Subzero: Support multiple fixups in one instruction.

tests_lit/llvm2ice_tests/ias-multi-reloc.ll

+; Tests the integrated assembler for instructions with multiple
+; relocations.
+
+; RUN: %p2i -i %s --args -O2 -ias=1 | FileCheck %s

[jvoung (off chromium), 2014/11/03 17:18:55]

nit: the other integrated assembler tests don't use "-ias=1"

[Jim Stichnoth, 2014/11/03 19:35:50]

OK, removed for consistency.

+
+@global_int = internal global [4 x i8] zeroinitializer, align 4
+@p_global_int = internal global [4 x i8] zeroinitializer, align 4
+declare void @dummy(i32)
+
+define internal void @store_immediate_to_global() {

[Jim Stichnoth, 2014/11/02 17:49:21]

This is the only test that actually triggered the bug.  The other two tests are
forward-looking for optimizations that O2 might/should do.

+entry:
+  %p_global_int.bc = bitcast [4 x i8]* @p_global_int to i32*
+  %expanded1 = ptrtoint [4 x i8]* @global_int to i32
+  store i32 %expanded1, i32* %p_global_int.bc, align 1
+  ret void
+}
+; CHECK-LABEL: store_immediate_to_global
+; CHECK: .long p_global_int
+; CHECK: .long global_int
+
+define internal void @add_in_place() {
+entry:
+  %global_int.bc = bitcast [4 x i8]* @global_int to i32*
+  %0 = load i32* %global_int.bc, align 1
+  %expanded1 = ptrtoint [4 x i8]* @p_global_int to i32
+  %add = add i32 %0, %expanded1
+  %global_int.bc3 = bitcast [4 x i8]* @global_int to i32*
+  store i32 %add, i32* %global_int.bc3, align 1
+  ret void
+}
+; CHECK-LABEL: add_in_place
+; CHECK: .long p_global_int

[jvoung (off chromium), 2014/11/03 17:18:55]

Hmm, I would have thought that "global_int" is the "displacement" here and comes
earlier, since it's used as the i32* for the load/store, and "p_global_int" is
the "immediate" being added to the existing value and comes later?

Oh, is the issue that it's not fused right now, so global_int shows up multiple
times and either order is statisfied?

Can the role of p_global_int and global_int be swapped to be consistent with the
other test cases?

[Jim Stichnoth, 2014/11/03 19:35:50]

I swapped the order to be consistent.  This also meant changing the int types to
char, to avoid multiplying the relocatable immediate value by 4.

When fusing is actually implemented for these two cases, I think we'll need to
swap the order of the CHECK lines.

+; CHECK: .long global_int
+
+define internal void @cmp_global_immediate() {
+entry:
+  %p_global_int.bc = bitcast [4 x i8]* @p_global_int to i32*
+  %0 = load i32* %p_global_int.bc, align 1
+  %expanded1 = ptrtoint [4 x i8]* @global_int to i32
+  %cmp = icmp eq i32 %0, %expanded1
+  br i1 %cmp, label %if.then, label %if.end
+
+if.then:                                          ; preds = %entry
+  %dummy.bc = bitcast void (i32)* @dummy to void ()*
+  tail call void %dummy.bc()
+  br label %if.end
+
+if.end:                                           ; preds = %if.then, %entry
+  ret void
+}
+; CHECK-LABEL: cmp_global_immediate
+; CHECK: .long p_global_int
+; CHECK: .long global_int
+; CHECK: call dummy