[llvm] [llvm][X86] Fix merging of large sp updates (PR #125007)

[Phoebe Wang via llvm-commits]

================
@@ -0,0 +1,31 @@
+; RUN: llc < %s -O3 -mtriple=x86_64-linux-unknown -verify-machineinstrs -o %t.s
+; RUN: FileCheck --input-file=%t.s %s
+
+; Double-check that we are able to assemble the generated '.s'. A symptom of the
+; problem that led to this test is an assembler failure when using
+; '-save-temps'. For example:
+;  
+; > ...s:683:7: error: invalid operand for instruction
+; >        addq    $2147483679, %rsp               # imm = 0x8000001F
+;
+; RUN: llvm-mc -triple x86_64-unknown-unknown %t.s
+
+; Check that the stack update after calling bar gets merged into the second add
+; and not the first which is already at the chunk size limit (0x7FFFFFFF).
+
+define void @foo(ptr %rhs) {
+; CHECK-LABEL: foo
+entry:
+  %lhs = alloca [5 x [5 x [3 x [162 x [161 x [161 x double]]]]]], align 16
+  store ptr %lhs, ptr %rhs, align 8
+  %0 = call i32 @baz()
+  call void @bar(i64 0, i64 0, i64 0, i64 0, i64 0, ptr null, ptr %rhs, ptr null, ptr %rhs)
+; CHECK: call{{.*}}bar
+; CHECK: addq{{.*}}$2147483647, %rsp
+; CHECK: addq{{.*}}$372037585, %rsp
----------------
phoebewang wrote:

I compared the assembly before and after this patch, but the CFI directive doesn't look correct to me after this patch.

Before:
```
        addq    $2147483679, %rsp               # imm = 0x8000001F
        .cfi_adjust_cfa_offset -2147483679
        addq    $372037553, %rsp                # imm = 0x162CD7B1
        .cfi_def_cfa_offset 16
        popq    %rbx
        .cfi_def_cfa_offset 8
        retq
```

After:
```
        addq    $2147483647, %rsp               # imm = 0x7FFFFFFF
        addq    $372037585, %rsp                # imm = 0x162CD7D1
        .cfi_adjust_cfa_offset -372037585
        popq    %rbx
        .cfi_def_cfa_offset 8
        retq
```

Can you take a look please?

https://github.com/llvm/llvm-project/pull/125007