http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45980
Summary: Use not in stead of add to generate new constant Product: gcc Version: 4.6.0 Status: UNCONFIRMED Severity: normal Priority: P3 Component: target AssignedTo: unassig...@gcc.gnu.org ReportedBy: car...@google.com CC: car...@google.com Host: i686-linux Target: arm-eabi Build: i686-linux Compile the following code: typedef struct { unsigned long state[5]; unsigned long count[2]; } SHA1_CTX; void SHA1Init(SHA1_CTX* context) { /* SHA1 initialization constants */ context->state[0] = 0x67452301; context->state[1] = 0xEFCDAB89; context->state[2] = 0x98BADCFE; context->state[3] = 0x10325476; context->state[4] = 0xC3D2E1F0; context->count[0] = context->count[1] = 0; } With options -march=armv7-a -mthumb -Os, gcc generates: SHA1Init: ldr r3, .L2 str r3, [r0, #0] add r3, r3, #-2004318072 str r3, [r0, #4] ldr r3, .L2+4 str r3, [r0, #8] sub r3, r3, #-2004318072 str r3, [r0, #12] ldr r3, .L2+8 str r3, [r0, #16] movs r3, #0 str r3, [r0, #24] str r3, [r0, #20] bx lr .L3: .align 2 .L2: .word 1732584193 .word -1732584194 .word -1009589776 This function needs to store 5 large constants to memory. Instead of load the 5 constants from constant pool, gcc found two of them can be computed out by a single add/sub constant instruction. But we can do better, notice that 0x67452301 + 0x98BADCFE = 0xFFFFFFFF 0xEFCDAB89 + 0x10325476 = 0xFFFFFFFF So if we have one such constant, the other one can be computed out by bitwise not. So a shorter result could be: SHA1Init: ldr r3, .L2 str r3, [r0, #0] add r2, r3, #-2004318072 str r2, [r0, #4] movns r3, r3 str r3, [r0, #8] movns r2, r2 str r2, [r0, #12] ldr r3, .L2+4 str r3, [r0, #16] movs r3, #0 str r3, [r0, #24] str r3, [r0, #20] bx lr .L3: .align 2 .L2: .word 1732584193 .word -1009589776