[llvm] e87a95b - [X86] Add test case for unfolding broadcast load from vpternlog.

[Craig Topper via llvm-commits]

Author: Craig Topper
Date: 2020-07-02T13:43:43-07:00
New Revision: e87a95b5c24bf2c47af3b90990ac559453266981

URL: https://github.com/llvm/llvm-project/commit/e87a95b5c24bf2c47af3b90990ac559453266981
DIFF: https://github.com/llvm/llvm-project/commit/e87a95b5c24bf2c47af3b90990ac559453266981.diff

LOG: [X86] Add test case for unfolding broadcast load from vpternlog.

Added: 
    

Modified: 
    llvm/test/CodeGen/X86/avx512-broadcast-unfold.ll

Removed: 
    


################################################################################
diff  --git a/llvm/test/CodeGen/X86/avx512-broadcast-unfold.ll b/llvm/test/CodeGen/X86/avx512-broadcast-unfold.ll
index aa07a4db8d59..6448b3c6ef46 100644
--- a/llvm/test/CodeGen/X86/avx512-broadcast-unfold.ll
+++ b/llvm/test/CodeGen/X86/avx512-broadcast-unfold.ll
@@ -4633,3 +4633,51 @@ bb1:                                              ; preds = %bb1, %bb
 bb10:                                             ; preds = %bb1
   ret void
 }
+
+; The or/and pattern here should be turned into vpternlog. The multiply is
+; there to increase the use count of the loads so they can't fold. We want to
+; unfold the broadcast and pull it out of the loop.
+define void @bcast_unfold_vpternlog_v16i32(i32* %arg, i32* %arg1) {
+; CHECK-LABEL: bcast_unfold_vpternlog_v16i32:
+; CHECK:       # %bb.0: # %bb
+; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; CHECK-NEXT:    .p2align 4, 0x90
+; CHECK-NEXT:  .LBB131_1: # %bb2
+; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    vmovdqu64 4096(%rdi,%rax), %zmm0
+; CHECK-NEXT:    vmovdqu64 4096(%rsi,%rax), %zmm1
+; CHECK-NEXT:    vpmulld %zmm1, %zmm0, %zmm2
+; CHECK-NEXT:    vpternlogd $216, {{.*}}(%rip){1to16}, %zmm0, %zmm1
+; CHECK-NEXT:    vpmulld %zmm2, %zmm1, %zmm0
+; CHECK-NEXT:    vmovdqu64 %zmm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    addq $64, %rax
+; CHECK-NEXT:    jne .LBB131_1
+; CHECK-NEXT:  # %bb.2: # %bb20
+; CHECK-NEXT:    vzeroupper
+; CHECK-NEXT:    retq
+bb:
+  br label %bb2
+
+bb2:                                              ; preds = %bb2, %bb
+  %tmp = phi i64 [ 0, %bb ], [ %tmp18, %bb2 ]
+  %tmp3 = getelementptr inbounds i32, i32* %arg, i64 %tmp
+  %tmp4 = bitcast i32* %tmp3 to <16 x i32>*
+  %tmp5 = load <16 x i32>, <16 x i32>* %tmp4, align 4
+  %tmp6 = getelementptr inbounds i32, i32* %arg1, i64 %tmp
+  %tmp10 = bitcast i32* %tmp6 to <16 x i32>*
+  %tmp11 = load <16 x i32>, <16 x i32>* %tmp10, align 4
+  %tmp12 = and <16 x i32> %tmp5, <i32 32767, i32 32767, i32 32767, i32 32767, i32 32767, i32 32767, i32 32767, i32 32767, i32 32767, i32 32767, i32 32767, i32 32767, i32 32767, i32 32767, i32 32767, i32 32767>
+  %tmp13 = and <16 x i32> %tmp11, <i32 -32768, i32 -32768, i32 -32768, i32 -32768, i32 -32768, i32 -32768, i32 -32768, i32 -32768, i32 -32768, i32 -32768, i32 -32768, i32 -32768, i32 -32768, i32 -32768, i32 -32768, i32 -32768>
+  %tmp14 = or <16 x i32> %tmp12, %tmp13
+  %tmp15 = mul <16 x i32> %tmp14, %tmp5
+  %tmp16 = mul <16 x i32> %tmp15, %tmp11
+  %tmp17 = bitcast i32* %tmp3 to <16 x i32>*
+  store <16 x i32> %tmp16, <16 x i32>* %tmp17, align 4
+  %tmp18 = add i64 %tmp, 16
+  %tmp19 = icmp eq i64 %tmp18, 1024
+  br i1 %tmp19, label %bb20, label %bb2
+
+bb20:                                             ; preds = %bb2
+  ret void
+}
+