[llvm] select m, sub/add(X, C), X --> sub/add (X, and(C, m)) (PR #82441)

[via llvm-commits]

https://github.com/elhewaty created https://github.com/llvm/llvm-project/pull/82441

- [DAG] Add tests for Folding select m, sub(X, C), X --> sub (X, and(C, m))(NFC)
- [DAG][X86] Fold select m, sub/add(X, C), X --> sub/add (X, and(C, m))
- Fixes: https://github.com/llvm/llvm-project/issues/66101


>From 2662cff2c70469f73f6cfabd54b8ee9cb2543f62 Mon Sep 17 00:00:00 2001
From: Mohamed Atef <mohamedatef1698 at gmail.com>
Date: Wed, 7 Feb 2024 02:18:26 +0200
Subject: [PATCH 1/2] [DAG] Add tests for Folding select m, sub(X, C), X -->
 sub (X, and(C, m))(NFC)

---
 llvm/test/CodeGen/X86/vselect.ll | 47 ++++++++++++++++++++++++++++++++
 1 file changed, 47 insertions(+)

diff --git a/llvm/test/CodeGen/X86/vselect.ll b/llvm/test/CodeGen/X86/vselect.ll
index ce3dc8cc873cc7..c688f56a4949ea 100644
--- a/llvm/test/CodeGen/X86/vselect.ll
+++ b/llvm/test/CodeGen/X86/vselect.ll
@@ -7,6 +7,53 @@
 ; Verify that we don't emit packed vector shifts instructions if the
 ; condition used by the vector select is a vector of constants.
 
+define <2 x i64> @masked_select_const(<2 x i64> %a, <2 x i64> %x, <2 x i64> %y) {
+; SSE2-LABEL: masked_select_const:
+; SSE2:       # %bb.0:
+; SSE2-NEXT:    movdqa {{.*#+}} xmm3 = [4294967272,4294967272,4294967272,4294967272]
+; SSE2-NEXT:    paddd %xmm0, %xmm3
+; SSE2-NEXT:    pcmpgtd %xmm2, %xmm1
+; SSE2-NEXT:    pand %xmm1, %xmm3
+; SSE2-NEXT:    pandn %xmm0, %xmm1
+; SSE2-NEXT:    por %xmm1, %xmm3
+; SSE2-NEXT:    movdqa %xmm3, %xmm0
+; SSE2-NEXT:    retq
+;
+; SSE41-LABEL: masked_select_const:
+; SSE41:       # %bb.0:
+; SSE41-NEXT:    movdqa %xmm0, %xmm3
+; SSE41-NEXT:    movdqa {{.*#+}} xmm4 = [4294967272,4294967272,4294967272,4294967272]
+; SSE41-NEXT:    paddd %xmm0, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm2, %xmm1
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    blendvps %xmm0, %xmm4, %xmm3
+; SSE41-NEXT:    movaps %xmm3, %xmm0
+; SSE41-NEXT:    retq
+;
+; AVX1-LABEL: masked_select_const:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    vpaddd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm3
+; AVX1-NEXT:    vpcmpgtd %xmm2, %xmm1, %xmm1
+; AVX1-NEXT:    vblendvps %xmm1, %xmm3, %xmm0, %xmm0
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: masked_select_const:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vpbroadcastd {{.*#+}} xmm3 = [4294967272,4294967272,4294967272,4294967272]
+; AVX2-NEXT:    vpaddd %xmm3, %xmm0, %xmm3
+; AVX2-NEXT:    vpcmpgtd %xmm2, %xmm1, %xmm1
+; AVX2-NEXT:    vblendvps %xmm1, %xmm3, %xmm0, %xmm0
+; AVX2-NEXT:    retq
+  %bit_a = bitcast <2 x i64> %a to <4 x i32>
+  %sub.i = add <4 x i32> %bit_a, <i32 -24, i32 -24, i32 -24, i32 -24>
+  %bit_x = bitcast <2 x i64> %x to <4 x i32>
+  %bit_y = bitcast <2 x i64> %y to <4 x i32>
+  %cmp.i = icmp sgt <4 x i32> %bit_x, %bit_y
+  %sel = select <4 x i1> %cmp.i, <4 x i32> %sub.i, <4 x i32> %bit_a
+  %bit_sel = bitcast <4 x i32> %sel to <2 x i64>
+  ret <2 x i64> %bit_sel
+}
+
 define <4 x float> @test1(<4 x float> %a, <4 x float> %b) {
 ; SSE2-LABEL: test1:
 ; SSE2:       # %bb.0:

>From 49213f9360afe63a272165d738883cf3f8fe1c67 Mon Sep 17 00:00:00 2001
From: Mohamed Atef <mohamedatef1698 at gmail.com>
Date: Wed, 21 Feb 2024 01:43:33 +0200
Subject: [PATCH 2/2] [DAG][X86] Fold select m, sub/add(X, C), X --> sub/add
 (X, and(C, m))

---
 llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp | 18 ++++++++++++++++++
 llvm/test/CodeGen/X86/vselect.ll              |  2 +-
 2 files changed, 19 insertions(+), 1 deletion(-)

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 2a09e44e192979..42b24efc81bc5e 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -11660,6 +11660,24 @@ SDValue DAGCombiner::visitSELECT(SDNode *N) {
     }
   }
 
+  // select m, sub(X, C), X --> sub (X, and(C, m))
+  if (N1.getOpcode() == ISD::SUB && N1.getOperand(0) == N2 && N1->hasOneUse() &&
+      DAG.isConstantIntBuildVectorOrConstantInt(N1.getOperand(1)) &&
+      N0.getScalarValueSizeInBits() == N1.getScalarValueSizeInBits()) {
+    return DAG.getNode(ISD::SUB, DL, N1.getValueType(), N2,
+                       DAG.getNode(ISD::AND, DL, N0.getValueType(), N1.getOperand(1),
+                                   N0));
+  }
+
+  // select (sext m), (add X, C), X --> (add X, (and C, (sext m))))
+  if (N1.getOpcode() == ISD::ADD && N1.getOperand(0) == N2 && N1->hasOneUse() &&
+      DAG.isConstantIntBuildVectorOrConstantInt(N1.getOperand(1)) && 
+      N0.getScalarValueSizeInBits() == N1.getScalarValueSizeInBits()) {
+    return DAG.getNode(ISD::ADD, DL, N1.getValueType(), N2,
+                       DAG.getNode(ISD::AND, DL, N0.getValueType(), N1.getOperand(1),
+                                   N0));
+  }
+
   // Fold selects based on a setcc into other things, such as min/max/abs.
   if (N0.getOpcode() == ISD::SETCC) {
     SDValue Cond0 = N0.getOperand(0), Cond1 = N0.getOperand(1);
diff --git a/llvm/test/CodeGen/X86/vselect.ll b/llvm/test/CodeGen/X86/vselect.ll
index c688f56a4949ea..962a72f2daa5d1 100644
--- a/llvm/test/CodeGen/X86/vselect.ll
+++ b/llvm/test/CodeGen/X86/vselect.ll
@@ -22,7 +22,7 @@ define <2 x i64> @masked_select_const(<2 x i64> %a, <2 x i64> %x, <2 x i64> %y)
 ; SSE41-LABEL: masked_select_const:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa %xmm0, %xmm3
-; SSE41-NEXT:    movdqa {{.*#+}} xmm4 = [4294967272,4294967272,4294967272,4294967272]
+; SSE41-NEXT:    pmovsxbd {{.*#+}} xmm4 = [4294967272,4294967272,4294967272,4294967272]
 ; SSE41-NEXT:    paddd %xmm0, %xmm4
 ; SSE41-NEXT:    pcmpgtd %xmm2, %xmm1
 ; SSE41-NEXT:    movdqa %xmm1, %xmm0