[llvm] [RISCV][GISEL] Fix legalization for G_MERGE/UNMERGE_VALUES (PR #75619)
via llvm-commits
llvm-commits at lists.llvm.org
Fri Dec 15 08:02:34 PST 2023
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-risc-v
Author: Michael Maitland (michaelmaitland)
<details>
<summary>Changes</summary>
There are two bugs fixed by this patch:
1. G_MERGE_VALUES is always true, causing IdxZeroTy and IdxOneTy to always be s64 and s32 respectively.
2. The legalFor check was for the type in index 0, instead of for index 0 and index 1, so we were allowing s64 or s32 in type 0 to be legal for both G_MERGE_VALUES and G_UNMERGE_VALUES.
The existing test was passing because (2) covered up (1).
---
Full diff: https://github.com/llvm/llvm-project/pull/75619.diff
1 Files Affected:
- (modified) llvm/lib/Target/RISCV/GISel/RISCVLegalizerInfo.cpp (+7-5)
``````````diff
diff --git a/llvm/lib/Target/RISCV/GISel/RISCVLegalizerInfo.cpp b/llvm/lib/Target/RISCV/GISel/RISCVLegalizerInfo.cpp
index 8f03a7ac41d37b..b46269fb9565fb 100644
--- a/llvm/lib/Target/RISCV/GISel/RISCVLegalizerInfo.cpp
+++ b/llvm/lib/Target/RISCV/GISel/RISCVLegalizerInfo.cpp
@@ -85,14 +85,16 @@ RISCVLegalizerInfo::RISCVLegalizerInfo(const RISCVSubtarget &ST)
// Merge/Unmerge
for (unsigned Op : {G_MERGE_VALUES, G_UNMERGE_VALUES}) {
- unsigned BigTyIdx = Op == G_MERGE_VALUES ? 0 : 1;
- unsigned LitTyIdx = Op == G_MERGE_VALUES ? 1 : 0;
auto &MergeUnmergeActions = getActionDefinitionsBuilder(Op);
if (XLen == 32 && ST.hasStdExtD()) {
- LLT IdxZeroTy = G_MERGE_VALUES ? s64 : s32;
- LLT IdxOneTy = G_MERGE_VALUES ? s32 : s64;
- MergeUnmergeActions.legalFor({IdxZeroTy, IdxOneTy});
+ // Only need to legalize on IdxZeroTy since there is no way to construct
+ // a valid MERGE/UNMERGE where the types don't combine/decompose as
+ // expected.
+ LLT IdxZeroTy = Op == G_MERGE_VALUES ? s64 : s32;
+ MergeUnmergeActions.legalFor({IdxZeroTy});
}
+ unsigned BigTyIdx = Op == G_MERGE_VALUES ? 0 : 1;
+ unsigned LitTyIdx = Op == G_MERGE_VALUES ? 1 : 0;
MergeUnmergeActions.widenScalarToNextPow2(LitTyIdx, XLen)
.widenScalarToNextPow2(BigTyIdx, XLen)
.clampScalar(LitTyIdx, sXLen, sXLen)
``````````
</details>
https://github.com/llvm/llvm-project/pull/75619
More information about the llvm-commits
mailing list