From 0e5d70c02c371c05c840e6c34b4d8a26aa071f40 Mon Sep 17 00:00:00 2001 From: Sanjay Patel Date: Tue, 15 Nov 2016 21:19:28 +0000 Subject: [PATCH] [x86] add tests for FP-logic equivalent instruction replacement The ANDN test needs at least 3 different fixes. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@287032 91177308-0d34-0410-b5e6-96231b3b80d8 --- test/CodeGen/X86/fp-logic-replace.ll | 91 ++++++++++++++++++++++++++++ 1 file changed, 91 insertions(+) create mode 100644 test/CodeGen/X86/fp-logic-replace.ll diff --git a/test/CodeGen/X86/fp-logic-replace.ll b/test/CodeGen/X86/fp-logic-replace.ll new file mode 100644 index 00000000000..8f2f649ce70 --- /dev/null +++ b/test/CodeGen/X86/fp-logic-replace.ll @@ -0,0 +1,91 @@ +; NOTE: Assertions have been autogenerated by utils/update_test_checks.py +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse2 | FileCheck %s --check-prefix=SSE +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefix=AVX + +; Test that we can replace "scalar" FP-bitwise-logic with the optimal instruction. +; Scalar x86 FP-logic instructions only exist in your imagination and/or the bowels +; of compilers, but float and double variants of FP-logic instructions are reality +; and float may be a shorter instruction depending on which flavor of vector ISA +; you have...so just prefer float all the time, ok? Yay, x86! + +define double @FsANDPSrr(double %x, double %y) { +; SSE-LABEL: FsANDPSrr: +; SSE: # BB#0: +; SSE-NEXT: andpd %xmm1, %xmm0 +; SSE-NEXT: retq +; +; AVX-LABEL: FsANDPSrr: +; AVX: # BB#0: +; AVX-NEXT: vandpd %xmm1, %xmm0, %xmm0 +; AVX-NEXT: retq +; + %bc1 = bitcast double %x to i64 + %bc2 = bitcast double %y to i64 + %and = and i64 %bc1, %bc2 + %bc3 = bitcast i64 %and to double + ret double %bc3 +} + +define double @FsANDNPSrr(double %x, double %y) { +; SSE-LABEL: FsANDNPSrr: +; SSE: # BB#0: +; SSE-NEXT: movd %xmm0, %rax +; SSE-NEXT: movd %xmm1, %rcx +; SSE-NEXT: notq %rcx +; SSE-NEXT: andq %rax, %rcx +; SSE-NEXT: movd %rcx, %xmm0 +; SSE-NEXT: retq +; +; AVX-LABEL: FsANDNPSrr: +; AVX: # BB#0: +; AVX-NEXT: vmovq %xmm0, %rax +; AVX-NEXT: vmovq %xmm1, %rcx +; AVX-NEXT: notq %rcx +; AVX-NEXT: andq %rax, %rcx +; AVX-NEXT: vmovq %rcx, %xmm0 +; AVX-NEXT: retq +; + %bc1 = bitcast double %x to i64 + %bc2 = bitcast double %y to i64 + %not = xor i64 %bc2, -1 + %and = and i64 %bc1, %not + %bc3 = bitcast i64 %and to double + ret double %bc3 +} + +define double @FsORPSrr(double %x, double %y) { +; SSE-LABEL: FsORPSrr: +; SSE: # BB#0: +; SSE-NEXT: orpd %xmm1, %xmm0 +; SSE-NEXT: retq +; +; AVX-LABEL: FsORPSrr: +; AVX: # BB#0: +; AVX-NEXT: vorpd %xmm1, %xmm0, %xmm0 +; AVX-NEXT: retq +; + %bc1 = bitcast double %x to i64 + %bc2 = bitcast double %y to i64 + %or = or i64 %bc1, %bc2 + %bc3 = bitcast i64 %or to double + ret double %bc3 +} + +define double @FsXORPSrr(double %x, double %y) { +; SSE-LABEL: FsXORPSrr: +; SSE: # BB#0: +; SSE-NEXT: xorpd %xmm1, %xmm0 +; SSE-NEXT: retq +; +; AVX-LABEL: FsXORPSrr: +; AVX: # BB#0: +; AVX-NEXT: vxorpd %xmm1, %xmm0, %xmm0 +; AVX-NEXT: retq +; + %bc1 = bitcast double %x to i64 + %bc2 = bitcast double %y to i64 + %xor = xor i64 %bc1, %bc2 + %bc3 = bitcast i64 %xor to double + ret double %bc3 +} + -- 2.40.0