[llvm] [DAGCombiner] Simplifying `{si|ui}tofp` when only signbit is needed (PR #85445)

[Phoebe Wang via llvm-commits]

================
@@ -0,0 +1,413 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 4
+; RUN: llc < %s -mtriple=i686-unknown-unknown | FileCheck %s --check-prefix=X86
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown | FileCheck %s --check-prefix=X64
+
+declare void @use.i1(i1)
+declare void @use.i32(i32)
+define i32 @sitofp_signbit_only(i32 %i_in) {
+; X86-LABEL: sitofp_signbit_only:
+; X86:       # %bb.0:
+; X86-NEXT:    movl $-2147483648, %eax # imm = 0x80000000
+; X86-NEXT:    andl {{[0-9]+}}(%esp), %eax
+; X86-NEXT:    retl
+;
+; X64-LABEL: sitofp_signbit_only:
+; X64:       # %bb.0:
+; X64-NEXT:    movd %edi, %xmm0
+; X64-NEXT:    movmskps %xmm0, %eax
+; X64-NEXT:    shll $31, %eax
+; X64-NEXT:    retq
+  %f = sitofp i32 %i_in to float
+  %i = bitcast float %f to i32
+  %r = and i32 %i, 2147483648
+  ret i32 %r
+}
+
+define i32 @sitofp_signbit_only_okay_width(i16 %i_in) {
+; X86-LABEL: sitofp_signbit_only_okay_width:
+; X86:       # %bb.0:
+; X86-NEXT:    subl $8, %esp
+; X86-NEXT:    .cfi_def_cfa_offset 12
----------------
phoebewang wrote:

Add `nounwind` to suppress the CFI directives.

https://github.com/llvm/llvm-project/pull/85445