[Codegen][X86] EltsFromConsecutiveLoads(): if only have AVX1, ensure that the "load" is actually foldable (PR51615)

[lebedev.ri]

Summary

This fixes another reproducer from https://bugs.llvm.org/show_bug.cgi?id=51615
And again, the fix lies not in the code added in D105390

In this case, we completely don't check that the "broadcast-from-mem" we create
can actually fold the load. In this case, it's operand was not a load at all:

Combining: t16: v8i32 = vector_shuffle<0,u,u,u,0,u,u,u> t14, undef:v8i32
Creating new node: t29: i32 = undef
RepeatLoad:
t8: i32 = truncate t7
  t7: i64 = extract_vector_elt t5, Constant:i64<0>
    t5: v2i64,ch = load<(load (s128) from %ir.arg)> t0, t2, undef:i64
      t2: i64,ch = CopyFromReg t0, Register:i64 %0
        t1: i64 = Register %0
      t4: i64 = undef
    t3: i64 = Constant<0>
Combining: t15: v8i32 = undef

[lebedev.ri]

Err, wrong snippet. now this actually shows the problem:

Combining: t16: v8i32 = vector_shuffle<0,u,u,u,0,u,u,u> t14, undef:v8i32
Creating new node: t29: i32 = undef
RepeatLoad:
t8: i32 = truncate t7
  t7: i64 = extract_vector_elt t5, Constant:i64<0>
    t5: v2i64,ch = load<(load (s128) from %ir.arg)> t0, t2, undef:i64
      t2: i64,ch = CopyFromReg t0, Register:i64 %0
        t1: i64 = Register %0
      t4: i64 = undef
    t3: i64 = Constant<0>
Combining: t15: v8i32 = undef

[RKSimon]

LGTM

[alexfh]

Thanks for the prompt fix! One nit.

llvm/lib/Target/X86/X86ISelLowering.cpp

5035–5036

nit: "Function names should be verb phrases (as they represent actions), and command-like function should be imperative. The name should be camel case, and start with a lower case letter (e.g. openFile() or isFoo())."

https://llvm.org/docs/CodingStandards.html#name-types-functions-variables-and-enumerators-properly

Same below.

llvm/lib/Target/X86/X86ISelLowering.cpp

X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
                                   const TargetLibraryInfo *libInfo) const {
   return X86::createFastISel(funcInfo, libInfo);
 }
 
 //===----------------------------------------------------------------------===//
 //                           Other Lowering Hooks
 //===----------------------------------------------------------------------===//
 
-static bool MayFoldLoad(SDValue Op) {
-  return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
+static bool MayFoldLoad(SDValue Op, bool AssumeSingleUse = false) {
+  return (AssumeSingleUse || Op.hasOneUse()) && ISD::isNormalLoad(Op.getNode());
 }
 
-static bool MayFoldLoadIntoBroadcastFromMem(SDValue Op, MVT EltVT) {
-  if (!MayFoldLoad(Op))
+static bool MayFoldLoadIntoBroadcastFromMem(SDValue Op, MVT EltVT,
+                                            bool AssumeSingleUse = false) {
+  if (!MayFoldLoad(Op, AssumeSingleUse))
     return false;
 
   // We can not replace a wide volatile load with a broadcast-from-memory,
   // because that would narrow the load, which isn't legal for volatiles.
   const LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op.getNode());
   return !Ld->isVolatile() ||
          Ld->getValueSizeInBits(0) == EltVT.getScalarSizeInBits();
 }
for (unsigned SubElems = 1; SubElems < NumElems; SubElems *= 2) {
       if (TLI.isTypeLegal(BroadcastVT)) {
         if (SDValue RepeatLoad = EltsFromConsecutiveLoads(
                 RepeatVT, RepeatedLoads, DL, DAG, Subtarget, IsAfterLegalize)) {
           SDValue Broadcast = RepeatLoad;
           if (RepeatSize > ScalarSize) {
             while (Broadcast.getValueSizeInBits() < VT.getSizeInBits())
               Broadcast = concatSubVectors(Broadcast, Broadcast, DAG, DL);
           } else {
+            if (!Subtarget.hasAVX2() &&
+                !MayFoldLoadIntoBroadcastFromMem(
+                    RepeatLoad, RepeatVT.getScalarType().getSimpleVT(),
+                    /*AssumeSingleUse=*/true))
+              return SDValue();
             Broadcast =
                 DAG.getNode(X86ISD::VBROADCAST, DL, BroadcastVT, RepeatLoad);
           }
           return DAG.getBitcast(VT, Broadcast);
         }
       }
     }
   }

llvm/test/CodeGen/X86/pr51615.ll

 ; ALL-NEXT:    retq
   %i0 = load volatile <2 x float>, <2 x float>* @g2, align 16
   %i = bitcast <2 x float> %i0 to <1 x double>
   %i1 = shufflevector <1 x double> %i, <1 x double> poison, <4 x i32> <i32 undef, i32 0, i32 undef, i32 0>
   %shuffle1 = shufflevector <4 x double> %i1, <4 x double> zeroinitializer, <8 x i32> <i32 6, i32 7, i32 3, i32 6, i32 7, i32 1, i32 7, i32 1>
   store volatile <8 x double> %shuffle1, <8 x double>* undef, align 64
   ret void
 }
+
+define void @elts_from_consecutive_loads(<2 x i64>* %arg, i32* %arg12, <8 x i32>* %arg13, float %arg14, i1 %arg15) {
+; AVX-LABEL: elts_from_consecutive_loads:
+; AVX:       # %bb.0: # %bb
+; AVX-NEXT:    vxorps %xmm1, %xmm1, %xmm1
+; AVX-NEXT:    .p2align 4, 0x90
+; AVX-NEXT:  .LBB3_1: # %bb16
+; AVX-NEXT:    # =>This Loop Header: Depth=1
+; AVX-NEXT:    # Child Loop BB3_2 Depth 2
+; AVX-NEXT:    testb $1, %cl
+; AVX-NEXT:    je .LBB3_1
+; AVX-NEXT:    .p2align 4, 0x90
+; AVX-NEXT:  .LBB3_2: # %bb17
+; AVX-NEXT:    # Parent Loop BB3_1 Depth=1
+; AVX-NEXT:    # => This Inner Loop Header: Depth=2
+; AVX-NEXT:    movl (%rdi), %eax
+; AVX-NEXT:    vbroadcastss (%rdi), %ymm2
+; AVX-NEXT:    movl %eax, (%rsi)
+; AVX-NEXT:    vmovaps %ymm2, (%rdx)
+; AVX-NEXT:    vucomiss %xmm1, %xmm0
+; AVX-NEXT:    jne .LBB3_2
+; AVX-NEXT:    jp .LBB3_2
+; AVX-NEXT:    jmp .LBB3_1
+;
+; AVX2-LABEL: elts_from_consecutive_loads:
+; AVX2:       # %bb.0: # %bb
+; AVX2-NEXT:    vxorps %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    .p2align 4, 0x90
+; AVX2-NEXT:  .LBB3_1: # %bb16
+; AVX2-NEXT:    # =>This Loop Header: Depth=1
+; AVX2-NEXT:    # Child Loop BB3_2 Depth 2
+; AVX2-NEXT:    testb $1, %cl
+; AVX2-NEXT:    je .LBB3_1
+; AVX2-NEXT:    .p2align 4, 0x90
+; AVX2-NEXT:  .LBB3_2: # %bb17
+; AVX2-NEXT:    # Parent Loop BB3_1 Depth=1
+; AVX2-NEXT:    # => This Inner Loop Header: Depth=2
+; AVX2-NEXT:    vmovaps (%rdi), %xmm2
+; AVX2-NEXT:    vmovss %xmm2, (%rsi)
+; AVX2-NEXT:    vbroadcastss %xmm2, %ymm2
+; AVX2-NEXT:    vmovaps %ymm2, (%rdx)
+; AVX2-NEXT:    vucomiss %xmm1, %xmm0
+; AVX2-NEXT:    jne .LBB3_2
+; AVX2-NEXT:    jp .LBB3_2
+; AVX2-NEXT:    jmp .LBB3_1
+bb:
+  br label %bb16
+
+bb16:                                             ; preds = %bb17, %bb16, %bb
+  br i1 %arg15, label %bb17, label %bb16
+
+bb17:                                             ; preds = %bb17, %bb16
+  %tmp = load <2 x i64>, <2 x i64>* %arg, align 16
+  %tmp18 = extractelement <2 x i64> %tmp, i32 0
+  %tmp19 = trunc i64 %tmp18 to i32
+  store i32 %tmp19, i32* %arg12, align 4
+  %tmp20 = insertelement <8 x i32> <i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1>, i32 %tmp19, i32 0
+  %tmp21 = shufflevector <8 x i32> %tmp20, <8 x i32> <i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1>, <8 x i32> <i32 0, i32 undef, i32 undef, i32 undef, i32 0, i32 undef, i32 undef, i32 undef>
+  store <8 x i32> %tmp21, <8 x i32>* %arg13, align 32
+  %tmp22 = fcmp une float %arg14, 0.000000e+00
+  br i1 %tmp22, label %bb17, label %bb16
+}