//===---------------------------------------------------------------------===//
-Compile this:
-_Bool f(_Bool a) { return a!=1; }
-
-into:
- movzbl %dil, %eax
- xorl $1, %eax
- ret
-
-(Although note that this isn't a legal way to express the code that llvm-gcc
-currently generates for that function.)
-
-//===---------------------------------------------------------------------===//
-
Some isel ideas:
1. Dynamic programming based approach when compile time if not an
//===---------------------------------------------------------------------===//
+This:
+
+void foo(void);
+void bar(int x, int *P) {
+ x >>= 2;
+ if (x)
+ foo();
+ *P = x;
+}
+
+compiles into:
+
+ movq %rsi, %rbx
+ movl %edi, %r14d
+ sarl $2, %r14d
+ testl %r14d, %r14d
+ je LBB0_2
+
+Instead of doing an explicit test, we can use the flags off the sar. This
+occurs in a bigger testcase like this, which is pretty common:
+
+#include <vector>
+int test1(std::vector<int> &X) {
+ int Sum = 0;
+ for (long i = 0, e = X.size(); i != e; ++i)
+ X[i] = 0;
+ return Sum;
+}
+
+//===---------------------------------------------------------------------===//
+
Only use inc/neg/not instructions on processors where they are faster than
add/sub/xor. They are slower on the P4 due to only updating some processor
flags.
//===---------------------------------------------------------------------===//
-Codegen:
-
-int f(int a, int b) {
- if (a == 4 || a == 6)
- b++;
- return b;
-}
-
-
-as:
-
-or eax, 2
-cmp eax, 6
-jz label
-
-//===---------------------------------------------------------------------===//
-
GCC's ix86_expand_int_movcc function (in i386.c) has a ton of interesting
-simplifications for integer "x cmp y ? a : b". For example, instead of:
-
-int G;
-void f(int X, int Y) {
- G = X < 0 ? 14 : 13;
-}
-
-compiling to:
-
-_f:
- movl $14, %eax
- movl $13, %ecx
- movl 4(%esp), %edx
- testl %edx, %edx
- cmovl %eax, %ecx
- movl %ecx, _G
- ret
-
-it could be:
-_f:
- movl 4(%esp), %eax
- sarl $31, %eax
- notl %eax
- addl $14, %eax
- movl %eax, _G
- ret
-
-etc.
-
-Another is:
-int usesbb(unsigned int a, unsigned int b) {
- return (a < b ? -1 : 0);
-}
-to:
-_usesbb:
- movl 8(%esp), %eax
- cmpl %eax, 4(%esp)
- sbbl %eax, %eax
- ret
-
-instead of:
-_usesbb:
- xorl %eax, %eax
- movl 8(%esp), %ecx
- cmpl %ecx, 4(%esp)
- movl $4294967295, %ecx
- cmovb %ecx, %eax
- ret
+simplifications for integer "x cmp y ? a : b".
//===---------------------------------------------------------------------===//
//===---------------------------------------------------------------------===//
-Consider:
-int test(unsigned long a, unsigned long b) { return -(a < b); }
-
-We currently compile this to:
-
-define i32 @test(i32 %a, i32 %b) nounwind {
- %tmp3 = icmp ult i32 %a, %b ; <i1> [#uses=1]
- %tmp34 = zext i1 %tmp3 to i32 ; <i32> [#uses=1]
- %tmp5 = sub i32 0, %tmp34 ; <i32> [#uses=1]
- ret i32 %tmp5
-}
-
-and
-
-_test:
- movl 8(%esp), %eax
- cmpl %eax, 4(%esp)
- setb %al
- movzbl %al, %eax
- negl %eax
- ret
-
-Several deficiencies here. First, we should instcombine zext+neg into sext:
-
-define i32 @test2(i32 %a, i32 %b) nounwind {
- %tmp3 = icmp ult i32 %a, %b ; <i1> [#uses=1]
- %tmp34 = sext i1 %tmp3 to i32 ; <i32> [#uses=1]
- ret i32 %tmp34
-}
-
-However, before we can do that, we have to fix the bad codegen that we get for
-sext from bool:
-
-_test2:
- movl 8(%esp), %eax
- cmpl %eax, 4(%esp)
- setb %al
- movzbl %al, %eax
- shll $31, %eax
- sarl $31, %eax
- ret
-
-This code should be at least as good as the code above. Once this is fixed, we
-can optimize this specific case even more to:
-
- movl 8(%esp), %eax
- xorl %ecx, %ecx
- cmpl %eax, 4(%esp)
- sbbl %ecx, %ecx
-
-//===---------------------------------------------------------------------===//
-
Take the following code (from
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16541):
RegsForValue::getCopyFromRegs doesn't know how to put an AssertSext on the
constructed BUILD_PAIR which represents the cast value.
+This can be handled by making CodeGenPrepare sink the cast.
+
//===---------------------------------------------------------------------===//
Test instructions can be eliminated by using EFLAGS values from arithmetic
//===---------------------------------------------------------------------===//
-Testcase:
-int x(int a) { return (a & 0x80) ? 0x100 : 0; }
-int y(int a) { return (a & 0x80) *2; }
-
-Current:
- testl $128, 4(%esp)
- setne %al
- movzbl %al, %eax
- shll $8, %eax
- ret
-
-Better:
- movl 4(%esp), %eax
- addl %eax, %eax
- andl $256, %eax
- ret
-
-This is another general instcombine transformation that is profitable on all
-targets. In LLVM IR, these functions look like this:
-
-define i32 @x(i32 %a) nounwind readnone {
-entry:
- %0 = and i32 %a, 128
- %1 = icmp eq i32 %0, 0
- %iftmp.0.0 = select i1 %1, i32 0, i32 256
- ret i32 %iftmp.0.0
-}
-
-define i32 @y(i32 %a) nounwind readnone {
-entry:
- %0 = shl i32 %a, 1
- %1 = and i32 %0, 256
- ret i32 %1
-}
-
-Replacing an icmp+select with a shift should always be considered profitable in
-instcombine.
-
-//===---------------------------------------------------------------------===//
-
Re-implement atomic builtins __sync_add_and_fetch() and __sync_sub_and_fetch
properly.
shows up with bitfields.
//===---------------------------------------------------------------------===//
+
+Take the following C code:
+int f(int a, int b) { return (unsigned char)a == (unsigned char)b; }
+
+We generate the following IR with clang:
+define i32 @f(i32 %a, i32 %b) nounwind readnone {
+entry:
+ %tmp = xor i32 %b, %a ; <i32> [#uses=1]
+ %tmp6 = and i32 %tmp, 255 ; <i32> [#uses=1]
+ %cmp = icmp eq i32 %tmp6, 0 ; <i1> [#uses=1]
+ %conv5 = zext i1 %cmp to i32 ; <i32> [#uses=1]
+ ret i32 %conv5
+}
+
+And the following x86 code:
+ xorl %esi, %edi
+ testb $-1, %dil
+ sete %al
+ movzbl %al, %eax
+ ret
+
+A cmpb instead of the xorl+testb would be one instruction shorter.
+
+//===---------------------------------------------------------------------===//
+
+Given the following C code:
+int f(int a, int b) { return (signed char)a == (signed char)b; }
+
+We generate the following IR with clang:
+define i32 @f(i32 %a, i32 %b) nounwind readnone {
+entry:
+ %sext = shl i32 %a, 24 ; <i32> [#uses=1]
+ %conv1 = ashr i32 %sext, 24 ; <i32> [#uses=1]
+ %sext6 = shl i32 %b, 24 ; <i32> [#uses=1]
+ %conv4 = ashr i32 %sext6, 24 ; <i32> [#uses=1]
+ %cmp = icmp eq i32 %conv1, %conv4 ; <i1> [#uses=1]
+ %conv5 = zext i1 %cmp to i32 ; <i32> [#uses=1]
+ ret i32 %conv5
+}
+
+And the following x86 code:
+ movsbl %sil, %eax
+ movsbl %dil, %ecx
+ cmpl %eax, %ecx
+ sete %al
+ movzbl %al, %eax
+ ret
+
+
+It should be possible to eliminate the sign extensions.
+
+//===---------------------------------------------------------------------===//
+
+LLVM misses a load+store narrowing opportunity in this code:
+
+%struct.bf = type { i64, i16, i16, i32 }
+
+@bfi = external global %struct.bf* ; <%struct.bf**> [#uses=2]
+
+define void @t1() nounwind ssp {
+entry:
+ %0 = load %struct.bf** @bfi, align 8 ; <%struct.bf*> [#uses=1]
+ %1 = getelementptr %struct.bf* %0, i64 0, i32 1 ; <i16*> [#uses=1]
+ %2 = bitcast i16* %1 to i32* ; <i32*> [#uses=2]
+ %3 = load i32* %2, align 1 ; <i32> [#uses=1]
+ %4 = and i32 %3, -65537 ; <i32> [#uses=1]
+ store i32 %4, i32* %2, align 1
+ %5 = load %struct.bf** @bfi, align 8 ; <%struct.bf*> [#uses=1]
+ %6 = getelementptr %struct.bf* %5, i64 0, i32 1 ; <i16*> [#uses=1]
+ %7 = bitcast i16* %6 to i32* ; <i32*> [#uses=2]
+ %8 = load i32* %7, align 1 ; <i32> [#uses=1]
+ %9 = and i32 %8, -131073 ; <i32> [#uses=1]
+ store i32 %9, i32* %7, align 1
+ ret void
+}
+
+LLVM currently emits this:
+
+ movq bfi(%rip), %rax
+ andl $-65537, 8(%rax)
+ movq bfi(%rip), %rax
+ andl $-131073, 8(%rax)
+ ret
+
+It could narrow the loads and stores to emit this:
+
+ movq bfi(%rip), %rax
+ andb $-2, 10(%rax)
+ movq bfi(%rip), %rax
+ andb $-3, 10(%rax)
+ ret
+
+The trouble is that there is a TokenFactor between the store and the
+load, making it non-trivial to determine if there's anything between
+the load and the store which would prohibit narrowing.
+
+//===---------------------------------------------------------------------===//
+
+This code:
+void foo(unsigned x) {
+ if (x == 0) bar();
+ else if (x == 1) qux();
+}
+
+currently compiles into:
+_foo:
+ movl 4(%esp), %eax
+ cmpl $1, %eax
+ je LBB0_3
+ testl %eax, %eax
+ jne LBB0_4
+
+the testl could be removed:
+_foo:
+ movl 4(%esp), %eax
+ cmpl $1, %eax
+ je LBB0_3
+ jb LBB0_4
+
+0 is the only unsigned number < 1.
+
+//===---------------------------------------------------------------------===//
+
+This code:
+
+%0 = type { i32, i1 }
+
+define i32 @add32carry(i32 %sum, i32 %x) nounwind readnone ssp {
+entry:
+ %uadd = tail call %0 @llvm.uadd.with.overflow.i32(i32 %sum, i32 %x)
+ %cmp = extractvalue %0 %uadd, 1
+ %inc = zext i1 %cmp to i32
+ %add = add i32 %x, %sum
+ %z.0 = add i32 %add, %inc
+ ret i32 %z.0
+}
+
+declare %0 @llvm.uadd.with.overflow.i32(i32, i32) nounwind readnone
+
+compiles to:
+
+_add32carry: ## @add32carry
+ addl %esi, %edi
+ sbbl %ecx, %ecx
+ movl %edi, %eax
+ subl %ecx, %eax
+ ret
+
+But it could be:
+
+_add32carry:
+ leal (%rsi,%rdi), %eax
+ cmpl %esi, %eax
+ adcl $0, %eax
+ ret
+
+//===---------------------------------------------------------------------===//