[PATCH] D128988: [GlobalStatus] Handle masked load intrinsics as special loads

[Tiehu Zhang via Phabricator via llvm-commits]

TiehuZhang created this revision.
TiehuZhang added reviewers: dmgreen, fhahn, sdesmalen.
Herald added a subscriber: hiraditya.
Herald added a project: All.
TiehuZhang requested review of this revision.
Herald added a project: LLVM.
Herald added a subscriber: llvm-commits.

GlobalOpt does not optimize global variables used by calls, even for masked load intrinsics. This patch treats masked load intrinsics as special loads so that global variables used by such calls can also be optimized later.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D128988

Files:
  llvm/lib/Transforms/Utils/GlobalStatus.cpp
  llvm/test/Transforms/GlobalOpt/masked-load-global.ll


Index: llvm/test/Transforms/GlobalOpt/masked-load-global.ll
===================================================================
--- llvm/test/Transforms/GlobalOpt/masked-load-global.ll
+++ llvm/test/Transforms/GlobalOpt/masked-load-global.ll
@@ -11,6 +11,7 @@
 
 ; Function Attrs: nofree norecurse nosync nounwind uwtable vscale_range(1,16)
 define dso_local i32 @main(i32 %argc, i8** nocapture readnone %argv) local_unnamed_addr #0 {
+;
 ; CHECK-LABEL: @main(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 2 x i64> @llvm.experimental.stepvector.nxv2i64()
@@ -50,24 +51,20 @@
 ; CHECK-NEXT:    [[ARRAYIDX12_I:%.*]] = getelementptr inbounds [16 x i32], [16 x i32]* @lPtr, i64 0, i64 [[INDVARS_IV_I]]
 ; CHECK-NEXT:    [[TMP13:%.*]] = load i32, i32* [[ARRAYIDX12_I]], align 4
 ; CHECK-NEXT:    [[IDXPROM13_I:%.*]] = sext i32 [[TMP13]] to i64
-; CHECK-NEXT:    [[ARRAYIDX14_I:%.*]] = getelementptr inbounds [16 x double], [16 x double]* @psiPtr, i64 0, i64 [[IDXPROM13_I]]
-; CHECK-NEXT:    [[TMP14:%.*]] = load double, double* [[ARRAYIDX14_I]], align 8
-; CHECK-NEXT:    [[MUL15_I:%.*]] = fmul contract double [[TMP12]], [[TMP14]]
+; CHECK-NEXT:    [[MUL15_I:%.*]] = fmul contract double [[TMP12]], 0.000000e+00
 ; CHECK-NEXT:    [[ARRAYIDX17_I:%.*]] = getelementptr inbounds [16 x i32], [16 x i32]* @uPtr, i64 0, i64 [[INDVARS_IV_I]]
-; CHECK-NEXT:    [[TMP15:%.*]] = load i32, i32* [[ARRAYIDX17_I]], align 4
-; CHECK-NEXT:    [[IDXPROM18_I:%.*]] = sext i32 [[TMP15]] to i64
+; CHECK-NEXT:    [[TMP14:%.*]] = load i32, i32* [[ARRAYIDX17_I]], align 4
+; CHECK-NEXT:    [[IDXPROM18_I:%.*]] = sext i32 [[TMP14]] to i64
 ; CHECK-NEXT:    [[ARRAYIDX19_I:%.*]] = getelementptr inbounds [16 x double], [16 x double]* @ApsiPtr, i64 0, i64 [[IDXPROM18_I]]
-; CHECK-NEXT:    [[TMP16:%.*]] = load double, double* [[ARRAYIDX19_I]], align 8
-; CHECK-NEXT:    [[ADD_I:%.*]] = fadd contract double [[MUL15_I]], [[TMP16]]
+; CHECK-NEXT:    [[TMP15:%.*]] = load double, double* [[ARRAYIDX19_I]], align 8
+; CHECK-NEXT:    [[ADD_I:%.*]] = fadd contract double [[MUL15_I]], [[TMP15]]
 ; CHECK-NEXT:    store double [[ADD_I]], double* [[ARRAYIDX19_I]], align 8
 ; CHECK-NEXT:    [[ARRAYIDX21_I:%.*]] = getelementptr inbounds [16 x double], [16 x double]* @upperPtr, i64 0, i64 [[INDVARS_IV_I]]
-; CHECK-NEXT:    [[TMP17:%.*]] = load double, double* [[ARRAYIDX21_I]], align 8
-; CHECK-NEXT:    [[ARRAYIDX25_I:%.*]] = getelementptr inbounds [16 x double], [16 x double]* @psiPtr, i64 0, i64 [[IDXPROM18_I]]
-; CHECK-NEXT:    [[TMP18:%.*]] = load double, double* [[ARRAYIDX25_I]], align 8
-; CHECK-NEXT:    [[MUL26_I:%.*]] = fmul contract double [[TMP17]], [[TMP18]]
+; CHECK-NEXT:    [[TMP16:%.*]] = load double, double* [[ARRAYIDX21_I]], align 8
+; CHECK-NEXT:    [[MUL26_I:%.*]] = fmul contract double [[TMP16]], 0.000000e+00
 ; CHECK-NEXT:    [[ARRAYIDX30_I:%.*]] = getelementptr inbounds [16 x double], [16 x double]* @ApsiPtr, i64 0, i64 [[IDXPROM13_I]]
-; CHECK-NEXT:    [[TMP19:%.*]] = load double, double* [[ARRAYIDX30_I]], align 8
-; CHECK-NEXT:    [[ADD31_I:%.*]] = fadd contract double [[TMP19]], [[MUL26_I]]
+; CHECK-NEXT:    [[TMP17:%.*]] = load double, double* [[ARRAYIDX30_I]], align 8
+; CHECK-NEXT:    [[ADD31_I:%.*]] = fadd contract double [[TMP17]], [[MUL26_I]]
 ; CHECK-NEXT:    store double [[ADD31_I]], double* [[ARRAYIDX30_I]], align 8
 ; CHECK-NEXT:    [[INDVARS_IV_NEXT_I]] = add nuw nsw i64 [[INDVARS_IV_I]], 1
 ; CHECK-NEXT:    [[EXITCOND_NOT_I:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT_I]], 16
Index: llvm/lib/Transforms/Utils/GlobalStatus.cpp
===================================================================
--- llvm/lib/Transforms/Utils/GlobalStatus.cpp
+++ llvm/lib/Transforms/Utils/GlobalStatus.cpp
@@ -172,9 +172,17 @@
           return true;
         GS.StoredType = GlobalStatus::Stored;
       } else if (const auto *CB = dyn_cast<CallBase>(I)) {
-        if (!CB->isCallee(&U))
-          return true;
-        GS.IsLoaded = true;
+        const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
+        if (II && II->getIntrinsicID() == Intrinsic::masked_load) {
+          if (II->isVolatile())
+            return true;
+          GS.Ordering =
+              strongerOrdering(GS.Ordering, AtomicOrdering::NotAtomic);
+        } else {
+          if (!CB->isCallee(&U))
+            return true;
+          GS.IsLoaded = true;
+        }
       } else {
         return true; // Any other non-load instruction might take address!
       }


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D128988.441673.patch
Type: text/x-patch
Size: 4484 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20220701/20336d2a/attachment.bin>