|
| 1 | +/*========================== begin_copyright_notice ============================ |
| 2 | + |
| 3 | +Copyright (C) 2023 Intel Corporation |
| 4 | + |
| 5 | +SPDX-License-Identifier: MIT |
| 6 | + |
| 7 | +============================= end_copyright_notice ===========================*/ |
| 8 | + |
| 9 | +From 6d6c78dd486c0ed1995bab48ac8cd7ecf1d926b8 Mon Sep 17 00:00:00 2001 |
| 10 | +From: Benjamin Kramer < [email protected]> |
| 11 | +Date: Mon, 13 Jun 2022 14:31:43 +0200 |
| 12 | +Subject: [InstCombine] Only fold trunc(ext) pairs to bitcast if the source and |
| 13 | + destination types are the same |
| 14 | + |
| 15 | +This used to be always the case, but the addition of bfloat to the type |
| 16 | +matrix makes this invalid. |
| 17 | +--- |
| 18 | + llvm/lib/IR/Instructions.cpp | 10 ++++++---- |
| 19 | + llvm/test/Transforms/InstCombine/fpextend.ll | 11 +++++++++++ |
| 20 | + 2 files changed, 17 insertions(+), 4 deletions(-) |
| 21 | + |
| 22 | +diff --git a/llvm/lib/IR/Instructions.cpp b/llvm/lib/IR/Instructions.cpp |
| 23 | +index 7798af3b19b9..2f68807bb8bc 100644 |
| 24 | +--- a/llvm/lib/IR/Instructions.cpp |
| 25 | ++++ b/llvm/lib/IR/Instructions.cpp |
| 26 | +@@ -3060,16 +3060,18 @@ unsigned CastInst::isEliminableCastPair( |
| 27 | + return 0; |
| 28 | + } |
| 29 | + case 8: { |
| 30 | +- // ext, trunc -> bitcast, if the SrcTy and DstTy are same size |
| 31 | ++ // ext, trunc -> bitcast, if the SrcTy and DstTy are the same |
| 32 | + // ext, trunc -> ext, if sizeof(SrcTy) < sizeof(DstTy) |
| 33 | + // ext, trunc -> trunc, if sizeof(SrcTy) > sizeof(DstTy) |
| 34 | + unsigned SrcSize = SrcTy->getScalarSizeInBits(); |
| 35 | + unsigned DstSize = DstTy->getScalarSizeInBits(); |
| 36 | +- if (SrcSize == DstSize) |
| 37 | ++ if (SrcTy == DstTy) |
| 38 | + return Instruction::BitCast; |
| 39 | +- else if (SrcSize < DstSize) |
| 40 | ++ if (SrcSize < DstSize) |
| 41 | + return firstOp; |
| 42 | +- return secondOp; |
| 43 | ++ if (SrcSize > DstSize) |
| 44 | ++ return secondOp; |
| 45 | ++ return 0; |
| 46 | + } |
| 47 | + case 9: |
| 48 | + // zext, sext -> zext, because sext can't sign extend after zext |
| 49 | +diff --git a/llvm/test/Transforms/InstCombine/fpextend.ll b/llvm/test/Transforms/InstCombine/fpextend.ll |
| 50 | +index 9fe85e983fb7..6b8fe928ff8a 100644 |
| 51 | +--- a/llvm/test/Transforms/InstCombine/fpextend.ll |
| 52 | ++++ b/llvm/test/Transforms/InstCombine/fpextend.ll |
| 53 | +@@ -429,3 +429,14 @@ define double @FtoItoFtoF_f32_su32_f32_f64(float %f) { |
| 54 | + %r = fpext float %x to double |
| 55 | + ret double %r |
| 56 | + } |
| 57 | ++ |
| 58 | ++define half @bf16_to_f32_to_f16(bfloat %a) nounwind { |
| 59 | ++; CHECK-LABEL: @bf16_to_f32_to_f16( |
| 60 | ++; CHECK-NEXT: [[Y:%.*]] = fpext bfloat [[A:%.*]] to float |
| 61 | ++; CHECK-NEXT: [[Z:%.*]] = fptrunc float [[Y]] to half |
| 62 | ++; CHECK-NEXT: ret half [[Z]] |
| 63 | ++; |
| 64 | ++ %y = fpext bfloat %a to float |
| 65 | ++ %z = fptrunc float %y to half |
| 66 | ++ ret half %z |
| 67 | ++} |
| 68 | +-- |
| 69 | +2.34.1 |
| 70 | + |
0 commit comments