aaron/ispc
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

patch for LLVM for fails at avx-x2

Browse Source
This commit is contained in:
Ilia Filippov
2013-10-11 11:39:02 +04:00
parent 2741e3c1d0
commit 7abbe97ee9
2 changed files with 78 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
fail_db.txt
View File

@@ -37,10 +37,6 @@
./tests/reduce-equal-13.ispc compfail x86-64 sse4-i8x16 Linux LLVM 3.3 g++4.4 -O2 *
./tests/reduce-equal-5.ispc compfail x86-64 sse4-i8x16 Linux LLVM 3.3 g++4.4 -O2 *
./tests/reduce-equal-6.ispc compfail x86-64 sse4-i8x16 Linux LLVM 3.3 g++4.4 -O2 *
./tests/avg-down-int8.ispc compfail x86 avx1-i32x16 Linux LLVM 3.3 g++4.4 -O2 *
./tests/avg-up-int8.ispc compfail x86 avx1-i32x16 Linux LLVM 3.3 g++4.4 -O2 *
./tests/avg-down-int8.ispc compfail x86-64 avx1-i32x16 Linux LLVM 3.3 g++4.4 -O2 *
./tests/avg-up-int8.ispc compfail x86-64 avx1-i32x16 Linux LLVM 3.3 g++4.4 -O2 *
./tests/atomics-4.ispc runfail x86-64 generic-4 Linux LLVM 3.3 g++4.4 -O2 *
./tests/atomics-6.ispc runfail x86-64 generic-4 Linux LLVM 3.3 g++4.4 -O2 *
./tests/atomics-swap.ispc runfail x86-64 generic-4 Linux LLVM 3.3 g++4.4 -O2 *
@@ -158,10 +154,6 @@
./tests/funcptr-null-4.ispc runfail x86-64 sse4-i8x16 Linux LLVM 3.4 g++4.4 -O2 *
./tests/funcptr-null-5.ispc runfail x86-64 sse4-i8x16 Linux LLVM 3.4 g++4.4 -O2 *
./tests/funcptr-null-6.ispc runfail x86-64 sse4-i8x16 Linux LLVM 3.4 g++4.4 -O2 *
./tests/avg-down-int8.ispc compfail x86 avx1-i32x16 Linux LLVM 3.4 g++4.4 -O2 *
./tests/avg-up-int8.ispc compfail x86 avx1-i32x16 Linux LLVM 3.4 g++4.4 -O2 *
./tests/avg-down-int8.ispc compfail x86-64 avx1-i32x16 Linux LLVM 3.4 g++4.4 -O2 *
./tests/avg-up-int8.ispc compfail x86-64 avx1-i32x16 Linux LLVM 3.4 g++4.4 -O2 *
./tests/array-gather-ifs.ispc runfail x86-64 generic-4 Linux LLVM 3.4 g++4.4 -O2 *
./tests/array-gather-multi-unif.ispc runfail x86-64 generic-4 Linux LLVM 3.4 g++4.4 -O2 *
./tests/array-gather-unif.ispc runfail x86-64 generic-4 Linux LLVM 3.4 g++4.4 -O2 *
@@ -362,10 +354,6 @@
./tests/reduce-equal-13.ispc compfail x86-64 sse4-i8x16 Linux LLVM 3.3 g++4.7 -O2 *
./tests/reduce-equal-5.ispc compfail x86-64 sse4-i8x16 Linux LLVM 3.3 g++4.7 -O2 *
./tests/reduce-equal-6.ispc compfail x86-64 sse4-i8x16 Linux LLVM 3.3 g++4.7 -O2 *
./tests/avg-down-int8.ispc compfail x86 avx1-i32x16 Linux LLVM 3.3 g++4.7 -O2 *
./tests/avg-up-int8.ispc compfail x86 avx1-i32x16 Linux LLVM 3.3 g++4.7 -O2 *
./tests/avg-down-int8.ispc compfail x86-64 avx1-i32x16 Linux LLVM 3.3 g++4.7 -O2 *
./tests/avg-up-int8.ispc compfail x86-64 avx1-i32x16 Linux LLVM 3.3 g++4.7 -O2 *
./tests/avg-down-int8.ispc compfail x86 avx1.1-i32x16 Linux LLVM 3.3 g++4.7 -O2 *
./tests/avg-up-int8.ispc compfail x86 avx1.1-i32x16 Linux LLVM 3.3 g++4.7 -O2 *
./tests/avg-down-int8.ispc compfail x86-64 avx1.1-i32x16 Linux LLVM 3.3 g++4.7 -O2 *
@@ -425,10 +413,6 @@
./tests/funcptr-null-4.ispc runfail x86-64 sse4-i8x16 Linux LLVM 3.4 g++4.7 -O2 *
./tests/funcptr-null-5.ispc runfail x86-64 sse4-i8x16 Linux LLVM 3.4 g++4.7 -O2 *
./tests/funcptr-null-6.ispc runfail x86-64 sse4-i8x16 Linux LLVM 3.4 g++4.7 -O2 *
./tests/avg-down-int8.ispc compfail x86 avx1-i32x16 Linux LLVM 3.4 g++4.7 -O2 *
./tests/avg-up-int8.ispc compfail x86 avx1-i32x16 Linux LLVM 3.4 g++4.7 -O2 *
./tests/avg-down-int8.ispc compfail x86-64 avx1-i32x16 Linux LLVM 3.4 g++4.7 -O2 *
./tests/avg-up-int8.ispc compfail x86-64 avx1-i32x16 Linux LLVM 3.4 g++4.7 -O2 *
./tests/avg-down-int8.ispc compfail x86 avx1.1-i32x16 Linux LLVM 3.4 g++4.7 -O2 *
./tests/avg-up-int8.ispc compfail x86 avx1.1-i32x16 Linux LLVM 3.4 g++4.7 -O2 *
./tests/avg-down-int8.ispc compfail x86-64 avx1.1-i32x16 Linux LLVM 3.4 g++4.7 -O2 *

78
llvm_patches/3_3_0001-Fix-PR16807.patch Executable file
View File

@@ -0,0 +1,78 @@
From b9c47f44691cb9a648b9fa1ae373f0defe53c757 Mon Sep 17 00:00:00 2001
From: Michael Liao <michael.hliao@gmail.com>
Date: Thu, 10 Oct 2013 16:47:00 -0700
Subject: [PATCH] Fix PR16807
- Lower signed division by constant powers-of-2 to target-independent
DAG operators instead of target-dependent ones to support them on
targets where vector types are legal but shift operators on that types
are illegal, e.g. on AVX, PSRAW is only available on <8 x i16> though
<16 x i16> is a legal type.
---
lib/Target/X86/X86ISelLowering.cpp | 22 ++++++++++++++++------
test/CodeGen/X86/pr16807.ll | 18 ++++++++++++++++++
2 files changed, 34 insertions(+), 6 deletions(-)
create mode 100644 test/CodeGen/X86/pr16807.ll
diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index bd5ad4e..518bb90 100644
--- lib/Target/X86/X86ISelLowering.cpp
+++ lib/Target/X86/X86ISelLowering.cpp
@@ -12462,14 +12462,24 @@ static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
(SplatValue.isPowerOf2() || (-SplatValue).isPowerOf2())) {
unsigned lg2 = SplatValue.countTrailingZeros();
// Splat the sign bit.
- SDValue Sz = DAG.getConstant(EltTy.getSizeInBits()-1, MVT::i32);
- SDValue SGN = getTargetVShiftNode(X86ISD::VSRAI, dl, VT, N0, Sz, DAG);
+ SmallVector<SDValue, 16> Sz(NumElts,
+ DAG.getConstant(EltTy.getSizeInBits() - 1,
+ EltTy));
+ SDValue SGN = DAG.getNode(ISD::SRA, dl, VT, N0,
+ DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &Sz[0],
+ NumElts));
// Add (N0 < 0) ? abs2 - 1 : 0;
- SDValue Amt = DAG.getConstant(EltTy.getSizeInBits() - lg2, MVT::i32);
- SDValue SRL = getTargetVShiftNode(X86ISD::VSRLI, dl, VT, SGN, Amt, DAG);
+ SmallVector<SDValue, 16> Amt(NumElts,
+ DAG.getConstant(EltTy.getSizeInBits() - lg2,
+ EltTy));
+ SDValue SRL = DAG.getNode(ISD::SRL, dl, VT, SGN,
+ DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &Amt[0],
+ NumElts));
SDValue ADD = DAG.getNode(ISD::ADD, dl, VT, N0, SRL);
- SDValue Lg2Amt = DAG.getConstant(lg2, MVT::i32);
- SDValue SRA = getTargetVShiftNode(X86ISD::VSRAI, dl, VT, ADD, Lg2Amt, DAG);
+ SmallVector<SDValue, 16> Lg2Amt(NumElts, DAG.getConstant(lg2, EltTy));
+ SDValue SRA = DAG.getNode(ISD::SRA, dl, VT, ADD,
+ DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &Lg2Amt[0],
+ NumElts));
// If we're dividing by a positive value, we're done. Otherwise, we must
// negate the result.
diff --git a/test/CodeGen/X86/pr16807.ll b/test/CodeGen/X86/pr16807.ll
new file mode 100644
index 0000000..6d55d99
--- /dev/null
+++ test/CodeGen/X86/pr16807.ll
@@ -0,0 +1,18 @@
+; RUN: llc < %s -mtriple=x86_64-linux-gnu -mcpu=core-avx-i | FileCheck %s
+
+define <16 x i16> @f_fu(<16 x i16> %bf) {
+allocas:
+ %avg.i.i = sdiv <16 x i16> %bf, <i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4>
+ ret <16 x i16> %avg.i.i
+}
+
+; CHECK: f_fu
+; CHECK: psraw
+; CHECK: psrlw
+; CHECK: paddw
+; CHECK: psraw
+; CHECK: psraw
+; CHECK: psrlw
+; CHECK: paddw
+; CHECK: psraw
+; CHECK: ret
--
1.8.1.2