Subzero: Fix address mode optimization involving phi temporaries.

tests_lit/llvm2ice_tests/balloc.ll

Index: tests_lit/llvm2ice_tests/balloc.ll
diff --git a/tests_lit/llvm2ice_tests/balloc.ll b/tests_lit/llvm2ice_tests/balloc.ll
new file mode 100644
index 0000000000000000000000000000000000000000..5d288f427a5006cfbcad639fc1f10a8ed5d7629d
--- /dev/null
+++ b/tests_lit/llvm2ice_tests/balloc.ll
@@ -0,0 +1,102 @@
+; This is a real function that led to a bug in the address mode
+; optimization code.  It followed assignment chains through non-SSA
+; temporaries created from Phi instruction lowering.
+;
+; This test depends to some degree on the stability of "--verbose
+; addropt" output format.
+
+; RUN: %llvm2ice -O2 --verbose addropt %s | FileCheck %s

[jvoung (off chromium), 2014/08/27 16:21:08]

It would probably be better to name the file after the bug instead of after the
function in which it occurred?

[Jim Stichnoth, 2014/08/27 20:23:15]

Done.

+
+declare i32 @_calloc_r(i32, i32, i32)
+
+define internal i32 @_Balloc(i32 %ptr, i32 %k) {
+entry:
+  %gep = add i32 %ptr, 76
+  %gep.asptr = inttoptr i32 %gep to i32*
+  %0 = load i32* %gep.asptr, align 1
+  %cmp = icmp eq i32 %0, 0
+  br i1 %cmp, label %if.then, label %if.end5
+
+if.then:                                          ; preds = %entry
+  %call = tail call i32 @_calloc_r(i32 %ptr, i32 4, i32 33)
+  %gep.asptr2 = inttoptr i32 %gep to i32*
+  store i32 %call, i32* %gep.asptr2, align 1
+  %cmp3 = icmp eq i32 %call, 0
+  br i1 %cmp3, label %return, label %if.end5
+
+if.end5:                                          ; preds = %if.then, %entry
+  %1 = phi i32 [ %call, %if.then ], [ %0, %entry ]
+  %gep_array = mul i32 %k, 4
+  %gep2 = add i32 %1, %gep_array
+  %gep2.asptr = inttoptr i32 %gep2 to i32*
+  %2 = load i32* %gep2.asptr, align 1
+; The above load instruction is a good target for address mode
+; optimization.  Correct analysis would lead to dump output like:
+;   Starting computeAddressOpt for instruction:
+;     [ 15]  %__13 = load i32* %gep2.asptr, align 1
+;   Instruction: [ 14]  %gep2.asptr = i32 %gep2
+;     results in Base=%gep2, Index=<null>, Shift=0, Offset=0
+;   Instruction: [ 13]  %gep2 = add i32 %__9, %gep_array
+;     results in Base=%__9, Index=%gep_array, Shift=0, Offset=0
+;   Instruction: [ 46]  %__9 = i32 %__9_phi
+;     results in Base=%__9_phi, Index=%gep_array, Shift=0, Offset=0
+;   Instruction: [ 12]  %gep_array = mul i32 %k, 4
+;     results in Base=%__9_phi, Index=%k, Shift=2, Offset=0
+;
+; Incorrect, overly-aggressive analysis would lead to output like:
+;   Starting computeAddressOpt for instruction:
+;     [ 15]  %__13 = load i32* %gep2.asptr, align 1
+;   Instruction: [ 14]  %gep2.asptr = i32 %gep2
+;     results in Base=%gep2, Index=<null>, Shift=0, Offset=0
+;   Instruction: [ 13]  %gep2 = add i32 %__9, %gep_array
+;     results in Base=%__9, Index=%gep_array, Shift=0, Offset=0
+;   Instruction: [ 46]  %__9 = i32 %__9_phi
+;     results in Base=%__9_phi, Index=%gep_array, Shift=0, Offset=0
+;   Instruction: [ 49]  %__9_phi = i32 %__4
+;     results in Base=%__4, Index=%gep_array, Shift=0, Offset=0
+;   Instruction: [ 12]  %gep_array = mul i32 %k, 4
+;     results in Base=%__4, Index=%k, Shift=2, Offset=0
+;
+; CHECK-NOT: results in Base=%__4,
+;
+  %cmp7 = icmp eq i32 %2, 0
+  br i1 %cmp7, label %if.else, label %if.then8
+
+if.then8:                                         ; preds = %if.end5
+  %.asptr = inttoptr i32 %2 to i32*

[jvoung (off chromium), 2014/08/27 16:21:08]

Does the rest of this test matter, or can it just return %2, instead of
branching into the if.then8, if.else ?

[Jim Stichnoth, 2014/08/27 20:23:15]

Done.

+  %3 = load i32* %.asptr, align 1
+  %gep2.asptr3 = inttoptr i32 %gep2 to i32*
+  store i32 %3, i32* %gep2.asptr3, align 1
+  br label %if.end15
+
+if.else:                                          ; preds = %if.end5
+  %shl = shl i32 1, %k
+  %sub = shl i32 %shl, 2
+  %add = add i32 %sub, 20
+  %call11 = tail call i32 @_calloc_r(i32 %ptr, i32 1, i32 %add)
+  %cmp12 = icmp eq i32 %call11, 0
+  br i1 %cmp12, label %return, label %if.end14
+
+if.end14:                                         ; preds = %if.else
+  %gep5 = add i32 %call11, 4
+  %gep5.asptr = inttoptr i32 %gep5 to i32*
+  store i32 %k, i32* %gep5.asptr, align 1
+  %gep7 = add i32 %call11, 8
+  %gep7.asptr = inttoptr i32 %gep7 to i32*
+  store i32 %shl, i32* %gep7.asptr, align 1
+  br label %if.end15
+
+if.end15:                                         ; preds = %if.end14, %if.then8
+  %rv.0 = phi i32 [ %2, %if.then8 ], [ %call11, %if.end14 ]
+  %gep9 = add i32 %rv.0, 16
+  %gep9.asptr = inttoptr i32 %gep9 to i32*
+  store i32 0, i32* %gep9.asptr, align 1
+  %gep11 = add i32 %rv.0, 12
+  %gep11.asptr = inttoptr i32 %gep11 to i32*
+  store i32 0, i32* %gep11.asptr, align 1
+  br label %return
+
+return:                                           ; preds = %if.end15, %if.else, %if.then
+  %retval.0 = phi i32 [ %rv.0, %if.end15 ], [ 0, %if.then ], [ 0, %if.else ]
+  ret i32 %retval.0
+}