[llvm] [InstCombine] Extend folding of aggregate construction to cases when source aggregates are partially available (PR #100828)

[via llvm-commits]

================
@@ -0,0 +1,353 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -S -passes=instcombine < %s | FileCheck %s
+
+declare {ptr, i64} @bar(i64)
+
+; Basic test.
+define {ptr, i64} @test1(i1 %cond1, ptr %p1, ptr %p2) {
+; CHECK-LABEL: define { ptr, i64 } @test1(
+; CHECK-SAME: i1 [[COND1:%.*]], ptr [[P1:%.*]], ptr [[P2:%.*]]) {
+; CHECK-NEXT:    br i1 [[COND1]], label %[[BBB1:.*]], label %[[BBB2:.*]]
+; CHECK:       [[BBB1]]:
+; CHECK-NEXT:    [[CALL1:%.*]] = call { ptr, i64 } @bar(i64 0)
+; CHECK-NEXT:    br label %[[EXIT:.*]]
+; CHECK:       [[BBB2]]:
+; CHECK-NEXT:    [[VAL21:%.*]] = load ptr, ptr [[P1]], align 8
+; CHECK-NEXT:    [[VAL22:%.*]] = load i64, ptr [[P2]], align 4
+; CHECK-NEXT:    [[TMP1:%.*]] = insertvalue { ptr, i64 } poison, ptr [[VAL21]], 0
+; CHECK-NEXT:    [[TMP2:%.*]] = insertvalue { ptr, i64 } [[TMP1]], i64 [[VAL22]], 1
+; CHECK-NEXT:    br label %[[EXIT]]
+; CHECK:       [[EXIT]]:
+; CHECK-NEXT:    [[RES_MERGED:%.*]] = phi { ptr, i64 } [ [[CALL1]], %[[BBB1]] ], [ [[TMP2]], %[[BBB2]] ]
+; CHECK-NEXT:    ret { ptr, i64 } [[RES_MERGED]]
+;
+  br i1 %cond1, label %bbb1, label %bbb2
+
+bbb1:
+  %call1 = call {ptr, i64} @bar(i64 0)
+  %val11 = extractvalue { ptr, i64 } %call1, 0
+  %val12 = extractvalue { ptr, i64 } %call1, 1
+  br label %exit
+
+bbb2:
+  %val21 = load ptr, ptr %p1
+  %val22 = load i64, ptr %p2
+  br label %exit
+
+exit:
+  %val1 = phi ptr [%val11, %bbb1], [%val21, %bbb2]
+  %val2 = phi i64 [%val12, %bbb1], [%val22, %bbb2]
+  %tmp = insertvalue { ptr, i64 } poison, ptr %val1, 0
+  %res = insertvalue { ptr, i64 } %tmp, i64 %val2, 1
+  ret {ptr, i64} %res
+}
----------------
weiguozhi wrote:

Add "tail" to the call instruction, then you can see the function call is changed to a tailcall, and function prolog and epilog is simplified(removed). In the real code the %exit block has 4 or 5 predecessors ending with tail call.

In our application it does improve tcmalloc's performance because tcmalloc_size_returning_operator_new is a hot function.

https://github.com/llvm/llvm-project/pull/100828