Subzero: Fix over-aggressive bool folding.

tests_lit/llvm2ice_tests/bool-folding.ll

 ; This tests the optimization where producers and consumers of i1 (bool)
 ; variables are combined to implicitly use flags instead of explicitly using
 ; stack or register variables.
 
 ; RUN: %p2i -i %s --filetype=obj --disassemble --args -O2 \
 ; RUN:   -allow-externally-defined-symbols | FileCheck %s
 
 ; RUN: %if --need=allow_dump --need=target_ARM32 --command %p2i --filetype=asm \
 ; RUN:   --target arm32 -i %s --args -O2 --skip-unimplemented \
 ; RUN:   -allow-externally-defined-symbols \
@@ skipping 251 matching lines @@
 ; ARM32: mov
 ; ARM32: tst
 ; ARM32: movne
 ; ARM32: mov
 ; ARM32: tst
 ; ARM32: movne
 ; ARM32: tst
 ; ARM32: movne
 ; ARM32: bx lr
 
+; Cmp/branch non-folding due to load folding and intervening store.
+define internal i32 @no_fold_cmp_br_store(i32 %arg2, i32 %argaddr) {

[Eric Holk, 2016/04/21 22:09:43]

These tests fail without this change, correct?

[Jim Stichnoth, 2016/04/21 22:50:29]

Yes, they all fail in the original.

One can easily test this by adding a "return;" statement to the very start of
BoolFolding<Traits>::invalidateProducersOnStore().

+entry:
+  %addr = inttoptr i32 %argaddr to i32*
+  %arg1 = load i32, i32* %addr, align 1
+  %cmp1 = icmp slt i32 %arg1, %arg2
+  store i32 1, i32* %addr, align 1
+  br i1 %cmp1, label %branch1, label %branch2
+branch1:
+  ret i32 1
+branch2:
+  ret i32 2
+}
+
+; CHECK-LABEL: no_fold_cmp_br_store
+; CHECK: cmp
+; CHECK: set
+; CHECK: cmp
+
+; Cmp/select non-folding due to load folding and intervening store.
+define internal i32 @no_fold_cmp_select_store(i32 %arg1, i32 %argaddr) {
+entry:
+  %addr = inttoptr i32 %argaddr to i32*
+  %arg2 = load i32, i32* %addr, align 1
+  %cmp1 = icmp slt i32 %arg1, %arg2
+  store i32 1, i32* %addr, align 1
+  %result = select i1 %cmp1, i32 %arg1, i32 %argaddr
+  ret i32 %result
+}
+
+; CHECK-LABEL: no_fold_cmp_select_store
+; CHECK: cmp
+; CHECK: setl
+; CHECK: mov DWORD PTR
+; CHECK: cmp
+; CHECK: cmovne
+
+; Cmp/select folding due to load folding and non-intervening store.
+define internal i32 @fold_cmp_select_store(i32 %arg1, i32 %argaddr) {
+entry:
+  %addr = inttoptr i32 %argaddr to i32*
+  %arg2 = load i32, i32* %addr, align 1
+  %cmp1 = icmp slt i32 %arg1, %arg2
+  %result = select i1 %cmp1, i32 %arg1, i32 %argaddr
+  store i32 1, i32* %addr, align 1
+  ret i32 %result
+}
+
+; CHECK-LABEL: fold_cmp_select_store
+; CHECK: cmp {{.*}},DWORD PTR
+; CHECK: cmovl
+
 ; Cmp/multi-select non-folding because of extra non-whitelisted uses.
 define internal i32 @no_fold_cmp_select_multi_non_whitelist(i32 %arg1,
                                                             i32 %arg2) {
 entry:
   %cmp1 = icmp slt i32 %arg1, %arg2
   %a = select i1 %cmp1, i32 %arg1, i32 %arg2
   %b = select i1 %cmp1, i32 %arg2, i32 %arg1
   %c = select i1 %cmp1, i32 123, i32 %arg1
   %ext = zext i1 %cmp1 to i32
   %partial1 = add i32 %a, %b
@@ skipping 138 matching lines @@
 ; ARM32: tst r0, #1
 ; ARM32: bne
 ; ARM32: tst r1, #1
 ; ARM32: beq
 ; ARM32: tst r2, #1
 ; ARM32: bne     {{.*}}target_true
 ; ARM32: tst     r3, #1
 ; ARM32: beq     {{.*}}target_false
 ; ARM32: tst     r4, #1
 ; ARM32: beq     {{.*}}target_false