[llvm] 7eea55f - LoopLoadElim: re-org tests after invalid #96656 (#97598)
via llvm-commits
llvm-commits at lists.llvm.org
Mon Sep 30 07:46:37 PDT 2024
Author: Ramkumar Ramachandra
Date: 2024-09-30T15:46:34+01:00
New Revision: 7eea55fd4bf197fea20b7c255febf287664dce36
URL: https://github.com/llvm/llvm-project/commit/7eea55fd4bf197fea20b7c255febf287664dce36
DIFF: https://github.com/llvm/llvm-project/commit/7eea55fd4bf197fea20b7c255febf287664dce36.diff
LOG: LoopLoadElim: re-org tests after invalid #96656 (#97598)
After pr96656.ll were added to LAA and LoopVersioning, it was decided
that the bug is in a caller of LoopVersioning, not in LAA or
LoopVersioning itself. The new candidate was LoopLoadElim, but #96656
has since been marked invalid. Hence, re-organize the added tests to
avoid confusion, and the testcase from the investigation to
LoopLoadElim.
Added:
llvm/test/Transforms/LoopLoadElim/unknown-stride-known-dep.ll
llvm/test/Transforms/LoopVersioning/single-iteration.ll
Modified:
llvm/test/Analysis/LoopAccessAnalysis/symbolic-stride.ll
Removed:
llvm/test/Analysis/LoopAccessAnalysis/pr96656.ll
llvm/test/Transforms/LoopVersioning/pr96656.ll
################################################################################
diff --git a/llvm/test/Analysis/LoopAccessAnalysis/pr96656.ll b/llvm/test/Analysis/LoopAccessAnalysis/pr96656.ll
deleted file mode 100644
index 5b9833553fa02c..00000000000000
--- a/llvm/test/Analysis/LoopAccessAnalysis/pr96656.ll
+++ /dev/null
@@ -1,49 +0,0 @@
-; NOTE: Assertions have been autogenerated by utils/update_analyze_test_checks.py UTC_ARGS: --version 5
-; RUN: opt -passes='print<access-info>' -disable-output %s 2>&1 | FileCheck %s
-
-define void @false.equal.predicate(ptr %arg, ptr %arg1, i1 %arg2) {
-; CHECK-LABEL: 'false.equal.predicate'
-; CHECK-NEXT: loop.body:
-; CHECK-NEXT: Memory dependences are safe
-; CHECK-NEXT: Dependences:
-; CHECK-NEXT: Run-time memory checks:
-; CHECK-NEXT: Grouped accesses:
-; CHECK-EMPTY:
-; CHECK-NEXT: Non vectorizable stores to invariant address were not found in loop.
-; CHECK-NEXT: SCEV assumptions:
-; CHECK-NEXT: Equal predicate: %load == 1
-; CHECK-EMPTY:
-; CHECK-NEXT: Expressions re-written:
-; CHECK-NEXT: [PSE] %gep10 = getelementptr double, ptr %gep8, i64 %mul:
-; CHECK-NEXT: {(8 + %arg1),+,(8 * (sext i32 %load to i64))<nsw>}<%loop.body>
-; CHECK-NEXT: --> {(8 + %arg1),+,8}<%loop.body>
-;
-entry:
- %load = load i32, ptr %arg, align 4
- br i1 %arg2, label %noloop.exit, label %loop.ph
-
-loop.ph: ; preds = %entry
- %sext7 = sext i32 %load to i64
- %gep8 = getelementptr i8, ptr %arg1, i64 8
- br label %loop.body
-
-loop.body: ; preds = %loop.body, %loop.ph
- %phi = phi i64 [ 0, %loop.ph ], [ %add, %loop.body ]
- %mul = mul i64 %phi, %sext7
- %gep10 = getelementptr double, ptr %gep8, i64 %mul
- %load11 = load double, ptr %gep10, align 8
- store double %load11, ptr %arg1, align 8
- %add = add i64 %phi, 1
- %icmp = icmp eq i64 %phi, 0
- br i1 %icmp, label %loop.exit, label %loop.body
-
-noloop.exit: ; preds = %entry
- %sext = sext i32 %load to i64
- %gep = getelementptr double, ptr %arg1, i64 %sext
- %load5 = load double, ptr %gep, align 8
- store double %load5, ptr %arg, align 8
- ret void
-
-loop.exit: ; preds = %loop.body
- ret void
-}
diff --git a/llvm/test/Analysis/LoopAccessAnalysis/symbolic-stride.ll b/llvm/test/Analysis/LoopAccessAnalysis/symbolic-stride.ll
index 1e12dbf3bbee31..1585c7b5628063 100644
--- a/llvm/test/Analysis/LoopAccessAnalysis/symbolic-stride.ll
+++ b/llvm/test/Analysis/LoopAccessAnalysis/symbolic-stride.ll
@@ -223,6 +223,54 @@ exit:
ret void
}
+define double @single_iteration_unknown_stride(i32 %x, ptr %y, i1 %cond) {
+; CHECK-LABEL: 'single_iteration_unknown_stride'
+; CHECK-NEXT: loop.body:
+; CHECK-NEXT: Memory dependences are safe
+; CHECK-NEXT: Dependences:
+; CHECK-NEXT: Run-time memory checks:
+; CHECK-NEXT: Grouped accesses:
+; CHECK-EMPTY:
+; CHECK-NEXT: Non vectorizable stores to invariant address were not found in loop.
+; CHECK-NEXT: SCEV assumptions:
+; CHECK-NEXT: Equal predicate: %x == 1
+; CHECK-EMPTY:
+; CHECK-NEXT: Expressions re-written:
+; CHECK-NEXT: [PSE] %gep10 = getelementptr double, ptr %gep8, i64 %mul:
+; CHECK-NEXT: {(8 + %y),+,(8 * (sext i32 %x to i64))<nsw>}<%loop.body>
+; CHECK-NEXT: --> {(8 + %y),+,8}<%loop.body>
+;
+entry:
+ br i1 %cond, label %noloop.exit, label %loop.ph
+
+loop.ph: ; preds = %entry
+ %sext7 = sext i32 %x to i64
+ %gep8 = getelementptr i8, ptr %y, i64 8
+ br label %loop.body
+
+loop.body: ; preds = %loop.body, %loop.ph
+ %iv = phi i64 [ 0, %loop.ph ], [ %iv.next, %loop.body ]
+ %mul = mul i64 %iv, %sext7
+ %gep10 = getelementptr double, ptr %gep8, i64 %mul
+ %load11 = load double, ptr %gep10, align 8
+ store double %load11, ptr %y, align 8
+ %iv.next = add i64 %iv, 1
+ %icmp = icmp eq i64 %iv, 0
+ br i1 %icmp, label %loop.exit, label %loop.body
+
+noloop.exit: ; preds = %entry
+ %sext = sext i32 %x to i64
+ %gep = getelementptr double, ptr %y, i64 %sext
+ %load5 = load double, ptr %gep, align 8
+ ret double %load5
+
+loop.exit: ; preds = %loop.body
+ %sext2 = sext i32 %x to i64
+ %gep2 = getelementptr double, ptr %y, i64 %sext2
+ %load6 = load double, ptr %gep2, align 8
+ ret double %load6
+}
+
; A loop with two symbolic strides.
define void @two_strides(ptr noalias %A, ptr noalias %B, i64 %N, i64 %stride.1, i64 %stride.2) {
; CHECK-LABEL: 'two_strides'
diff --git a/llvm/test/Transforms/LoopLoadElim/unknown-stride-known-dep.ll b/llvm/test/Transforms/LoopLoadElim/unknown-stride-known-dep.ll
new file mode 100644
index 00000000000000..e7b0968c8e8264
--- /dev/null
+++ b/llvm/test/Transforms/LoopLoadElim/unknown-stride-known-dep.ll
@@ -0,0 +1,91 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -passes=loop-load-elim -S %s | FileCheck %s
+
+; The test was originally written as part of the investigation of #96656.
+; The bug has now been marked as invalid, and we keep the test to show
+; LLE's operation on known dependence returned by LAA.
+
+define void @unknown_stride_known_dependence(ptr %x, ptr %y, i1 %cond) {
+; CHECK-LABEL: define void @unknown_stride_known_dependence(
+; CHECK-SAME: ptr [[X:%.*]], ptr [[Y:%.*]], i1 [[COND:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[LOAD:%.*]] = load i32, ptr [[X]], align 4
+; CHECK-NEXT: br i1 [[COND]], label %[[NOLOOP_EXIT:.*]], label %[[LOOP_LVER_CHECK:.*]]
+; CHECK: [[LOOP_LVER_CHECK]]:
+; CHECK-NEXT: [[SEXT_X:%.*]] = sext i32 [[LOAD]] to i64
+; CHECK-NEXT: [[GEP_8:%.*]] = getelementptr i8, ptr [[Y]], i64 8
+; CHECK-NEXT: [[GEP_16:%.*]] = getelementptr i8, ptr [[Y]], i64 16
+; CHECK-NEXT: [[IDENT_CHECK:%.*]] = icmp ne i32 [[LOAD]], 1
+; CHECK-NEXT: br i1 [[IDENT_CHECK]], label %[[LOOP_PH_LVER_ORIG:.*]], label %[[LOOP_PH:.*]]
+; CHECK: [[LOOP_PH_LVER_ORIG]]:
+; CHECK-NEXT: br label %[[LOOP_LVER_ORIG:.*]]
+; CHECK: [[LOOP_LVER_ORIG]]:
+; CHECK-NEXT: [[IV_LVER_ORIG:%.*]] = phi i64 [ 0, %[[LOOP_PH_LVER_ORIG]] ], [ [[IV_NEXT_LVER_ORIG:%.*]], %[[LOOP_LVER_ORIG]] ]
+; CHECK-NEXT: [[MUL_LVER_ORIG:%.*]] = mul i64 [[IV_LVER_ORIG]], [[SEXT_X]]
+; CHECK-NEXT: [[GEP_8_MUL_LVER_ORIG:%.*]] = getelementptr double, ptr [[GEP_8]], i64 [[MUL_LVER_ORIG]]
+; CHECK-NEXT: [[LOAD_8_LVER_ORIG:%.*]] = load double, ptr [[GEP_8_MUL_LVER_ORIG]], align 8
+; CHECK-NEXT: [[GEP_16_MUL_LVER_ORIG:%.*]] = getelementptr double, ptr [[GEP_16]], i64 [[MUL_LVER_ORIG]]
+; CHECK-NEXT: store double [[LOAD_8_LVER_ORIG]], ptr [[GEP_16_MUL_LVER_ORIG]], align 8
+; CHECK-NEXT: [[IV_NEXT_LVER_ORIG]] = add i64 [[IV_LVER_ORIG]], 1
+; CHECK-NEXT: [[ICMP_LVER_ORIG:%.*]] = icmp eq i64 [[IV_LVER_ORIG]], 1
+; CHECK-NEXT: br i1 [[ICMP_LVER_ORIG]], label %[[EXIT_LOOPEXIT_LOOPEXIT:.*]], label %[[LOOP_LVER_ORIG]]
+; CHECK: [[LOOP_PH]]:
+; CHECK-NEXT: [[LOAD_INITIAL:%.*]] = load double, ptr [[GEP_8]], align 8
+; CHECK-NEXT: br label %[[LOOP:.*]]
+; CHECK: [[LOOP]]:
+; CHECK-NEXT: [[STORE_FORWARDED:%.*]] = phi double [ [[LOAD_INITIAL]], %[[LOOP_PH]] ], [ [[STORE_FORWARDED]], %[[LOOP]] ]
+; CHECK-NEXT: [[IV:%.*]] = phi i64 [ 0, %[[LOOP_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT: [[MUL:%.*]] = mul i64 [[IV]], [[SEXT_X]]
+; CHECK-NEXT: [[GEP_8_MUL:%.*]] = getelementptr double, ptr [[GEP_8]], i64 [[MUL]]
+; CHECK-NEXT: [[LOAD_8:%.*]] = load double, ptr [[GEP_8_MUL]], align 8
+; CHECK-NEXT: [[GEP_16_MUL:%.*]] = getelementptr double, ptr [[GEP_16]], i64 [[MUL]]
+; CHECK-NEXT: store double [[STORE_FORWARDED]], ptr [[GEP_16_MUL]], align 8
+; CHECK-NEXT: [[IV_NEXT]] = add i64 [[IV]], 1
+; CHECK-NEXT: [[ICMP:%.*]] = icmp eq i64 [[IV]], 1
+; CHECK-NEXT: br i1 [[ICMP]], label %[[EXIT_LOOPEXIT_LOOPEXIT1:.*]], label %[[LOOP]]
+; CHECK: [[NOLOOP_EXIT]]:
+; CHECK-NEXT: [[SEXT:%.*]] = sext i32 [[LOAD]] to i64
+; CHECK-NEXT: [[GEP_Y:%.*]] = getelementptr double, ptr [[Y]], i64 [[SEXT]]
+; CHECK-NEXT: [[LOAD_Y:%.*]] = load double, ptr [[GEP_Y]], align 8
+; CHECK-NEXT: store double [[LOAD_Y]], ptr [[X]], align 8
+; CHECK-NEXT: br label %[[EXIT:.*]]
+; CHECK: [[EXIT_LOOPEXIT_LOOPEXIT]]:
+; CHECK-NEXT: br label %[[EXIT_LOOPEXIT:.*]]
+; CHECK: [[EXIT_LOOPEXIT_LOOPEXIT1]]:
+; CHECK-NEXT: br label %[[EXIT_LOOPEXIT]]
+; CHECK: [[EXIT_LOOPEXIT]]:
+; CHECK-NEXT: br label %[[EXIT]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret void
+;
+entry:
+ %load = load i32, ptr %x, align 4
+ br i1 %cond, label %noloop.exit, label %loop.ph
+
+loop.ph: ; preds = %entry
+ %sext.x = sext i32 %load to i64
+ %gep.8 = getelementptr i8, ptr %y, i64 8
+ %gep.16 = getelementptr i8, ptr %y, i64 16
+ br label %loop
+
+loop: ; preds = %loop, %loop.ph
+ %iv = phi i64 [ 0, %loop.ph ], [ %iv.next, %loop ]
+ %mul = mul i64 %iv, %sext.x
+ %gep.8.mul = getelementptr double, ptr %gep.8, i64 %mul
+ %load.8 = load double, ptr %gep.8.mul, align 8
+ %gep.16.mul = getelementptr double, ptr %gep.16, i64 %mul
+ store double %load.8, ptr %gep.16.mul
+ %iv.next = add i64 %iv, 1
+ %icmp = icmp eq i64 %iv, 1
+ br i1 %icmp, label %exit, label %loop
+
+noloop.exit: ; preds = %loop.ph
+ %sext = sext i32 %load to i64
+ %gep.y = getelementptr double, ptr %y, i64 %sext
+ %load.y = load double, ptr %gep.y
+ store double %load.y, ptr %x
+ br label %exit
+
+exit: ; preds = %loop.body
+ ret void
+}
diff --git a/llvm/test/Transforms/LoopVersioning/pr96656.ll b/llvm/test/Transforms/LoopVersioning/single-iteration.ll
similarity index 59%
rename from llvm/test/Transforms/LoopVersioning/pr96656.ll
rename to llvm/test/Transforms/LoopVersioning/single-iteration.ll
index 0264fe40a94302..ffb9c7d4cd5ab0 100644
--- a/llvm/test/Transforms/LoopVersioning/pr96656.ll
+++ b/llvm/test/Transforms/LoopVersioning/single-iteration.ll
@@ -1,16 +1,20 @@
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
; RUN: opt -passes=loop-versioning -S %s | FileCheck %s
-define void @lver.check.unnecessary(ptr %arg, ptr %arg1, i1 %arg2) {
-; CHECK-LABEL: define void @lver.check.unnecessary(
-; CHECK-SAME: ptr [[ARG:%.*]], ptr [[ARG1:%.*]], i1 [[ARG2:%.*]]) {
+; Callers should not call LoopVersioning on single-iteration loops, as it
+; is very likely not profitable.
+; LoopVersioning faithfully versions single-iteration loops when the stride
+; is unknown.
+
+define double @single_iteration_unknown_stride(i32 %x, ptr %y, i1 %cond) {
+; CHECK-LABEL: define double @single_iteration_unknown_stride(
+; CHECK-SAME: i32 [[X:%.*]], ptr [[Y:%.*]], i1 [[COND:%.*]]) {
; CHECK-NEXT: [[ENTRY:.*:]]
-; CHECK-NEXT: [[LOAD:%.*]] = load i32, ptr [[ARG]], align 4
-; CHECK-NEXT: br i1 [[ARG2]], label %[[NOLOOP_EXIT:.*]], label %[[LOOP_BODY_LVER_CHECK:.*]]
+; CHECK-NEXT: br i1 [[COND]], label %[[NOLOOP_EXIT:.*]], label %[[LOOP_BODY_LVER_CHECK:.*]]
; CHECK: [[LOOP_BODY_LVER_CHECK]]:
-; CHECK-NEXT: [[SEXT7:%.*]] = sext i32 [[LOAD]] to i64
-; CHECK-NEXT: [[GEP8:%.*]] = getelementptr i8, ptr [[ARG1]], i64 8
-; CHECK-NEXT: [[IDENT_CHECK:%.*]] = icmp ne i32 [[LOAD]], 1
+; CHECK-NEXT: [[SEXT7:%.*]] = sext i32 [[X]] to i64
+; CHECK-NEXT: [[GEP8:%.*]] = getelementptr i8, ptr [[Y]], i64 8
+; CHECK-NEXT: [[IDENT_CHECK:%.*]] = icmp ne i32 [[X]], 1
; CHECK-NEXT: br i1 [[IDENT_CHECK]], label %[[LOOP_BODY_PH_LVER_ORIG:.*]], label %[[LOOP_BODY_PH:.*]]
; CHECK: [[LOOP_BODY_PH_LVER_ORIG]]:
; CHECK-NEXT: br label %[[LOOP_BODY_LVER_ORIG:.*]]
@@ -19,7 +23,7 @@ define void @lver.check.unnecessary(ptr %arg, ptr %arg1, i1 %arg2) {
; CHECK-NEXT: [[MUL_LVER_ORIG:%.*]] = mul i64 [[PHI_LVER_ORIG]], [[SEXT7]]
; CHECK-NEXT: [[GEP10_LVER_ORIG:%.*]] = getelementptr double, ptr [[GEP8]], i64 [[MUL_LVER_ORIG]]
; CHECK-NEXT: [[LOAD11_LVER_ORIG:%.*]] = load double, ptr [[GEP10_LVER_ORIG]], align 8
-; CHECK-NEXT: store double [[LOAD11_LVER_ORIG]], ptr [[ARG1]], align 8
+; CHECK-NEXT: store double [[LOAD11_LVER_ORIG]], ptr [[Y]], align 8
; CHECK-NEXT: [[ADD_LVER_ORIG]] = add i64 [[PHI_LVER_ORIG]], 1
; CHECK-NEXT: [[ICMP_LVER_ORIG:%.*]] = icmp eq i64 [[PHI_LVER_ORIG]], 0
; CHECK-NEXT: br i1 [[ICMP_LVER_ORIG]], label %[[LOOP_EXIT_LOOPEXIT:.*]], label %[[LOOP_BODY_LVER_ORIG]]
@@ -30,49 +34,52 @@ define void @lver.check.unnecessary(ptr %arg, ptr %arg1, i1 %arg2) {
; CHECK-NEXT: [[MUL:%.*]] = mul i64 [[PHI]], [[SEXT7]]
; CHECK-NEXT: [[GEP10:%.*]] = getelementptr double, ptr [[GEP8]], i64 [[MUL]]
; CHECK-NEXT: [[LOAD11:%.*]] = load double, ptr [[GEP10]], align 8
-; CHECK-NEXT: store double [[LOAD11]], ptr [[ARG1]], align 8
+; CHECK-NEXT: store double [[LOAD11]], ptr [[Y]], align 8
; CHECK-NEXT: [[ADD]] = add i64 [[PHI]], 1
; CHECK-NEXT: [[ICMP:%.*]] = icmp eq i64 [[PHI]], 0
; CHECK-NEXT: br i1 [[ICMP]], label %[[LOOP_EXIT_LOOPEXIT1:.*]], label %[[LOOP_BODY]]
; CHECK: [[NOLOOP_EXIT]]:
-; CHECK-NEXT: [[SEXT:%.*]] = sext i32 [[LOAD]] to i64
-; CHECK-NEXT: [[GEP:%.*]] = getelementptr double, ptr [[ARG1]], i64 [[SEXT]]
+; CHECK-NEXT: [[SEXT:%.*]] = sext i32 [[X]] to i64
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr double, ptr [[Y]], i64 [[SEXT]]
; CHECK-NEXT: [[LOAD5:%.*]] = load double, ptr [[GEP]], align 8
-; CHECK-NEXT: store double [[LOAD5]], ptr [[ARG]], align 8
-; CHECK-NEXT: ret void
+; CHECK-NEXT: ret double [[LOAD5]]
; CHECK: [[LOOP_EXIT_LOOPEXIT]]:
; CHECK-NEXT: br label %[[LOOP_EXIT:.*]]
; CHECK: [[LOOP_EXIT_LOOPEXIT1]]:
; CHECK-NEXT: br label %[[LOOP_EXIT]]
; CHECK: [[LOOP_EXIT]]:
-; CHECK-NEXT: ret void
+; CHECK-NEXT: [[SEXT2:%.*]] = sext i32 [[X]] to i64
+; CHECK-NEXT: [[GEP2:%.*]] = getelementptr double, ptr [[Y]], i64 [[SEXT2]]
+; CHECK-NEXT: [[LOAD6:%.*]] = load double, ptr [[GEP2]], align 8
+; CHECK-NEXT: ret double [[LOAD6]]
;
entry:
- %load = load i32, ptr %arg, align 4
- br i1 %arg2, label %noloop.exit, label %loop.ph
+ br i1 %cond, label %noloop.exit, label %loop.ph
loop.ph: ; preds = %entry
- %sext7 = sext i32 %load to i64
- %gep8 = getelementptr i8, ptr %arg1, i64 8
+ %sext7 = sext i32 %x to i64
+ %gep8 = getelementptr i8, ptr %y, i64 8
br label %loop.body
loop.body: ; preds = %loop.body, %loop.ph
- %phi = phi i64 [ 0, %loop.ph ], [ %add, %loop.body ]
- %mul = mul i64 %phi, %sext7
+ %iv = phi i64 [ 0, %loop.ph ], [ %iv.next, %loop.body ]
+ %mul = mul i64 %iv, %sext7
%gep10 = getelementptr double, ptr %gep8, i64 %mul
%load11 = load double, ptr %gep10, align 8
- store double %load11, ptr %arg1, align 8
- %add = add i64 %phi, 1
- %icmp = icmp eq i64 %phi, 0
+ store double %load11, ptr %y, align 8
+ %iv.next = add i64 %iv, 1
+ %icmp = icmp eq i64 %iv, 0
br i1 %icmp, label %loop.exit, label %loop.body
noloop.exit: ; preds = %entry
- %sext = sext i32 %load to i64
- %gep = getelementptr double, ptr %arg1, i64 %sext
+ %sext = sext i32 %x to i64
+ %gep = getelementptr double, ptr %y, i64 %sext
%load5 = load double, ptr %gep, align 8
- store double %load5, ptr %arg, align 8
- ret void
+ ret double %load5
loop.exit: ; preds = %loop.body
- ret void
+ %sext2 = sext i32 %x to i64
+ %gep2 = getelementptr double, ptr %y, i64 %sext2
+ %load6 = load double, ptr %gep2, align 8
+ ret double %load6
}
More information about the llvm-commits
mailing list