[llvm] 690bed0 - [X86][AMX] Fix infinite loop of getShape.

[via llvm-commits]

Author: Luo, Yuanke
Date: 2022-04-10T14:44:51+08:00
New Revision: 690bed0cec7e3359b0136acd98c0940aed5d0722

URL: https://github.com/llvm/llvm-project/commit/690bed0cec7e3359b0136acd98c0940aed5d0722
DIFF: https://github.com/llvm/llvm-project/commit/690bed0cec7e3359b0136acd98c0940aed5d0722.diff

LOG: [X86][AMX] Fix infinite loop of getShape.

When walk the user chain to get the shape of a phi node. If it is phi
node in the chain, we should walk to the user of this phi node instead
of the original phi node.

Added: 
    

Modified: 
    llvm/lib/Target/X86/X86LowerAMXType.cpp
    llvm/test/CodeGen/X86/AMX/amx-combine-undef.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/X86/X86LowerAMXType.cpp b/llvm/lib/Target/X86/X86LowerAMXType.cpp
index 0fb8d07b6b61e..1596f3cb0bb5e 100644
--- a/llvm/lib/Target/X86/X86LowerAMXType.cpp
+++ b/llvm/lib/Target/X86/X86LowerAMXType.cpp
@@ -200,7 +200,7 @@ static std::pair<Value *, Value *> getShape(PHINode *Phi) {
     } else if (isa<PHINode>(V)) {
       if (V->use_empty())
         break;
-      Use &U = *(Phi->use_begin());
+      Use &U = *(V->use_begin());
       V = U.getUser();
     } else {
       break;

diff  --git a/llvm/test/CodeGen/X86/AMX/amx-combine-undef.ll b/llvm/test/CodeGen/X86/AMX/amx-combine-undef.ll
index 67e4f3eea42bc..104fa5703e370 100644
--- a/llvm/test/CodeGen/X86/AMX/amx-combine-undef.ll
+++ b/llvm/test/CodeGen/X86/AMX/amx-combine-undef.ll
@@ -1,6 +1,46 @@
 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt --codegen-opt-level=2 -mtriple=x86_64 -lower-amx-type %s -S | FileCheck %s
 
+define void @undef_2phi(i8 *%buf) {
+; CHECK-LABEL: @undef_2phi(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = call x86_amx @llvm.x86.tilezero.internal(i16 8, i16 32)
+; CHECK-NEXT:    br i1 undef, label [[L1:%.*]], label [[L2:%.*]]
+; CHECK:       l1:
+; CHECK-NEXT:    [[T1:%.*]] = call x86_amx @llvm.x86.tilezero.internal(i16 8, i16 32)
+; CHECK-NEXT:    br i1 undef, label [[L2]], label [[L3:%.*]]
+; CHECK:       l2:
+; CHECK-NEXT:    [[TMP1:%.*]] = phi x86_amx [ [[TMP0]], [[ENTRY:%.*]] ], [ [[T1]], [[L1]] ]
+; CHECK-NEXT:    br i1 undef, label [[L3]], label [[EXIT:%.*]]
+; CHECK:       l3:
+; CHECK-NEXT:    [[TMP2:%.*]] = phi x86_amx [ [[TMP1]], [[L2]] ], [ [[T1]], [[L1]] ]
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 8, i16 32, i8* [[BUF:%.*]], i64 1024, x86_amx [[TMP2]])
+; CHECK-NEXT:    br label [[EXIT]]
+; CHECK:       exit:
+; CHECK-NEXT:    ret void
+;
+entry:
+  br i1 undef, label %l1, label %l2
+
+l1:
+  %t1 = call x86_amx @llvm.x86.tilezero.internal(i16 8, i16 32)
+  %t2 = call <256 x i32> @llvm.x86.cast.tile.to.vector.v256i32(x86_amx %t1)
+  br i1 undef, label %l2, label %l3
+
+l2:
+  %t3 = phi <256 x i32> [ undef, %entry ], [ %t2, %l1 ]
+  br i1 undef, label %l3, label %exit
+
+l3:
+  %t4 = phi <256 x i32> [ %t3, %l2], [ %t2, %l1 ]
+  %t5 = call x86_amx @llvm.x86.cast.vector.to.tile.v256i32(<256 x i32> %t4)
+  call void @llvm.x86.tilestored64.internal(i16 8, i16 32, i8* %buf, i64 1024, x86_amx %t5)
+  br label %exit
+
+exit:
+  ret void
+}
+
 define void @foo_undef(i8 *%buf) {
 ; CHECK-LABEL: @foo_undef(
 ; CHECK-NEXT:  entry: