[llvm] ffb5bea - [X86] LowerBITREVERSE - support SSE-only GFNI i32/i64 bitreverse

[Simon Pilgrim via llvm-commits]

Author: Simon Pilgrim
Date: 2024-04-11T16:33:15+01:00
New Revision: ffb5bea2be9f966a39f243a7d8c2f48a1343cb4c

URL: https://github.com/llvm/llvm-project/commit/ffb5bea2be9f966a39f243a7d8c2f48a1343cb4c
DIFF: https://github.com/llvm/llvm-project/commit/ffb5bea2be9f966a39f243a7d8c2f48a1343cb4c.diff

LOG: [X86] LowerBITREVERSE - support SSE-only GFNI i32/i64 bitreverse

Support Tremont CPUs which don't have AVX but do have GFNI.

Noticed while trying to workout how to clean up the costmodel for GFNI bitreverse

Added: 
    

Modified: 
    llvm/lib/Target/X86/X86ISelLowering.cpp
    llvm/test/CodeGen/X86/vector-bitreverse.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 52be35aafb0f57..f274da6f6f7767 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -1286,6 +1286,11 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
       setOperationAction(ISD::CTLZ,             VT, Custom);
     }
 
+    if (Subtarget.hasGFNI()) {
+      setOperationAction(ISD::BITREVERSE,       MVT::i32, Custom);
+      setOperationAction(ISD::BITREVERSE,       MVT::i64, Custom);
+    }
+
     // These might be better off as horizontal vector ops.
     setOperationAction(ISD::ADD,                MVT::i16, Custom);
     setOperationAction(ISD::ADD,                MVT::i32, Custom);
@@ -1496,11 +1501,6 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
     setOperationAction(ISD::TRUNCATE,          MVT::v32i32, Custom);
     setOperationAction(ISD::TRUNCATE,          MVT::v32i64, Custom);
 
-    if (Subtarget.hasGFNI()) {
-      setOperationAction(ISD::BITREVERSE, MVT::i32, Custom);
-      setOperationAction(ISD::BITREVERSE, MVT::i64, Custom);
-    }
-
     for (auto VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32, MVT::v4i64 }) {
       setOperationAction(ISD::SETCC,           VT, Custom);
       setOperationAction(ISD::CTPOP,           VT, Custom);
@@ -31337,12 +31337,9 @@ static SDValue LowerBITREVERSE(SDValue Op, const X86Subtarget &Subtarget,
   if (VT.is256BitVector() && !Subtarget.hasInt256())
     return splitVectorIntUnary(Op, DAG, DL);
 
-  // Lower i32/i64 to GFNI as vXi8 BITREVERSE + BSWAP
+  // Lower i32/i64 as vXi8 BITREVERSE + BSWAP
   if (!VT.isVector()) {
-
-    assert((VT.getScalarType() == MVT::i32) ||
-           (VT.getScalarType() == MVT::i64));
-
+    assert((VT == MVT::i32 || VT == MVT::i64) && "Only tested for i32/i64");
     MVT VecVT = MVT::getVectorVT(VT, 128 / VT.getSizeInBits());
     SDValue Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, In);
     Res = DAG.getNode(ISD::BITREVERSE, DL, MVT::v16i8,

diff  --git a/llvm/test/CodeGen/X86/vector-bitreverse.ll b/llvm/test/CodeGen/X86/vector-bitreverse.ll
index 1c5326d35bb00a..b22b508db8b280 100644
--- a/llvm/test/CodeGen/X86/vector-bitreverse.ll
+++ b/llvm/test/CodeGen/X86/vector-bitreverse.ll
@@ -254,24 +254,10 @@ define i32 @test_bitreverse_i32(i32 %a) nounwind {
 ;
 ; GFNISSE-LABEL: test_bitreverse_i32:
 ; GFNISSE:       # %bb.0:
-; GFNISSE-NEXT:    # kill: def $edi killed $edi def $rdi
-; GFNISSE-NEXT:    bswapl %edi
-; GFNISSE-NEXT:    movl %edi, %eax
-; GFNISSE-NEXT:    andl $252645135, %eax # imm = 0xF0F0F0F
-; GFNISSE-NEXT:    shll $4, %eax
-; GFNISSE-NEXT:    shrl $4, %edi
-; GFNISSE-NEXT:    andl $252645135, %edi # imm = 0xF0F0F0F
-; GFNISSE-NEXT:    orl %eax, %edi
-; GFNISSE-NEXT:    movl %edi, %eax
-; GFNISSE-NEXT:    andl $858993459, %eax # imm = 0x33333333
-; GFNISSE-NEXT:    shrl $2, %edi
-; GFNISSE-NEXT:    andl $858993459, %edi # imm = 0x33333333
-; GFNISSE-NEXT:    leal (%rdi,%rax,4), %eax
-; GFNISSE-NEXT:    movl %eax, %ecx
-; GFNISSE-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
-; GFNISSE-NEXT:    shrl %eax
-; GFNISSE-NEXT:    andl $1431655765, %eax # imm = 0x55555555
-; GFNISSE-NEXT:    leal (%rax,%rcx,2), %eax
+; GFNISSE-NEXT:    movd %edi, %xmm0
+; GFNISSE-NEXT:    gf2p8affineqb $0, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; GFNISSE-NEXT:    movd %xmm0, %eax
+; GFNISSE-NEXT:    bswapl %eax
 ; GFNISSE-NEXT:    retq
 ;
 ; GFNIAVX-LABEL: test_bitreverse_i32:
@@ -343,26 +329,10 @@ define i64 @test_bitreverse_i64(i64 %a) nounwind {
 ;
 ; GFNISSE-LABEL: test_bitreverse_i64:
 ; GFNISSE:       # %bb.0:
-; GFNISSE-NEXT:    bswapq %rdi
-; GFNISSE-NEXT:    movq %rdi, %rax
-; GFNISSE-NEXT:    shrq $4, %rax
-; GFNISSE-NEXT:    movabsq $1085102592571150095, %rcx # imm = 0xF0F0F0F0F0F0F0F
-; GFNISSE-NEXT:    andq %rcx, %rax
-; GFNISSE-NEXT:    andq %rcx, %rdi
-; GFNISSE-NEXT:    shlq $4, %rdi
-; GFNISSE-NEXT:    orq %rax, %rdi
-; GFNISSE-NEXT:    movabsq $3689348814741910323, %rax # imm = 0x3333333333333333
-; GFNISSE-NEXT:    movq %rdi, %rcx
-; GFNISSE-NEXT:    andq %rax, %rcx
-; GFNISSE-NEXT:    shrq $2, %rdi
-; GFNISSE-NEXT:    andq %rax, %rdi
-; GFNISSE-NEXT:    leaq (%rdi,%rcx,4), %rax
-; GFNISSE-NEXT:    movabsq $6148914691236517205, %rcx # imm = 0x5555555555555555
-; GFNISSE-NEXT:    movq %rax, %rdx
-; GFNISSE-NEXT:    andq %rcx, %rdx
-; GFNISSE-NEXT:    shrq %rax
-; GFNISSE-NEXT:    andq %rcx, %rax
-; GFNISSE-NEXT:    leaq (%rax,%rdx,2), %rax
+; GFNISSE-NEXT:    movq %rdi, %xmm0
+; GFNISSE-NEXT:    gf2p8affineqb $0, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; GFNISSE-NEXT:    movq %xmm0, %rax
+; GFNISSE-NEXT:    bswapq %rax
 ; GFNISSE-NEXT:    retq
 ;
 ; GFNIAVX-LABEL: test_bitreverse_i64: