[X86] Loosen memory folding requirements for cvtdq2pd and cvtps2pd instructions

[aturetsk]

Summary

According to spec cvtdq2pd and cvtps2pd instructions don't require memory operand to be aligned to 16 bytes. This patch removes this requirement from the memory folding table.

[RKSimon]

test/CodeGen/X86/peephole-cvt-sse.ll

2

Please can you test on a 32-bit target as well and use utils/update_llc_test_checks.py if you can.

[aturetsk]

Improve the test.

[aturetsk]

Hi Simon,

Thanks for the review.

test/CodeGen/X86/peephole-cvt-sse.ll

2

Both are done.

[RKSimon]

test/CodeGen/X86/peephole-cvt-sse.ll

2

Thanks, its also better to use -mattr=+sse4.2 (or similar) instead of a cpu target unless you are specifically testings for that cpu.

[aturetsk]

Change -mcpu=slm with -mattr=+sse4.2.

[RKSimon]

LGTM

lib/Target/X86/X86InstrInfo.cpp

static const X86MemoryFoldTableEntry MemoryFoldTable1[] = {
     { X86::IMUL64rri32,     X86::IMUL64rmi32,         0 },
     { X86::IMUL64rri8,      X86::IMUL64rmi8,          0 },
     { X86::Int_COMISDrr,    X86::Int_COMISDrm,        0 },
     { X86::Int_COMISSrr,    X86::Int_COMISSrm,        0 },
     { X86::CVTSD2SI64rr,    X86::CVTSD2SI64rm,        0 },
     { X86::CVTSD2SIrr,      X86::CVTSD2SIrm,          0 },
     { X86::CVTSS2SI64rr,    X86::CVTSS2SI64rm,        0 },
     { X86::CVTSS2SIrr,      X86::CVTSS2SIrm,          0 },
-    { X86::CVTDQ2PDrr,      X86::CVTDQ2PDrm,          TB_ALIGN_16 },
+    { X86::CVTDQ2PDrr,      X86::CVTDQ2PDrm,          0 },
     { X86::CVTDQ2PSrr,      X86::CVTDQ2PSrm,          TB_ALIGN_16 },
     { X86::CVTPD2DQrr,      X86::CVTPD2DQrm,          TB_ALIGN_16 },
     { X86::CVTPD2PSrr,      X86::CVTPD2PSrm,          TB_ALIGN_16 },
     { X86::CVTPS2DQrr,      X86::CVTPS2DQrm,          TB_ALIGN_16 },
-    { X86::CVTPS2PDrr,      X86::CVTPS2PDrm,          TB_ALIGN_16 },
+    { X86::CVTPS2PDrr,      X86::CVTPS2PDrm,          0 },
     { X86::CVTTPD2DQrr,     X86::CVTTPD2DQrm,         TB_ALIGN_16 },
     { X86::CVTTPS2DQrr,     X86::CVTTPS2DQrm,         TB_ALIGN_16 },
     { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm,  0 },
     { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm,     0 },
     { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm,  0 },
     { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm,     0 },
     { X86::Int_UCOMISDrr,   X86::Int_UCOMISDrm,       0 },
     { X86::Int_UCOMISSrr,   X86::Int_UCOMISSrm,       0 },

test/CodeGen/X86/peephole-cvt-sse.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=x86_64-pc-linux -mcpu=slm < %s | FileCheck %s --check-prefix=X86-64
+; RUN: llc -mtriple=i386-pc-linux -mcpu=slm < %s | FileCheck %s --check-prefix=I386
+
+; Check that unaligned loads merge with cvtdq2pd and cvtps2pd.
+
+define <2 x double> @peephole_cvtps2pd(<4 x float>* %a0) {
+; X86-64-LABEL: peephole_cvtps2pd:
+; X86-64:       # BB#0:
+; X86-64-NEXT:    cvtps2pd (%rdi), %xmm0
+; X86-64-NEXT:    retq
+;
+; I386-LABEL: peephole_cvtps2pd:
+; I386:       # BB#0:
+; I386-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; I386-NEXT:    cvtps2pd (%eax), %xmm0
+; I386-NEXT:    retl
+  %1 = load <4 x float>, <4 x float>* %a0, align 1
+  %2 = shufflevector <4 x float> %1, <4 x float> undef, <2 x i32> <i32 0, i32 1>
+  %3 = fpext <2 x float> %2 to <2 x double>
+  ret <2 x double> %3
+}
+
+define <2 x double> @peephole_cvtdq2pd(<4 x i32>* %a0) {
+; X86-64-LABEL: peephole_cvtdq2pd:
+; X86-64:       # BB#0:
+; X86-64-NEXT:    cvtdq2pd (%rdi), %xmm0
+; X86-64-NEXT:    retq
+;
+; I386-LABEL: peephole_cvtdq2pd:
+; I386:       # BB#0:
+; I386-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; I386-NEXT:    cvtdq2pd (%eax), %xmm0
+; I386-NEXT:    retl
+  %1 = load <4 x i32>, <4 x i32>* %a0, align 1
+  %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <2 x i32> <i32 0, i32 1>
+  %3 = sitofp <2 x i32> %2 to <2 x double>
+  ret <2 x double> %3
+}