[DAG] Legalize abs(x) -> umin(x,sub(0,x)) iff umin/sub are legal

If umin() is legal, this is likely to result in smaller codegen expansion for abs(x) than the xor(add,ashr) method.

Followup to D92095

Alive2: https://alive2.llvm.org/ce/z/8nuX6s  https://alive2.llvm.org/ce/z/q2hB9w

diff --git a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
index 0b1c257..3464fe8 100644 (file)
--- a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -6830,6 +6830,14 @@ bool TargetLowering::expandABS(SDNode *N, SDValue &Result,
     return true;
   }
 
+  // abs(x) -> umin(x,sub(0,x))
+  if (isOperationLegal(ISD::SUB, VT) && isOperationLegal(ISD::UMIN, VT)) {
+    SDValue Zero = DAG.getConstant(0, dl, VT);
+    Result = DAG.getNode(ISD::UMIN, dl, VT, Op,
+                         DAG.getNode(ISD::SUB, dl, VT, Zero, Op));
+    return true;
+  }
+
   // Only expand vector types if we have the appropriate vector operations.
   if (VT.isVector() && (!isOperationLegalOrCustom(ISD::SRA, VT) ||
                         !isOperationLegalOrCustom(ISD::ADD, VT) ||

diff --git a/llvm/test/CodeGen/X86/abs.ll b/llvm/test/CodeGen/X86/abs.ll
index 942f09f..0bbc785 100644 (file)
--- a/llvm/test/CodeGen/X86/abs.ll
+++ b/llvm/test/CodeGen/X86/abs.ll
@@ -474,9 +474,8 @@ define <16 x i8> @test_v16i8(<16 x i8> %a) nounwind {
 ; SSE-LABEL: test_v16i8:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    pxor %xmm1, %xmm1
-; SSE-NEXT:    pcmpgtb %xmm0, %xmm1
-; SSE-NEXT:    paddb %xmm1, %xmm0
-; SSE-NEXT:    pxor %xmm1, %xmm0
+; SSE-NEXT:    psubb %xmm0, %xmm1
+; SSE-NEXT:    pminub %xmm1, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: test_v16i8:

diff --git a/llvm/test/CodeGen/X86/combine-abs.ll b/llvm/test/CodeGen/X86/combine-abs.ll
index da59da4..4ff9fef 100644 (file)
--- a/llvm/test/CodeGen/X86/combine-abs.ll
+++ b/llvm/test/CodeGen/X86/combine-abs.ll
@@ -81,12 +81,10 @@ define <32 x i8> @combine_v32i8_abs_abs(<32 x i8> %a) {
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    pxor %xmm2, %xmm2
 ; SSE2-NEXT:    pxor %xmm3, %xmm3
-; SSE2-NEXT:    pcmpgtb %xmm0, %xmm3
-; SSE2-NEXT:    paddb %xmm3, %xmm0
-; SSE2-NEXT:    pxor %xmm3, %xmm0
-; SSE2-NEXT:    pcmpgtb %xmm1, %xmm2
-; SSE2-NEXT:    paddb %xmm2, %xmm1
-; SSE2-NEXT:    pxor %xmm2, %xmm1
+; SSE2-NEXT:    psubb %xmm0, %xmm3
+; SSE2-NEXT:    pminub %xmm3, %xmm0
+; SSE2-NEXT:    psubb %xmm1, %xmm2
+; SSE2-NEXT:    pminub %xmm2, %xmm1
 ; SSE2-NEXT:    retq
 ;
 ; SSE42-LABEL: combine_v32i8_abs_abs:

diff --git a/llvm/test/CodeGen/X86/viabs.ll b/llvm/test/CodeGen/X86/viabs.ll
index 873dea2..405d9ea 100644 (file)
--- a/llvm/test/CodeGen/X86/viabs.ll
+++ b/llvm/test/CodeGen/X86/viabs.ll
@@ -132,9 +132,8 @@ define <16 x i8> @test_abs_lt_v16i8(<16 x i8> %a) nounwind {
 ; SSE2-LABEL: test_abs_lt_v16i8:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    pxor %xmm1, %xmm1
-; SSE2-NEXT:    pcmpgtb %xmm0, %xmm1
-; SSE2-NEXT:    paddb %xmm1, %xmm0
-; SSE2-NEXT:    pxor %xmm1, %xmm0
+; SSE2-NEXT:    psubb %xmm0, %xmm1
+; SSE2-NEXT:    pminub %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: test_abs_lt_v16i8:
@@ -363,12 +362,10 @@ define <32 x i8> @test_abs_lt_v32i8(<32 x i8> %a) nounwind {
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    pxor %xmm2, %xmm2
 ; SSE2-NEXT:    pxor %xmm3, %xmm3
-; SSE2-NEXT:    pcmpgtb %xmm0, %xmm3
-; SSE2-NEXT:    paddb %xmm3, %xmm0
-; SSE2-NEXT:    pxor %xmm3, %xmm0
-; SSE2-NEXT:    pcmpgtb %xmm1, %xmm2
-; SSE2-NEXT:    paddb %xmm2, %xmm1
-; SSE2-NEXT:    pxor %xmm2, %xmm1
+; SSE2-NEXT:    psubb %xmm0, %xmm3
+; SSE2-NEXT:    pminub %xmm3, %xmm0
+; SSE2-NEXT:    psubb %xmm1, %xmm2
+; SSE2-NEXT:    pminub %xmm2, %xmm1
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: test_abs_lt_v32i8:
@@ -874,20 +871,16 @@ define <64 x i8> @test_abs_lt_v64i8(<64 x i8> %a) nounwind {
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    pxor %xmm4, %xmm4
 ; SSE2-NEXT:    pxor %xmm5, %xmm5
-; SSE2-NEXT:    pcmpgtb %xmm0, %xmm5
-; SSE2-NEXT:    paddb %xmm5, %xmm0
-; SSE2-NEXT:    pxor %xmm5, %xmm0
+; SSE2-NEXT:    psubb %xmm0, %xmm5
+; SSE2-NEXT:    pminub %xmm5, %xmm0
 ; SSE2-NEXT:    pxor %xmm5, %xmm5
-; SSE2-NEXT:    pcmpgtb %xmm1, %xmm5
-; SSE2-NEXT:    paddb %xmm5, %xmm1
-; SSE2-NEXT:    pxor %xmm5, %xmm1
+; SSE2-NEXT:    psubb %xmm1, %xmm5
+; SSE2-NEXT:    pminub %xmm5, %xmm1
 ; SSE2-NEXT:    pxor %xmm5, %xmm5
-; SSE2-NEXT:    pcmpgtb %xmm2, %xmm5
-; SSE2-NEXT:    paddb %xmm5, %xmm2
-; SSE2-NEXT:    pxor %xmm5, %xmm2
-; SSE2-NEXT:    pcmpgtb %xmm3, %xmm4
-; SSE2-NEXT:    paddb %xmm4, %xmm3
-; SSE2-NEXT:    pxor %xmm4, %xmm3
+; SSE2-NEXT:    psubb %xmm2, %xmm5
+; SSE2-NEXT:    pminub %xmm5, %xmm2
+; SSE2-NEXT:    psubb %xmm3, %xmm4
+; SSE2-NEXT:    pminub %xmm4, %xmm3
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: test_abs_lt_v64i8: