[llvm] [X86][CodeGen] Support using NF instructions for flag copy lowering (PR #93508)

[Phoebe Wang via llvm-commits]

================
@@ -0,0 +1,90 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+nf | FileCheck %s
+
+define i32 @flag_copy_1(i32 %x, i32 %y, ptr %pz) nounwind {
+; CHECK-LABEL: flag_copy_1:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    movq %rdx, %rcx
+; CHECK-NEXT:    movl %edi, %eax
+; CHECK-NEXT:    mull %esi
+; CHECK-NEXT:    movl (%rcx), %ecx
+; CHECK-NEXT:    {nf} addl %eax, %ecx
+; CHECK-NEXT:    cmovol %ecx, %eax
+; CHECK-NEXT:    retq
+  %o = tail call { i32, i1 } @llvm.umul.with.overflow.i32(i32 %x, i32 %y)
+  %v1 = extractvalue { i32, i1 } %o, 1
+  %v2 = extractvalue { i32, i1 } %o, 0
+  %z = load i32, ptr %pz
+  %a = add i32 %v2, %z
+  %r = select i1 %v1, i32 %a, i32 %v2
+  ret i32 %r
+}
+
+declare <2 x i128> @llvm.ssub.sat.v2i128(<2 x i128>, <2 x i128>)
+
+define <2 x i128> @flag_copy_2(<2 x i128> %x, <2 x i128> %y) nounwind {
+; CHECK-LABEL: flag_copy_2:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    movq %rdi, %rax
+; CHECK-NEXT:    subq {{[0-9]+}}(%rsp), %rcx
+; CHECK-NEXT:    sbbq {{[0-9]+}}(%rsp), %r8
+; CHECK-NEXT:    movq %r8, %rdi
+; CHECK-NEXT:    {nf} sarq $63, %rdi
+; CHECK-NEXT:    cmovoq %rdi, %rcx
+; CHECK-NEXT:    movabsq $-9223372036854775808, %r9 # imm = 0x8000000000000000
+; CHECK-NEXT:    {nf} xorq %r9, %rdi
+; CHECK-NEXT:    cmovnoq %r8, %rdi
+; CHECK-NEXT:    subq {{[0-9]+}}(%rsp), %rsi
+; CHECK-NEXT:    sbbq {{[0-9]+}}(%rsp), %rdx
+; CHECK-NEXT:    movq %rdx, %r8
+; CHECK-NEXT:    {nf} sarq $63, %r8
+; CHECK-NEXT:    cmovoq %r8, %rsi
+; CHECK-NEXT:    {nf} xorq %r9, %r8
+; CHECK-NEXT:    cmovnoq %rdx, %r8
+; CHECK-NEXT:    movq %rcx, 16(%rax)
+; CHECK-NEXT:    movq %rsi, (%rax)
+; CHECK-NEXT:    movq %rdi, 24(%rax)
+; CHECK-NEXT:    movq %r8, 8(%rax)
+; CHECK-NEXT:    retq
+  %z = call <2 x i128> @llvm.ssub.sat.v2i128(<2 x i128> %x, <2 x i128> %y)
+  ret <2 x i128> %z
+}
+
+; TODO: Remove the 2nd cmpl by using NF imul.
----------------
phoebewang wrote:

I checked with this, firstly this was done by `sinkCmpExpression` in CodeGenPrepare.
But even we don't do the sink, e.g., https://godbolt.org/z/fMbsnrYdo
We still cannot pass it through `EFLAGS`.

That means, We can always assume the def and use of `EFLAGS` occur within the same BB. So you don't need to traverse BBs to search the dead def EFLAGS. Searching inside the BB is enough. Or if you have concern, you can add an assert to make sure of def/use of `EFLAGS` are in the same BB.

https://github.com/llvm/llvm-project/pull/93508