https://llvm.org/bugs/show_bug.cgi?id=31719
Bug ID: 31719
Summary: LLVM maintain useless state when using adc
Product: libraries
Version: trunk
Hardware: PC
OS: Windows NT
Status: NEW
Severity: normal
Priority: P
Component: Backend: X86
Assignee: unassignedb...@nondot.org
Reporter: deadal...@gmail.com
CC: llvm-bugs@lists.llvm.org
Classification: Unclassified
Sample IR (optimized):
; Function Attrs: norecurse nounwind readonly
define %scalar @foo(%scalar* nocapture readonly %this, %scalar %arg.b)
local_unnamed_addr #1 {
entry:
%0 = extractvalue %scalar %arg.b, 0
%.elt = extractvalue [4 x i64] %0, 0
%.elt24 = extractvalue [4 x i64] %0, 1
%.elt26 = extractvalue [4 x i64] %0, 2
%.elt28 = extractvalue [4 x i64] %0, 3
%1 = getelementptr inbounds %scalar , %scalar* %this, i64 0, i32 0, i64 0
%2 = load i64, i64* %1, align 8
%3 = zext i64 %2 to i128
%4 = zext i64 %.elt to i128
%5 = add nuw nsw i128 %3, %4
%6 = trunc i128 %5 to i64
%7 = lshr i128 %5, 64
%8 = getelementptr inbounds %scalar , %scalar * %this, i64 0, i32 0, i64 1
%9 = load i64, i64* %8, align 8
%10 = zext i64 %9 to i128
%11 = zext i64 %.elt24 to i128
%12 = add nuw nsw i128 %10, %11
%13 = add nuw nsw i128 %12, %7
%14 = trunc i128 %13 to i64
%15 = lshr i128 %13, 64
%16 = getelementptr inbounds %scalar , %scalar* %this, i64 0, i32 0, i64 2
%17 = load i64, i64* %16, align 8
%18 = zext i64 %17 to i128
%19 = zext i64 %.elt26 to i128
%20 = add nuw nsw i128 %18, %19
%21 = add nuw nsw i128 %20, %15
%22 = trunc i128 %21 to i64
%23 = lshr i128 %21, 64
%24 = getelementptr inbounds %scalar , %scalar* %this, i64 0, i32 0, i64 3
%25 = load i64, i64* %24, align 8
%26 = zext i64 %25 to i128
%27 = zext i64 %.elt28 to i128
%28 = add nuw nsw i128 %26, %27
%29 = add nuw nsw i128 %28, %23
%30 = trunc i128 %29 to i64
%31 = insertvalue [4 x i64] undef, i64 %6, 0
%32 = insertvalue [4 x i64] %31, i64 %14, 1
%33 = insertvalue [4 x i64] %32, i64 %22, 2
%34 = insertvalue [4 x i64] %33, i64 %30, 3
%35 = insertvalue %S6crypto5field6Scalar undef, [4 x i64] %34, 0
ret %scalar%35
}
attributes #0 = { norecurse nounwind readnone }
attributes #1 = { norecurse nounwind readonly }
Codegen:
foo:
addq (%rsi), %rdx
sbbq %r10, %r10
andl $1, %r10d
addq 8(%rsi), %rcx
sbbq %r11, %r11
andl $1, %r11d
addq %r10, %rcx
adcq $0, %r11
addq 16(%rsi), %r8
sbbq %rax, %rax
andl $1, %eax
addq %r11, %r8
adcq $0, %rax
addq 24(%rsi), %r9
addq %rax, %r9
movq %rdx, (%rdi)
movq %rcx, 8(%rdi)
movq %r8, 16(%rdi)
movq %r9, 24(%rdi)
movq %rdi, %rax
retq
While LLVM is able to leverage the use of the adc instruction (good) it is
unclear why it is doing so in RAX and then adding RAX rather than using ADC
right away. See for instance:
adcq $0, %rax
addq 24(%rsi), %r9
addq %rax, %r9
Uses adc to store the carry in RAX and then perform 2 additions, when it could
simply do
adcq 24(%rsi), %r9
These routine are at the core of various crypto libraries and need to be fast.
Any chance to get better codegen here ?
--
You are receiving this mail because:
You are on the CC list for the bug.
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
