ManuelJBrito added inline comments.
================ Comment at: llvm/lib/Target/X86/X86InstCombineIntrinsic.cpp:3102-3105 case Intrinsic::x86_sse4a_extrqi: case Intrinsic::x86_sse4a_insertq: case Intrinsic::x86_sse4a_insertqi: + PoisonElts.setHighBits(VWidth / 2); ---------------- nlopes wrote: > this change doesn't look correct. The bits are undefined, doesn't seem like > you can use poison here. Thanks, I'll fix it. This will cause some regressions in the cases where only the upper half is demanded. Eventually we could fold those cases to freeze(poison). ================ Comment at: llvm/test/Transforms/InstCombine/vec_shuffle.ll:1301 ; CHECK-LABEL: @fmul_splat_constant( -; CHECK-NEXT: [[TMP1:%.*]] = fmul <2 x float> [[X:%.*]], <float 4.200000e+01, float poison> -; CHECK-NEXT: [[R:%.*]] = shufflevector <2 x float> [[TMP1]], <2 x float> poison, <2 x i32> zeroinitializer + ; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <2 x float> [[X:%.*]], <2 x float> undef, <2 x i32> zeroinitializer +; CHECK-NEXT: [[R:%.*]] = fmul <2 x float> [[TMP1]], <float 4.200000e+01, float 4.200000e+01> ---------------- nlopes wrote: > Why does it regress here (goes from poison to undef)? This regression is due to the change in foldVectorBinop in InstructionCombining.cpp It seems that without this canonicalization, demanded elements doesn't trigger causing the regression. I changed it to match poison instead of undef, because of the following: shufflevector < 4 x i32 > %1, < 4 x i32 > undef , <1, 2, 6, 7> --> shufflevector < 4 x i32 > %1, < 4 x i32 > undef , <1, 2, undef, undef> Now we don't do this because it would insert poison. Not simplifying the mask causes the following assert to fail: ``` assert(ShMask[I] < (int)NumElts && "Not expecting narrowing shuffle") ``` Repository: rG LLVM Github Monorepo CHANGES SINCE LAST ACTION https://reviews.llvm.org/D152658/new/ https://reviews.llvm.org/D152658 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
