[llvm] r360526 - [X86][SSE] Add SimplifyDemandedVectorElts HADD/HSUB handling.
Simon Pilgrim via llvm-commits
llvm-commits at lists.llvm.org
Sat May 11 09:07:12 PDT 2019
Author: rksimon
Date: Sat May 11 09:07:12 2019
New Revision: 360526
URL: http://llvm.org/viewvc/llvm-project?rev=360526&view=rev
Log:
[X86][SSE] Add SimplifyDemandedVectorElts HADD/HSUB handling.
Still missing PHADDW/PHSUBW tests because PEXTRW doesn't call SimplifyDemandedVectorElts
Modified:
llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
llvm/trunk/test/CodeGen/X86/horizontal-shuffle-demanded.ll
Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=360526&r1=360525&r2=360526&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Sat May 11 09:07:12 2019
@@ -6152,6 +6152,34 @@ static void getPackDemandedElts(EVT VT,
}
}
+// Split the demanded elts of a HADD/HSUB node between its operands.
+static void getHorizDemandedElts(EVT VT, const APInt &DemandedElts,
+ APInt &DemandedLHS, APInt &DemandedRHS) {
+ int NumLanes = VT.getSizeInBits() / 128;
+ int NumElts = DemandedElts.getBitWidth();
+ int NumEltsPerLane = NumElts / NumLanes;
+ int HalfEltsPerLane = NumEltsPerLane / 2;
+
+ DemandedLHS = APInt::getNullValue(NumElts);
+ DemandedRHS = APInt::getNullValue(NumElts);
+
+ // Map DemandedElts to the horizontal operands.
+ for (int Idx = 0; Idx != NumElts; ++Idx) {
+ if (!DemandedElts[Idx])
+ continue;
+ int LaneIdx = (Idx / NumEltsPerLane) * NumEltsPerLane;
+ int LocalIdx = Idx % NumEltsPerLane;
+ if (LocalIdx < HalfEltsPerLane) {
+ DemandedLHS.setBit(LaneIdx + 2 * LocalIdx + 0);
+ DemandedLHS.setBit(LaneIdx + 2 * LocalIdx + 1);
+ } else {
+ LocalIdx -= HalfEltsPerLane;
+ DemandedRHS.setBit(LaneIdx + 2 * LocalIdx + 0);
+ DemandedRHS.setBit(LaneIdx + 2 * LocalIdx + 1);
+ }
+ }
+}
+
/// Calculates the shuffle mask corresponding to the target-specific opcode.
/// If the mask could be calculated, returns it in \p Mask, returns the shuffle
/// operands in \p Ops, and returns true.
@@ -33433,6 +33461,23 @@ bool X86TargetLowering::SimplifyDemanded
return true;
break;
}
+ case X86ISD::HADD:
+ case X86ISD::HSUB:
+ case X86ISD::FHADD:
+ case X86ISD::FHSUB: {
+ APInt DemandedLHS, DemandedRHS;
+ getHorizDemandedElts(VT, DemandedElts, DemandedLHS, DemandedRHS);
+
+ APInt LHSUndef, LHSZero;
+ if (SimplifyDemandedVectorElts(Op.getOperand(0), DemandedLHS, LHSUndef,
+ LHSZero, TLO, Depth + 1))
+ return true;
+ APInt RHSUndef, RHSZero;
+ if (SimplifyDemandedVectorElts(Op.getOperand(1), DemandedRHS, RHSUndef,
+ RHSZero, TLO, Depth + 1))
+ return true;
+ break;
+ }
case X86ISD::VTRUNC:
case X86ISD::VTRUNCS:
case X86ISD::VTRUNCUS: {
Modified: llvm/trunk/test/CodeGen/X86/horizontal-shuffle-demanded.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/horizontal-shuffle-demanded.ll?rev=360526&r1=360525&r2=360526&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/horizontal-shuffle-demanded.ll (original)
+++ llvm/trunk/test/CodeGen/X86/horizontal-shuffle-demanded.ll Sat May 11 09:07:12 2019
@@ -10,15 +10,13 @@ define void @test_demanded_haddps_128(<4
; X86-LABEL: test_demanded_haddps_128:
; X86: ## %bb.0:
; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
-; X86-NEXT: vbroadcastss %xmm1, %xmm1
-; X86-NEXT: vhaddps %xmm1, %xmm0, %xmm0
+; X86-NEXT: vhaddps %xmm0, %xmm0, %xmm0
; X86-NEXT: vmovss %xmm0, (%eax)
; X86-NEXT: retl
;
; X64-LABEL: test_demanded_haddps_128:
; X64: ## %bb.0:
-; X64-NEXT: vbroadcastss %xmm1, %xmm1
-; X64-NEXT: vhaddps %xmm1, %xmm0, %xmm0
+; X64-NEXT: vhaddps %xmm0, %xmm0, %xmm0
; X64-NEXT: vmovss %xmm0, (%rdi)
; X64-NEXT: retq
%1 = shufflevector <4 x float> %a1, <4 x float> undef, <4 x i32> zeroinitializer
@@ -32,14 +30,12 @@ define void @test_demanded_hsubps_128(<4
; X86-LABEL: test_demanded_hsubps_128:
; X86: ## %bb.0:
; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
-; X86-NEXT: vbroadcastss %xmm0, %xmm0
; X86-NEXT: vhsubps %xmm1, %xmm0, %xmm0
; X86-NEXT: vextractps $2, %xmm0, (%eax)
; X86-NEXT: retl
;
; X64-LABEL: test_demanded_hsubps_128:
; X64: ## %bb.0:
-; X64-NEXT: vbroadcastss %xmm0, %xmm0
; X64-NEXT: vhsubps %xmm1, %xmm0, %xmm0
; X64-NEXT: vextractps $2, %xmm0, (%rdi)
; X64-NEXT: retq
@@ -54,15 +50,13 @@ define void @test_demanded_haddpd_128(<2
; X86-LABEL: test_demanded_haddpd_128:
; X86: ## %bb.0:
; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
-; X86-NEXT: vmovddup {{.*#+}} xmm1 = xmm1[0,0]
-; X86-NEXT: vhaddpd %xmm1, %xmm0, %xmm0
+; X86-NEXT: vhaddpd %xmm0, %xmm0, %xmm0
; X86-NEXT: vmovlpd %xmm0, (%eax)
; X86-NEXT: retl
;
; X64-LABEL: test_demanded_haddpd_128:
; X64: ## %bb.0:
-; X64-NEXT: vmovddup {{.*#+}} xmm1 = xmm1[0,0]
-; X64-NEXT: vhaddpd %xmm1, %xmm0, %xmm0
+; X64-NEXT: vhaddpd %xmm0, %xmm0, %xmm0
; X64-NEXT: vmovlpd %xmm0, (%rdi)
; X64-NEXT: retq
%1 = shufflevector <2 x double> %a1, <2 x double> undef, <2 x i32> zeroinitializer
@@ -76,15 +70,13 @@ define void @test_demanded_hsubpd_128(<2
; X86-LABEL: test_demanded_hsubpd_128:
; X86: ## %bb.0:
; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
-; X86-NEXT: vmovddup {{.*#+}} xmm1 = xmm1[0,0]
-; X86-NEXT: vhsubpd %xmm1, %xmm0, %xmm0
+; X86-NEXT: vhsubpd %xmm0, %xmm0, %xmm0
; X86-NEXT: vmovlpd %xmm0, (%eax)
; X86-NEXT: retl
;
; X64-LABEL: test_demanded_hsubpd_128:
; X64: ## %bb.0:
-; X64-NEXT: vmovddup {{.*#+}} xmm1 = xmm1[0,0]
-; X64-NEXT: vhsubpd %xmm1, %xmm0, %xmm0
+; X64-NEXT: vhsubpd %xmm0, %xmm0, %xmm0
; X64-NEXT: vmovlpd %xmm0, (%rdi)
; X64-NEXT: retq
%1 = shufflevector <2 x double> %a1, <2 x double> undef, <2 x i32> zeroinitializer
@@ -98,15 +90,13 @@ define void @test_demanded_phaddd_128(<4
; X86-LABEL: test_demanded_phaddd_128:
; X86: ## %bb.0:
; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
-; X86-NEXT: vpbroadcastd %xmm1, %xmm1
-; X86-NEXT: vphaddd %xmm1, %xmm0, %xmm0
+; X86-NEXT: vphaddd %xmm0, %xmm0, %xmm0
; X86-NEXT: vmovd %xmm0, (%eax)
; X86-NEXT: retl
;
; X64-LABEL: test_demanded_phaddd_128:
; X64: ## %bb.0:
-; X64-NEXT: vpbroadcastd %xmm1, %xmm1
-; X64-NEXT: vphaddd %xmm1, %xmm0, %xmm0
+; X64-NEXT: vphaddd %xmm0, %xmm0, %xmm0
; X64-NEXT: vmovd %xmm0, (%rdi)
; X64-NEXT: retq
%1 = shufflevector <4 x i32> %a1, <4 x i32> undef, <4 x i32> zeroinitializer
@@ -120,15 +110,13 @@ define void @test_demanded_phsubd_128(<4
; X86-LABEL: test_demanded_phsubd_128:
; X86: ## %bb.0:
; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
-; X86-NEXT: vpbroadcastd %xmm1, %xmm1
-; X86-NEXT: vphsubd %xmm1, %xmm0, %xmm0
+; X86-NEXT: vphsubd %xmm0, %xmm0, %xmm0
; X86-NEXT: vpextrd $1, %xmm0, (%eax)
; X86-NEXT: retl
;
; X64-LABEL: test_demanded_phsubd_128:
; X64: ## %bb.0:
-; X64-NEXT: vpbroadcastd %xmm1, %xmm1
-; X64-NEXT: vphsubd %xmm1, %xmm0, %xmm0
+; X64-NEXT: vphsubd %xmm0, %xmm0, %xmm0
; X64-NEXT: vpextrd $1, %xmm0, (%rdi)
; X64-NEXT: retq
%1 = shufflevector <4 x i32> %a1, <4 x i32> undef, <4 x i32> zeroinitializer
@@ -190,8 +178,7 @@ define void @test_demanded_haddps_256(<8
; X86-LABEL: test_demanded_haddps_256:
; X86: ## %bb.0:
; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
-; X86-NEXT: vbroadcastss %xmm1, %ymm1
-; X86-NEXT: vhaddps %ymm1, %ymm0, %ymm0
+; X86-NEXT: vhaddps %ymm0, %ymm0, %ymm0
; X86-NEXT: vextractf128 $1, %ymm0, %xmm0
; X86-NEXT: vmovss %xmm0, (%eax)
; X86-NEXT: vzeroupper
@@ -199,8 +186,7 @@ define void @test_demanded_haddps_256(<8
;
; X64-LABEL: test_demanded_haddps_256:
; X64: ## %bb.0:
-; X64-NEXT: vbroadcastss %xmm1, %ymm1
-; X64-NEXT: vhaddps %ymm1, %ymm0, %ymm0
+; X64-NEXT: vhaddps %ymm0, %ymm0, %ymm0
; X64-NEXT: vextractf128 $1, %ymm0, %xmm0
; X64-NEXT: vmovss %xmm0, (%rdi)
; X64-NEXT: vzeroupper
@@ -216,7 +202,6 @@ define void @test_demanded_hsubps_256(<8
; X86-LABEL: test_demanded_hsubps_256:
; X86: ## %bb.0:
; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
-; X86-NEXT: vbroadcastss %xmm0, %ymm0
; X86-NEXT: vhsubps %ymm1, %ymm0, %ymm0
; X86-NEXT: vextractf128 $1, %ymm0, %xmm0
; X86-NEXT: vextractps $3, %xmm0, (%eax)
@@ -225,7 +210,6 @@ define void @test_demanded_hsubps_256(<8
;
; X64-LABEL: test_demanded_hsubps_256:
; X64: ## %bb.0:
-; X64-NEXT: vbroadcastss %xmm0, %ymm0
; X64-NEXT: vhsubps %ymm1, %ymm0, %ymm0
; X64-NEXT: vextractf128 $1, %ymm0, %xmm0
; X64-NEXT: vextractps $3, %xmm0, (%rdi)
@@ -242,8 +226,7 @@ define void @test_demanded_haddpd_256(<4
; X86-LABEL: test_demanded_haddpd_256:
; X86: ## %bb.0:
; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
-; X86-NEXT: vbroadcastsd %xmm1, %ymm1
-; X86-NEXT: vhaddpd %ymm1, %ymm0, %ymm0
+; X86-NEXT: vhaddpd %ymm0, %ymm0, %ymm0
; X86-NEXT: vextractf128 $1, %ymm0, %xmm0
; X86-NEXT: vmovlpd %xmm0, (%eax)
; X86-NEXT: vzeroupper
@@ -251,8 +234,7 @@ define void @test_demanded_haddpd_256(<4
;
; X64-LABEL: test_demanded_haddpd_256:
; X64: ## %bb.0:
-; X64-NEXT: vbroadcastsd %xmm1, %ymm1
-; X64-NEXT: vhaddpd %ymm1, %ymm0, %ymm0
+; X64-NEXT: vhaddpd %ymm0, %ymm0, %ymm0
; X64-NEXT: vextractf128 $1, %ymm0, %xmm0
; X64-NEXT: vmovlpd %xmm0, (%rdi)
; X64-NEXT: vzeroupper
@@ -268,8 +250,7 @@ define void @test_demanded_hsubpd_256(<4
; X86-LABEL: test_demanded_hsubpd_256:
; X86: ## %bb.0:
; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
-; X86-NEXT: vbroadcastsd %xmm1, %ymm1
-; X86-NEXT: vhsubpd %ymm1, %ymm0, %ymm0
+; X86-NEXT: vhsubpd %ymm0, %ymm0, %ymm0
; X86-NEXT: vextractf128 $1, %ymm0, %xmm0
; X86-NEXT: vmovlpd %xmm0, (%eax)
; X86-NEXT: vzeroupper
@@ -277,8 +258,7 @@ define void @test_demanded_hsubpd_256(<4
;
; X64-LABEL: test_demanded_hsubpd_256:
; X64: ## %bb.0:
-; X64-NEXT: vbroadcastsd %xmm1, %ymm1
-; X64-NEXT: vhsubpd %ymm1, %ymm0, %ymm0
+; X64-NEXT: vhsubpd %ymm0, %ymm0, %ymm0
; X64-NEXT: vextractf128 $1, %ymm0, %xmm0
; X64-NEXT: vmovlpd %xmm0, (%rdi)
; X64-NEXT: vzeroupper
@@ -294,7 +274,6 @@ define void @test_demanded_phaddd_256(<8
; X86-LABEL: test_demanded_phaddd_256:
; X86: ## %bb.0:
; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
-; X86-NEXT: vpbroadcastd %xmm0, %ymm0
; X86-NEXT: vphaddd %ymm1, %ymm0, %ymm0
; X86-NEXT: vextracti128 $1, %ymm0, %xmm0
; X86-NEXT: vpextrd $3, %xmm0, (%eax)
@@ -303,7 +282,6 @@ define void @test_demanded_phaddd_256(<8
;
; X64-LABEL: test_demanded_phaddd_256:
; X64: ## %bb.0:
-; X64-NEXT: vpbroadcastd %xmm0, %ymm0
; X64-NEXT: vphaddd %ymm1, %ymm0, %ymm0
; X64-NEXT: vextracti128 $1, %ymm0, %xmm0
; X64-NEXT: vpextrd $3, %xmm0, (%rdi)
@@ -320,8 +298,7 @@ define void @test_demanded_phsubd_256(<8
; X86-LABEL: test_demanded_phsubd_256:
; X86: ## %bb.0:
; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
-; X86-NEXT: vpbroadcastd %xmm1, %ymm1
-; X86-NEXT: vphsubd %ymm1, %ymm0, %ymm0
+; X86-NEXT: vphsubd %ymm0, %ymm0, %ymm0
; X86-NEXT: vextracti128 $1, %ymm0, %xmm0
; X86-NEXT: vpextrd $1, %xmm0, (%eax)
; X86-NEXT: vzeroupper
@@ -329,8 +306,7 @@ define void @test_demanded_phsubd_256(<8
;
; X64-LABEL: test_demanded_phsubd_256:
; X64: ## %bb.0:
-; X64-NEXT: vpbroadcastd %xmm1, %ymm1
-; X64-NEXT: vphsubd %ymm1, %ymm0, %ymm0
+; X64-NEXT: vphsubd %ymm0, %ymm0, %ymm0
; X64-NEXT: vextracti128 $1, %ymm0, %xmm0
; X64-NEXT: vpextrd $1, %xmm0, (%rdi)
; X64-NEXT: vzeroupper
More information about the llvm-commits
mailing list