//===---------------------------------------------------------------------===//
-mov $reg, 48(%esp)
-...
-leal 48(%esp), %eax
-mov %eax, (%esp)
-call _foo
-
-Obviously it would have been better for the first mov (or any op) to store
-directly %esp[0] if there are no other uses.
-
-//===---------------------------------------------------------------------===//
-
Adding to the list of cmp / test poor codegen issues:
int test(__m128 *A, __m128 *B) {
//===---------------------------------------------------------------------===//
-If shorter, we should use things like:
-movzwl %ax, %eax
-instead of:
-andl $65535, %EAX
-
-The former can also be used when the two-addressy nature of the 'and' would
-require a copy to be inserted (in X86InstrInfo::convertToThreeAddress).
-
-//===---------------------------------------------------------------------===//
-
-Another instruction selector deficiency:
-
-void %bar() {
- %tmp = load int (int)** %foo
- %tmp = tail call int %tmp( int 3 )
- ret void
-}
-
-_bar:
- subl $12, %esp
- movl L_foo$non_lazy_ptr, %eax
- movl (%eax), %eax
- call *%eax
- addl $12, %esp
- ret
-
-The current isel scheme will not allow the load to be folded in the call since
-the load's chain result is read by the callseq_start.
-
-//===---------------------------------------------------------------------===//
-
For this:
int test(int a)
However, if we care more about code size, then imull is better. It's two bytes
shorter than movl + leal.
+On a Pentium M, both variants have the same characteristics with regard
+to throughput; however, the multiplication has a latency of four cycles, as
+opposed to two cycles for the movl+lea variant.
+
//===---------------------------------------------------------------------===//
__builtin_ffs codegen is messy.
to grab the bytes from the next cacheline.
532 532 0x3cfc movb (1809(%esp, %esi), %bl <<<--- spans 2 64 byte lines
-942 942 0x3d03 movl %dh, (1809(%esp, %esi)
-937 937 0x3d0a incl %esi
-3 3 0x3d0b cmpb %bl, %dl
+942 942 0x3d03 movl %dh, (1809(%esp, %esi)
+937 937 0x3d0a incl %esi
+3 3 0x3d0b cmpb %bl, %dl
27 27 0x3d0d jnz 0x000062db <main+11707>
//===---------------------------------------------------------------------===//
//===---------------------------------------------------------------------===//
-For this code:
-
-cond_next603: ; preds = %bb493, %cond_true336, %cond_next599
- %v.21050.1 = phi i32 [ %v.21050.0, %cond_next599 ], [ %tmp344, %cond_true336 ], [ %v.2, %bb493 ] ; <i32> [#uses=1]
- %maxz.21051.1 = phi i32 [ %maxz.21051.0, %cond_next599 ], [ 0, %cond_true336 ], [ %maxz.2, %bb493 ] ; <i32> [#uses=2]
- %cnt.01055.1 = phi i32 [ %cnt.01055.0, %cond_next599 ], [ 0, %cond_true336 ], [ %cnt.0, %bb493 ] ; <i32> [#uses=2]
- %byteptr.9 = phi i8* [ %byteptr.12, %cond_next599 ], [ %byteptr.0, %cond_true336 ], [ %byteptr.10, %bb493 ] ; <i8*> [#uses=9]
- %bitptr.6 = phi i32 [ %tmp5571104.1, %cond_next599 ], [ %tmp4921049, %cond_true336 ], [ %bitptr.7, %bb493 ] ; <i32> [#uses=4]
- %source.5 = phi i32 [ %tmp602, %cond_next599 ], [ %source.0, %cond_true336 ], [ %source.6, %bb493 ] ; <i32> [#uses=7]
- %tmp606 = getelementptr %struct.const_tables* @tables, i32 0, i32 0, i32 %cnt.01055.1 ; <i8*> [#uses=1]
- %tmp607 = load i8* %tmp606, align 1 ; <i8> [#uses=1]
-
-We produce this:
-
-LBB4_70: # cond_next603
- movl -20(%ebp), %esi
- movl L_tables$non_lazy_ptr-"L4$pb"(%esi), %esi
-
-However, ICC caches this information before the loop and produces this:
-
- movl 88(%esp), %eax #481.12
-
-//===---------------------------------------------------------------------===//
-
This code:
%tmp659 = icmp slt i16 %tmp654, 0 ; <i1> [#uses=1]
//===---------------------------------------------------------------------===//
-These two functions perform identical operations:
+See rdar://4653682.
+
+From flops:
+
+LBB1_15: # bb310
+ cvtss2sd LCPI1_0, %xmm1
+ addsd %xmm1, %xmm0
+ movsd 176(%esp), %xmm2
+ mulsd %xmm0, %xmm2
+ movapd %xmm2, %xmm3
+ mulsd %xmm3, %xmm3
+ movapd %xmm3, %xmm4
+ mulsd LCPI1_23, %xmm4
+ addsd LCPI1_24, %xmm4
+ mulsd %xmm3, %xmm4
+ addsd LCPI1_25, %xmm4
+ mulsd %xmm3, %xmm4
+ addsd LCPI1_26, %xmm4
+ mulsd %xmm3, %xmm4
+ addsd LCPI1_27, %xmm4
+ mulsd %xmm3, %xmm4
+ addsd LCPI1_28, %xmm4
+ mulsd %xmm3, %xmm4
+ addsd %xmm1, %xmm4
+ mulsd %xmm2, %xmm4
+ movsd 152(%esp), %xmm1
+ addsd %xmm4, %xmm1
+ movsd %xmm1, 152(%esp)
+ incl %eax
+ cmpl %eax, %esi
+ jge LBB1_15 # bb310
+LBB1_16: # bb358.loopexit
+ movsd 152(%esp), %xmm0
+ addsd %xmm0, %xmm0
+ addsd LCPI1_22, %xmm0
+ movsd %xmm0, 152(%esp)
+
+Rather than spilling the result of the last addsd in the loop, we should have
+insert a copy to split the interval (one for the duration of the loop, one
+extending to the fall through). The register pressure in the loop isn't high
+enough to warrant the spill.
+
+Also check why xmm7 is not used at all in the function.
+
+//===---------------------------------------------------------------------===//
+
+Legalize loses track of the fact that bools are always zero extended when in
+memory. This causes us to compile abort_gzip (from 164.gzip) from:
-define i32 @test(i32 %f12) {
- %tmp7.25 = lshr i32 %f12, 16
- %tmp7.26 = trunc i32 %tmp7.25 to i8
- %tmp78.2 = sext i8 %tmp7.26 to i32
- ret i32 %tmp78.2
+target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:128:128"
+target triple = "i386-apple-darwin8"
+@in_exit.4870.b = internal global i1 false ; <i1*> [#uses=2]
+define fastcc void @abort_gzip() noreturn nounwind {
+entry:
+ %tmp.b.i = load i1* @in_exit.4870.b ; <i1> [#uses=1]
+ br i1 %tmp.b.i, label %bb.i, label %bb4.i
+bb.i: ; preds = %entry
+ tail call void @exit( i32 1 ) noreturn nounwind
+ unreachable
+bb4.i: ; preds = %entry
+ store i1 true, i1* @in_exit.4870.b
+ tail call void @exit( i32 1 ) noreturn nounwind
+ unreachable
}
+declare void @exit(i32) noreturn nounwind
-define i32 @test2(i32 %f12) {
- %f11 = shl i32 %f12, 8
- %tmp7.25 = ashr i32 %f11, 24
- ret i32 %tmp7.25
+into:
+
+_abort_gzip:
+ subl $12, %esp
+ movb _in_exit.4870.b, %al
+ notb %al
+ testb $1, %al
+ jne LBB1_2 ## bb4.i
+LBB1_1: ## bb.i
+ ...
+
+//===---------------------------------------------------------------------===//
+
+We compile:
+
+int test(int x, int y) {
+ return x-y-1;
}
-but the first compiles into significantly better code on x86-32:
+into (-m64):
_test:
- movsbl 6(%esp), %eax
- ret
-_test2:
- movl 4(%esp), %eax
- shll $8, %eax
- sarl $24, %eax
+ decl %edi
+ movl %edi, %eax
+ subl %esi, %eax
ret
-
-and on x86-64:
-_test:
- shrl $16, %edi
- movsbl %dil, %eax
+it would be better to codegen as: x+~y (notl+addl)
+
+//===---------------------------------------------------------------------===//
+
+This code:
+
+int foo(const char *str,...)
+{
+ __builtin_va_list a; int x;
+ __builtin_va_start(a,str); x = __builtin_va_arg(a,int); __builtin_va_end(a);
+ return x;
+}
+
+gets compiled into this on x86-64:
+ subq $200, %rsp
+ movaps %xmm7, 160(%rsp)
+ movaps %xmm6, 144(%rsp)
+ movaps %xmm5, 128(%rsp)
+ movaps %xmm4, 112(%rsp)
+ movaps %xmm3, 96(%rsp)
+ movaps %xmm2, 80(%rsp)
+ movaps %xmm1, 64(%rsp)
+ movaps %xmm0, 48(%rsp)
+ movq %r9, 40(%rsp)
+ movq %r8, 32(%rsp)
+ movq %rcx, 24(%rsp)
+ movq %rdx, 16(%rsp)
+ movq %rsi, 8(%rsp)
+ leaq (%rsp), %rax
+ movq %rax, 192(%rsp)
+ leaq 208(%rsp), %rax
+ movq %rax, 184(%rsp)
+ movl $48, 180(%rsp)
+ movl $8, 176(%rsp)
+ movl 176(%rsp), %eax
+ cmpl $47, %eax
+ jbe .LBB1_3 # bb
+.LBB1_1: # bb3
+ movq 184(%rsp), %rcx
+ leaq 8(%rcx), %rax
+ movq %rax, 184(%rsp)
+.LBB1_2: # bb4
+ movl (%rcx), %eax
+ addq $200, %rsp
+ ret
+.LBB1_3: # bb
+ movl %eax, %ecx
+ addl $8, %eax
+ addq 192(%rsp), %rcx
+ movl %eax, 176(%rsp)
+ jmp .LBB1_2 # bb4
+
+gcc 4.3 generates:
+ subq $96, %rsp
+.LCFI0:
+ leaq 104(%rsp), %rax
+ movq %rsi, -80(%rsp)
+ movl $8, -120(%rsp)
+ movq %rax, -112(%rsp)
+ leaq -88(%rsp), %rax
+ movq %rax, -104(%rsp)
+ movl $8, %eax
+ cmpl $48, %eax
+ jb .L6
+ movq -112(%rsp), %rdx
+ movl (%rdx), %eax
+ addq $96, %rsp
+ ret
+ .p2align 4,,10
+ .p2align 3
+.L6:
+ mov %eax, %edx
+ addq -104(%rsp), %rdx
+ addl $8, %eax
+ movl %eax, -120(%rsp)
+ movl (%rdx), %eax
+ addq $96, %rsp
+ ret
+
+and it gets compiled into this on x86:
+ pushl %ebp
+ movl %esp, %ebp
+ subl $4, %esp
+ leal 12(%ebp), %eax
+ movl %eax, -4(%ebp)
+ leal 16(%ebp), %eax
+ movl %eax, -4(%ebp)
+ movl 12(%ebp), %eax
+ addl $4, %esp
+ popl %ebp
+ ret
+
+gcc 4.3 generates:
+ pushl %ebp
+ movl %esp, %ebp
+ movl 12(%ebp), %eax
+ popl %ebp
+ ret
+
+//===---------------------------------------------------------------------===//
+
+Teach tblgen not to check bitconvert source type in some cases. This allows us
+to consolidate the following patterns in X86InstrMMX.td:
+
+def : Pat<(v2i32 (bitconvert (i64 (vector_extract (v2i64 VR128:$src),
+ (iPTR 0))))),
+ (v2i32 (MMX_MOVDQ2Qrr VR128:$src))>;
+def : Pat<(v4i16 (bitconvert (i64 (vector_extract (v2i64 VR128:$src),
+ (iPTR 0))))),
+ (v4i16 (MMX_MOVDQ2Qrr VR128:$src))>;
+def : Pat<(v8i8 (bitconvert (i64 (vector_extract (v2i64 VR128:$src),
+ (iPTR 0))))),
+ (v8i8 (MMX_MOVDQ2Qrr VR128:$src))>;
+
+There are other cases in various td files.
+
+//===---------------------------------------------------------------------===//
+
+Take something like the following on x86-32:
+unsigned a(unsigned long long x, unsigned y) {return x % y;}
+
+We currently generate a libcall, but we really shouldn't: the expansion is
+shorter and likely faster than the libcall. The expected code is something
+like the following:
+
+ movl 12(%ebp), %eax
+ movl 16(%ebp), %ecx
+ xorl %edx, %edx
+ divl %ecx
+ movl 8(%ebp), %eax
+ divl %ecx
+ movl %edx, %eax
ret
-_test2:
- shll $8, %edi
- movl %edi, %eax
- sarl $24, %eax
+
+A similar code sequence works for division.
+
+//===---------------------------------------------------------------------===//
+
+These should compile to the same code, but the later codegen's to useless
+instructions on X86. This may be a trivial dag combine (GCC PR7061):
+
+struct s1 { unsigned char a, b; };
+unsigned long f1(struct s1 x) {
+ return x.a + x.b;
+}
+struct s2 { unsigned a: 8, b: 8; };
+unsigned long f2(struct s2 x) {
+ return x.a + x.b;
+}
+
+//===---------------------------------------------------------------------===//
+
+We currently compile this:
+
+define i32 @func1(i32 %v1, i32 %v2) nounwind {
+entry:
+ %t = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %v1, i32 %v2)
+ %sum = extractvalue {i32, i1} %t, 0
+ %obit = extractvalue {i32, i1} %t, 1
+ br i1 %obit, label %overflow, label %normal
+normal:
+ ret i32 %sum
+overflow:
+ call void @llvm.trap()
+ unreachable
+}
+declare {i32, i1} @llvm.sadd.with.overflow.i32(i32, i32)
+declare void @llvm.trap()
+
+to:
+
+_func1:
+ movl 4(%esp), %eax
+ addl 8(%esp), %eax
+ jo LBB1_2 ## overflow
+LBB1_1: ## normal
ret
+LBB1_2: ## overflow
+ ud2
-I would like instcombine to canonicalize the first into the second (since it is
-shorter and doesn't involve type width changes) but the x86 backend needs to do
-the right thing with the later sequence first.
+it would be nice to produce "into" someday.
//===---------------------------------------------------------------------===//
projects / oota-llvm.git / blobdiff