOLD-NEXT: [[ADD:%.*]] = add i32 [[V1]], [[V2]]
; THRESHOLD-NEXT: [[INS]] = insertvalue { i32, i8 } undef, i32 [[ADD]], 0
; THRESHOLD-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
; THRESHOLD-NEXT: [[ARRAYIDX_NEXT:%.*]] = getelementptr inbounds { i32, i8 }, ptr [[A]], i64 [[INDVARS_IV_NEXT]]
; THRESHOLD-NEXT: store { i32, i8 } [[INS]], ptr [[ARRAYIDX_NEXT]], align 4
; THRESHOLD-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[N]]
; THRESHOLD-NEXT: br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT1:%.*]], label [[FOR_BODY]]
; THRESHOLD: for.end.loopexit:
; THRESHOLD-NEXT: br label [[FOR_END:%.*]]
; THRESHOLD: for.end.loopexit1:
; THRESHOLD-NEXT: br label [[FOR_END]]
; THRESHOLD: for.end:
; THRESHOLD-NEXT: ret void
;
i64 %stride) {
entry:
br label %for.body
for.body: ; preds = %for.body, %entry
%indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
%mul = mul i64 %indvars.iv, %stride
%arrayidx = getelementptr inbounds { i32, i8 }, ptr %A, i64 %mul
%load = load { i32, i8 }, ptr %arrayidx, align 4
%arrayidx2 = getelementptr inbounds { i32, i8 }, ptr %B, i64 %indvars.iv
%load_1 = load { i32, i8 }, ptr %arrayidx2, align 4
%v1 = extractvalue { i32, i8 } %load, 0
%v2 = extractvalue { i32, i8} %load_1, 0
%add = add i32 %v1, %v2
%ins = insertvalue { i32, i8 } undef, i32 %add, 0
%indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
%arrayidx_next = getelementptr inbounds { i32, i8 }, ptr %A, i64 %indvars.iv.next
store { i32, i8 } %ins, ptr %arrayidx_next, align 4
%exitcond = icmp eq i64 %indvars.iv.next, %N
br i1 %exitcond, label %for.end, label %for.body
for.end: ; preds = %for.body
ret void
}
; With two symbolic strides:
;
; for (unsigned i = 0; i < 100; i++)
; A[Stride2 * (i + 1)] = A[Stride1 * i] + B[i];
define void @two_strides(ptr noalias nocapture %A, ptr noalias nocapture readonly %B, i64 %N,
;
;
;
;
;
;
; DEFAULT-LABEL: @two_strides(
; DEFAULT-NEXT: for.body.lver.check:
; DEFAULT-NEXT: [[IDENT_CHECK:%.*]] = icmp ne i64 [[STRIDE_2:%.*]], 1
; DEFAULT-NEXT: [[IDENT_CHECK1:%.*]] = icmp ne i64 [[STRIDE_1:%.*]], 1
; DEFAULT-NEXT: [[TMP0:%.*]] = or i1 [[IDENT_CHECK]], [[IDENT_CHECK1]]
; DEFAULT-NEXT: br i1 [[TMP0]], label [[FOR_BODY_PH_LVER_ORIG:%.*]], label [[FOR_BODY_PH:%.*]]
; DEFAULT: for.body.ph.lver.orig:
; DEFAULT-NEXT: br label [[FOR_BODY_LVER_ORIG:%.*]]
; DEFAULT: for.body.lver.orig:
; DEFAULT-NEXT: [[INDVARS_IV_LVER_ORIG:%.*]] = phi i64 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[INDVARS_IV_NEXT_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ]
; DEFAULT-NEXT: [[MUL_LVER_ORIG:%.*]] = mul i64 [[INDVARS_IV_LVER_ORIG]], [[STRIDE_1]]
; DEFAULT-NEXT: [[ARRAYIDX_LVER_ORIG:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[MUL_LVER_ORIG]]
; DEFAULT-NEXT: [[LOAD_LVER_ORIG:%.*]] = load i32, ptr [[ARRAYIDX_LVER_ORIG]], align 4
; DEFAULT-NEXT: [[ARRAYIDX2_LVER_ORIG:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[INDVARS_IV_LVER_ORIG]]
; DEFAULT-NEXT: [[LOAD_1_LVER_ORIG:%.*]] = load i32, ptr [[ARRAYIDX2_LVER_ORIG]], align 4
; DEFAULT-NEXT: [[ADD_LVER_ORIG:%.*]] = add i32 [[LOAD_1_LVER_ORIG]], [[LOAD_LVER_ORIG]]
; DEFAULT-NEXT: [[INDVARS_IV_NEXT_LVER_ORIG]] = add nuw nsw i64 [[INDVARS_IV_LVER_ORIG]], 1
; DEFAULT-NEXT: [[MUL_2_LVER_ORIG:%.*]] = mul i64 [[INDVARS_IV_NEXT_LVER_ORIG]], [[STRIDE_2]]
; DEFAULT-NEXT: [[ARRAYIDX_NEXT_LVER_ORIG:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[MUL_2_LVER_ORIG]]
; DEFAULT-NEXT: store i32 [[ADD_LVER_ORIG]], ptr [[ARRAYIDX_NEXT_LVER_ORIG]], align 4
; DEFAULT-NEXT: [[EXITCOND_LVER_ORIG:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT_LVER_ORIG]], [[N:%.*]]
; DEFAULT-NEXT: br i1 [[EXITCOND_LVER_ORIG]], label [[FOR_END_LOOPEXIT:%.*]], label [[FOR_BODY_LVER_ORIG]]
; DEFAULT: for.body.ph:
; DEFAULT-NEXT: [[LOAD_INITIAL:%.*]] = load i32, ptr [[A]], align 4
; DEFAULT-NEXT: br label [[FOR_BODY:%.*]]
; DEFAULT: for.body:
; DEFAULT-NEXT: [[STORE_FORWARDED:%.*]] = phi i32 [ [[LOAD_INITIAL]], [[FOR_BODY_PH]] ], [ [[ADD:%.*]], [[FOR_BODY]] ]
; DEFAULT-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ 0, [[FOR_BODY_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
; DEFAULT-NEXT: [[MUL:%.*]] = mul i64 [[INDVARS_IV]], [[STRIDE_1]]
; DEFAULT-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[MUL]]
; DEFAULT-NEXT: [[LOAD:%.*]] = load i32, ptr [[ARRAYIDX]], align 4
; DEFAULT-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[INDVARS_IV]]
; DEFAULT-NEXT: [[LOAD_1:%.*]] = load i32, ptr [[ARRAYIDX2]], align 4
; DEFAULT-NEXT: [[ADD]] = add i32 [[LOAD_1]], [[STORE_FORWARDED]]
; DEFAULT-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
; DEFAULT-NEXT: [[MUL_2:%.*]] = mul i64 [[INDVARS_IV_NEXT]], [[STRIDE_2]]
; DEFAULT-NEXT: [[ARRAYIDX_NEXT:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[MUL_2]]
; DEFAULT-NEXT: store i32 [[ADD]], ptr [[ARRAYIDX_NEXT]], align 4
; DEFAULT-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[N]]
; DEFAULT-NEXT: br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT2:%.*]], label [[FOR_BODY]]
; DEFAULT: for.end.loopexit:
; DEFAULT-NEXT: br label [[FOR_END:%.*]]
; DEFAULT: for.end.loopexit2:
; DEFAULT-NEXT: br label [[FOR_END]]
; DEFAULT: for.end:
; DEFAULT-NEXT: ret void
;
; NO-VERSION-LABEL: @two_strides(
; NO-VERSION-NEXT: entry:
; NO-VERSION-NEXT: br label [[FOR_BODY:%.*]]
; NO-VERSION: for.body:
; NO-VERSION-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
; NO-VERSION-NEXT: [[MUL:%.*]] = mul i64 [[INDVARS_IV]], [[STRIDE_1:%.*]]
; NO-VERSION-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[MUL]]
; NO-VERSION-NEXT: [[LOAD:%.*]] = load i32, ptr [[ARRAYIDX]], align 4
; NO-VERSION-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[INDVARS_IV]]
; NO-VERSION-NEXT: [[LOAD_1:%.*]] = load i32, ptr [[ARRAYIDX2]], align 4
; NO-VERSION-NEXT: [[ADD:%.*]] = add i32 [[LOAD_1]], [[LOAD]]
; NO-VERSION-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
; NO-VERSION-NEXT: [[MUL_2:%.*]] = mul i64 [[INDVARS_IV_NEXT]], [[STRIDE_2:%.*]]
; NO-VERSION-NEXT: [[ARRAYIDX_NEXT:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[MUL_2]]
; NO-VERSION-NEXT: store i32 [[ADD]], ptr [[ARRAYIDX_NEXT]], align 4
; NO-VERSION-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[N:%.*]]
; NO-VERSION-NEXT: br i1 [[EXITCOND]], label [[FOR_END:%.*]], label [[FOR_BODY]]
; NO-VERSION: for.end:
; NO-VERSION-NEXT: ret void
;
; THRESHOLD-LABEL: @two_strides(
; THRESHOLD-NEXT: entry:
; THRESHOLD-NEXT: br label [[FOR_BODY:%.*]]
; THRESHOLD: for.body:
; THRESHOLD-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
; THRESHOLD-NEXT: [[MUL:%.*]] = mul i64 [[INDVARS_IV]], [[STRIDE_1:%.*]]
; THRESHOLD-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[MUL]]
; THRESHOLD-NEXT: [[LOAD:%.*]] = load i32, ptr [[ARRAYIDX]], align 4
; THRESHOLD-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[INDVARS_IV]]
; THRESHOLD-NEXT: [[LOAD_1:%.*]] = load i32, ptr [[ARRAYIDX2]], align 4
; THRESHOLD-NEXT: [[ADD:%.*]] = add i32 [[LOAD_1]], [[LOAD]]
; THRESHOLD-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
; THRESHOLD-NEXT: [[MUL_2:%.*]] = mul i64 [[INDVARS_IV_NEXT]], [[STRIDE_2:%.*]]
; THRESHOLD-NEXT: [[ARRAYIDX_NEXT:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[MUL_2]]
; THRESHOLD-NEXT: store i32 [[ADD]], ptr [[ARRAYIDX_NEXT]], align 4
; THRESHOLD-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[N:%.*]]
; THRESHOLD-NEXT: br i1 [[EXITCOND]], label [[FOR_END:%.*]], label [[FOR_BODY]]
; THRESHOLD: for.end:
; THRESHOLD-NEXT: ret void
;
i64 %stride.1, i64 %stride.2) {
entry:
br label %for.body
for.body: ; preds = %for.body, %entry
%indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
%mul = mul i64 %indvars.iv, %stride.1
%arrayidx = getelementptr inbounds i32, ptr %A, i64 %mul
%load = load i32, ptr %arrayidx, align 4
%arrayidx2 = getelementptr inbounds i32, ptr %B, i64 %indvars.iv
%load_1 = load i32, ptr %arrayidx2, align 4
%add = add i32 %load_1, %load
%indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
%mul.2 = mul i64 %indvars.iv.next, %stride.2
%arrayidx_next = getelementptr inbounds i32, ptr %A, i64 %mul.2
store i32 %add, ptr %arrayidx_next, align 4
%exitcond = icmp eq i64 %indvars.iv.next, %N
br i1 %exitcond, label %for.end, label %for.body
for.end: ; preds = %for.body
ret void
}
Codebase Browser
llvm