[X86][SSE] Add some nt-store test cases inspired by PR42123

llvm-svn: 362594

1 files changed, 163 insertions, 0 deletions

diff --git a/llvm/test/CodeGen/X86/merge-consecutive-stores-nt.ll b/llvm/test/CodeGen/X86/merge-consecutive-stores-nt.ll
new file mode 100644
index 00000000000..0461008f84d
--- /dev/null
+++ b/llvm/test/CodeGen/X86/merge-consecutive-stores-nt.ll
@@ -0,0 +1,163 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+sse2 | FileCheck %s --check-prefixes=X86
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown | FileCheck %s --check-prefixes=X64,X64-SSE,X64-SSE2
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4a | FileCheck %s --check-prefixes=X64,X64-SSE,X64-SSE4A
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefixes=X64,X64-AVX
+
+;
+; PR42123
+;
+
+; FIXME: AVX doesn't retain NT flag on store.
+; Should be VMOVNTPS ymm.
+define void @merge_2_v4f32_align32(<4 x float>* %a0, <4 x float>* %a1)  {
+; X86-LABEL: merge_2_v4f32_align32:
+; X86:       # %bb.0:
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT:    movaps (%ecx), %xmm0
+; X86-NEXT:    movaps 16(%ecx), %xmm1
+; X86-NEXT:    movntps %xmm0, (%eax)
+; X86-NEXT:    movntps %xmm1, 16(%eax)
+; X86-NEXT:    retl
+;
+; X64-SSE-LABEL: merge_2_v4f32_align32:
+; X64-SSE:       # %bb.0:
+; X64-SSE-NEXT:    movaps (%rdi), %xmm0
+; X64-SSE-NEXT:    movaps 16(%rdi), %xmm1
+; X64-SSE-NEXT:    movntps %xmm0, (%rsi)
+; X64-SSE-NEXT:    movntps %xmm1, 16(%rsi)
+; X64-SSE-NEXT:    retq
+;
+; X64-AVX-LABEL: merge_2_v4f32_align32:
+; X64-AVX:       # %bb.0:
+; X64-AVX-NEXT:    vmovaps (%rdi), %ymm0
+; X64-AVX-NEXT:    vmovaps %ymm0, (%rsi)
+; X64-AVX-NEXT:    vzeroupper
+; X64-AVX-NEXT:    retq
+  %1 = getelementptr inbounds <4 x float>, <4 x float>* %a0, i64 1, i64 0
+  %2 = bitcast float* %1 to <4 x float>*
+  %3 = load <4 x float>, <4 x float>* %a0, align 32
+  %4 = load <4 x float>, <4 x float>* %2, align 16
+  %5 = getelementptr inbounds <4 x float>, <4 x float>* %a1, i64 1, i64 0
+  %6 = bitcast float* %5 to <4 x float>*
+  store <4 x float> %3, <4 x float>* %a1, align 32, !nontemporal !0
+  store <4 x float> %4, <4 x float>* %6, align 16, !nontemporal !0
+  ret void
+}
+
+; FIXME: shouldn't attempt to merge nt and non-nt stores even if aligned.
+; Must be kept seperate as VMOVNTPS xmm + VMOVAPS xmm.
+define void @merge_2_v4f32_align32_mix(<4 x float>* %a0, <4 x float>* %a1)  {
+; X86-LABEL: merge_2_v4f32_align32_mix:
+; X86:       # %bb.0:
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT:    movaps (%ecx), %xmm0
+; X86-NEXT:    movaps 16(%ecx), %xmm1
+; X86-NEXT:    movntps %xmm0, (%eax)
+; X86-NEXT:    movaps %xmm1, 16(%eax)
+; X86-NEXT:    retl
+;
+; X64-SSE-LABEL: merge_2_v4f32_align32_mix:
+; X64-SSE:       # %bb.0:
+; X64-SSE-NEXT:    movaps (%rdi), %xmm0
+; X64-SSE-NEXT:    movaps 16(%rdi), %xmm1
+; X64-SSE-NEXT:    movntps %xmm0, (%rsi)
+; X64-SSE-NEXT:    movaps %xmm1, 16(%rsi)
+; X64-SSE-NEXT:    retq
+;
+; X64-AVX-LABEL: merge_2_v4f32_align32_mix:
+; X64-AVX:       # %bb.0:
+; X64-AVX-NEXT:    vmovaps (%rdi), %ymm0
+; X64-AVX-NEXT:    vmovaps %ymm0, (%rsi)
+; X64-AVX-NEXT:    vzeroupper
+; X64-AVX-NEXT:    retq
+  %1 = getelementptr inbounds <4 x float>, <4 x float>* %a0, i64 1, i64 0
+  %2 = bitcast float* %1 to <4 x float>*
+  %3 = load <4 x float>, <4 x float>* %a0, align 32
+  %4 = load <4 x float>, <4 x float>* %2, align 16
+  %5 = getelementptr inbounds <4 x float>, <4 x float>* %a1, i64 1, i64 0
+  %6 = bitcast float* %5 to <4 x float>*
+  store <4 x float> %3, <4 x float>* %a1, align 32, !nontemporal !0
+  store <4 x float> %4, <4 x float>* %6, align 16
+  ret void
+}
+
+; FIXME: AVX can't perform NT-store-ymm on 16-byte aligned memory.
+; Must be kept seperate as VMOVNTPS xmm.
+define void @merge_2_v4f32_align16(<4 x float>* %a0, <4 x float>* %a1)  {
+; X86-LABEL: merge_2_v4f32_align16:
+; X86:       # %bb.0:
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT:    movaps (%ecx), %xmm0
+; X86-NEXT:    movaps 16(%ecx), %xmm1
+; X86-NEXT:    movntps %xmm0, (%eax)
+; X86-NEXT:    movntps %xmm1, 16(%eax)
+; X86-NEXT:    retl
+;
+; X64-SSE-LABEL: merge_2_v4f32_align16:
+; X64-SSE:       # %bb.0:
+; X64-SSE-NEXT:    movaps (%rdi), %xmm0
+; X64-SSE-NEXT:    movaps 16(%rdi), %xmm1
+; X64-SSE-NEXT:    movntps %xmm0, (%rsi)
+; X64-SSE-NEXT:    movntps %xmm1, 16(%rsi)
+; X64-SSE-NEXT:    retq
+;
+; X64-AVX-LABEL: merge_2_v4f32_align16:
+; X64-AVX:       # %bb.0:
+; X64-AVX-NEXT:    vmovups (%rdi), %ymm0
+; X64-AVX-NEXT:    vmovups %ymm0, (%rsi)
+; X64-AVX-NEXT:    vzeroupper
+; X64-AVX-NEXT:    retq
+  %1 = getelementptr inbounds <4 x float>, <4 x float>* %a0, i64 1, i64 0
+  %2 = bitcast float* %1 to <4 x float>*
+  %3 = load <4 x float>, <4 x float>* %a0, align 16
+  %4 = load <4 x float>, <4 x float>* %2, align 16
+  %5 = getelementptr inbounds <4 x float>, <4 x float>* %a1, i64 1, i64 0
+  %6 = bitcast float* %5 to <4 x float>*
+  store <4 x float> %3, <4 x float>* %a1, align 16, !nontemporal !0
+  store <4 x float> %4, <4 x float>* %6, align 16, !nontemporal !0
+  ret void
+}
+
+; FIXME: Nothing can perform NT-store-vector on 1-byte aligned memory.
+; Must be scalarized to use MOVTNI/MOVNTSD.
+define void @merge_2_v4f32_align1(<4 x float>* %a0, <4 x float>* %a1)  {
+; X86-LABEL: merge_2_v4f32_align1:
+; X86:       # %bb.0:
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT:    movups (%ecx), %xmm0
+; X86-NEXT:    movups 16(%ecx), %xmm1
+; X86-NEXT:    movups %xmm0, (%eax)
+; X86-NEXT:    movups %xmm1, 16(%eax)
+; X86-NEXT:    retl
+;
+; X64-SSE-LABEL: merge_2_v4f32_align1:
+; X64-SSE:       # %bb.0:
+; X64-SSE-NEXT:    movups (%rdi), %xmm0
+; X64-SSE-NEXT:    movups 16(%rdi), %xmm1
+; X64-SSE-NEXT:    movups %xmm0, (%rsi)
+; X64-SSE-NEXT:    movups %xmm1, 16(%rsi)
+; X64-SSE-NEXT:    retq
+;
+; X64-AVX-LABEL: merge_2_v4f32_align1:
+; X64-AVX:       # %bb.0:
+; X64-AVX-NEXT:    vmovups (%rdi), %ymm0
+; X64-AVX-NEXT:    vmovups %ymm0, (%rsi)
+; X64-AVX-NEXT:    vzeroupper
+; X64-AVX-NEXT:    retq
+  %1 = getelementptr inbounds <4 x float>, <4 x float>* %a0, i64 1, i64 0
+  %2 = bitcast float* %1 to <4 x float>*
+  %3 = load <4 x float>, <4 x float>* %a0, align 1
+  %4 = load <4 x float>, <4 x float>* %2, align 1
+  %5 = getelementptr inbounds <4 x float>, <4 x float>* %a1, i64 1, i64 0
+  %6 = bitcast float* %5 to <4 x float>*
+  store <4 x float> %3, <4 x float>* %a1, align 1, !nontemporal !0
+  store <4 x float> %4, <4 x float>* %6, align 1, !nontemporal !0
+  ret void
+}
+
+!0 = !{i32 1}