[llvm-branch-commits] [llvm] [ConstantTime][RISCV] Add comprehensive tests for ct.select (PR #166708)
Julius Alexandre via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Thu Nov 6 12:15:23 PST 2025
https://github.com/wizardengineer updated https://github.com/llvm/llvm-project/pull/166708
>From 7aec58aa6f8029c514857a755b5a381e6a6b22af Mon Sep 17 00:00:00 2001
From: wizardengineer <juliuswoosebert at gmail.com>
Date: Wed, 5 Nov 2025 11:01:00 -0500
Subject: [PATCH] [ConstantTime][RISCV] Add comprehensive tests for ct.select
Add comprehensive test suite for RISC-V fallback implementation:
- Edge cases (zero conditions, large integers, sign extension)
- Pattern matching (nested selects, chains)
- Vector support with RVV extensions
- Side effects and memory operations
The basic fallback test is in the core infrastructure PR.
---
.../RISCV/ctselect-fallback-edge-cases.ll | 214 +++++
.../RISCV/ctselect-fallback-patterns.ll | 383 +++++++++
.../RISCV/ctselect-fallback-vector-rvv.ll | 804 ++++++++++++++++++
.../CodeGen/RISCV/ctselect-side-effects.ll | 176 ++++
4 files changed, 1577 insertions(+)
create mode 100644 llvm/test/CodeGen/RISCV/ctselect-fallback-edge-cases.ll
create mode 100644 llvm/test/CodeGen/RISCV/ctselect-fallback-patterns.ll
create mode 100644 llvm/test/CodeGen/RISCV/ctselect-fallback-vector-rvv.ll
create mode 100644 llvm/test/CodeGen/RISCV/ctselect-side-effects.ll
diff --git a/llvm/test/CodeGen/RISCV/ctselect-fallback-edge-cases.ll b/llvm/test/CodeGen/RISCV/ctselect-fallback-edge-cases.ll
new file mode 100644
index 0000000000000..af1be0c8f3ddc
--- /dev/null
+++ b/llvm/test/CodeGen/RISCV/ctselect-fallback-edge-cases.ll
@@ -0,0 +1,214 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc < %s -mtriple=riscv64 -O3 | FileCheck %s --check-prefix=RV64
+; RUN: llc < %s -mtriple=riscv32 -O3 | FileCheck %s --check-prefix=RV32
+
+; Test with small integer types
+define i1 @test_ctselect_i1(i1 %cond, i1 %a, i1 %b) {
+; RV64-LABEL: test_ctselect_i1:
+; RV64: # %bb.0:
+; RV64-NEXT: and a1, a0, a1
+; RV64-NEXT: xori a0, a0, 1
+; RV64-NEXT: and a0, a0, a2
+; RV64-NEXT: or a0, a1, a0
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_ctselect_i1:
+; RV32: # %bb.0:
+; RV32-NEXT: and a1, a0, a1
+; RV32-NEXT: xori a0, a0, 1
+; RV32-NEXT: and a0, a0, a2
+; RV32-NEXT: or a0, a1, a0
+; RV32-NEXT: ret
+ %result = call i1 @llvm.ct.select.i1(i1 %cond, i1 %a, i1 %b)
+ ret i1 %result
+}
+
+; Test with extremal values
+define i32 @test_ctselect_extremal_values(i1 %cond) {
+; RV64-LABEL: test_ctselect_extremal_values:
+; RV64: # %bb.0:
+; RV64-NEXT: andi a0, a0, 1
+; RV64-NEXT: lui a1, 524288
+; RV64-NEXT: subw a0, a1, a0
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_ctselect_extremal_values:
+; RV32: # %bb.0:
+; RV32-NEXT: andi a0, a0, 1
+; RV32-NEXT: lui a1, 524288
+; RV32-NEXT: addi a2, a0, -1
+; RV32-NEXT: neg a0, a0
+; RV32-NEXT: and a1, a2, a1
+; RV32-NEXT: slli a0, a0, 1
+; RV32-NEXT: srli a0, a0, 1
+; RV32-NEXT: or a0, a0, a1
+; RV32-NEXT: ret
+ %result = call i32 @llvm.ct.select.i32(i1 %cond, i32 2147483647, i32 -2147483648)
+ ret i32 %result
+}
+
+; Test with null pointers
+define ptr @test_ctselect_null_ptr(i1 %cond, ptr %ptr) {
+; RV64-LABEL: test_ctselect_null_ptr:
+; RV64: # %bb.0:
+; RV64-NEXT: slli a0, a0, 63
+; RV64-NEXT: srai a0, a0, 63
+; RV64-NEXT: and a0, a0, a1
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_ctselect_null_ptr:
+; RV32: # %bb.0:
+; RV32-NEXT: slli a0, a0, 31
+; RV32-NEXT: srai a0, a0, 31
+; RV32-NEXT: and a0, a0, a1
+; RV32-NEXT: ret
+ %result = call ptr @llvm.ct.select.p0(i1 %cond, ptr %ptr, ptr null)
+ ret ptr %result
+}
+
+; Test with function pointers
+define ptr @test_ctselect_function_ptr(i1 %cond, ptr %func1, ptr %func2) {
+; RV64-LABEL: test_ctselect_function_ptr:
+; RV64: # %bb.0:
+; RV64-NEXT: andi a0, a0, 1
+; RV64-NEXT: neg a3, a0
+; RV64-NEXT: addi a0, a0, -1
+; RV64-NEXT: and a1, a3, a1
+; RV64-NEXT: and a0, a0, a2
+; RV64-NEXT: or a0, a1, a0
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_ctselect_function_ptr:
+; RV32: # %bb.0:
+; RV32-NEXT: andi a0, a0, 1
+; RV32-NEXT: neg a3, a0
+; RV32-NEXT: addi a0, a0, -1
+; RV32-NEXT: and a1, a3, a1
+; RV32-NEXT: and a0, a0, a2
+; RV32-NEXT: or a0, a1, a0
+; RV32-NEXT: ret
+ %result = call ptr @llvm.ct.select.p0(i1 %cond, ptr %func1, ptr %func2)
+ ret ptr %result
+}
+
+; Test with condition from icmp on pointers
+define ptr @test_ctselect_ptr_cmp(ptr %p1, ptr %p2, ptr %a, ptr %b) {
+; RV64-LABEL: test_ctselect_ptr_cmp:
+; RV64: # %bb.0:
+; RV64-NEXT: xor a0, a0, a1
+; RV64-NEXT: snez a0, a0
+; RV64-NEXT: addi a0, a0, -1
+; RV64-NEXT: and a2, a0, a2
+; RV64-NEXT: not a0, a0
+; RV64-NEXT: and a0, a0, a3
+; RV64-NEXT: or a0, a2, a0
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_ctselect_ptr_cmp:
+; RV32: # %bb.0:
+; RV32-NEXT: xor a0, a0, a1
+; RV32-NEXT: snez a0, a0
+; RV32-NEXT: addi a0, a0, -1
+; RV32-NEXT: and a2, a0, a2
+; RV32-NEXT: not a0, a0
+; RV32-NEXT: and a0, a0, a3
+; RV32-NEXT: or a0, a2, a0
+; RV32-NEXT: ret
+ %cmp = icmp eq ptr %p1, %p2
+ %result = call ptr @llvm.ct.select.p0(i1 %cmp, ptr %a, ptr %b)
+ ret ptr %result
+}
+
+; Test with struct pointer types
+%struct.pair = type { i32, i32 }
+
+define ptr @test_ctselect_struct_ptr(i1 %cond, ptr %a, ptr %b) {
+; RV64-LABEL: test_ctselect_struct_ptr:
+; RV64: # %bb.0:
+; RV64-NEXT: andi a0, a0, 1
+; RV64-NEXT: neg a3, a0
+; RV64-NEXT: addi a0, a0, -1
+; RV64-NEXT: and a1, a3, a1
+; RV64-NEXT: and a0, a0, a2
+; RV64-NEXT: or a0, a1, a0
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_ctselect_struct_ptr:
+; RV32: # %bb.0:
+; RV32-NEXT: andi a0, a0, 1
+; RV32-NEXT: neg a3, a0
+; RV32-NEXT: addi a0, a0, -1
+; RV32-NEXT: and a1, a3, a1
+; RV32-NEXT: and a0, a0, a2
+; RV32-NEXT: or a0, a1, a0
+; RV32-NEXT: ret
+ %result = call ptr @llvm.ct.select.p0(i1 %cond, ptr %a, ptr %b)
+ ret ptr %result
+}
+
+; Test with deeply nested conditions
+define i32 @test_ctselect_deeply_nested(i1 %c1, i1 %c2, i1 %c3, i1 %c4, i32 %a, i32 %b, i32 %c, i32 %d, i32 %e) {
+; RV64-LABEL: test_ctselect_deeply_nested:
+; RV64: # %bb.0:
+; RV64-NEXT: lw t0, 0(sp)
+; RV64-NEXT: xor a4, a4, a5
+; RV64-NEXT: slli a0, a0, 63
+; RV64-NEXT: xor a5, a5, a6
+; RV64-NEXT: slli a1, a1, 63
+; RV64-NEXT: xor a6, a6, a7
+; RV64-NEXT: slli a2, a2, 63
+; RV64-NEXT: slli a3, a3, 63
+; RV64-NEXT: srai a0, a0, 63
+; RV64-NEXT: srai a1, a1, 63
+; RV64-NEXT: srai a2, a2, 63
+; RV64-NEXT: and a0, a4, a0
+; RV64-NEXT: xor a0, a0, a5
+; RV64-NEXT: and a0, a0, a1
+; RV64-NEXT: xor a1, a7, t0
+; RV64-NEXT: xor a0, a0, a6
+; RV64-NEXT: and a0, a0, a2
+; RV64-NEXT: xor a0, a0, a1
+; RV64-NEXT: srai a3, a3, 63
+; RV64-NEXT: and a0, a0, a3
+; RV64-NEXT: xor a0, a0, t0
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_ctselect_deeply_nested:
+; RV32: # %bb.0:
+; RV32-NEXT: lw t0, 0(sp)
+; RV32-NEXT: andi a0, a0, 1
+; RV32-NEXT: andi a1, a1, 1
+; RV32-NEXT: andi a2, a2, 1
+; RV32-NEXT: andi a3, a3, 1
+; RV32-NEXT: neg t1, a0
+; RV32-NEXT: addi a0, a0, -1
+; RV32-NEXT: and a4, t1, a4
+; RV32-NEXT: neg t1, a1
+; RV32-NEXT: addi a1, a1, -1
+; RV32-NEXT: and a0, a0, a5
+; RV32-NEXT: neg a5, a2
+; RV32-NEXT: addi a2, a2, -1
+; RV32-NEXT: and a1, a1, a6
+; RV32-NEXT: neg a6, a3
+; RV32-NEXT: addi a3, a3, -1
+; RV32-NEXT: and a2, a2, a7
+; RV32-NEXT: or a0, a4, a0
+; RV32-NEXT: and a0, t1, a0
+; RV32-NEXT: or a0, a0, a1
+; RV32-NEXT: and a0, a5, a0
+; RV32-NEXT: or a0, a0, a2
+; RV32-NEXT: and a0, a6, a0
+; RV32-NEXT: and a1, a3, t0
+; RV32-NEXT: or a0, a0, a1
+; RV32-NEXT: ret
+ %sel1 = call i32 @llvm.ct.select.i32(i1 %c1, i32 %a, i32 %b)
+ %sel2 = call i32 @llvm.ct.select.i32(i1 %c2, i32 %sel1, i32 %c)
+ %sel3 = call i32 @llvm.ct.select.i32(i1 %c3, i32 %sel2, i32 %d)
+ %sel4 = call i32 @llvm.ct.select.i32(i1 %c4, i32 %sel3, i32 %e)
+ ret i32 %sel4
+}
+
+; Declare the intrinsics
+declare i1 @llvm.ct.select.i1(i1, i1, i1)
+declare i32 @llvm.ct.select.i32(i1, i32, i32)
+declare ptr @llvm.ct.select.p0(i1, ptr, ptr)
diff --git a/llvm/test/CodeGen/RISCV/ctselect-fallback-patterns.ll b/llvm/test/CodeGen/RISCV/ctselect-fallback-patterns.ll
new file mode 100644
index 0000000000000..1149971fd090e
--- /dev/null
+++ b/llvm/test/CodeGen/RISCV/ctselect-fallback-patterns.ll
@@ -0,0 +1,383 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc < %s -mtriple=riscv64 -O3 | FileCheck %s --check-prefix=RV64
+; RUN: llc < %s -mtriple=riscv32 -O3 | FileCheck %s --check-prefix=RV32
+
+; Test smin(x, 0) pattern
+define i32 @test_ctselect_smin_zero(i32 %x) {
+; RV64-LABEL: test_ctselect_smin_zero:
+; RV64: # %bb.0:
+; RV64-NEXT: sraiw a1, a0, 31
+; RV64-NEXT: and a0, a1, a0
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_ctselect_smin_zero:
+; RV32: # %bb.0:
+; RV32-NEXT: srai a1, a0, 31
+; RV32-NEXT: and a0, a1, a0
+; RV32-NEXT: ret
+ %cmp = icmp slt i32 %x, 0
+ %result = call i32 @llvm.ct.select.i32(i1 %cmp, i32 %x, i32 0)
+ ret i32 %result
+}
+
+; Test smax(x, 0) pattern
+define i32 @test_ctselect_smax_zero(i32 %x) {
+; RV64-LABEL: test_ctselect_smax_zero:
+; RV64: # %bb.0:
+; RV64-NEXT: sext.w a1, a0
+; RV64-NEXT: sgtz a1, a1
+; RV64-NEXT: neg a1, a1
+; RV64-NEXT: and a0, a1, a0
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_ctselect_smax_zero:
+; RV32: # %bb.0:
+; RV32-NEXT: sgtz a1, a0
+; RV32-NEXT: neg a1, a1
+; RV32-NEXT: and a0, a1, a0
+; RV32-NEXT: ret
+ %cmp = icmp sgt i32 %x, 0
+ %result = call i32 @llvm.ct.select.i32(i1 %cmp, i32 %x, i32 0)
+ ret i32 %result
+}
+
+; Test generic smin pattern
+define i32 @test_ctselect_smin_generic(i32 %x, i32 %y) {
+; RV64-LABEL: test_ctselect_smin_generic:
+; RV64: # %bb.0:
+; RV64-NEXT: sext.w a2, a1
+; RV64-NEXT: sext.w a3, a0
+; RV64-NEXT: slt a2, a3, a2
+; RV64-NEXT: xor a0, a0, a1
+; RV64-NEXT: neg a2, a2
+; RV64-NEXT: and a0, a0, a2
+; RV64-NEXT: xor a0, a0, a1
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_ctselect_smin_generic:
+; RV32: # %bb.0:
+; RV32-NEXT: slt a2, a0, a1
+; RV32-NEXT: neg a3, a2
+; RV32-NEXT: addi a2, a2, -1
+; RV32-NEXT: and a0, a3, a0
+; RV32-NEXT: and a1, a2, a1
+; RV32-NEXT: or a0, a0, a1
+; RV32-NEXT: ret
+ %cmp = icmp slt i32 %x, %y
+ %result = call i32 @llvm.ct.select.i32(i1 %cmp, i32 %x, i32 %y)
+ ret i32 %result
+}
+
+; Test generic smax pattern
+define i32 @test_ctselect_smax_generic(i32 %x, i32 %y) {
+; RV64-LABEL: test_ctselect_smax_generic:
+; RV64: # %bb.0:
+; RV64-NEXT: sext.w a2, a0
+; RV64-NEXT: sext.w a3, a1
+; RV64-NEXT: slt a2, a3, a2
+; RV64-NEXT: xor a0, a0, a1
+; RV64-NEXT: neg a2, a2
+; RV64-NEXT: and a0, a0, a2
+; RV64-NEXT: xor a0, a0, a1
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_ctselect_smax_generic:
+; RV32: # %bb.0:
+; RV32-NEXT: slt a2, a1, a0
+; RV32-NEXT: neg a3, a2
+; RV32-NEXT: addi a2, a2, -1
+; RV32-NEXT: and a0, a3, a0
+; RV32-NEXT: and a1, a2, a1
+; RV32-NEXT: or a0, a0, a1
+; RV32-NEXT: ret
+ %cmp = icmp sgt i32 %x, %y
+ %result = call i32 @llvm.ct.select.i32(i1 %cmp, i32 %x, i32 %y)
+ ret i32 %result
+}
+
+; Test umin pattern
+define i32 @test_ctselect_umin_generic(i32 %x, i32 %y) {
+; RV64-LABEL: test_ctselect_umin_generic:
+; RV64: # %bb.0:
+; RV64-NEXT: sext.w a2, a1
+; RV64-NEXT: sext.w a3, a0
+; RV64-NEXT: sltu a2, a3, a2
+; RV64-NEXT: xor a0, a0, a1
+; RV64-NEXT: neg a2, a2
+; RV64-NEXT: and a0, a0, a2
+; RV64-NEXT: xor a0, a0, a1
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_ctselect_umin_generic:
+; RV32: # %bb.0:
+; RV32-NEXT: sltu a2, a0, a1
+; RV32-NEXT: neg a3, a2
+; RV32-NEXT: addi a2, a2, -1
+; RV32-NEXT: and a0, a3, a0
+; RV32-NEXT: and a1, a2, a1
+; RV32-NEXT: or a0, a0, a1
+; RV32-NEXT: ret
+ %cmp = icmp ult i32 %x, %y
+ %result = call i32 @llvm.ct.select.i32(i1 %cmp, i32 %x, i32 %y)
+ ret i32 %result
+}
+
+; Test umax pattern
+define i32 @test_ctselect_umax_generic(i32 %x, i32 %y) {
+; RV64-LABEL: test_ctselect_umax_generic:
+; RV64: # %bb.0:
+; RV64-NEXT: sext.w a2, a0
+; RV64-NEXT: sext.w a3, a1
+; RV64-NEXT: sltu a2, a3, a2
+; RV64-NEXT: xor a0, a0, a1
+; RV64-NEXT: neg a2, a2
+; RV64-NEXT: and a0, a0, a2
+; RV64-NEXT: xor a0, a0, a1
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_ctselect_umax_generic:
+; RV32: # %bb.0:
+; RV32-NEXT: sltu a2, a1, a0
+; RV32-NEXT: neg a3, a2
+; RV32-NEXT: addi a2, a2, -1
+; RV32-NEXT: and a0, a3, a0
+; RV32-NEXT: and a1, a2, a1
+; RV32-NEXT: or a0, a0, a1
+; RV32-NEXT: ret
+ %cmp = icmp ugt i32 %x, %y
+ %result = call i32 @llvm.ct.select.i32(i1 %cmp, i32 %x, i32 %y)
+ ret i32 %result
+}
+
+; Test abs pattern
+define i32 @test_ctselect_abs(i32 %x) {
+; RV64-LABEL: test_ctselect_abs:
+; RV64: # %bb.0:
+; RV64-NEXT: negw a1, a0
+; RV64-NEXT: xor a1, a1, a0
+; RV64-NEXT: sraiw a2, a0, 31
+; RV64-NEXT: and a1, a1, a2
+; RV64-NEXT: xor a0, a1, a0
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_ctselect_abs:
+; RV32: # %bb.0:
+; RV32-NEXT: neg a1, a0
+; RV32-NEXT: srai a2, a0, 31
+; RV32-NEXT: and a1, a2, a1
+; RV32-NEXT: not a2, a2
+; RV32-NEXT: and a0, a2, a0
+; RV32-NEXT: or a0, a1, a0
+; RV32-NEXT: ret
+ %neg = sub i32 0, %x
+ %cmp = icmp slt i32 %x, 0
+ %result = call i32 @llvm.ct.select.i32(i1 %cmp, i32 %neg, i32 %x)
+ ret i32 %result
+}
+
+; Test nabs pattern (negative abs)
+define i32 @test_ctselect_nabs(i32 %x) {
+; RV64-LABEL: test_ctselect_nabs:
+; RV64: # %bb.0:
+; RV64-NEXT: negw a1, a0
+; RV64-NEXT: xor a2, a0, a1
+; RV64-NEXT: sraiw a0, a0, 31
+; RV64-NEXT: and a0, a2, a0
+; RV64-NEXT: xor a0, a0, a1
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_ctselect_nabs:
+; RV32: # %bb.0:
+; RV32-NEXT: neg a1, a0
+; RV32-NEXT: srai a2, a0, 31
+; RV32-NEXT: and a0, a2, a0
+; RV32-NEXT: not a2, a2
+; RV32-NEXT: and a1, a2, a1
+; RV32-NEXT: or a0, a0, a1
+; RV32-NEXT: ret
+ %neg = sub i32 0, %x
+ %cmp = icmp slt i32 %x, 0
+ %result = call i32 @llvm.ct.select.i32(i1 %cmp, i32 %x, i32 %neg)
+ ret i32 %result
+}
+
+; Test sign extension pattern
+define i32 @test_ctselect_sign_extend(i32 %x) {
+; RV64-LABEL: test_ctselect_sign_extend:
+; RV64: # %bb.0:
+; RV64-NEXT: sraiw a0, a0, 31
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_ctselect_sign_extend:
+; RV32: # %bb.0:
+; RV32-NEXT: srai a0, a0, 31
+; RV32-NEXT: ret
+ %cmp = icmp slt i32 %x, 0
+ %result = call i32 @llvm.ct.select.i32(i1 %cmp, i32 -1, i32 0)
+ ret i32 %result
+}
+
+; Test zero extension pattern
+define i32 @test_ctselect_zero_extend(i32 %x) {
+; RV64-LABEL: test_ctselect_zero_extend:
+; RV64: # %bb.0:
+; RV64-NEXT: sext.w a0, a0
+; RV64-NEXT: snez a0, a0
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_ctselect_zero_extend:
+; RV32: # %bb.0:
+; RV32-NEXT: snez a0, a0
+; RV32-NEXT: ret
+ %cmp = icmp ne i32 %x, 0
+ %result = call i32 @llvm.ct.select.i32(i1 %cmp, i32 1, i32 0)
+ ret i32 %result
+}
+
+; Test constant folding with known condition
+define i32 @test_ctselect_constant_folding_true(i32 %a, i32 %b) {
+; RV64-LABEL: test_ctselect_constant_folding_true:
+; RV64: # %bb.0:
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_ctselect_constant_folding_true:
+; RV32: # %bb.0:
+; RV32-NEXT: ret
+ %result = call i32 @llvm.ct.select.i32(i1 true, i32 %a, i32 %b)
+ ret i32 %result
+}
+
+define i32 @test_ctselect_constant_folding_false(i32 %a, i32 %b) {
+; RV64-LABEL: test_ctselect_constant_folding_false:
+; RV64: # %bb.0:
+; RV64-NEXT: mv a0, a1
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_ctselect_constant_folding_false:
+; RV32: # %bb.0:
+; RV32-NEXT: mv a0, a1
+; RV32-NEXT: ret
+ %result = call i32 @llvm.ct.select.i32(i1 false, i32 %a, i32 %b)
+ ret i32 %result
+}
+
+; Test with identical operands
+define i32 @test_ctselect_identical_operands(i1 %cond, i32 %x) {
+; RV64-LABEL: test_ctselect_identical_operands:
+; RV64: # %bb.0:
+; RV64-NEXT: mv a0, a1
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_ctselect_identical_operands:
+; RV32: # %bb.0:
+; RV32-NEXT: andi a0, a0, 1
+; RV32-NEXT: neg a2, a0
+; RV32-NEXT: addi a0, a0, -1
+; RV32-NEXT: and a2, a2, a1
+; RV32-NEXT: and a0, a0, a1
+; RV32-NEXT: or a0, a2, a0
+; RV32-NEXT: ret
+ %result = call i32 @llvm.ct.select.i32(i1 %cond, i32 %x, i32 %x)
+ ret i32 %result
+}
+
+; Test with inverted condition
+define i32 @test_ctselect_inverted_condition(i32 %x, i32 %y, i32 %a, i32 %b) {
+; RV64-LABEL: test_ctselect_inverted_condition:
+; RV64: # %bb.0:
+; RV64-NEXT: sext.w a1, a1
+; RV64-NEXT: sext.w a0, a0
+; RV64-NEXT: xor a0, a0, a1
+; RV64-NEXT: seqz a0, a0
+; RV64-NEXT: xor a2, a2, a3
+; RV64-NEXT: addi a0, a0, -1
+; RV64-NEXT: and a0, a2, a0
+; RV64-NEXT: xor a0, a0, a3
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_ctselect_inverted_condition:
+; RV32: # %bb.0:
+; RV32-NEXT: xor a0, a0, a1
+; RV32-NEXT: seqz a0, a0
+; RV32-NEXT: addi a0, a0, -1
+; RV32-NEXT: and a2, a0, a2
+; RV32-NEXT: not a0, a0
+; RV32-NEXT: and a0, a0, a3
+; RV32-NEXT: or a0, a2, a0
+; RV32-NEXT: ret
+ %cmp = icmp eq i32 %x, %y
+ %not_cmp = xor i1 %cmp, true
+ %result = call i32 @llvm.ct.select.i32(i1 %not_cmp, i32 %a, i32 %b)
+ ret i32 %result
+}
+
+; Test chain of ct.select operations
+define i32 @test_ctselect_chain(i1 %c1, i1 %c2, i1 %c3, i32 %a, i32 %b, i32 %c, i32 %d) {
+; RV64-LABEL: test_ctselect_chain:
+; RV64: # %bb.0:
+; RV64-NEXT: xor a3, a3, a4
+; RV64-NEXT: slli a0, a0, 63
+; RV64-NEXT: xor a4, a4, a5
+; RV64-NEXT: slli a1, a1, 63
+; RV64-NEXT: xor a5, a5, a6
+; RV64-NEXT: slli a2, a2, 63
+; RV64-NEXT: srai a0, a0, 63
+; RV64-NEXT: srai a1, a1, 63
+; RV64-NEXT: and a0, a3, a0
+; RV64-NEXT: xor a0, a0, a4
+; RV64-NEXT: and a0, a0, a1
+; RV64-NEXT: xor a0, a0, a5
+; RV64-NEXT: srai a2, a2, 63
+; RV64-NEXT: and a0, a0, a2
+; RV64-NEXT: xor a0, a0, a6
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_ctselect_chain:
+; RV32: # %bb.0:
+; RV32-NEXT: andi a0, a0, 1
+; RV32-NEXT: andi a1, a1, 1
+; RV32-NEXT: andi a2, a2, 1
+; RV32-NEXT: neg a7, a0
+; RV32-NEXT: addi a0, a0, -1
+; RV32-NEXT: and a3, a7, a3
+; RV32-NEXT: neg a7, a1
+; RV32-NEXT: addi a1, a1, -1
+; RV32-NEXT: and a0, a0, a4
+; RV32-NEXT: neg a4, a2
+; RV32-NEXT: addi a2, a2, -1
+; RV32-NEXT: and a1, a1, a5
+; RV32-NEXT: or a0, a3, a0
+; RV32-NEXT: and a0, a7, a0
+; RV32-NEXT: or a0, a0, a1
+; RV32-NEXT: and a0, a4, a0
+; RV32-NEXT: and a1, a2, a6
+; RV32-NEXT: or a0, a0, a1
+; RV32-NEXT: ret
+ %sel1 = call i32 @llvm.ct.select.i32(i1 %c1, i32 %a, i32 %b)
+ %sel2 = call i32 @llvm.ct.select.i32(i1 %c2, i32 %sel1, i32 %c)
+ %sel3 = call i32 @llvm.ct.select.i32(i1 %c3, i32 %sel2, i32 %d)
+ ret i32 %sel3
+}
+
+; Test for 64-bit operations (supported on all 64-bit architectures)
+define i64 @test_ctselect_i64_smin_zero(i64 %x) {
+; RV64-LABEL: test_ctselect_i64_smin_zero:
+; RV64: # %bb.0:
+; RV64-NEXT: srai a1, a0, 63
+; RV64-NEXT: and a0, a1, a0
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_ctselect_i64_smin_zero:
+; RV32: # %bb.0:
+; RV32-NEXT: srai a2, a1, 31
+; RV32-NEXT: and a0, a2, a0
+; RV32-NEXT: and a1, a2, a1
+; RV32-NEXT: ret
+ %cmp = icmp slt i64 %x, 0
+ %result = call i64 @llvm.ct.select.i64(i1 %cmp, i64 %x, i64 0)
+ ret i64 %result
+}
+
+; Declare the intrinsics
+declare i32 @llvm.ct.select.i32(i1, i32, i32)
+declare i64 @llvm.ct.select.i64(i1, i64, i64)
diff --git a/llvm/test/CodeGen/RISCV/ctselect-fallback-vector-rvv.ll b/llvm/test/CodeGen/RISCV/ctselect-fallback-vector-rvv.ll
new file mode 100644
index 0000000000000..a02e1e4749443
--- /dev/null
+++ b/llvm/test/CodeGen/RISCV/ctselect-fallback-vector-rvv.ll
@@ -0,0 +1,804 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc < %s -mtriple=riscv64 -mattr=+v -O3 | FileCheck %s --check-prefix=RV64
+; RUN: llc < %s -mtriple=riscv32 -mattr=+v -O3 | FileCheck %s --check-prefix=RV32
+; RUN: llc < %s -mtriple=riscv32 -mattr=+v,+zvl128b -O3 | FileCheck %s --check-prefix=RV32-V128
+; RUN: llc < %s -mtriple=riscv64 -mattr=+v,+zvl256b -O3 | FileCheck %s --check-prefix=RV64-V256
+
+
+; Basic pass-through select on nxv4i32
+define <vscale x 4 x i32> @ctsel_nxv4i32_basic(i1 %cond, <vscale x 4 x i32> %a, <vscale x 4 x i32> %b) {
+; RV64-LABEL: ctsel_nxv4i32_basic:
+; RV64: # %bb.0:
+; RV64-NEXT: andi a0, a0, 1
+; RV64-NEXT: vsetvli a1, zero, e8, mf2, ta, ma
+; RV64-NEXT: vmv.v.x v12, a0
+; RV64-NEXT: vmsne.vi v0, v12, 0
+; RV64-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; RV64-NEXT: vmv.v.i v12, 0
+; RV64-NEXT: vmerge.vim v12, v12, -1, v0
+; RV64-NEXT: vand.vv v8, v12, v8
+; RV64-NEXT: vnot.v v12, v12
+; RV64-NEXT: vand.vv v10, v12, v10
+; RV64-NEXT: vor.vv v8, v8, v10
+; RV64-NEXT: ret
+;
+; RV32-LABEL: ctsel_nxv4i32_basic:
+; RV32: # %bb.0:
+; RV32-NEXT: andi a0, a0, 1
+; RV32-NEXT: vsetvli a1, zero, e8, mf2, ta, ma
+; RV32-NEXT: vmv.v.x v12, a0
+; RV32-NEXT: vmsne.vi v0, v12, 0
+; RV32-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; RV32-NEXT: vmv.v.i v12, 0
+; RV32-NEXT: vmerge.vim v12, v12, -1, v0
+; RV32-NEXT: vand.vv v8, v12, v8
+; RV32-NEXT: vnot.v v12, v12
+; RV32-NEXT: vand.vv v10, v12, v10
+; RV32-NEXT: vor.vv v8, v8, v10
+; RV32-NEXT: ret
+;
+; RV32-V128-LABEL: ctsel_nxv4i32_basic:
+; RV32-V128: # %bb.0:
+; RV32-V128-NEXT: andi a0, a0, 1
+; RV32-V128-NEXT: vsetvli a1, zero, e8, mf2, ta, ma
+; RV32-V128-NEXT: vmv.v.x v12, a0
+; RV32-V128-NEXT: vmsne.vi v0, v12, 0
+; RV32-V128-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; RV32-V128-NEXT: vmv.v.i v12, 0
+; RV32-V128-NEXT: vmerge.vim v12, v12, -1, v0
+; RV32-V128-NEXT: vand.vv v8, v12, v8
+; RV32-V128-NEXT: vnot.v v12, v12
+; RV32-V128-NEXT: vand.vv v10, v12, v10
+; RV32-V128-NEXT: vor.vv v8, v8, v10
+; RV32-V128-NEXT: ret
+;
+; RV64-V256-LABEL: ctsel_nxv4i32_basic:
+; RV64-V256: # %bb.0:
+; RV64-V256-NEXT: andi a0, a0, 1
+; RV64-V256-NEXT: vsetvli a1, zero, e8, mf2, ta, ma
+; RV64-V256-NEXT: vmv.v.x v12, a0
+; RV64-V256-NEXT: vmsne.vi v0, v12, 0
+; RV64-V256-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; RV64-V256-NEXT: vmv.v.i v12, 0
+; RV64-V256-NEXT: vmerge.vim v12, v12, -1, v0
+; RV64-V256-NEXT: vand.vv v8, v12, v8
+; RV64-V256-NEXT: vnot.v v12, v12
+; RV64-V256-NEXT: vand.vv v10, v12, v10
+; RV64-V256-NEXT: vor.vv v8, v8, v10
+; RV64-V256-NEXT: ret
+ %r = call <vscale x 4 x i32> @llvm.ct.select.nxv4i32(i1 %cond, <vscale x 4 x i32> %a, <vscale x 4 x i32> %b)
+ ret <vscale x 4 x i32> %r
+}
+
+; Select with loads (aligned)
+define <vscale x 4 x i32> @ctsel_nxv4i32_load(i1 %cond, ptr %p1, ptr %p2) {
+; RV64-LABEL: ctsel_nxv4i32_load:
+; RV64: # %bb.0:
+; RV64-NEXT: vl2re32.v v8, (a1)
+; RV64-NEXT: vl2re32.v v10, (a2)
+; RV64-NEXT: andi a0, a0, 1
+; RV64-NEXT: vsetvli a1, zero, e8, mf2, ta, ma
+; RV64-NEXT: vmv.v.x v12, a0
+; RV64-NEXT: vmsne.vi v0, v12, 0
+; RV64-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; RV64-NEXT: vmv.v.i v12, 0
+; RV64-NEXT: vmerge.vim v12, v12, -1, v0
+; RV64-NEXT: vand.vv v8, v12, v8
+; RV64-NEXT: vnot.v v12, v12
+; RV64-NEXT: vand.vv v10, v12, v10
+; RV64-NEXT: vor.vv v8, v8, v10
+; RV64-NEXT: ret
+;
+; RV32-LABEL: ctsel_nxv4i32_load:
+; RV32: # %bb.0:
+; RV32-NEXT: vl2re32.v v8, (a1)
+; RV32-NEXT: vl2re32.v v10, (a2)
+; RV32-NEXT: andi a0, a0, 1
+; RV32-NEXT: vsetvli a1, zero, e8, mf2, ta, ma
+; RV32-NEXT: vmv.v.x v12, a0
+; RV32-NEXT: vmsne.vi v0, v12, 0
+; RV32-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; RV32-NEXT: vmv.v.i v12, 0
+; RV32-NEXT: vmerge.vim v12, v12, -1, v0
+; RV32-NEXT: vand.vv v8, v12, v8
+; RV32-NEXT: vnot.v v12, v12
+; RV32-NEXT: vand.vv v10, v12, v10
+; RV32-NEXT: vor.vv v8, v8, v10
+; RV32-NEXT: ret
+;
+; RV32-V128-LABEL: ctsel_nxv4i32_load:
+; RV32-V128: # %bb.0:
+; RV32-V128-NEXT: vl2re32.v v8, (a1)
+; RV32-V128-NEXT: vl2re32.v v10, (a2)
+; RV32-V128-NEXT: andi a0, a0, 1
+; RV32-V128-NEXT: vsetvli a1, zero, e8, mf2, ta, ma
+; RV32-V128-NEXT: vmv.v.x v12, a0
+; RV32-V128-NEXT: vmsne.vi v0, v12, 0
+; RV32-V128-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; RV32-V128-NEXT: vmv.v.i v12, 0
+; RV32-V128-NEXT: vmerge.vim v12, v12, -1, v0
+; RV32-V128-NEXT: vand.vv v8, v12, v8
+; RV32-V128-NEXT: vnot.v v12, v12
+; RV32-V128-NEXT: vand.vv v10, v12, v10
+; RV32-V128-NEXT: vor.vv v8, v8, v10
+; RV32-V128-NEXT: ret
+;
+; RV64-V256-LABEL: ctsel_nxv4i32_load:
+; RV64-V256: # %bb.0:
+; RV64-V256-NEXT: vl2re32.v v8, (a1)
+; RV64-V256-NEXT: vl2re32.v v10, (a2)
+; RV64-V256-NEXT: andi a0, a0, 1
+; RV64-V256-NEXT: vsetvli a1, zero, e8, mf2, ta, ma
+; RV64-V256-NEXT: vmv.v.x v12, a0
+; RV64-V256-NEXT: vmsne.vi v0, v12, 0
+; RV64-V256-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; RV64-V256-NEXT: vmv.v.i v12, 0
+; RV64-V256-NEXT: vmerge.vim v12, v12, -1, v0
+; RV64-V256-NEXT: vand.vv v8, v12, v8
+; RV64-V256-NEXT: vnot.v v12, v12
+; RV64-V256-NEXT: vand.vv v10, v12, v10
+; RV64-V256-NEXT: vor.vv v8, v8, v10
+; RV64-V256-NEXT: ret
+ %a = load <vscale x 4 x i32>, ptr %p1, align 16
+ %b = load <vscale x 4 x i32>, ptr %p2, align 16
+ %r = call <vscale x 4 x i32> @llvm.ct.select.nxv4i32(i1 %cond, <vscale x 4 x i32> %a, <vscale x 4 x i32> %b)
+ ret <vscale x 4 x i32> %r
+}
+
+; Mixed: do arithmetic first, then select, then store
+define void @ctsel_nxv4i32_mixed(i1 %cond, ptr %p1, ptr %p2, ptr %out) {
+; RV64-LABEL: ctsel_nxv4i32_mixed:
+; RV64: # %bb.0:
+; RV64-NEXT: vl2re32.v v8, (a1)
+; RV64-NEXT: vl2re32.v v10, (a2)
+; RV64-NEXT: andi a0, a0, 1
+; RV64-NEXT: vsetvli a1, zero, e8, mf2, ta, ma
+; RV64-NEXT: vmv.v.x v12, a0
+; RV64-NEXT: vmsne.vi v0, v12, 0
+; RV64-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; RV64-NEXT: vmv.v.i v12, 0
+; RV64-NEXT: vmerge.vim v12, v12, -1, v0
+; RV64-NEXT: vadd.vv v8, v8, v8
+; RV64-NEXT: vadd.vv v10, v10, v10
+; RV64-NEXT: vand.vv v8, v12, v8
+; RV64-NEXT: vnot.v v12, v12
+; RV64-NEXT: vand.vv v10, v12, v10
+; RV64-NEXT: vor.vv v8, v8, v10
+; RV64-NEXT: vs2r.v v8, (a3)
+; RV64-NEXT: ret
+;
+; RV32-LABEL: ctsel_nxv4i32_mixed:
+; RV32: # %bb.0:
+; RV32-NEXT: vl2re32.v v8, (a1)
+; RV32-NEXT: vl2re32.v v10, (a2)
+; RV32-NEXT: andi a0, a0, 1
+; RV32-NEXT: vsetvli a1, zero, e8, mf2, ta, ma
+; RV32-NEXT: vmv.v.x v12, a0
+; RV32-NEXT: vmsne.vi v0, v12, 0
+; RV32-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; RV32-NEXT: vmv.v.i v12, 0
+; RV32-NEXT: vmerge.vim v12, v12, -1, v0
+; RV32-NEXT: vadd.vv v8, v8, v8
+; RV32-NEXT: vadd.vv v10, v10, v10
+; RV32-NEXT: vand.vv v8, v12, v8
+; RV32-NEXT: vnot.v v12, v12
+; RV32-NEXT: vand.vv v10, v12, v10
+; RV32-NEXT: vor.vv v8, v8, v10
+; RV32-NEXT: vs2r.v v8, (a3)
+; RV32-NEXT: ret
+;
+; RV32-V128-LABEL: ctsel_nxv4i32_mixed:
+; RV32-V128: # %bb.0:
+; RV32-V128-NEXT: vl2re32.v v8, (a1)
+; RV32-V128-NEXT: vl2re32.v v10, (a2)
+; RV32-V128-NEXT: andi a0, a0, 1
+; RV32-V128-NEXT: vsetvli a1, zero, e8, mf2, ta, ma
+; RV32-V128-NEXT: vmv.v.x v12, a0
+; RV32-V128-NEXT: vmsne.vi v0, v12, 0
+; RV32-V128-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; RV32-V128-NEXT: vmv.v.i v12, 0
+; RV32-V128-NEXT: vmerge.vim v12, v12, -1, v0
+; RV32-V128-NEXT: vadd.vv v8, v8, v8
+; RV32-V128-NEXT: vadd.vv v10, v10, v10
+; RV32-V128-NEXT: vand.vv v8, v12, v8
+; RV32-V128-NEXT: vnot.v v12, v12
+; RV32-V128-NEXT: vand.vv v10, v12, v10
+; RV32-V128-NEXT: vor.vv v8, v8, v10
+; RV32-V128-NEXT: vs2r.v v8, (a3)
+; RV32-V128-NEXT: ret
+;
+; RV64-V256-LABEL: ctsel_nxv4i32_mixed:
+; RV64-V256: # %bb.0:
+; RV64-V256-NEXT: vl2re32.v v8, (a1)
+; RV64-V256-NEXT: vl2re32.v v10, (a2)
+; RV64-V256-NEXT: andi a0, a0, 1
+; RV64-V256-NEXT: vsetvli a1, zero, e8, mf2, ta, ma
+; RV64-V256-NEXT: vmv.v.x v12, a0
+; RV64-V256-NEXT: vmsne.vi v0, v12, 0
+; RV64-V256-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; RV64-V256-NEXT: vmv.v.i v12, 0
+; RV64-V256-NEXT: vmerge.vim v12, v12, -1, v0
+; RV64-V256-NEXT: vadd.vv v8, v8, v8
+; RV64-V256-NEXT: vadd.vv v10, v10, v10
+; RV64-V256-NEXT: vand.vv v8, v12, v8
+; RV64-V256-NEXT: vnot.v v12, v12
+; RV64-V256-NEXT: vand.vv v10, v12, v10
+; RV64-V256-NEXT: vor.vv v8, v8, v10
+; RV64-V256-NEXT: vs2r.v v8, (a3)
+; RV64-V256-NEXT: ret
+ %a = load <vscale x 4 x i32>, ptr %p1, align 16
+ %b = load <vscale x 4 x i32>, ptr %p2, align 16
+ ; avoid scalable vector constants: use %a+%a and %b+%b
+ %a2 = add <vscale x 4 x i32> %a, %a
+ %b2 = add <vscale x 4 x i32> %b, %b
+ %r = call <vscale x 4 x i32> @llvm.ct.select.nxv4i32(i1 %cond, <vscale x 4 x i32> %a2, <vscale x 4 x i32> %b2)
+ store <vscale x 4 x i32> %r, ptr %out, align 16
+ ret void
+}
+
+; Const-true/false fold smoke tests
+define <vscale x 4 x i32> @ctsel_nxv4i32_true(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b) {
+; RV64-LABEL: ctsel_nxv4i32_true:
+; RV64: # %bb.0:
+; RV64-NEXT: ret
+;
+; RV32-LABEL: ctsel_nxv4i32_true:
+; RV32: # %bb.0:
+; RV32-NEXT: ret
+;
+; RV32-V128-LABEL: ctsel_nxv4i32_true:
+; RV32-V128: # %bb.0:
+; RV32-V128-NEXT: ret
+;
+; RV64-V256-LABEL: ctsel_nxv4i32_true:
+; RV64-V256: # %bb.0:
+; RV64-V256-NEXT: ret
+ %r = call <vscale x 4 x i32> @llvm.ct.select.nxv4i32(i1 true, <vscale x 4 x i32> %a, <vscale x 4 x i32> %b)
+ ret <vscale x 4 x i32> %r
+}
+
+define <vscale x 4 x i32> @ctsel_nxv4i32_false(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b) {
+; RV64-LABEL: ctsel_nxv4i32_false:
+; RV64: # %bb.0:
+; RV64-NEXT: vsetivli zero, 1, e8, m1, ta, ma
+; RV64-NEXT: vmv2r.v v8, v10
+; RV64-NEXT: ret
+;
+; RV32-LABEL: ctsel_nxv4i32_false:
+; RV32: # %bb.0:
+; RV32-NEXT: vsetivli zero, 1, e8, m1, ta, ma
+; RV32-NEXT: vmv2r.v v8, v10
+; RV32-NEXT: ret
+;
+; RV32-V128-LABEL: ctsel_nxv4i32_false:
+; RV32-V128: # %bb.0:
+; RV32-V128-NEXT: vsetivli zero, 1, e8, m1, ta, ma
+; RV32-V128-NEXT: vmv2r.v v8, v10
+; RV32-V128-NEXT: ret
+;
+; RV64-V256-LABEL: ctsel_nxv4i32_false:
+; RV64-V256: # %bb.0:
+; RV64-V256-NEXT: vsetivli zero, 1, e8, m1, ta, ma
+; RV64-V256-NEXT: vmv2r.v v8, v10
+; RV64-V256-NEXT: ret
+ %r = call <vscale x 4 x i32> @llvm.ct.select.nxv4i32(i1 false, <vscale x 4 x i32> %a, <vscale x 4 x i32> %b)
+ ret <vscale x 4 x i32> %r
+}
+
+; Chain two selects to ensure masks don’t get merged away
+define <vscale x 4 x i32> @ctsel_nxv4i32_chain(i1 %c1, i1 %c2,
+; RV64-LABEL: ctsel_nxv4i32_chain:
+; RV64: # %bb.0:
+; RV64-NEXT: andi a0, a0, 1
+; RV64-NEXT: vsetvli a2, zero, e32, m2, ta, ma
+; RV64-NEXT: vmv.v.i v14, 0
+; RV64-NEXT: andi a1, a1, 1
+; RV64-NEXT: vsetvli zero, zero, e8, mf2, ta, ma
+; RV64-NEXT: vmv.v.x v16, a0
+; RV64-NEXT: vmsne.vi v0, v16, 0
+; RV64-NEXT: vmv.v.x v18, a1
+; RV64-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; RV64-NEXT: vmerge.vim v16, v14, -1, v0
+; RV64-NEXT: vsetvli zero, zero, e8, mf2, ta, ma
+; RV64-NEXT: vmsne.vi v0, v18, 0
+; RV64-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; RV64-NEXT: vmerge.vim v14, v14, -1, v0
+; RV64-NEXT: vand.vv v8, v16, v8
+; RV64-NEXT: vnot.v v16, v16
+; RV64-NEXT: vand.vv v10, v16, v10
+; RV64-NEXT: vnot.v v16, v14
+; RV64-NEXT: vor.vv v8, v8, v10
+; RV64-NEXT: vand.vv v8, v14, v8
+; RV64-NEXT: vand.vv v10, v16, v12
+; RV64-NEXT: vor.vv v8, v8, v10
+; RV64-NEXT: ret
+;
+; RV32-LABEL: ctsel_nxv4i32_chain:
+; RV32: # %bb.0:
+; RV32-NEXT: andi a0, a0, 1
+; RV32-NEXT: vsetvli a2, zero, e32, m2, ta, ma
+; RV32-NEXT: vmv.v.i v14, 0
+; RV32-NEXT: andi a1, a1, 1
+; RV32-NEXT: vsetvli zero, zero, e8, mf2, ta, ma
+; RV32-NEXT: vmv.v.x v16, a0
+; RV32-NEXT: vmsne.vi v0, v16, 0
+; RV32-NEXT: vmv.v.x v18, a1
+; RV32-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; RV32-NEXT: vmerge.vim v16, v14, -1, v0
+; RV32-NEXT: vsetvli zero, zero, e8, mf2, ta, ma
+; RV32-NEXT: vmsne.vi v0, v18, 0
+; RV32-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; RV32-NEXT: vmerge.vim v14, v14, -1, v0
+; RV32-NEXT: vand.vv v8, v16, v8
+; RV32-NEXT: vnot.v v16, v16
+; RV32-NEXT: vand.vv v10, v16, v10
+; RV32-NEXT: vnot.v v16, v14
+; RV32-NEXT: vor.vv v8, v8, v10
+; RV32-NEXT: vand.vv v8, v14, v8
+; RV32-NEXT: vand.vv v10, v16, v12
+; RV32-NEXT: vor.vv v8, v8, v10
+; RV32-NEXT: ret
+;
+; RV32-V128-LABEL: ctsel_nxv4i32_chain:
+; RV32-V128: # %bb.0:
+; RV32-V128-NEXT: andi a0, a0, 1
+; RV32-V128-NEXT: vsetvli a2, zero, e32, m2, ta, ma
+; RV32-V128-NEXT: vmv.v.i v14, 0
+; RV32-V128-NEXT: andi a1, a1, 1
+; RV32-V128-NEXT: vsetvli zero, zero, e8, mf2, ta, ma
+; RV32-V128-NEXT: vmv.v.x v16, a0
+; RV32-V128-NEXT: vmsne.vi v0, v16, 0
+; RV32-V128-NEXT: vmv.v.x v18, a1
+; RV32-V128-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; RV32-V128-NEXT: vmerge.vim v16, v14, -1, v0
+; RV32-V128-NEXT: vsetvli zero, zero, e8, mf2, ta, ma
+; RV32-V128-NEXT: vmsne.vi v0, v18, 0
+; RV32-V128-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; RV32-V128-NEXT: vmerge.vim v14, v14, -1, v0
+; RV32-V128-NEXT: vand.vv v8, v16, v8
+; RV32-V128-NEXT: vnot.v v16, v16
+; RV32-V128-NEXT: vand.vv v10, v16, v10
+; RV32-V128-NEXT: vnot.v v16, v14
+; RV32-V128-NEXT: vor.vv v8, v8, v10
+; RV32-V128-NEXT: vand.vv v8, v14, v8
+; RV32-V128-NEXT: vand.vv v10, v16, v12
+; RV32-V128-NEXT: vor.vv v8, v8, v10
+; RV32-V128-NEXT: ret
+;
+; RV64-V256-LABEL: ctsel_nxv4i32_chain:
+; RV64-V256: # %bb.0:
+; RV64-V256-NEXT: andi a0, a0, 1
+; RV64-V256-NEXT: vsetvli a2, zero, e32, m2, ta, ma
+; RV64-V256-NEXT: vmv.v.i v14, 0
+; RV64-V256-NEXT: andi a1, a1, 1
+; RV64-V256-NEXT: vsetvli zero, zero, e8, mf2, ta, ma
+; RV64-V256-NEXT: vmv.v.x v16, a0
+; RV64-V256-NEXT: vmsne.vi v0, v16, 0
+; RV64-V256-NEXT: vmv.v.x v18, a1
+; RV64-V256-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; RV64-V256-NEXT: vmerge.vim v16, v14, -1, v0
+; RV64-V256-NEXT: vsetvli zero, zero, e8, mf2, ta, ma
+; RV64-V256-NEXT: vmsne.vi v0, v18, 0
+; RV64-V256-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; RV64-V256-NEXT: vmerge.vim v14, v14, -1, v0
+; RV64-V256-NEXT: vand.vv v8, v16, v8
+; RV64-V256-NEXT: vnot.v v16, v16
+; RV64-V256-NEXT: vand.vv v10, v16, v10
+; RV64-V256-NEXT: vnot.v v16, v14
+; RV64-V256-NEXT: vor.vv v8, v8, v10
+; RV64-V256-NEXT: vand.vv v8, v14, v8
+; RV64-V256-NEXT: vand.vv v10, v16, v12
+; RV64-V256-NEXT: vor.vv v8, v8, v10
+; RV64-V256-NEXT: ret
+ <vscale x 4 x i32> %a,
+ <vscale x 4 x i32> %b,
+ <vscale x 4 x i32> %c) {
+ %t = call <vscale x 4 x i32> @llvm.ct.select.nxv4i32(i1 %c1, <vscale x 4 x i32> %a, <vscale x 4 x i32> %b)
+ %r = call <vscale x 4 x i32> @llvm.ct.select.nxv4i32(i1 %c2, <vscale x 4 x i32> %t, <vscale x 4 x i32> %c)
+ ret <vscale x 4 x i32> %r
+}
+
+; A different element width
+define <vscale x 8 x i16> @ctsel_nxv8i16_basic(i1 %cond, <vscale x 8 x i16> %a, <vscale x 8 x i16> %b) {
+; RV64-LABEL: ctsel_nxv8i16_basic:
+; RV64: # %bb.0:
+; RV64-NEXT: andi a0, a0, 1
+; RV64-NEXT: vsetvli a1, zero, e8, m1, ta, ma
+; RV64-NEXT: vmv.v.x v12, a0
+; RV64-NEXT: vmsne.vi v0, v12, 0
+; RV64-NEXT: vsetvli zero, zero, e16, m2, ta, ma
+; RV64-NEXT: vmv.v.i v12, 0
+; RV64-NEXT: vmerge.vim v12, v12, -1, v0
+; RV64-NEXT: vand.vv v8, v12, v8
+; RV64-NEXT: vnot.v v12, v12
+; RV64-NEXT: vand.vv v10, v12, v10
+; RV64-NEXT: vor.vv v8, v8, v10
+; RV64-NEXT: ret
+;
+; RV32-LABEL: ctsel_nxv8i16_basic:
+; RV32: # %bb.0:
+; RV32-NEXT: andi a0, a0, 1
+; RV32-NEXT: vsetvli a1, zero, e8, m1, ta, ma
+; RV32-NEXT: vmv.v.x v12, a0
+; RV32-NEXT: vmsne.vi v0, v12, 0
+; RV32-NEXT: vsetvli zero, zero, e16, m2, ta, ma
+; RV32-NEXT: vmv.v.i v12, 0
+; RV32-NEXT: vmerge.vim v12, v12, -1, v0
+; RV32-NEXT: vand.vv v8, v12, v8
+; RV32-NEXT: vnot.v v12, v12
+; RV32-NEXT: vand.vv v10, v12, v10
+; RV32-NEXT: vor.vv v8, v8, v10
+; RV32-NEXT: ret
+;
+; RV32-V128-LABEL: ctsel_nxv8i16_basic:
+; RV32-V128: # %bb.0:
+; RV32-V128-NEXT: andi a0, a0, 1
+; RV32-V128-NEXT: vsetvli a1, zero, e8, m1, ta, ma
+; RV32-V128-NEXT: vmv.v.x v12, a0
+; RV32-V128-NEXT: vmsne.vi v0, v12, 0
+; RV32-V128-NEXT: vsetvli zero, zero, e16, m2, ta, ma
+; RV32-V128-NEXT: vmv.v.i v12, 0
+; RV32-V128-NEXT: vmerge.vim v12, v12, -1, v0
+; RV32-V128-NEXT: vand.vv v8, v12, v8
+; RV32-V128-NEXT: vnot.v v12, v12
+; RV32-V128-NEXT: vand.vv v10, v12, v10
+; RV32-V128-NEXT: vor.vv v8, v8, v10
+; RV32-V128-NEXT: ret
+;
+; RV64-V256-LABEL: ctsel_nxv8i16_basic:
+; RV64-V256: # %bb.0:
+; RV64-V256-NEXT: andi a0, a0, 1
+; RV64-V256-NEXT: vsetvli a1, zero, e8, m1, ta, ma
+; RV64-V256-NEXT: vmv.v.x v12, a0
+; RV64-V256-NEXT: vmsne.vi v0, v12, 0
+; RV64-V256-NEXT: vsetvli zero, zero, e16, m2, ta, ma
+; RV64-V256-NEXT: vmv.v.i v12, 0
+; RV64-V256-NEXT: vmerge.vim v12, v12, -1, v0
+; RV64-V256-NEXT: vand.vv v8, v12, v8
+; RV64-V256-NEXT: vnot.v v12, v12
+; RV64-V256-NEXT: vand.vv v10, v12, v10
+; RV64-V256-NEXT: vor.vv v8, v8, v10
+; RV64-V256-NEXT: ret
+ %r = call <vscale x 8 x i16> @llvm.ct.select.nxv8i16(i1 %cond, <vscale x 8 x i16> %a, <vscale x 8 x i16> %b)
+ ret <vscale x 8 x i16> %r
+}
+
+define <vscale x 16 x i8> @ctsel_nxv16i8_basic(i1 %cond, <vscale x 16 x i8> %a, <vscale x 16 x i8> %b) {
+; RV64-LABEL: ctsel_nxv16i8_basic:
+; RV64: # %bb.0:
+; RV64-NEXT: andi a0, a0, 1
+; RV64-NEXT: vsetvli a1, zero, e8, m2, ta, ma
+; RV64-NEXT: vmv.v.x v12, a0
+; RV64-NEXT: vmsne.vi v0, v12, 0
+; RV64-NEXT: vmv.v.i v12, 0
+; RV64-NEXT: vmerge.vim v12, v12, -1, v0
+; RV64-NEXT: vand.vv v8, v12, v8
+; RV64-NEXT: vnot.v v12, v12
+; RV64-NEXT: vand.vv v10, v12, v10
+; RV64-NEXT: vor.vv v8, v8, v10
+; RV64-NEXT: ret
+;
+; RV32-LABEL: ctsel_nxv16i8_basic:
+; RV32: # %bb.0:
+; RV32-NEXT: andi a0, a0, 1
+; RV32-NEXT: vsetvli a1, zero, e8, m2, ta, ma
+; RV32-NEXT: vmv.v.x v12, a0
+; RV32-NEXT: vmsne.vi v0, v12, 0
+; RV32-NEXT: vmv.v.i v12, 0
+; RV32-NEXT: vmerge.vim v12, v12, -1, v0
+; RV32-NEXT: vand.vv v8, v12, v8
+; RV32-NEXT: vnot.v v12, v12
+; RV32-NEXT: vand.vv v10, v12, v10
+; RV32-NEXT: vor.vv v8, v8, v10
+; RV32-NEXT: ret
+;
+; RV32-V128-LABEL: ctsel_nxv16i8_basic:
+; RV32-V128: # %bb.0:
+; RV32-V128-NEXT: andi a0, a0, 1
+; RV32-V128-NEXT: vsetvli a1, zero, e8, m2, ta, ma
+; RV32-V128-NEXT: vmv.v.x v12, a0
+; RV32-V128-NEXT: vmsne.vi v0, v12, 0
+; RV32-V128-NEXT: vmv.v.i v12, 0
+; RV32-V128-NEXT: vmerge.vim v12, v12, -1, v0
+; RV32-V128-NEXT: vand.vv v8, v12, v8
+; RV32-V128-NEXT: vnot.v v12, v12
+; RV32-V128-NEXT: vand.vv v10, v12, v10
+; RV32-V128-NEXT: vor.vv v8, v8, v10
+; RV32-V128-NEXT: ret
+;
+; RV64-V256-LABEL: ctsel_nxv16i8_basic:
+; RV64-V256: # %bb.0:
+; RV64-V256-NEXT: andi a0, a0, 1
+; RV64-V256-NEXT: vsetvli a1, zero, e8, m2, ta, ma
+; RV64-V256-NEXT: vmv.v.x v12, a0
+; RV64-V256-NEXT: vmsne.vi v0, v12, 0
+; RV64-V256-NEXT: vmv.v.i v12, 0
+; RV64-V256-NEXT: vmerge.vim v12, v12, -1, v0
+; RV64-V256-NEXT: vand.vv v8, v12, v8
+; RV64-V256-NEXT: vnot.v v12, v12
+; RV64-V256-NEXT: vand.vv v10, v12, v10
+; RV64-V256-NEXT: vor.vv v8, v8, v10
+; RV64-V256-NEXT: ret
+ %r = call <vscale x 16 x i8> @llvm.ct.select.nxv16i8(i1 %cond, <vscale x 16 x i8> %a, <vscale x 16 x i8> %b)
+ ret <vscale x 16 x i8> %r
+}
+
+; 64-bit elements (useful on RV64)
+define <vscale x 2 x i64> @ctsel_nxv2i64_basic(i1 %cond, <vscale x 2 x i64> %a, <vscale x 2 x i64> %b) {
+; RV64-LABEL: ctsel_nxv2i64_basic:
+; RV64: # %bb.0:
+; RV64-NEXT: andi a0, a0, 1
+; RV64-NEXT: vsetvli a1, zero, e8, mf4, ta, ma
+; RV64-NEXT: vmv.v.x v12, a0
+; RV64-NEXT: vmsne.vi v0, v12, 0
+; RV64-NEXT: vsetvli zero, zero, e64, m2, ta, ma
+; RV64-NEXT: vmv.v.i v12, 0
+; RV64-NEXT: vmerge.vim v12, v12, -1, v0
+; RV64-NEXT: vand.vv v8, v12, v8
+; RV64-NEXT: vnot.v v12, v12
+; RV64-NEXT: vand.vv v10, v12, v10
+; RV64-NEXT: vor.vv v8, v8, v10
+; RV64-NEXT: ret
+;
+; RV32-LABEL: ctsel_nxv2i64_basic:
+; RV32: # %bb.0:
+; RV32-NEXT: andi a0, a0, 1
+; RV32-NEXT: vsetvli a1, zero, e8, mf4, ta, ma
+; RV32-NEXT: vmv.v.x v12, a0
+; RV32-NEXT: vmsne.vi v0, v12, 0
+; RV32-NEXT: vsetvli zero, zero, e64, m2, ta, ma
+; RV32-NEXT: vmv.v.i v12, 0
+; RV32-NEXT: vmerge.vim v12, v12, -1, v0
+; RV32-NEXT: vand.vv v8, v12, v8
+; RV32-NEXT: vnot.v v12, v12
+; RV32-NEXT: vand.vv v10, v12, v10
+; RV32-NEXT: vor.vv v8, v8, v10
+; RV32-NEXT: ret
+;
+; RV32-V128-LABEL: ctsel_nxv2i64_basic:
+; RV32-V128: # %bb.0:
+; RV32-V128-NEXT: andi a0, a0, 1
+; RV32-V128-NEXT: vsetvli a1, zero, e8, mf4, ta, ma
+; RV32-V128-NEXT: vmv.v.x v12, a0
+; RV32-V128-NEXT: vmsne.vi v0, v12, 0
+; RV32-V128-NEXT: vsetvli zero, zero, e64, m2, ta, ma
+; RV32-V128-NEXT: vmv.v.i v12, 0
+; RV32-V128-NEXT: vmerge.vim v12, v12, -1, v0
+; RV32-V128-NEXT: vand.vv v8, v12, v8
+; RV32-V128-NEXT: vnot.v v12, v12
+; RV32-V128-NEXT: vand.vv v10, v12, v10
+; RV32-V128-NEXT: vor.vv v8, v8, v10
+; RV32-V128-NEXT: ret
+;
+; RV64-V256-LABEL: ctsel_nxv2i64_basic:
+; RV64-V256: # %bb.0:
+; RV64-V256-NEXT: andi a0, a0, 1
+; RV64-V256-NEXT: vsetvli a1, zero, e8, mf4, ta, ma
+; RV64-V256-NEXT: vmv.v.x v12, a0
+; RV64-V256-NEXT: vmsne.vi v0, v12, 0
+; RV64-V256-NEXT: vsetvli zero, zero, e64, m2, ta, ma
+; RV64-V256-NEXT: vmv.v.i v12, 0
+; RV64-V256-NEXT: vmerge.vim v12, v12, -1, v0
+; RV64-V256-NEXT: vand.vv v8, v12, v8
+; RV64-V256-NEXT: vnot.v v12, v12
+; RV64-V256-NEXT: vand.vv v10, v12, v10
+; RV64-V256-NEXT: vor.vv v8, v8, v10
+; RV64-V256-NEXT: ret
+ %r = call <vscale x 2 x i64> @llvm.ct.select.nxv2i64(i1 %cond, <vscale x 2 x i64> %a, <vscale x 2 x i64> %b)
+ ret <vscale x 2 x i64> %r
+}
+
+; Floating-point scalable vectors (bitcasted in your fallback)
+define <vscale x 4 x float> @ctsel_nxv4f32_basic(i1 %cond, <vscale x 4 x float> %a, <vscale x 4 x float> %b) {
+; RV64-LABEL: ctsel_nxv4f32_basic:
+; RV64: # %bb.0:
+; RV64-NEXT: andi a0, a0, 1
+; RV64-NEXT: vsetvli a1, zero, e8, mf2, ta, ma
+; RV64-NEXT: vmv.v.x v12, a0
+; RV64-NEXT: vmsne.vi v0, v12, 0
+; RV64-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; RV64-NEXT: vmv.v.i v12, 0
+; RV64-NEXT: vmerge.vim v12, v12, -1, v0
+; RV64-NEXT: vand.vv v8, v12, v8
+; RV64-NEXT: vnot.v v12, v12
+; RV64-NEXT: vand.vv v10, v12, v10
+; RV64-NEXT: vor.vv v8, v8, v10
+; RV64-NEXT: ret
+;
+; RV32-LABEL: ctsel_nxv4f32_basic:
+; RV32: # %bb.0:
+; RV32-NEXT: andi a0, a0, 1
+; RV32-NEXT: vsetvli a1, zero, e8, mf2, ta, ma
+; RV32-NEXT: vmv.v.x v12, a0
+; RV32-NEXT: vmsne.vi v0, v12, 0
+; RV32-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; RV32-NEXT: vmv.v.i v12, 0
+; RV32-NEXT: vmerge.vim v12, v12, -1, v0
+; RV32-NEXT: vand.vv v8, v12, v8
+; RV32-NEXT: vnot.v v12, v12
+; RV32-NEXT: vand.vv v10, v12, v10
+; RV32-NEXT: vor.vv v8, v8, v10
+; RV32-NEXT: ret
+;
+; RV32-V128-LABEL: ctsel_nxv4f32_basic:
+; RV32-V128: # %bb.0:
+; RV32-V128-NEXT: andi a0, a0, 1
+; RV32-V128-NEXT: vsetvli a1, zero, e8, mf2, ta, ma
+; RV32-V128-NEXT: vmv.v.x v12, a0
+; RV32-V128-NEXT: vmsne.vi v0, v12, 0
+; RV32-V128-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; RV32-V128-NEXT: vmv.v.i v12, 0
+; RV32-V128-NEXT: vmerge.vim v12, v12, -1, v0
+; RV32-V128-NEXT: vand.vv v8, v12, v8
+; RV32-V128-NEXT: vnot.v v12, v12
+; RV32-V128-NEXT: vand.vv v10, v12, v10
+; RV32-V128-NEXT: vor.vv v8, v8, v10
+; RV32-V128-NEXT: ret
+;
+; RV64-V256-LABEL: ctsel_nxv4f32_basic:
+; RV64-V256: # %bb.0:
+; RV64-V256-NEXT: andi a0, a0, 1
+; RV64-V256-NEXT: vsetvli a1, zero, e8, mf2, ta, ma
+; RV64-V256-NEXT: vmv.v.x v12, a0
+; RV64-V256-NEXT: vmsne.vi v0, v12, 0
+; RV64-V256-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; RV64-V256-NEXT: vmv.v.i v12, 0
+; RV64-V256-NEXT: vmerge.vim v12, v12, -1, v0
+; RV64-V256-NEXT: vand.vv v8, v12, v8
+; RV64-V256-NEXT: vnot.v v12, v12
+; RV64-V256-NEXT: vand.vv v10, v12, v10
+; RV64-V256-NEXT: vor.vv v8, v8, v10
+; RV64-V256-NEXT: ret
+ %r = call <vscale x 4 x float> @llvm.ct.select.nxv4f32(i1 %cond, <vscale x 4 x float> %a, <vscale x 4 x float> %b)
+ ret <vscale x 4 x float> %r
+}
+
+; FP arithmetic around select
+define <vscale x 4 x float> @ctsel_nxv4f32_arith(i1 %cond, <vscale x 4 x float> %x, <vscale x 4 x float> %y) {
+; RV64-LABEL: ctsel_nxv4f32_arith:
+; RV64: # %bb.0:
+; RV64-NEXT: vsetvli a1, zero, e32, m2, ta, ma
+; RV64-NEXT: vfadd.vv v12, v8, v10
+; RV64-NEXT: vfsub.vv v8, v8, v10
+; RV64-NEXT: andi a0, a0, 1
+; RV64-NEXT: vsetvli zero, zero, e8, mf2, ta, ma
+; RV64-NEXT: vmv.v.x v10, a0
+; RV64-NEXT: vmsne.vi v0, v10, 0
+; RV64-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; RV64-NEXT: vmv.v.i v10, 0
+; RV64-NEXT: vmerge.vim v10, v10, -1, v0
+; RV64-NEXT: vand.vv v12, v10, v12
+; RV64-NEXT: vnot.v v10, v10
+; RV64-NEXT: vand.vv v8, v10, v8
+; RV64-NEXT: vor.vv v8, v12, v8
+; RV64-NEXT: ret
+;
+; RV32-LABEL: ctsel_nxv4f32_arith:
+; RV32: # %bb.0:
+; RV32-NEXT: vsetvli a1, zero, e32, m2, ta, ma
+; RV32-NEXT: vfadd.vv v12, v8, v10
+; RV32-NEXT: vfsub.vv v8, v8, v10
+; RV32-NEXT: andi a0, a0, 1
+; RV32-NEXT: vsetvli zero, zero, e8, mf2, ta, ma
+; RV32-NEXT: vmv.v.x v10, a0
+; RV32-NEXT: vmsne.vi v0, v10, 0
+; RV32-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; RV32-NEXT: vmv.v.i v10, 0
+; RV32-NEXT: vmerge.vim v10, v10, -1, v0
+; RV32-NEXT: vand.vv v12, v10, v12
+; RV32-NEXT: vnot.v v10, v10
+; RV32-NEXT: vand.vv v8, v10, v8
+; RV32-NEXT: vor.vv v8, v12, v8
+; RV32-NEXT: ret
+;
+; RV32-V128-LABEL: ctsel_nxv4f32_arith:
+; RV32-V128: # %bb.0:
+; RV32-V128-NEXT: vsetvli a1, zero, e32, m2, ta, ma
+; RV32-V128-NEXT: vfadd.vv v12, v8, v10
+; RV32-V128-NEXT: vfsub.vv v8, v8, v10
+; RV32-V128-NEXT: andi a0, a0, 1
+; RV32-V128-NEXT: vsetvli zero, zero, e8, mf2, ta, ma
+; RV32-V128-NEXT: vmv.v.x v10, a0
+; RV32-V128-NEXT: vmsne.vi v0, v10, 0
+; RV32-V128-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; RV32-V128-NEXT: vmv.v.i v10, 0
+; RV32-V128-NEXT: vmerge.vim v10, v10, -1, v0
+; RV32-V128-NEXT: vand.vv v12, v10, v12
+; RV32-V128-NEXT: vnot.v v10, v10
+; RV32-V128-NEXT: vand.vv v8, v10, v8
+; RV32-V128-NEXT: vor.vv v8, v12, v8
+; RV32-V128-NEXT: ret
+;
+; RV64-V256-LABEL: ctsel_nxv4f32_arith:
+; RV64-V256: # %bb.0:
+; RV64-V256-NEXT: vsetvli a1, zero, e32, m2, ta, ma
+; RV64-V256-NEXT: vfadd.vv v12, v8, v10
+; RV64-V256-NEXT: vfsub.vv v8, v8, v10
+; RV64-V256-NEXT: andi a0, a0, 1
+; RV64-V256-NEXT: vsetvli zero, zero, e8, mf2, ta, ma
+; RV64-V256-NEXT: vmv.v.x v10, a0
+; RV64-V256-NEXT: vmsne.vi v0, v10, 0
+; RV64-V256-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; RV64-V256-NEXT: vmv.v.i v10, 0
+; RV64-V256-NEXT: vmerge.vim v10, v10, -1, v0
+; RV64-V256-NEXT: vand.vv v12, v10, v12
+; RV64-V256-NEXT: vnot.v v10, v10
+; RV64-V256-NEXT: vand.vv v8, v10, v8
+; RV64-V256-NEXT: vor.vv v8, v12, v8
+; RV64-V256-NEXT: ret
+ %sum = fadd <vscale x 4 x float> %x, %y
+ %diff = fsub <vscale x 4 x float> %x, %y
+ %r = call <vscale x 4 x float> @llvm.ct.select.nxv4f32(i1 %cond, <vscale x 4 x float> %sum, <vscale x 4 x float> %diff)
+ ret <vscale x 4 x float> %r
+}
+
+define <vscale x 2 x double> @ctsel_nxv2f64_basic(i1 %cond, <vscale x 2 x double> %a, <vscale x 2 x double> %b) {
+; RV64-LABEL: ctsel_nxv2f64_basic:
+; RV64: # %bb.0:
+; RV64-NEXT: andi a0, a0, 1
+; RV64-NEXT: vsetvli a1, zero, e8, mf4, ta, ma
+; RV64-NEXT: vmv.v.x v12, a0
+; RV64-NEXT: vmsne.vi v0, v12, 0
+; RV64-NEXT: vsetvli zero, zero, e64, m2, ta, ma
+; RV64-NEXT: vmv.v.i v12, 0
+; RV64-NEXT: vmerge.vim v12, v12, -1, v0
+; RV64-NEXT: vand.vv v8, v12, v8
+; RV64-NEXT: vnot.v v12, v12
+; RV64-NEXT: vand.vv v10, v12, v10
+; RV64-NEXT: vor.vv v8, v8, v10
+; RV64-NEXT: ret
+;
+; RV32-LABEL: ctsel_nxv2f64_basic:
+; RV32: # %bb.0:
+; RV32-NEXT: andi a0, a0, 1
+; RV32-NEXT: vsetvli a1, zero, e8, mf4, ta, ma
+; RV32-NEXT: vmv.v.x v12, a0
+; RV32-NEXT: vmsne.vi v0, v12, 0
+; RV32-NEXT: vsetvli zero, zero, e64, m2, ta, ma
+; RV32-NEXT: vmv.v.i v12, 0
+; RV32-NEXT: vmerge.vim v12, v12, -1, v0
+; RV32-NEXT: vand.vv v8, v12, v8
+; RV32-NEXT: vnot.v v12, v12
+; RV32-NEXT: vand.vv v10, v12, v10
+; RV32-NEXT: vor.vv v8, v8, v10
+; RV32-NEXT: ret
+;
+; RV32-V128-LABEL: ctsel_nxv2f64_basic:
+; RV32-V128: # %bb.0:
+; RV32-V128-NEXT: andi a0, a0, 1
+; RV32-V128-NEXT: vsetvli a1, zero, e8, mf4, ta, ma
+; RV32-V128-NEXT: vmv.v.x v12, a0
+; RV32-V128-NEXT: vmsne.vi v0, v12, 0
+; RV32-V128-NEXT: vsetvli zero, zero, e64, m2, ta, ma
+; RV32-V128-NEXT: vmv.v.i v12, 0
+; RV32-V128-NEXT: vmerge.vim v12, v12, -1, v0
+; RV32-V128-NEXT: vand.vv v8, v12, v8
+; RV32-V128-NEXT: vnot.v v12, v12
+; RV32-V128-NEXT: vand.vv v10, v12, v10
+; RV32-V128-NEXT: vor.vv v8, v8, v10
+; RV32-V128-NEXT: ret
+;
+; RV64-V256-LABEL: ctsel_nxv2f64_basic:
+; RV64-V256: # %bb.0:
+; RV64-V256-NEXT: andi a0, a0, 1
+; RV64-V256-NEXT: vsetvli a1, zero, e8, mf4, ta, ma
+; RV64-V256-NEXT: vmv.v.x v12, a0
+; RV64-V256-NEXT: vmsne.vi v0, v12, 0
+; RV64-V256-NEXT: vsetvli zero, zero, e64, m2, ta, ma
+; RV64-V256-NEXT: vmv.v.i v12, 0
+; RV64-V256-NEXT: vmerge.vim v12, v12, -1, v0
+; RV64-V256-NEXT: vand.vv v8, v12, v8
+; RV64-V256-NEXT: vnot.v v12, v12
+; RV64-V256-NEXT: vand.vv v10, v12, v10
+; RV64-V256-NEXT: vor.vv v8, v8, v10
+; RV64-V256-NEXT: ret
+ %r = call <vscale x 2 x double> @llvm.ct.select.nxv2f64(i1 %cond, <vscale x 2 x double> %a, <vscale x 2 x double> %b)
+ ret <vscale x 2 x double> %r
+}
+
+declare <vscale x 4 x i32> @llvm.ct.select.nxv4i32(i1, <vscale x 4 x i32>, <vscale x 4 x i32>)
+declare <vscale x 8 x i16> @llvm.ct.select.nxv8i16(i1, <vscale x 8 x i16>, <vscale x 8 x i16>)
+declare <vscale x 16 x i8> @llvm.ct.select.nxv16i8(i1, <vscale x 16 x i8>, <vscale x 16 x i8>)
+declare <vscale x 2 x i64> @llvm.ct.select.nxv2i64(i1, <vscale x 2 x i64>, <vscale x 2 x i64>)
+declare <vscale x 4 x float> @llvm.ct.select.nxv4f32(i1, <vscale x 4 x float>, <vscale x 4 x float>)
+declare <vscale x 2 x double>@llvm.ct.select.nxv2f64(i1, <vscale x 2 x double>,<vscale x 2 x double>)
diff --git a/llvm/test/CodeGen/RISCV/ctselect-side-effects.ll b/llvm/test/CodeGen/RISCV/ctselect-side-effects.ll
new file mode 100644
index 0000000000000..255d575ca8f9f
--- /dev/null
+++ b/llvm/test/CodeGen/RISCV/ctselect-side-effects.ll
@@ -0,0 +1,176 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc < %s -mtriple=riscv64 -O3 -filetype=asm | FileCheck %s --check-prefix=RV64
+; RUN: llc < %s -mtriple=riscv32 -O3 -filetype=asm | FileCheck %s --check-prefix=RV32
+
+; Test 1: Basic optimizations should still work
+define i32 @test_basic_opts(i32 %x) {
+; RV64-LABEL: test_basic_opts:
+; RV64: # %bb.0:
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_basic_opts:
+; RV32: # %bb.0:
+; RV32-NEXT: ret
+ %a = or i32 %x, 0 ; Should eliminate
+ %b = and i32 %a, -1 ; Should eliminate
+ %c = xor i32 %b, 0 ; Should eliminate
+ ret i32 %c
+}
+
+; Test 2: Constant folding should work
+define i32 @test_constant_fold() {
+; RV64-LABEL: test_constant_fold:
+; RV64: # %bb.0:
+; RV64-NEXT: li a0, 0
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_constant_fold:
+; RV32: # %bb.0:
+; RV32-NEXT: li a0, 0
+; RV32-NEXT: ret
+ %a = xor i32 -1, -1 ; Should fold to 0
+ ret i32 %a
+}
+
+; Test 3: Protected pattern should NOT have branches
+define i32 @test_protected_no_branch(i1 %cond, i32 %a, i32 %b) {
+; RV64-LABEL: test_protected_no_branch:
+; RV64: # %bb.0:
+; RV64-NEXT: xor a1, a1, a2
+; RV64-NEXT: slli a0, a0, 63
+; RV64-NEXT: srai a0, a0, 63
+; RV64-NEXT: and a0, a1, a0
+; RV64-NEXT: xor a0, a0, a2
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_protected_no_branch:
+; RV32: # %bb.0:
+; RV32-NEXT: andi a0, a0, 1
+; RV32-NEXT: neg a3, a0
+; RV32-NEXT: addi a0, a0, -1
+; RV32-NEXT: and a1, a3, a1
+; RV32-NEXT: and a0, a0, a2
+; RV32-NEXT: or a0, a1, a0
+; RV32-NEXT: ret
+ %result = call i32 @llvm.ct.select.i32(i1 %cond, i32 %a, i32 %b)
+ ret i32 %result
+}
+
+; Test 4: Explicit branch should still generate branches
+define i32 @test_explicit_branch(i1 %cond, i32 %a, i32 %b) {
+; RV64-LABEL: test_explicit_branch:
+; RV64: # %bb.0:
+; RV64-NEXT: andi a0, a0, 1
+; RV64-NEXT: beqz a0, .LBB3_2
+; RV64-NEXT: # %bb.1: # %true
+; RV64-NEXT: mv a0, a1
+; RV64-NEXT: ret
+; RV64-NEXT: .LBB3_2: # %false
+; RV64-NEXT: mv a0, a2
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_explicit_branch:
+; RV32: # %bb.0:
+; RV32-NEXT: andi a0, a0, 1
+; RV32-NEXT: beqz a0, .LBB3_2
+; RV32-NEXT: # %bb.1: # %true
+; RV32-NEXT: mv a0, a1
+; RV32-NEXT: ret
+; RV32-NEXT: .LBB3_2: # %false
+; RV32-NEXT: mv a0, a2
+; RV32-NEXT: ret
+ br i1 %cond, label %true, label %false
+true:
+ ret i32 %a
+false:
+ ret i32 %b
+}
+
+; Test 5: Regular select (not ct.select) - whatever wasm wants to do
+define i32 @test_regular_select(i1 %cond, i32 %a, i32 %b) {
+; RV64-LABEL: test_regular_select:
+; RV64: # %bb.0:
+; RV64-NEXT: andi a3, a0, 1
+; RV64-NEXT: mv a0, a1
+; RV64-NEXT: bnez a3, .LBB4_2
+; RV64-NEXT: # %bb.1:
+; RV64-NEXT: mv a0, a2
+; RV64-NEXT: .LBB4_2:
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_regular_select:
+; RV32: # %bb.0:
+; RV32-NEXT: andi a3, a0, 1
+; RV32-NEXT: mv a0, a1
+; RV32-NEXT: bnez a3, .LBB4_2
+; RV32-NEXT: # %bb.1:
+; RV32-NEXT: mv a0, a2
+; RV32-NEXT: .LBB4_2:
+; RV32-NEXT: ret
+ %result = select i1 %cond, i32 %a, i32 %b
+ ret i32 %result
+}
+
+; Test if XOR with all-ones still gets optimized
+define i32 @test_xor_all_ones() {
+; RV64-LABEL: test_xor_all_ones:
+; RV64: # %bb.0:
+; RV64-NEXT: li a0, 0
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_xor_all_ones:
+; RV32: # %bb.0:
+; RV32-NEXT: li a0, 0
+; RV32-NEXT: ret
+ %xor1 = xor i32 -1, -1 ; Should optimize to 0
+ ret i32 %xor1
+}
+
+define i32 @test_xor_same_value(i32 %x) {
+; RV64-LABEL: test_xor_same_value:
+; RV64: # %bb.0:
+; RV64-NEXT: li a0, 0
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_xor_same_value:
+; RV32: # %bb.0:
+; RV32-NEXT: li a0, 0
+; RV32-NEXT: ret
+ %xor2 = xor i32 %x, %x ; Should optimize to 0
+ ret i32 %xor2
+}
+
+define i32 @test_normal_ops(i32 %x) {
+; RV64-LABEL: test_normal_ops:
+; RV64: # %bb.0:
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_normal_ops:
+; RV32: # %bb.0:
+; RV32-NEXT: ret
+ %or1 = or i32 %x, 0 ; Should optimize to %x
+ %and1 = and i32 %or1, -1 ; Should optimize to %x
+ %xor1 = xor i32 %and1, 0 ; Should optimize to %x
+ ret i32 %xor1
+}
+
+; This simulates what the reviewer is worried about
+define i32 @test_xor_with_const_operands() {
+; RV64-LABEL: test_xor_with_const_operands:
+; RV64: # %bb.0:
+; RV64-NEXT: li a0, 0
+; RV64-NEXT: ret
+;
+; RV32-LABEL: test_xor_with_const_operands:
+; RV32: # %bb.0:
+; RV32-NEXT: li a0, 0
+; RV32-NEXT: ret
+ %a = xor i32 -1, -1 ; -1 ^ -1 should become 0
+ %b = xor i32 0, 0 ; 0 ^ 0 should become 0
+ %c = xor i32 42, 42 ; 42 ^ 42 should become 0
+ %result = or i32 %a, %b
+ %final = or i32 %result, %c
+ ret i32 %final ; Should optimize to 0
+}
+
+declare i32 @llvm.ct.select.i32(i1, i32, i32)
More information about the llvm-branch-commits
mailing list